4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/perf_event.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/kthread.h>
59 #include <linux/proc_fs.h>
60 #include <linux/seq_file.h>
61 #include <linux/sysctl.h>
62 #include <linux/syscalls.h>
63 #include <linux/times.h>
64 #include <linux/tsacct_kern.h>
65 #include <linux/kprobes.h>
66 #include <linux/delayacct.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
73 #include <linux/ftrace.h>
76 #include <asm/irq_regs.h>
78 #include "sched_cpupri.h"
80 #define CREATE_TRACE_POINTS
81 #include <trace/events/sched.h>
84 * Convert user-nice values [ -20 ... 0 ... 19 ]
85 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
88 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
89 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
90 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
93 * 'User priority' is the nice value converted to something we
94 * can work with better when scaling various scheduler parameters,
95 * it's a [ 0 ... 39 ] range.
97 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
98 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
99 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
102 * Helpers for converting nanosecond timing to jiffy resolution
104 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
106 #define NICE_0_LOAD SCHED_LOAD_SCALE
107 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
110 * These are the 'tuning knobs' of the scheduler:
112 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
113 * Timeslices get refilled after they expire.
115 #define DEF_TIMESLICE (100 * HZ / 1000)
118 * single value that denotes runtime == period, ie unlimited time.
120 #define RUNTIME_INF ((u64)~0ULL)
122 static inline int rt_policy(int policy)
124 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
129 static inline int task_has_rt_policy(struct task_struct *p)
131 return rt_policy(p->policy);
135 * This is the priority-queue data structure of the RT scheduling class:
137 struct rt_prio_array {
138 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
139 struct list_head queue[MAX_RT_PRIO];
142 struct rt_bandwidth {
143 /* nests inside the rq lock: */
144 spinlock_t rt_runtime_lock;
147 struct hrtimer rt_period_timer;
150 static struct rt_bandwidth def_rt_bandwidth;
152 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
154 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
156 struct rt_bandwidth *rt_b =
157 container_of(timer, struct rt_bandwidth, rt_period_timer);
163 now = hrtimer_cb_get_time(timer);
164 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
169 idle = do_sched_rt_period_timer(rt_b, overrun);
172 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
176 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
178 rt_b->rt_period = ns_to_ktime(period);
179 rt_b->rt_runtime = runtime;
181 spin_lock_init(&rt_b->rt_runtime_lock);
183 hrtimer_init(&rt_b->rt_period_timer,
184 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
185 rt_b->rt_period_timer.function = sched_rt_period_timer;
188 static inline int rt_bandwidth_enabled(void)
190 return sysctl_sched_rt_runtime >= 0;
193 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
197 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
200 if (hrtimer_active(&rt_b->rt_period_timer))
203 spin_lock(&rt_b->rt_runtime_lock);
208 if (hrtimer_active(&rt_b->rt_period_timer))
211 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
212 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
214 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
215 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
216 delta = ktime_to_ns(ktime_sub(hard, soft));
217 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
218 HRTIMER_MODE_ABS_PINNED, 0);
220 spin_unlock(&rt_b->rt_runtime_lock);
223 #ifdef CONFIG_RT_GROUP_SCHED
224 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
226 hrtimer_cancel(&rt_b->rt_period_timer);
231 * sched_domains_mutex serializes calls to arch_init_sched_domains,
232 * detach_destroy_domains and partition_sched_domains.
234 static DEFINE_MUTEX(sched_domains_mutex);
236 #ifdef CONFIG_CGROUP_SCHED
238 #include <linux/cgroup.h>
242 static LIST_HEAD(task_groups);
244 /* task group related information */
246 struct cgroup_subsys_state css;
248 #ifdef CONFIG_FAIR_GROUP_SCHED
249 /* schedulable entities of this group on each cpu */
250 struct sched_entity **se;
251 /* runqueue "owned" by this group on each cpu */
252 struct cfs_rq **cfs_rq;
253 unsigned long shares;
256 #ifdef CONFIG_RT_GROUP_SCHED
257 struct sched_rt_entity **rt_se;
258 struct rt_rq **rt_rq;
260 struct rt_bandwidth rt_bandwidth;
264 struct list_head list;
266 struct task_group *parent;
267 struct list_head siblings;
268 struct list_head children;
271 #define root_task_group init_task_group
273 /* task_group_lock serializes add/remove of task groups and also changes to
274 * a task group's cpu shares.
276 static DEFINE_SPINLOCK(task_group_lock);
278 #ifdef CONFIG_FAIR_GROUP_SCHED
281 static int root_task_group_empty(void)
283 return list_empty(&root_task_group.children);
287 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
290 * A weight of 0 or 1 can cause arithmetics problems.
291 * A weight of a cfs_rq is the sum of weights of which entities
292 * are queued on this cfs_rq, so a weight of a entity should not be
293 * too large, so as the shares value of a task group.
294 * (The default weight is 1024 - so there's no practical
295 * limitation from this.)
298 #define MAX_SHARES (1UL << 18)
300 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
303 /* Default task group.
304 * Every task in system belong to this group at bootup.
306 struct task_group init_task_group;
308 /* return group to which a task belongs */
309 static inline struct task_group *task_group(struct task_struct *p)
311 struct task_group *tg;
313 #ifdef CONFIG_CGROUP_SCHED
314 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
315 struct task_group, css);
317 tg = &init_task_group;
322 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
323 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
325 #ifdef CONFIG_FAIR_GROUP_SCHED
326 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
327 p->se.parent = task_group(p)->se[cpu];
330 #ifdef CONFIG_RT_GROUP_SCHED
331 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
332 p->rt.parent = task_group(p)->rt_se[cpu];
338 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
339 static inline struct task_group *task_group(struct task_struct *p)
344 #endif /* CONFIG_CGROUP_SCHED */
346 /* CFS-related fields in a runqueue */
348 struct load_weight load;
349 unsigned long nr_running;
354 struct rb_root tasks_timeline;
355 struct rb_node *rb_leftmost;
357 struct list_head tasks;
358 struct list_head *balance_iterator;
361 * 'curr' points to currently running entity on this cfs_rq.
362 * It is set to NULL otherwise (i.e when none are currently running).
364 struct sched_entity *curr, *next, *last;
366 unsigned int nr_spread_over;
368 #ifdef CONFIG_FAIR_GROUP_SCHED
369 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
372 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
373 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
374 * (like users, containers etc.)
376 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
377 * list is used during load balance.
379 struct list_head leaf_cfs_rq_list;
380 struct task_group *tg; /* group that "owns" this runqueue */
384 * the part of load.weight contributed by tasks
386 unsigned long task_weight;
389 * h_load = weight * f(tg)
391 * Where f(tg) is the recursive weight fraction assigned to
394 unsigned long h_load;
397 * this cpu's part of tg->shares
399 unsigned long shares;
402 * load.weight at the time we set shares
404 unsigned long rq_weight;
409 /* Real-Time classes' related field in a runqueue: */
411 struct rt_prio_array active;
412 unsigned long rt_nr_running;
413 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
415 int curr; /* highest queued rt task prio */
417 int next; /* next highest */
422 unsigned long rt_nr_migratory;
423 unsigned long rt_nr_total;
425 struct plist_head pushable_tasks;
430 /* Nests inside the rq lock: */
431 spinlock_t rt_runtime_lock;
433 #ifdef CONFIG_RT_GROUP_SCHED
434 unsigned long rt_nr_boosted;
437 struct list_head leaf_rt_rq_list;
438 struct task_group *tg;
439 struct sched_rt_entity *rt_se;
446 * We add the notion of a root-domain which will be used to define per-domain
447 * variables. Each exclusive cpuset essentially defines an island domain by
448 * fully partitioning the member cpus from any other cpuset. Whenever a new
449 * exclusive cpuset is created, we also create and attach a new root-domain
456 cpumask_var_t online;
459 * The "RT overload" flag: it gets set if a CPU has more than
460 * one runnable RT task.
462 cpumask_var_t rto_mask;
465 struct cpupri cpupri;
470 * By default the system creates a single root-domain with all cpus as
471 * members (mimicking the global state we have today).
473 static struct root_domain def_root_domain;
478 * This is the main, per-CPU runqueue data structure.
480 * Locking rule: those places that want to lock multiple runqueues
481 * (such as the load balancing or the thread migration code), lock
482 * acquire operations must be ordered by ascending &runqueue.
489 * nr_running and cpu_load should be in the same cacheline because
490 * remote CPUs use both these fields when doing load calculation.
492 unsigned long nr_running;
493 #define CPU_LOAD_IDX_MAX 5
494 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
496 unsigned long last_tick_seen;
497 unsigned char in_nohz_recently;
499 /* capture load from *all* tasks on this cpu: */
500 struct load_weight load;
501 unsigned long nr_load_updates;
507 #ifdef CONFIG_FAIR_GROUP_SCHED
508 /* list of leaf cfs_rq on this cpu: */
509 struct list_head leaf_cfs_rq_list;
511 #ifdef CONFIG_RT_GROUP_SCHED
512 struct list_head leaf_rt_rq_list;
516 * This is part of a global counter where only the total sum
517 * over all CPUs matters. A task can increase this counter on
518 * one CPU and if it got migrated afterwards it may decrease
519 * it on another CPU. Always updated under the runqueue lock:
521 unsigned long nr_uninterruptible;
523 struct task_struct *curr, *idle;
524 unsigned long next_balance;
525 struct mm_struct *prev_mm;
533 struct root_domain *rd;
534 struct sched_domain *sd;
536 unsigned char idle_at_tick;
537 /* For active balancing */
541 /* cpu of this runqueue: */
545 unsigned long avg_load_per_task;
547 struct task_struct *migration_thread;
548 struct list_head migration_queue;
556 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
560 /* calc_load related fields */
561 unsigned long calc_load_update;
562 long calc_load_active;
564 #ifdef CONFIG_SCHED_HRTICK
566 int hrtick_csd_pending;
567 struct call_single_data hrtick_csd;
569 struct hrtimer hrtick_timer;
572 #ifdef CONFIG_SCHEDSTATS
574 struct sched_info rq_sched_info;
575 unsigned long long rq_cpu_time;
576 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
578 /* sys_sched_yield() stats */
579 unsigned int yld_count;
581 /* schedule() stats */
582 unsigned int sched_switch;
583 unsigned int sched_count;
584 unsigned int sched_goidle;
586 /* try_to_wake_up() stats */
587 unsigned int ttwu_count;
588 unsigned int ttwu_local;
591 unsigned int bkl_count;
595 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
597 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags);
599 static inline int cpu_of(struct rq *rq)
609 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
610 * See detach_destroy_domains: synchronize_sched for details.
612 * The domain tree of any CPU may only be accessed from within
613 * preempt-disabled sections.
615 #define for_each_domain(cpu, __sd) \
616 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
618 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
619 #define this_rq() (&__get_cpu_var(runqueues))
620 #define task_rq(p) cpu_rq(task_cpu(p))
621 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
622 #define raw_rq() (&__raw_get_cpu_var(runqueues))
624 static u64 irq_time_cpu(int cpu);
625 static void sched_irq_time_avg_update(struct rq *rq, u64 irq_time);
627 inline void update_rq_clock(struct rq *rq)
629 int cpu = cpu_of(rq);
632 rq->clock = sched_clock_cpu(cpu_of(rq));
633 irq_time = irq_time_cpu(cpu);
634 if (rq->clock - irq_time > rq->clock_task)
635 rq->clock_task = rq->clock - irq_time;
637 sched_irq_time_avg_update(rq, irq_time);
641 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
643 #ifdef CONFIG_SCHED_DEBUG
644 # define const_debug __read_mostly
646 # define const_debug static const
651 * @cpu: the processor in question.
653 * Returns true if the current cpu runqueue is locked.
654 * This interface allows printk to be called with the runqueue lock
655 * held and know whether or not it is OK to wake up the klogd.
657 int runqueue_is_locked(int cpu)
659 return spin_is_locked(&cpu_rq(cpu)->lock);
663 * Debugging: various feature bits
666 #define SCHED_FEAT(name, enabled) \
667 __SCHED_FEAT_##name ,
670 #include "sched_features.h"
675 #define SCHED_FEAT(name, enabled) \
676 (1UL << __SCHED_FEAT_##name) * enabled |
678 const_debug unsigned int sysctl_sched_features =
679 #include "sched_features.h"
684 #ifdef CONFIG_SCHED_DEBUG
685 #define SCHED_FEAT(name, enabled) \
688 static __read_mostly char *sched_feat_names[] = {
689 #include "sched_features.h"
695 static int sched_feat_show(struct seq_file *m, void *v)
699 for (i = 0; sched_feat_names[i]; i++) {
700 if (!(sysctl_sched_features & (1UL << i)))
702 seq_printf(m, "%s ", sched_feat_names[i]);
710 sched_feat_write(struct file *filp, const char __user *ubuf,
711 size_t cnt, loff_t *ppos)
721 if (copy_from_user(&buf, ubuf, cnt))
727 if (strncmp(buf, "NO_", 3) == 0) {
732 for (i = 0; sched_feat_names[i]; i++) {
733 if (strcmp(cmp, sched_feat_names[i]) == 0) {
735 sysctl_sched_features &= ~(1UL << i);
737 sysctl_sched_features |= (1UL << i);
742 if (!sched_feat_names[i])
750 static int sched_feat_open(struct inode *inode, struct file *filp)
752 return single_open(filp, sched_feat_show, NULL);
755 static const struct file_operations sched_feat_fops = {
756 .open = sched_feat_open,
757 .write = sched_feat_write,
760 .release = single_release,
763 static __init int sched_init_debug(void)
765 debugfs_create_file("sched_features", 0644, NULL, NULL,
770 late_initcall(sched_init_debug);
774 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
777 * Number of tasks to iterate in a single balance run.
778 * Limited because this is done with IRQs disabled.
780 const_debug unsigned int sysctl_sched_nr_migrate = 32;
783 * ratelimit for updating the group shares.
786 unsigned int sysctl_sched_shares_ratelimit = 250000;
787 unsigned int normalized_sysctl_sched_shares_ratelimit = 250000;
790 * Inject some fuzzyness into changing the per-cpu group shares
791 * this avoids remote rq-locks at the expense of fairness.
794 unsigned int sysctl_sched_shares_thresh = 4;
797 * period over which we average the RT time consumption, measured
802 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
805 * period over which we measure -rt task cpu usage in us.
808 unsigned int sysctl_sched_rt_period = 1000000;
810 static __read_mostly int scheduler_running;
813 * part of the period that we allow rt tasks to run in us.
816 int sysctl_sched_rt_runtime = 950000;
818 static inline u64 global_rt_period(void)
820 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
823 static inline u64 global_rt_runtime(void)
825 if (sysctl_sched_rt_runtime < 0)
828 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
831 #ifndef prepare_arch_switch
832 # define prepare_arch_switch(next) do { } while (0)
834 #ifndef finish_arch_switch
835 # define finish_arch_switch(prev) do { } while (0)
838 static inline int task_current(struct rq *rq, struct task_struct *p)
840 return rq->curr == p;
843 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
844 static inline int task_running(struct rq *rq, struct task_struct *p)
846 return task_current(rq, p);
849 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
853 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
855 #ifdef CONFIG_DEBUG_SPINLOCK
856 /* this is a valid case when another task releases the spinlock */
857 rq->lock.owner = current;
860 * If we are tracking spinlock dependencies then we have to
861 * fix up the runqueue lock - which gets 'carried over' from
864 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
866 spin_unlock_irq(&rq->lock);
869 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
870 static inline int task_running(struct rq *rq, struct task_struct *p)
875 return task_current(rq, p);
879 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
883 * We can optimise this out completely for !SMP, because the
884 * SMP rebalancing from interrupt is the only thing that cares
889 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
890 spin_unlock_irq(&rq->lock);
892 spin_unlock(&rq->lock);
896 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
900 * After ->oncpu is cleared, the task can be moved to a different CPU.
901 * We must ensure this doesn't happen until the switch is completely
907 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
911 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
914 * Check whether the task is waking, we use this to synchronize ->cpus_allowed
917 static inline int task_is_waking(struct task_struct *p)
919 return unlikely(p->state == TASK_WAKING);
923 * __task_rq_lock - lock the runqueue a given task resides on.
924 * Must be called interrupts disabled.
926 static inline struct rq *__task_rq_lock(struct task_struct *p)
933 spin_lock(&rq->lock);
934 if (likely(rq == task_rq(p)))
936 spin_unlock(&rq->lock);
941 * task_rq_lock - lock the runqueue a given task resides on and disable
942 * interrupts. Note the ordering: we can safely lookup the task_rq without
943 * explicitly disabling preemption.
945 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
951 local_irq_save(*flags);
953 spin_lock(&rq->lock);
954 if (likely(rq == task_rq(p)))
956 spin_unlock_irqrestore(&rq->lock, *flags);
960 void task_rq_unlock_wait(struct task_struct *p)
962 struct rq *rq = task_rq(p);
964 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
965 spin_unlock_wait(&rq->lock);
968 static void __task_rq_unlock(struct rq *rq)
971 spin_unlock(&rq->lock);
974 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
977 spin_unlock_irqrestore(&rq->lock, *flags);
981 * this_rq_lock - lock this runqueue and disable interrupts.
983 static struct rq *this_rq_lock(void)
990 spin_lock(&rq->lock);
995 #ifdef CONFIG_SCHED_HRTICK
997 * Use HR-timers to deliver accurate preemption points.
999 * Its all a bit involved since we cannot program an hrt while holding the
1000 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1003 * When we get rescheduled we reprogram the hrtick_timer outside of the
1009 * - enabled by features
1010 * - hrtimer is actually high res
1012 static inline int hrtick_enabled(struct rq *rq)
1014 if (!sched_feat(HRTICK))
1016 if (!cpu_active(cpu_of(rq)))
1018 return hrtimer_is_hres_active(&rq->hrtick_timer);
1021 static void hrtick_clear(struct rq *rq)
1023 if (hrtimer_active(&rq->hrtick_timer))
1024 hrtimer_cancel(&rq->hrtick_timer);
1028 * High-resolution timer tick.
1029 * Runs from hardirq context with interrupts disabled.
1031 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1033 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1035 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1037 spin_lock(&rq->lock);
1038 update_rq_clock(rq);
1039 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1040 spin_unlock(&rq->lock);
1042 return HRTIMER_NORESTART;
1047 * called from hardirq (IPI) context
1049 static void __hrtick_start(void *arg)
1051 struct rq *rq = arg;
1053 spin_lock(&rq->lock);
1054 hrtimer_restart(&rq->hrtick_timer);
1055 rq->hrtick_csd_pending = 0;
1056 spin_unlock(&rq->lock);
1060 * Called to set the hrtick timer state.
1062 * called with rq->lock held and irqs disabled
1064 static void hrtick_start(struct rq *rq, u64 delay)
1066 struct hrtimer *timer = &rq->hrtick_timer;
1067 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1069 hrtimer_set_expires(timer, time);
1071 if (rq == this_rq()) {
1072 hrtimer_restart(timer);
1073 } else if (!rq->hrtick_csd_pending) {
1074 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1075 rq->hrtick_csd_pending = 1;
1080 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1082 int cpu = (int)(long)hcpu;
1085 case CPU_UP_CANCELED:
1086 case CPU_UP_CANCELED_FROZEN:
1087 case CPU_DOWN_PREPARE:
1088 case CPU_DOWN_PREPARE_FROZEN:
1090 case CPU_DEAD_FROZEN:
1091 hrtick_clear(cpu_rq(cpu));
1098 static __init void init_hrtick(void)
1100 hotcpu_notifier(hotplug_hrtick, 0);
1104 * Called to set the hrtick timer state.
1106 * called with rq->lock held and irqs disabled
1108 static void hrtick_start(struct rq *rq, u64 delay)
1110 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1111 HRTIMER_MODE_REL_PINNED, 0);
1114 static inline void init_hrtick(void)
1117 #endif /* CONFIG_SMP */
1119 static void init_rq_hrtick(struct rq *rq)
1122 rq->hrtick_csd_pending = 0;
1124 rq->hrtick_csd.flags = 0;
1125 rq->hrtick_csd.func = __hrtick_start;
1126 rq->hrtick_csd.info = rq;
1129 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1130 rq->hrtick_timer.function = hrtick;
1132 #else /* CONFIG_SCHED_HRTICK */
1133 static inline void hrtick_clear(struct rq *rq)
1137 static inline void init_rq_hrtick(struct rq *rq)
1141 static inline void init_hrtick(void)
1144 #endif /* CONFIG_SCHED_HRTICK */
1147 * resched_task - mark a task 'to be rescheduled now'.
1149 * On UP this means the setting of the need_resched flag, on SMP it
1150 * might also involve a cross-CPU call to trigger the scheduler on
1155 #ifndef tsk_is_polling
1156 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1159 static void resched_task(struct task_struct *p)
1163 assert_spin_locked(&task_rq(p)->lock);
1165 if (test_tsk_need_resched(p))
1168 set_tsk_need_resched(p);
1171 if (cpu == smp_processor_id())
1174 /* NEED_RESCHED must be visible before we test polling */
1176 if (!tsk_is_polling(p))
1177 smp_send_reschedule(cpu);
1180 static void resched_cpu(int cpu)
1182 struct rq *rq = cpu_rq(cpu);
1183 unsigned long flags;
1185 if (!spin_trylock_irqsave(&rq->lock, flags))
1187 resched_task(cpu_curr(cpu));
1188 spin_unlock_irqrestore(&rq->lock, flags);
1193 * When add_timer_on() enqueues a timer into the timer wheel of an
1194 * idle CPU then this timer might expire before the next timer event
1195 * which is scheduled to wake up that CPU. In case of a completely
1196 * idle system the next event might even be infinite time into the
1197 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1198 * leaves the inner idle loop so the newly added timer is taken into
1199 * account when the CPU goes back to idle and evaluates the timer
1200 * wheel for the next timer event.
1202 void wake_up_idle_cpu(int cpu)
1204 struct rq *rq = cpu_rq(cpu);
1206 if (cpu == smp_processor_id())
1210 * This is safe, as this function is called with the timer
1211 * wheel base lock of (cpu) held. When the CPU is on the way
1212 * to idle and has not yet set rq->curr to idle then it will
1213 * be serialized on the timer wheel base lock and take the new
1214 * timer into account automatically.
1216 if (rq->curr != rq->idle)
1220 * We can set TIF_RESCHED on the idle task of the other CPU
1221 * lockless. The worst case is that the other CPU runs the
1222 * idle task through an additional NOOP schedule()
1224 set_tsk_need_resched(rq->idle);
1226 /* NEED_RESCHED must be visible before we test polling */
1228 if (!tsk_is_polling(rq->idle))
1229 smp_send_reschedule(cpu);
1231 #endif /* CONFIG_NO_HZ */
1233 static u64 sched_avg_period(void)
1235 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1238 static void sched_avg_update(struct rq *rq)
1240 s64 period = sched_avg_period();
1242 while ((s64)(rq->clock - rq->age_stamp) > period) {
1244 * Inline assembly required to prevent the compiler
1245 * optimising this loop into a divmod call.
1246 * See __iter_div_u64_rem() for another example of this.
1248 asm("" : "+rm" (rq->age_stamp));
1249 rq->age_stamp += period;
1254 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1256 rq->rt_avg += rt_delta;
1257 sched_avg_update(rq);
1260 #else /* !CONFIG_SMP */
1261 static void resched_task(struct task_struct *p)
1263 assert_spin_locked(&task_rq(p)->lock);
1264 set_tsk_need_resched(p);
1267 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1271 static void sched_avg_update(struct rq *rq)
1274 #endif /* CONFIG_SMP */
1276 #if BITS_PER_LONG == 32
1277 # define WMULT_CONST (~0UL)
1279 # define WMULT_CONST (1UL << 32)
1282 #define WMULT_SHIFT 32
1285 * Shift right and round:
1287 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1290 * delta *= weight / lw
1292 static unsigned long
1293 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1294 struct load_weight *lw)
1298 if (!lw->inv_weight) {
1299 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1302 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1306 tmp = (u64)delta_exec * weight;
1308 * Check whether we'd overflow the 64-bit multiplication:
1310 if (unlikely(tmp > WMULT_CONST))
1311 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1314 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1316 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1319 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1325 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1332 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1333 * of tasks with abnormal "nice" values across CPUs the contribution that
1334 * each task makes to its run queue's load is weighted according to its
1335 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1336 * scaled version of the new time slice allocation that they receive on time
1340 #define WEIGHT_IDLEPRIO 3
1341 #define WMULT_IDLEPRIO 1431655765
1344 * Nice levels are multiplicative, with a gentle 10% change for every
1345 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1346 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1347 * that remained on nice 0.
1349 * The "10% effect" is relative and cumulative: from _any_ nice level,
1350 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1351 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1352 * If a task goes up by ~10% and another task goes down by ~10% then
1353 * the relative distance between them is ~25%.)
1355 static const int prio_to_weight[40] = {
1356 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1357 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1358 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1359 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1360 /* 0 */ 1024, 820, 655, 526, 423,
1361 /* 5 */ 335, 272, 215, 172, 137,
1362 /* 10 */ 110, 87, 70, 56, 45,
1363 /* 15 */ 36, 29, 23, 18, 15,
1367 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1369 * In cases where the weight does not change often, we can use the
1370 * precalculated inverse to speed up arithmetics by turning divisions
1371 * into multiplications:
1373 static const u32 prio_to_wmult[40] = {
1374 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1375 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1376 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1377 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1378 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1379 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1380 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1381 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1384 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1387 * runqueue iterator, to support SMP load-balancing between different
1388 * scheduling classes, without having to expose their internal data
1389 * structures to the load-balancing proper:
1391 struct rq_iterator {
1393 struct task_struct *(*start)(void *);
1394 struct task_struct *(*next)(void *);
1398 static unsigned long
1399 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1400 unsigned long max_load_move, struct sched_domain *sd,
1401 enum cpu_idle_type idle, int *all_pinned,
1402 int *this_best_prio, struct rq_iterator *iterator);
1405 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1406 struct sched_domain *sd, enum cpu_idle_type idle,
1407 struct rq_iterator *iterator);
1410 /* Time spent by the tasks of the cpu accounting group executing in ... */
1411 enum cpuacct_stat_index {
1412 CPUACCT_STAT_USER, /* ... user mode */
1413 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1415 CPUACCT_STAT_NSTATS,
1418 #ifdef CONFIG_CGROUP_CPUACCT
1419 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1420 static void cpuacct_update_stats(struct task_struct *tsk,
1421 enum cpuacct_stat_index idx, cputime_t val);
1423 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1424 static inline void cpuacct_update_stats(struct task_struct *tsk,
1425 enum cpuacct_stat_index idx, cputime_t val) {}
1428 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1430 update_load_add(&rq->load, load);
1433 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1435 update_load_sub(&rq->load, load);
1438 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1439 typedef int (*tg_visitor)(struct task_group *, void *);
1442 * Iterate the full tree, calling @down when first entering a node and @up when
1443 * leaving it for the final time.
1445 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1447 struct task_group *parent, *child;
1451 parent = &root_task_group;
1453 ret = (*down)(parent, data);
1456 list_for_each_entry_rcu(child, &parent->children, siblings) {
1463 ret = (*up)(parent, data);
1468 parent = parent->parent;
1477 static int tg_nop(struct task_group *tg, void *data)
1484 /* Used instead of source_load when we know the type == 0 */
1485 static unsigned long weighted_cpuload(const int cpu)
1487 return cpu_rq(cpu)->load.weight;
1491 * Return a low guess at the load of a migration-source cpu weighted
1492 * according to the scheduling class and "nice" value.
1494 * We want to under-estimate the load of migration sources, to
1495 * balance conservatively.
1497 static unsigned long source_load(int cpu, int type)
1499 struct rq *rq = cpu_rq(cpu);
1500 unsigned long total = weighted_cpuload(cpu);
1502 if (type == 0 || !sched_feat(LB_BIAS))
1505 return min(rq->cpu_load[type-1], total);
1509 * Return a high guess at the load of a migration-target cpu weighted
1510 * according to the scheduling class and "nice" value.
1512 static unsigned long target_load(int cpu, int type)
1514 struct rq *rq = cpu_rq(cpu);
1515 unsigned long total = weighted_cpuload(cpu);
1517 if (type == 0 || !sched_feat(LB_BIAS))
1520 return max(rq->cpu_load[type-1], total);
1523 static struct sched_group *group_of(int cpu)
1525 struct sched_domain *sd = rcu_dereference(cpu_rq(cpu)->sd);
1533 static unsigned long power_of(int cpu)
1535 struct sched_group *group = group_of(cpu);
1538 return SCHED_LOAD_SCALE;
1540 return group->cpu_power;
1543 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1545 static unsigned long cpu_avg_load_per_task(int cpu)
1547 struct rq *rq = cpu_rq(cpu);
1548 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1551 rq->avg_load_per_task = rq->load.weight / nr_running;
1553 rq->avg_load_per_task = 0;
1555 return rq->avg_load_per_task;
1558 #ifdef CONFIG_FAIR_GROUP_SCHED
1560 static __read_mostly unsigned long *update_shares_data;
1562 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1565 * Calculate and set the cpu's group shares.
1567 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1568 unsigned long sd_shares,
1569 unsigned long sd_rq_weight,
1570 unsigned long *usd_rq_weight)
1572 unsigned long shares, rq_weight;
1575 rq_weight = usd_rq_weight[cpu];
1578 rq_weight = NICE_0_LOAD;
1582 * \Sum_j shares_j * rq_weight_i
1583 * shares_i = -----------------------------
1584 * \Sum_j rq_weight_j
1586 shares = (sd_shares * rq_weight) / sd_rq_weight;
1587 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1589 if (abs(shares - tg->se[cpu]->load.weight) >
1590 sysctl_sched_shares_thresh) {
1591 struct rq *rq = cpu_rq(cpu);
1592 unsigned long flags;
1594 spin_lock_irqsave(&rq->lock, flags);
1595 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1596 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1597 __set_se_shares(tg->se[cpu], shares);
1598 spin_unlock_irqrestore(&rq->lock, flags);
1603 * Re-compute the task group their per cpu shares over the given domain.
1604 * This needs to be done in a bottom-up fashion because the rq weight of a
1605 * parent group depends on the shares of its child groups.
1607 static int tg_shares_up(struct task_group *tg, void *data)
1609 unsigned long weight, rq_weight = 0, sum_weight = 0, shares = 0;
1610 unsigned long *usd_rq_weight;
1611 struct sched_domain *sd = data;
1612 unsigned long flags;
1618 local_irq_save(flags);
1619 usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id());
1621 for_each_cpu(i, sched_domain_span(sd)) {
1622 weight = tg->cfs_rq[i]->load.weight;
1623 usd_rq_weight[i] = weight;
1625 rq_weight += weight;
1627 * If there are currently no tasks on the cpu pretend there
1628 * is one of average load so that when a new task gets to
1629 * run here it will not get delayed by group starvation.
1632 weight = NICE_0_LOAD;
1634 sum_weight += weight;
1635 shares += tg->cfs_rq[i]->shares;
1639 rq_weight = sum_weight;
1641 if ((!shares && rq_weight) || shares > tg->shares)
1642 shares = tg->shares;
1644 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1645 shares = tg->shares;
1647 for_each_cpu(i, sched_domain_span(sd))
1648 update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight);
1650 local_irq_restore(flags);
1656 * Compute the cpu's hierarchical load factor for each task group.
1657 * This needs to be done in a top-down fashion because the load of a child
1658 * group is a fraction of its parents load.
1660 static int tg_load_down(struct task_group *tg, void *data)
1663 long cpu = (long)data;
1666 load = cpu_rq(cpu)->load.weight;
1668 load = tg->parent->cfs_rq[cpu]->h_load;
1669 load *= tg->cfs_rq[cpu]->shares;
1670 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1673 tg->cfs_rq[cpu]->h_load = load;
1678 static void update_shares(struct sched_domain *sd)
1683 if (root_task_group_empty())
1686 now = cpu_clock(raw_smp_processor_id());
1687 elapsed = now - sd->last_update;
1689 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1690 sd->last_update = now;
1691 walk_tg_tree(tg_nop, tg_shares_up, sd);
1695 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1697 if (root_task_group_empty())
1700 spin_unlock(&rq->lock);
1702 spin_lock(&rq->lock);
1705 static void update_h_load(long cpu)
1707 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1712 static inline void update_shares(struct sched_domain *sd)
1716 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1722 #ifdef CONFIG_PREEMPT
1724 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1727 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1728 * way at the expense of forcing extra atomic operations in all
1729 * invocations. This assures that the double_lock is acquired using the
1730 * same underlying policy as the spinlock_t on this architecture, which
1731 * reduces latency compared to the unfair variant below. However, it
1732 * also adds more overhead and therefore may reduce throughput.
1734 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1735 __releases(this_rq->lock)
1736 __acquires(busiest->lock)
1737 __acquires(this_rq->lock)
1739 spin_unlock(&this_rq->lock);
1740 double_rq_lock(this_rq, busiest);
1747 * Unfair double_lock_balance: Optimizes throughput at the expense of
1748 * latency by eliminating extra atomic operations when the locks are
1749 * already in proper order on entry. This favors lower cpu-ids and will
1750 * grant the double lock to lower cpus over higher ids under contention,
1751 * regardless of entry order into the function.
1753 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1754 __releases(this_rq->lock)
1755 __acquires(busiest->lock)
1756 __acquires(this_rq->lock)
1760 if (unlikely(!spin_trylock(&busiest->lock))) {
1761 if (busiest < this_rq) {
1762 spin_unlock(&this_rq->lock);
1763 spin_lock(&busiest->lock);
1764 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
1767 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
1772 #endif /* CONFIG_PREEMPT */
1775 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1777 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1779 if (unlikely(!irqs_disabled())) {
1780 /* printk() doesn't work good under rq->lock */
1781 spin_unlock(&this_rq->lock);
1785 return _double_lock_balance(this_rq, busiest);
1788 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1789 __releases(busiest->lock)
1791 spin_unlock(&busiest->lock);
1792 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1796 #ifdef CONFIG_FAIR_GROUP_SCHED
1797 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1800 cfs_rq->shares = shares;
1805 static void calc_load_account_active(struct rq *this_rq);
1806 static void update_sysctl(void);
1808 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1810 set_task_rq(p, cpu);
1813 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1814 * successfuly executed on another CPU. We must ensure that updates of
1815 * per-task data have been completed by this moment.
1818 task_thread_info(p)->cpu = cpu;
1822 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1825 * There are no locks covering percpu hardirq/softirq time.
1826 * They are only modified in account_system_vtime, on corresponding CPU
1827 * with interrupts disabled. So, writes are safe.
1828 * They are read and saved off onto struct rq in update_rq_clock().
1829 * This may result in other CPU reading this CPU's irq time and can
1830 * race with irq/account_system_vtime on this CPU. We would either get old
1831 * or new value (or semi updated value on 32 bit) with a side effect of
1832 * accounting a slice of irq time to wrong task when irq is in progress
1833 * while we read rq->clock. That is a worthy compromise in place of having
1834 * locks on each irq in account_system_time.
1836 static DEFINE_PER_CPU(u64, cpu_hardirq_time);
1837 static DEFINE_PER_CPU(u64, cpu_softirq_time);
1839 static DEFINE_PER_CPU(u64, irq_start_time);
1840 static int sched_clock_irqtime;
1842 void enable_sched_clock_irqtime(void)
1844 sched_clock_irqtime = 1;
1847 void disable_sched_clock_irqtime(void)
1849 sched_clock_irqtime = 0;
1852 static u64 irq_time_cpu(int cpu)
1854 if (!sched_clock_irqtime)
1857 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
1860 void account_system_vtime(struct task_struct *curr)
1862 unsigned long flags;
1866 if (!sched_clock_irqtime)
1869 local_irq_save(flags);
1871 cpu = smp_processor_id();
1872 now = sched_clock_cpu(cpu);
1873 delta = now - per_cpu(irq_start_time, cpu);
1874 per_cpu(irq_start_time, cpu) = now;
1876 * We do not account for softirq time from ksoftirqd here.
1877 * We want to continue accounting softirq time to ksoftirqd thread
1878 * in that case, so as not to confuse scheduler with a special task
1879 * that do not consume any time, but still wants to run.
1881 if (hardirq_count())
1882 per_cpu(cpu_hardirq_time, cpu) += delta;
1883 else if (in_serving_softirq() && !(curr->flags & PF_KSOFTIRQD))
1884 per_cpu(cpu_softirq_time, cpu) += delta;
1886 local_irq_restore(flags);
1888 EXPORT_SYMBOL_GPL(account_system_vtime);
1890 static void sched_irq_time_avg_update(struct rq *rq, u64 curr_irq_time)
1892 if (sched_clock_irqtime && sched_feat(NONIRQ_POWER)) {
1893 u64 delta_irq = curr_irq_time - rq->prev_irq_time;
1894 rq->prev_irq_time = curr_irq_time;
1895 sched_rt_avg_update(rq, delta_irq);
1901 static u64 irq_time_cpu(int cpu)
1906 static void sched_irq_time_avg_update(struct rq *rq, u64 curr_irq_time) { }
1910 #include "sched_stats.h"
1911 #include "sched_idletask.c"
1912 #include "sched_fair.c"
1913 #include "sched_rt.c"
1914 #ifdef CONFIG_SCHED_DEBUG
1915 # include "sched_debug.c"
1918 #define sched_class_highest (&rt_sched_class)
1919 #define for_each_class(class) \
1920 for (class = sched_class_highest; class; class = class->next)
1922 static void inc_nr_running(struct rq *rq)
1927 static void dec_nr_running(struct rq *rq)
1932 static void set_load_weight(struct task_struct *p)
1934 if (task_has_rt_policy(p)) {
1935 p->se.load.weight = prio_to_weight[0] * 2;
1936 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1941 * SCHED_IDLE tasks get minimal weight:
1943 if (p->policy == SCHED_IDLE) {
1944 p->se.load.weight = WEIGHT_IDLEPRIO;
1945 p->se.load.inv_weight = WMULT_IDLEPRIO;
1949 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1950 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1953 static void update_avg(u64 *avg, u64 sample)
1955 s64 diff = sample - *avg;
1960 enqueue_task(struct rq *rq, struct task_struct *p, int wakeup, bool head)
1963 p->se.start_runtime = p->se.sum_exec_runtime;
1965 sched_info_queued(p);
1966 p->sched_class->enqueue_task(rq, p, wakeup, head);
1970 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1973 if (p->se.last_wakeup) {
1974 update_avg(&p->se.avg_overlap,
1975 p->se.sum_exec_runtime - p->se.last_wakeup);
1976 p->se.last_wakeup = 0;
1978 update_avg(&p->se.avg_wakeup,
1979 sysctl_sched_wakeup_granularity);
1983 sched_info_dequeued(p);
1984 p->sched_class->dequeue_task(rq, p, sleep);
1989 * __normal_prio - return the priority that is based on the static prio
1991 static inline int __normal_prio(struct task_struct *p)
1993 return p->static_prio;
1997 * Calculate the expected normal priority: i.e. priority
1998 * without taking RT-inheritance into account. Might be
1999 * boosted by interactivity modifiers. Changes upon fork,
2000 * setprio syscalls, and whenever the interactivity
2001 * estimator recalculates.
2003 static inline int normal_prio(struct task_struct *p)
2007 if (task_has_rt_policy(p))
2008 prio = MAX_RT_PRIO-1 - p->rt_priority;
2010 prio = __normal_prio(p);
2015 * Calculate the current priority, i.e. the priority
2016 * taken into account by the scheduler. This value might
2017 * be boosted by RT tasks, or might be boosted by
2018 * interactivity modifiers. Will be RT if the task got
2019 * RT-boosted. If not then it returns p->normal_prio.
2021 static int effective_prio(struct task_struct *p)
2023 p->normal_prio = normal_prio(p);
2025 * If we are RT tasks or we were boosted to RT priority,
2026 * keep the priority unchanged. Otherwise, update priority
2027 * to the normal priority:
2029 if (!rt_prio(p->prio))
2030 return p->normal_prio;
2035 * activate_task - move a task to the runqueue.
2037 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
2039 if (task_contributes_to_load(p))
2040 rq->nr_uninterruptible--;
2042 enqueue_task(rq, p, wakeup, false);
2047 * deactivate_task - remove a task from the runqueue.
2049 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
2051 if (task_contributes_to_load(p))
2052 rq->nr_uninterruptible++;
2054 dequeue_task(rq, p, sleep);
2059 * task_curr - is this task currently executing on a CPU?
2060 * @p: the task in question.
2062 inline int task_curr(const struct task_struct *p)
2064 return cpu_curr(task_cpu(p)) == p;
2067 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2068 const struct sched_class *prev_class,
2069 int oldprio, int running)
2071 if (prev_class != p->sched_class) {
2072 if (prev_class->switched_from)
2073 prev_class->switched_from(rq, p, running);
2074 p->sched_class->switched_to(rq, p, running);
2076 p->sched_class->prio_changed(rq, p, oldprio, running);
2080 * kthread_bind - bind a just-created kthread to a cpu.
2081 * @p: thread created by kthread_create().
2082 * @cpu: cpu (might not be online, must be possible) for @k to run on.
2084 * Description: This function is equivalent to set_cpus_allowed(),
2085 * except that @cpu doesn't need to be online, and the thread must be
2086 * stopped (i.e., just returned from kthread_create()).
2088 * Function lives here instead of kthread.c because it messes with
2089 * scheduler internals which require locking.
2091 void kthread_bind(struct task_struct *p, unsigned int cpu)
2093 /* Must have done schedule() in kthread() before we set_task_cpu */
2094 if (!wait_task_inactive(p, TASK_UNINTERRUPTIBLE)) {
2099 p->cpus_allowed = cpumask_of_cpu(cpu);
2100 p->rt.nr_cpus_allowed = 1;
2101 p->flags |= PF_THREAD_BOUND;
2103 EXPORT_SYMBOL(kthread_bind);
2107 * Is this task likely cache-hot:
2110 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2114 if (p->sched_class != &fair_sched_class)
2117 if (unlikely(p->policy == SCHED_IDLE))
2121 * Buddy candidates are cache hot:
2123 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2124 (&p->se == cfs_rq_of(&p->se)->next ||
2125 &p->se == cfs_rq_of(&p->se)->last))
2128 if (sysctl_sched_migration_cost == -1)
2130 if (sysctl_sched_migration_cost == 0)
2133 delta = now - p->se.exec_start;
2135 return delta < (s64)sysctl_sched_migration_cost;
2139 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2141 int old_cpu = task_cpu(p);
2143 #ifdef CONFIG_SCHED_DEBUG
2145 * We should never call set_task_cpu() on a blocked task,
2146 * ttwu() will sort out the placement.
2148 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2149 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2152 trace_sched_migrate_task(p, new_cpu);
2154 if (old_cpu != new_cpu) {
2155 p->se.nr_migrations++;
2156 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS,
2160 __set_task_cpu(p, new_cpu);
2163 struct migration_req {
2164 struct list_head list;
2166 struct task_struct *task;
2169 struct completion done;
2173 * The task's runqueue lock must be held.
2174 * Returns true if you have to wait for migration thread.
2177 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2179 struct rq *rq = task_rq(p);
2182 * If the task is not on a runqueue (and not running), then
2183 * the next wake-up will properly place the task.
2185 if (!p->se.on_rq && !task_running(rq, p))
2188 init_completion(&req->done);
2190 req->dest_cpu = dest_cpu;
2191 list_add(&req->list, &rq->migration_queue);
2197 * wait_task_context_switch - wait for a thread to complete at least one
2200 * @p must not be current.
2202 void wait_task_context_switch(struct task_struct *p)
2204 unsigned long nvcsw, nivcsw, flags;
2212 * The runqueue is assigned before the actual context
2213 * switch. We need to take the runqueue lock.
2215 * We could check initially without the lock but it is
2216 * very likely that we need to take the lock in every
2219 rq = task_rq_lock(p, &flags);
2220 running = task_running(rq, p);
2221 task_rq_unlock(rq, &flags);
2223 if (likely(!running))
2226 * The switch count is incremented before the actual
2227 * context switch. We thus wait for two switches to be
2228 * sure at least one completed.
2230 if ((p->nvcsw - nvcsw) > 1)
2232 if ((p->nivcsw - nivcsw) > 1)
2240 * wait_task_inactive - wait for a thread to unschedule.
2242 * If @match_state is nonzero, it's the @p->state value just checked and
2243 * not expected to change. If it changes, i.e. @p might have woken up,
2244 * then return zero. When we succeed in waiting for @p to be off its CPU,
2245 * we return a positive number (its total switch count). If a second call
2246 * a short while later returns the same number, the caller can be sure that
2247 * @p has remained unscheduled the whole time.
2249 * The caller must ensure that the task *will* unschedule sometime soon,
2250 * else this function might spin for a *long* time. This function can't
2251 * be called with interrupts off, or it may introduce deadlock with
2252 * smp_call_function() if an IPI is sent by the same process we are
2253 * waiting to become inactive.
2255 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2257 unsigned long flags;
2264 * We do the initial early heuristics without holding
2265 * any task-queue locks at all. We'll only try to get
2266 * the runqueue lock when things look like they will
2272 * If the task is actively running on another CPU
2273 * still, just relax and busy-wait without holding
2276 * NOTE! Since we don't hold any locks, it's not
2277 * even sure that "rq" stays as the right runqueue!
2278 * But we don't care, since "task_running()" will
2279 * return false if the runqueue has changed and p
2280 * is actually now running somewhere else!
2282 while (task_running(rq, p)) {
2283 if (match_state && unlikely(p->state != match_state))
2289 * Ok, time to look more closely! We need the rq
2290 * lock now, to be *sure*. If we're wrong, we'll
2291 * just go back and repeat.
2293 rq = task_rq_lock(p, &flags);
2294 trace_sched_wait_task(rq, p);
2295 running = task_running(rq, p);
2296 on_rq = p->se.on_rq;
2298 if (!match_state || p->state == match_state)
2299 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2300 task_rq_unlock(rq, &flags);
2303 * If it changed from the expected state, bail out now.
2305 if (unlikely(!ncsw))
2309 * Was it really running after all now that we
2310 * checked with the proper locks actually held?
2312 * Oops. Go back and try again..
2314 if (unlikely(running)) {
2320 * It's not enough that it's not actively running,
2321 * it must be off the runqueue _entirely_, and not
2324 * So if it was still runnable (but just not actively
2325 * running right now), it's preempted, and we should
2326 * yield - it could be a while.
2328 if (unlikely(on_rq)) {
2329 schedule_timeout_uninterruptible(1);
2334 * Ahh, all good. It wasn't running, and it wasn't
2335 * runnable, which means that it will never become
2336 * running in the future either. We're all done!
2345 * kick_process - kick a running thread to enter/exit the kernel
2346 * @p: the to-be-kicked thread
2348 * Cause a process which is running on another CPU to enter
2349 * kernel-mode, without any delay. (to get signals handled.)
2351 * NOTE: this function doesnt have to take the runqueue lock,
2352 * because all it wants to ensure is that the remote task enters
2353 * the kernel. If the IPI races and the task has been migrated
2354 * to another CPU then no harm is done and the purpose has been
2357 void kick_process(struct task_struct *p)
2363 if ((cpu != smp_processor_id()) && task_curr(p))
2364 smp_send_reschedule(cpu);
2367 EXPORT_SYMBOL_GPL(kick_process);
2368 #endif /* CONFIG_SMP */
2371 * task_oncpu_function_call - call a function on the cpu on which a task runs
2372 * @p: the task to evaluate
2373 * @func: the function to be called
2374 * @info: the function call argument
2376 * Calls the function @func when the task is currently running. This might
2377 * be on the current CPU, which just calls the function directly
2379 void task_oncpu_function_call(struct task_struct *p,
2380 void (*func) (void *info), void *info)
2387 smp_call_function_single(cpu, func, info, 1);
2391 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2393 const struct sched_class *class;
2395 if (p->sched_class == rq->curr->sched_class) {
2396 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2398 for_each_class(class) {
2399 if (class == rq->curr->sched_class)
2401 if (class == p->sched_class) {
2402 resched_task(rq->curr);
2411 * ->cpus_allowed is protected by either TASK_WAKING or rq->lock held.
2413 static int select_fallback_rq(int cpu, struct task_struct *p)
2416 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2418 /* Look for allowed, online CPU in same node. */
2419 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2420 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2423 /* Any allowed, online CPU? */
2424 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2425 if (dest_cpu < nr_cpu_ids)
2428 /* No more Mr. Nice Guy. */
2429 if (unlikely(dest_cpu >= nr_cpu_ids)) {
2430 dest_cpu = cpuset_cpus_allowed_fallback(p);
2432 * Don't tell them about moving exiting tasks or
2433 * kernel threads (both mm NULL), since they never
2436 if (p->mm && printk_ratelimit()) {
2437 printk(KERN_INFO "process %d (%s) no "
2438 "longer affine to cpu%d\n",
2439 task_pid_nr(p), p->comm, cpu);
2447 * The caller (fork, wakeup) owns TASK_WAKING, ->cpus_allowed is stable.
2450 int select_task_rq(struct rq *rq, struct task_struct *p, int sd_flags, int wake_flags)
2452 int cpu = p->sched_class->select_task_rq(rq, p, sd_flags, wake_flags);
2455 * In order not to call set_task_cpu() on a blocking task we need
2456 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2459 * Since this is common to all placement strategies, this lives here.
2461 * [ this allows ->select_task() to simply return task_cpu(p) and
2462 * not worry about this generic constraint ]
2464 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2466 cpu = select_fallback_rq(task_cpu(p), p);
2473 * try_to_wake_up - wake up a thread
2474 * @p: the to-be-woken-up thread
2475 * @state: the mask of task states that can be woken
2476 * @sync: do a synchronous wakeup?
2478 * Put it on the run-queue if it's not already there. The "current"
2479 * thread is always on the run-queue (except when the actual
2480 * re-schedule is in progress), and as such you're allowed to do
2481 * the simpler "current->state = TASK_RUNNING" to mark yourself
2482 * runnable without the overhead of this.
2484 * returns failure only if the task is already active.
2486 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2489 int cpu, orig_cpu, this_cpu, success = 0;
2490 unsigned long flags;
2491 struct rq *rq, *orig_rq;
2493 if (!sched_feat(SYNC_WAKEUPS))
2494 wake_flags &= ~WF_SYNC;
2496 this_cpu = get_cpu();
2499 rq = orig_rq = task_rq_lock(p, &flags);
2500 update_rq_clock(rq);
2501 if (!(p->state & state))
2511 if (unlikely(task_running(rq, p)))
2515 * In order to handle concurrent wakeups and release the rq->lock
2516 * we put the task in TASK_WAKING state.
2518 * First fix up the nr_uninterruptible count:
2520 if (task_contributes_to_load(p)) {
2521 if (likely(cpu_online(orig_cpu)))
2522 rq->nr_uninterruptible--;
2524 this_rq()->nr_uninterruptible--;
2526 p->state = TASK_WAKING;
2528 if (p->sched_class->task_waking)
2529 p->sched_class->task_waking(rq, p);
2531 cpu = select_task_rq(rq, p, SD_BALANCE_WAKE, wake_flags);
2532 if (cpu != orig_cpu)
2533 set_task_cpu(p, cpu);
2534 __task_rq_unlock(rq);
2537 spin_lock(&rq->lock);
2538 update_rq_clock(rq);
2541 * We migrated the task without holding either rq->lock, however
2542 * since the task is not on the task list itself, nobody else
2543 * will try and migrate the task, hence the rq should match the
2544 * cpu we just moved it to.
2546 WARN_ON(task_cpu(p) != cpu);
2547 WARN_ON(p->state != TASK_WAKING);
2549 #ifdef CONFIG_SCHEDSTATS
2550 schedstat_inc(rq, ttwu_count);
2551 if (cpu == this_cpu)
2552 schedstat_inc(rq, ttwu_local);
2554 struct sched_domain *sd;
2555 for_each_domain(this_cpu, sd) {
2556 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2557 schedstat_inc(sd, ttwu_wake_remote);
2562 #endif /* CONFIG_SCHEDSTATS */
2565 #endif /* CONFIG_SMP */
2566 schedstat_inc(p, se.nr_wakeups);
2567 if (wake_flags & WF_SYNC)
2568 schedstat_inc(p, se.nr_wakeups_sync);
2569 if (orig_cpu != cpu)
2570 schedstat_inc(p, se.nr_wakeups_migrate);
2571 if (cpu == this_cpu)
2572 schedstat_inc(p, se.nr_wakeups_local);
2574 schedstat_inc(p, se.nr_wakeups_remote);
2575 activate_task(rq, p, 1);
2579 * Only attribute actual wakeups done by this task.
2581 if (!in_interrupt()) {
2582 struct sched_entity *se = ¤t->se;
2583 u64 sample = se->sum_exec_runtime;
2585 if (se->last_wakeup)
2586 sample -= se->last_wakeup;
2588 sample -= se->start_runtime;
2589 update_avg(&se->avg_wakeup, sample);
2591 se->last_wakeup = se->sum_exec_runtime;
2595 trace_sched_wakeup(rq, p, success);
2596 check_preempt_curr(rq, p, wake_flags);
2598 p->state = TASK_RUNNING;
2600 if (p->sched_class->task_woken)
2601 p->sched_class->task_woken(rq, p);
2603 if (unlikely(rq->idle_stamp)) {
2604 u64 delta = rq->clock - rq->idle_stamp;
2605 u64 max = 2*sysctl_sched_migration_cost;
2610 update_avg(&rq->avg_idle, delta);
2615 task_rq_unlock(rq, &flags);
2622 * wake_up_process - Wake up a specific process
2623 * @p: The process to be woken up.
2625 * Attempt to wake up the nominated process and move it to the set of runnable
2626 * processes. Returns 1 if the process was woken up, 0 if it was already
2629 * It may be assumed that this function implies a write memory barrier before
2630 * changing the task state if and only if any tasks are woken up.
2632 int wake_up_process(struct task_struct *p)
2634 return try_to_wake_up(p, TASK_ALL, 0);
2636 EXPORT_SYMBOL(wake_up_process);
2638 int wake_up_state(struct task_struct *p, unsigned int state)
2640 return try_to_wake_up(p, state, 0);
2644 * Perform scheduler related setup for a newly forked process p.
2645 * p is forked by current.
2647 * __sched_fork() is basic setup used by init_idle() too:
2649 static void __sched_fork(struct task_struct *p)
2651 p->se.exec_start = 0;
2652 p->se.sum_exec_runtime = 0;
2653 p->se.prev_sum_exec_runtime = 0;
2654 p->se.nr_migrations = 0;
2655 p->se.last_wakeup = 0;
2656 p->se.avg_overlap = 0;
2657 p->se.start_runtime = 0;
2658 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2659 p->se.avg_running = 0;
2661 #ifdef CONFIG_SCHEDSTATS
2662 p->se.wait_start = 0;
2664 p->se.wait_count = 0;
2667 p->se.sleep_start = 0;
2668 p->se.sleep_max = 0;
2669 p->se.sum_sleep_runtime = 0;
2671 p->se.block_start = 0;
2672 p->se.block_max = 0;
2674 p->se.slice_max = 0;
2676 p->se.nr_migrations_cold = 0;
2677 p->se.nr_failed_migrations_affine = 0;
2678 p->se.nr_failed_migrations_running = 0;
2679 p->se.nr_failed_migrations_hot = 0;
2680 p->se.nr_forced_migrations = 0;
2682 p->se.nr_wakeups = 0;
2683 p->se.nr_wakeups_sync = 0;
2684 p->se.nr_wakeups_migrate = 0;
2685 p->se.nr_wakeups_local = 0;
2686 p->se.nr_wakeups_remote = 0;
2687 p->se.nr_wakeups_affine = 0;
2688 p->se.nr_wakeups_affine_attempts = 0;
2689 p->se.nr_wakeups_passive = 0;
2690 p->se.nr_wakeups_idle = 0;
2694 INIT_LIST_HEAD(&p->rt.run_list);
2696 INIT_LIST_HEAD(&p->se.group_node);
2698 #ifdef CONFIG_PREEMPT_NOTIFIERS
2699 INIT_HLIST_HEAD(&p->preempt_notifiers);
2704 * fork()/clone()-time setup:
2706 void sched_fork(struct task_struct *p, int clone_flags)
2708 int cpu = get_cpu();
2712 * We mark the process as running here. This guarantees that
2713 * nobody will actually run it, and a signal or other external
2714 * event cannot wake it up and insert it on the runqueue either.
2716 p->state = TASK_RUNNING;
2719 * Revert to default priority/policy on fork if requested.
2721 if (unlikely(p->sched_reset_on_fork)) {
2722 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2723 p->policy = SCHED_NORMAL;
2724 p->normal_prio = p->static_prio;
2727 if (PRIO_TO_NICE(p->static_prio) < 0) {
2728 p->static_prio = NICE_TO_PRIO(0);
2729 p->normal_prio = p->static_prio;
2734 * We don't need the reset flag anymore after the fork. It has
2735 * fulfilled its duty:
2737 p->sched_reset_on_fork = 0;
2741 * Make sure we do not leak PI boosting priority to the child.
2743 p->prio = current->normal_prio;
2745 if (!rt_prio(p->prio))
2746 p->sched_class = &fair_sched_class;
2748 if (p->sched_class->task_fork)
2749 p->sched_class->task_fork(p);
2751 set_task_cpu(p, cpu);
2753 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2754 if (likely(sched_info_on()))
2755 memset(&p->sched_info, 0, sizeof(p->sched_info));
2757 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2760 #ifdef CONFIG_PREEMPT
2761 /* Want to start with kernel preemption disabled. */
2762 task_thread_info(p)->preempt_count = 1;
2764 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2770 * wake_up_new_task - wake up a newly created task for the first time.
2772 * This function will do some initial scheduler statistics housekeeping
2773 * that must be done for every newly created context, then puts the task
2774 * on the runqueue and wakes it.
2776 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2778 unsigned long flags;
2780 int cpu = get_cpu();
2783 rq = task_rq_lock(p, &flags);
2784 p->state = TASK_WAKING;
2787 * Fork balancing, do it here and not earlier because:
2788 * - cpus_allowed can change in the fork path
2789 * - any previously selected cpu might disappear through hotplug
2791 * We set TASK_WAKING so that select_task_rq() can drop rq->lock
2792 * without people poking at ->cpus_allowed.
2794 cpu = select_task_rq(rq, p, SD_BALANCE_FORK, 0);
2795 set_task_cpu(p, cpu);
2797 p->state = TASK_RUNNING;
2798 task_rq_unlock(rq, &flags);
2801 rq = task_rq_lock(p, &flags);
2802 update_rq_clock(rq);
2803 activate_task(rq, p, 0);
2804 trace_sched_wakeup_new(rq, p, 1);
2805 check_preempt_curr(rq, p, WF_FORK);
2807 if (p->sched_class->task_woken)
2808 p->sched_class->task_woken(rq, p);
2810 task_rq_unlock(rq, &flags);
2814 #ifdef CONFIG_PREEMPT_NOTIFIERS
2817 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2818 * @notifier: notifier struct to register
2820 void preempt_notifier_register(struct preempt_notifier *notifier)
2822 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2824 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2827 * preempt_notifier_unregister - no longer interested in preemption notifications
2828 * @notifier: notifier struct to unregister
2830 * This is safe to call from within a preemption notifier.
2832 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2834 hlist_del(¬ifier->link);
2836 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2838 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2840 struct preempt_notifier *notifier;
2841 struct hlist_node *node;
2843 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2844 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2848 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2849 struct task_struct *next)
2851 struct preempt_notifier *notifier;
2852 struct hlist_node *node;
2854 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2855 notifier->ops->sched_out(notifier, next);
2858 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2860 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2865 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2866 struct task_struct *next)
2870 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2873 * prepare_task_switch - prepare to switch tasks
2874 * @rq: the runqueue preparing to switch
2875 * @prev: the current task that is being switched out
2876 * @next: the task we are going to switch to.
2878 * This is called with the rq lock held and interrupts off. It must
2879 * be paired with a subsequent finish_task_switch after the context
2882 * prepare_task_switch sets up locking and calls architecture specific
2886 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2887 struct task_struct *next)
2889 fire_sched_out_preempt_notifiers(prev, next);
2890 prepare_lock_switch(rq, next);
2891 prepare_arch_switch(next);
2895 * finish_task_switch - clean up after a task-switch
2896 * @rq: runqueue associated with task-switch
2897 * @prev: the thread we just switched away from.
2899 * finish_task_switch must be called after the context switch, paired
2900 * with a prepare_task_switch call before the context switch.
2901 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2902 * and do any other architecture-specific cleanup actions.
2904 * Note that we may have delayed dropping an mm in context_switch(). If
2905 * so, we finish that here outside of the runqueue lock. (Doing it
2906 * with the lock held can cause deadlocks; see schedule() for
2909 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2910 __releases(rq->lock)
2912 struct mm_struct *mm = rq->prev_mm;
2918 * A task struct has one reference for the use as "current".
2919 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2920 * schedule one last time. The schedule call will never return, and
2921 * the scheduled task must drop that reference.
2922 * The test for TASK_DEAD must occur while the runqueue locks are
2923 * still held, otherwise prev could be scheduled on another cpu, die
2924 * there before we look at prev->state, and then the reference would
2926 * Manfred Spraul <manfred@colorfullife.com>
2928 prev_state = prev->state;
2929 finish_arch_switch(prev);
2930 perf_event_task_sched_in(current, cpu_of(rq));
2931 finish_lock_switch(rq, prev);
2933 fire_sched_in_preempt_notifiers(current);
2936 if (unlikely(prev_state == TASK_DEAD)) {
2938 * Remove function-return probe instances associated with this
2939 * task and put them back on the free list.
2941 kprobe_flush_task(prev);
2942 put_task_struct(prev);
2948 /* assumes rq->lock is held */
2949 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2951 if (prev->sched_class->pre_schedule)
2952 prev->sched_class->pre_schedule(rq, prev);
2955 /* rq->lock is NOT held, but preemption is disabled */
2956 static inline void post_schedule(struct rq *rq)
2958 if (rq->post_schedule) {
2959 unsigned long flags;
2961 spin_lock_irqsave(&rq->lock, flags);
2962 if (rq->curr->sched_class->post_schedule)
2963 rq->curr->sched_class->post_schedule(rq);
2964 spin_unlock_irqrestore(&rq->lock, flags);
2966 rq->post_schedule = 0;
2972 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2976 static inline void post_schedule(struct rq *rq)
2983 * schedule_tail - first thing a freshly forked thread must call.
2984 * @prev: the thread we just switched away from.
2986 asmlinkage void schedule_tail(struct task_struct *prev)
2987 __releases(rq->lock)
2989 struct rq *rq = this_rq();
2991 finish_task_switch(rq, prev);
2994 * FIXME: do we need to worry about rq being invalidated by the
2999 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
3000 /* In this case, finish_task_switch does not reenable preemption */
3003 if (current->set_child_tid)
3004 put_user(task_pid_vnr(current), current->set_child_tid);
3008 * context_switch - switch to the new MM and the new
3009 * thread's register state.
3012 context_switch(struct rq *rq, struct task_struct *prev,
3013 struct task_struct *next)
3015 struct mm_struct *mm, *oldmm;
3017 prepare_task_switch(rq, prev, next);
3018 trace_sched_switch(rq, prev, next);
3020 oldmm = prev->active_mm;
3022 * For paravirt, this is coupled with an exit in switch_to to
3023 * combine the page table reload and the switch backend into
3026 arch_start_context_switch(prev);
3028 if (unlikely(!mm)) {
3029 next->active_mm = oldmm;
3030 atomic_inc(&oldmm->mm_count);
3031 enter_lazy_tlb(oldmm, next);
3033 switch_mm(oldmm, mm, next);
3035 if (unlikely(!prev->mm)) {
3036 prev->active_mm = NULL;
3037 rq->prev_mm = oldmm;
3040 * Since the runqueue lock will be released by the next
3041 * task (which is an invalid locking op but in the case
3042 * of the scheduler it's an obvious special-case), so we
3043 * do an early lockdep release here:
3045 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
3046 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3049 /* Here we just switch the register state and the stack. */
3050 switch_to(prev, next, prev);
3054 * this_rq must be evaluated again because prev may have moved
3055 * CPUs since it called schedule(), thus the 'rq' on its stack
3056 * frame will be invalid.
3058 finish_task_switch(this_rq(), prev);
3062 * nr_running, nr_uninterruptible and nr_context_switches:
3064 * externally visible scheduler statistics: current number of runnable
3065 * threads, current number of uninterruptible-sleeping threads, total
3066 * number of context switches performed since bootup.
3068 unsigned long nr_running(void)
3070 unsigned long i, sum = 0;
3072 for_each_online_cpu(i)
3073 sum += cpu_rq(i)->nr_running;
3078 unsigned long nr_uninterruptible(void)
3080 unsigned long i, sum = 0;
3082 for_each_possible_cpu(i)
3083 sum += cpu_rq(i)->nr_uninterruptible;
3086 * Since we read the counters lockless, it might be slightly
3087 * inaccurate. Do not allow it to go below zero though:
3089 if (unlikely((long)sum < 0))
3095 unsigned long long nr_context_switches(void)
3098 unsigned long long sum = 0;
3100 for_each_possible_cpu(i)
3101 sum += cpu_rq(i)->nr_switches;
3106 unsigned long nr_iowait(void)
3108 unsigned long i, sum = 0;
3110 for_each_possible_cpu(i)
3111 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3116 unsigned long nr_iowait_cpu(void)
3118 struct rq *this = this_rq();
3119 return atomic_read(&this->nr_iowait);
3122 unsigned long this_cpu_load(void)
3124 struct rq *this = this_rq();
3125 return this->cpu_load[0];
3129 /* Variables and functions for calc_load */
3130 static atomic_long_t calc_load_tasks;
3131 static unsigned long calc_load_update;
3132 unsigned long avenrun[3];
3133 EXPORT_SYMBOL(avenrun);
3136 * get_avenrun - get the load average array
3137 * @loads: pointer to dest load array
3138 * @offset: offset to add
3139 * @shift: shift count to shift the result left
3141 * These values are estimates at best, so no need for locking.
3143 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3145 loads[0] = (avenrun[0] + offset) << shift;
3146 loads[1] = (avenrun[1] + offset) << shift;
3147 loads[2] = (avenrun[2] + offset) << shift;
3150 static unsigned long
3151 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3154 load += active * (FIXED_1 - exp);
3155 return load >> FSHIFT;
3159 * calc_load - update the avenrun load estimates 10 ticks after the
3160 * CPUs have updated calc_load_tasks.
3162 void calc_global_load(void)
3164 unsigned long upd = calc_load_update + 10;
3167 if (time_before(jiffies, upd))
3170 active = atomic_long_read(&calc_load_tasks);
3171 active = active > 0 ? active * FIXED_1 : 0;
3173 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3174 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3175 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3177 calc_load_update += LOAD_FREQ;
3181 * Either called from update_cpu_load() or from a cpu going idle
3183 static void calc_load_account_active(struct rq *this_rq)
3185 long nr_active, delta;
3187 nr_active = this_rq->nr_running;
3188 nr_active += (long) this_rq->nr_uninterruptible;
3190 if (nr_active != this_rq->calc_load_active) {
3191 delta = nr_active - this_rq->calc_load_active;
3192 this_rq->calc_load_active = nr_active;
3193 atomic_long_add(delta, &calc_load_tasks);
3198 * Update rq->cpu_load[] statistics. This function is usually called every
3199 * scheduler tick (TICK_NSEC).
3201 static void update_cpu_load(struct rq *this_rq)
3203 unsigned long this_load = this_rq->load.weight;
3206 this_rq->nr_load_updates++;
3208 /* Update our load: */
3209 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3210 unsigned long old_load, new_load;
3212 /* scale is effectively 1 << i now, and >> i divides by scale */
3214 old_load = this_rq->cpu_load[i];
3215 new_load = this_load;
3217 * Round up the averaging division if load is increasing. This
3218 * prevents us from getting stuck on 9 if the load is 10, for
3221 if (new_load > old_load)
3222 new_load += scale-1;
3223 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3226 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3227 this_rq->calc_load_update += LOAD_FREQ;
3228 calc_load_account_active(this_rq);
3231 sched_avg_update(this_rq);
3237 * double_rq_lock - safely lock two runqueues
3239 * Note this does not disable interrupts like task_rq_lock,
3240 * you need to do so manually before calling.
3242 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
3243 __acquires(rq1->lock)
3244 __acquires(rq2->lock)
3246 BUG_ON(!irqs_disabled());
3248 spin_lock(&rq1->lock);
3249 __acquire(rq2->lock); /* Fake it out ;) */
3252 spin_lock(&rq1->lock);
3253 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
3255 spin_lock(&rq2->lock);
3256 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
3259 update_rq_clock(rq1);
3260 update_rq_clock(rq2);
3264 * double_rq_unlock - safely unlock two runqueues
3266 * Note this does not restore interrupts like task_rq_unlock,
3267 * you need to do so manually after calling.
3269 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
3270 __releases(rq1->lock)
3271 __releases(rq2->lock)
3273 spin_unlock(&rq1->lock);
3275 spin_unlock(&rq2->lock);
3277 __release(rq2->lock);
3281 * sched_exec - execve() is a valuable balancing opportunity, because at
3282 * this point the task has the smallest effective memory and cache footprint.
3284 void sched_exec(void)
3286 struct task_struct *p = current;
3287 struct migration_req req;
3288 unsigned long flags;
3292 rq = task_rq_lock(p, &flags);
3293 dest_cpu = p->sched_class->select_task_rq(rq, p, SD_BALANCE_EXEC, 0);
3294 if (dest_cpu == smp_processor_id())
3298 * select_task_rq() can race against ->cpus_allowed
3300 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed) &&
3301 likely(cpu_active(dest_cpu)) &&
3302 migrate_task(p, dest_cpu, &req)) {
3303 /* Need to wait for migration thread (might exit: take ref). */
3304 struct task_struct *mt = rq->migration_thread;
3306 get_task_struct(mt);
3307 task_rq_unlock(rq, &flags);
3308 wake_up_process(mt);
3309 put_task_struct(mt);
3310 wait_for_completion(&req.done);
3315 task_rq_unlock(rq, &flags);
3319 * pull_task - move a task from a remote runqueue to the local runqueue.
3320 * Both runqueues must be locked.
3322 static void pull_task(struct rq *src_rq, struct task_struct *p,
3323 struct rq *this_rq, int this_cpu)
3325 deactivate_task(src_rq, p, 0);
3326 set_task_cpu(p, this_cpu);
3327 activate_task(this_rq, p, 0);
3328 check_preempt_curr(this_rq, p, 0);
3332 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3335 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3336 struct sched_domain *sd, enum cpu_idle_type idle,
3339 int tsk_cache_hot = 0;
3341 * We do not migrate tasks that are:
3342 * 1) running (obviously), or
3343 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3344 * 3) are cache-hot on their current CPU.
3346 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
3347 schedstat_inc(p, se.nr_failed_migrations_affine);
3352 if (task_running(rq, p)) {
3353 schedstat_inc(p, se.nr_failed_migrations_running);
3358 * Aggressive migration if:
3359 * 1) task is cache cold, or
3360 * 2) too many balance attempts have failed.
3363 tsk_cache_hot = task_hot(p, rq->clock_task, sd);
3364 if (!tsk_cache_hot ||
3365 sd->nr_balance_failed > sd->cache_nice_tries) {
3366 #ifdef CONFIG_SCHEDSTATS
3367 if (tsk_cache_hot) {
3368 schedstat_inc(sd, lb_hot_gained[idle]);
3369 schedstat_inc(p, se.nr_forced_migrations);
3375 if (tsk_cache_hot) {
3376 schedstat_inc(p, se.nr_failed_migrations_hot);
3382 static unsigned long
3383 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3384 unsigned long max_load_move, struct sched_domain *sd,
3385 enum cpu_idle_type idle, int *all_pinned,
3386 int *this_best_prio, struct rq_iterator *iterator)
3388 int loops = 0, pulled = 0, pinned = 0;
3389 struct task_struct *p;
3390 long rem_load_move = max_load_move;
3392 if (max_load_move == 0)
3398 * Start the load-balancing iterator:
3400 p = iterator->start(iterator->arg);
3402 if (!p || loops++ > sysctl_sched_nr_migrate)
3405 if ((p->se.load.weight >> 1) > rem_load_move ||
3406 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3407 p = iterator->next(iterator->arg);
3411 pull_task(busiest, p, this_rq, this_cpu);
3413 rem_load_move -= p->se.load.weight;
3415 #ifdef CONFIG_PREEMPT
3417 * NEWIDLE balancing is a source of latency, so preemptible kernels
3418 * will stop after the first task is pulled to minimize the critical
3421 if (idle == CPU_NEWLY_IDLE)
3426 * We only want to steal up to the prescribed amount of weighted load.
3428 if (rem_load_move > 0) {
3429 if (p->prio < *this_best_prio)
3430 *this_best_prio = p->prio;
3431 p = iterator->next(iterator->arg);
3436 * Right now, this is one of only two places pull_task() is called,
3437 * so we can safely collect pull_task() stats here rather than
3438 * inside pull_task().
3440 schedstat_add(sd, lb_gained[idle], pulled);
3443 *all_pinned = pinned;
3445 return max_load_move - rem_load_move;
3449 * move_tasks tries to move up to max_load_move weighted load from busiest to
3450 * this_rq, as part of a balancing operation within domain "sd".
3451 * Returns 1 if successful and 0 otherwise.
3453 * Called with both runqueues locked.
3455 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3456 unsigned long max_load_move,
3457 struct sched_domain *sd, enum cpu_idle_type idle,
3460 const struct sched_class *class = sched_class_highest;
3461 unsigned long total_load_moved = 0;
3462 int this_best_prio = this_rq->curr->prio;
3466 class->load_balance(this_rq, this_cpu, busiest,
3467 max_load_move - total_load_moved,
3468 sd, idle, all_pinned, &this_best_prio);
3469 class = class->next;
3471 #ifdef CONFIG_PREEMPT
3473 * NEWIDLE balancing is a source of latency, so preemptible
3474 * kernels will stop after the first task is pulled to minimize
3475 * the critical section.
3477 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3480 } while (class && max_load_move > total_load_moved);
3482 return total_load_moved > 0;
3486 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3487 struct sched_domain *sd, enum cpu_idle_type idle,
3488 struct rq_iterator *iterator)
3490 struct task_struct *p = iterator->start(iterator->arg);
3494 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3495 pull_task(busiest, p, this_rq, this_cpu);
3497 * Right now, this is only the second place pull_task()
3498 * is called, so we can safely collect pull_task()
3499 * stats here rather than inside pull_task().
3501 schedstat_inc(sd, lb_gained[idle]);
3505 p = iterator->next(iterator->arg);
3512 * move_one_task tries to move exactly one task from busiest to this_rq, as
3513 * part of active balancing operations within "domain".
3514 * Returns 1 if successful and 0 otherwise.
3516 * Called with both runqueues locked.
3518 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3519 struct sched_domain *sd, enum cpu_idle_type idle)
3521 const struct sched_class *class;
3523 for_each_class(class) {
3524 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3530 /********** Helpers for find_busiest_group ************************/
3532 * sd_lb_stats - Structure to store the statistics of a sched_domain
3533 * during load balancing.
3535 struct sd_lb_stats {
3536 struct sched_group *busiest; /* Busiest group in this sd */
3537 struct sched_group *this; /* Local group in this sd */
3538 unsigned long total_load; /* Total load of all groups in sd */
3539 unsigned long total_pwr; /* Total power of all groups in sd */
3540 unsigned long avg_load; /* Average load across all groups in sd */
3542 /** Statistics of this group */
3543 unsigned long this_load;
3544 unsigned long this_load_per_task;
3545 unsigned long this_nr_running;
3546 unsigned long this_has_capacity;
3547 unsigned int this_idle_cpus;
3549 /* Statistics of the busiest group */
3550 unsigned int busiest_idle_cpus;
3551 unsigned long max_load;
3552 unsigned long busiest_load_per_task;
3553 unsigned long busiest_nr_running;
3554 unsigned long busiest_group_capacity;
3555 unsigned long busiest_has_capacity;
3556 unsigned int busiest_group_weight;
3558 int group_imb; /* Is there imbalance in this sd */
3559 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3560 int power_savings_balance; /* Is powersave balance needed for this sd */
3561 struct sched_group *group_min; /* Least loaded group in sd */
3562 struct sched_group *group_leader; /* Group which relieves group_min */
3563 unsigned long min_load_per_task; /* load_per_task in group_min */
3564 unsigned long leader_nr_running; /* Nr running of group_leader */
3565 unsigned long min_nr_running; /* Nr running of group_min */
3570 * sg_lb_stats - stats of a sched_group required for load_balancing
3572 struct sg_lb_stats {
3573 unsigned long avg_load; /*Avg load across the CPUs of the group */
3574 unsigned long group_load; /* Total load over the CPUs of the group */
3575 unsigned long sum_nr_running; /* Nr tasks running in the group */
3576 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3577 unsigned long group_capacity;
3578 unsigned long idle_cpus;
3579 unsigned long group_weight;
3580 int group_imb; /* Is there an imbalance in the group ? */
3581 int group_has_capacity; /* Is there extra capacity in the group? */
3585 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3586 * @group: The group whose first cpu is to be returned.
3588 static inline unsigned int group_first_cpu(struct sched_group *group)
3590 return cpumask_first(sched_group_cpus(group));
3594 * get_sd_load_idx - Obtain the load index for a given sched domain.
3595 * @sd: The sched_domain whose load_idx is to be obtained.
3596 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3598 static inline int get_sd_load_idx(struct sched_domain *sd,
3599 enum cpu_idle_type idle)
3605 load_idx = sd->busy_idx;
3608 case CPU_NEWLY_IDLE:
3609 load_idx = sd->newidle_idx;
3612 load_idx = sd->idle_idx;
3620 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3622 * init_sd_power_savings_stats - Initialize power savings statistics for
3623 * the given sched_domain, during load balancing.
3625 * @sd: Sched domain whose power-savings statistics are to be initialized.
3626 * @sds: Variable containing the statistics for sd.
3627 * @idle: Idle status of the CPU at which we're performing load-balancing.
3629 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3630 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3633 * Busy processors will not participate in power savings
3636 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3637 sds->power_savings_balance = 0;
3639 sds->power_savings_balance = 1;
3640 sds->min_nr_running = ULONG_MAX;
3641 sds->leader_nr_running = 0;
3646 * update_sd_power_savings_stats - Update the power saving stats for a
3647 * sched_domain while performing load balancing.
3649 * @group: sched_group belonging to the sched_domain under consideration.
3650 * @sds: Variable containing the statistics of the sched_domain
3651 * @local_group: Does group contain the CPU for which we're performing
3653 * @sgs: Variable containing the statistics of the group.
3655 static inline void update_sd_power_savings_stats(struct sched_group *group,
3656 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3659 if (!sds->power_savings_balance)
3663 * If the local group is idle or completely loaded
3664 * no need to do power savings balance at this domain
3666 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3667 !sds->this_nr_running))
3668 sds->power_savings_balance = 0;
3671 * If a group is already running at full capacity or idle,
3672 * don't include that group in power savings calculations
3674 if (!sds->power_savings_balance ||
3675 sgs->sum_nr_running >= sgs->group_capacity ||
3676 !sgs->sum_nr_running)
3680 * Calculate the group which has the least non-idle load.
3681 * This is the group from where we need to pick up the load
3684 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3685 (sgs->sum_nr_running == sds->min_nr_running &&
3686 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3687 sds->group_min = group;
3688 sds->min_nr_running = sgs->sum_nr_running;
3689 sds->min_load_per_task = sgs->sum_weighted_load /
3690 sgs->sum_nr_running;
3694 * Calculate the group which is almost near its
3695 * capacity but still has some space to pick up some load
3696 * from other group and save more power
3698 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
3701 if (sgs->sum_nr_running > sds->leader_nr_running ||
3702 (sgs->sum_nr_running == sds->leader_nr_running &&
3703 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3704 sds->group_leader = group;
3705 sds->leader_nr_running = sgs->sum_nr_running;
3710 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3711 * @sds: Variable containing the statistics of the sched_domain
3712 * under consideration.
3713 * @this_cpu: Cpu at which we're currently performing load-balancing.
3714 * @imbalance: Variable to store the imbalance.
3717 * Check if we have potential to perform some power-savings balance.
3718 * If yes, set the busiest group to be the least loaded group in the
3719 * sched_domain, so that it's CPUs can be put to idle.
3721 * Returns 1 if there is potential to perform power-savings balance.
3724 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3725 int this_cpu, unsigned long *imbalance)
3727 if (!sds->power_savings_balance)
3730 if (sds->this != sds->group_leader ||
3731 sds->group_leader == sds->group_min)
3734 *imbalance = sds->min_load_per_task;
3735 sds->busiest = sds->group_min;
3740 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3741 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3742 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3747 static inline void update_sd_power_savings_stats(struct sched_group *group,
3748 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3753 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3754 int this_cpu, unsigned long *imbalance)
3758 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3761 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
3763 return SCHED_LOAD_SCALE;
3766 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
3768 return default_scale_freq_power(sd, cpu);
3771 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
3773 unsigned long weight = sd->span_weight;
3774 unsigned long smt_gain = sd->smt_gain;
3781 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3783 return default_scale_smt_power(sd, cpu);
3786 unsigned long scale_rt_power(int cpu)
3788 struct rq *rq = cpu_rq(cpu);
3789 u64 total, available;
3791 total = sched_avg_period() + (rq->clock - rq->age_stamp);
3793 if (unlikely(total < rq->rt_avg)) {
3794 /* Ensures that power won't end up being negative */
3797 available = total - rq->rt_avg;
3800 if (unlikely((s64)total < SCHED_LOAD_SCALE))
3801 total = SCHED_LOAD_SCALE;
3803 total >>= SCHED_LOAD_SHIFT;
3805 return div_u64(available, total);
3808 static void update_cpu_power(struct sched_domain *sd, int cpu)
3810 unsigned long weight = sd->span_weight;
3811 unsigned long power = SCHED_LOAD_SCALE;
3812 struct sched_group *sdg = sd->groups;
3814 if (sched_feat(ARCH_POWER))
3815 power *= arch_scale_freq_power(sd, cpu);
3817 power *= default_scale_freq_power(sd, cpu);
3819 power >>= SCHED_LOAD_SHIFT;
3821 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
3822 if (sched_feat(ARCH_POWER))
3823 power *= arch_scale_smt_power(sd, cpu);
3825 power *= default_scale_smt_power(sd, cpu);
3827 power >>= SCHED_LOAD_SHIFT;
3830 power *= scale_rt_power(cpu);
3831 power >>= SCHED_LOAD_SHIFT;
3836 sdg->cpu_power = power;
3839 static void update_group_power(struct sched_domain *sd, int cpu)
3841 struct sched_domain *child = sd->child;
3842 struct sched_group *group, *sdg = sd->groups;
3843 unsigned long power;
3846 update_cpu_power(sd, cpu);
3852 group = child->groups;
3854 power += group->cpu_power;
3855 group = group->next;
3856 } while (group != child->groups);
3858 sdg->cpu_power = power;
3862 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3863 * @sd: The sched_domain whose statistics are to be updated.
3864 * @group: sched_group whose statistics are to be updated.
3865 * @this_cpu: Cpu for which load balance is currently performed.
3866 * @idle: Idle status of this_cpu
3867 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3868 * @sd_idle: Idle status of the sched_domain containing group.
3869 * @local_group: Does group contain this_cpu.
3870 * @cpus: Set of cpus considered for load balancing.
3871 * @balance: Should we balance.
3872 * @sgs: variable to hold the statistics for this group.
3874 static inline void update_sg_lb_stats(struct sched_domain *sd,
3875 struct sched_group *group, int this_cpu,
3876 enum cpu_idle_type idle, int load_idx, int *sd_idle,
3877 int local_group, const struct cpumask *cpus,
3878 int *balance, struct sg_lb_stats *sgs)
3880 unsigned long load, max_cpu_load, min_cpu_load, max_nr_running;
3882 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3883 unsigned long avg_load_per_task = 0;
3886 balance_cpu = group_first_cpu(group);
3887 if (balance_cpu == this_cpu)
3888 update_group_power(sd, this_cpu);
3891 /* Tally up the load of all CPUs in the group */
3893 min_cpu_load = ~0UL;
3896 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3897 struct rq *rq = cpu_rq(i);
3899 if (*sd_idle && rq->nr_running)
3902 /* Bias balancing toward cpus of our domain */
3904 if (idle_cpu(i) && !first_idle_cpu) {
3909 load = target_load(i, load_idx);
3911 load = source_load(i, load_idx);
3912 if (load > max_cpu_load) {
3913 max_cpu_load = load;
3914 max_nr_running = rq->nr_running;
3916 if (min_cpu_load > load)
3917 min_cpu_load = load;
3920 sgs->group_load += load;
3921 sgs->sum_nr_running += rq->nr_running;
3922 sgs->sum_weighted_load += weighted_cpuload(i);
3928 * First idle cpu or the first cpu(busiest) in this sched group
3929 * is eligible for doing load balancing at this and above
3930 * domains. In the newly idle case, we will allow all the cpu's
3931 * to do the newly idle load balance.
3933 if (idle != CPU_NEWLY_IDLE && local_group &&
3934 balance_cpu != this_cpu && balance) {
3939 /* Adjust by relative CPU power of the group */
3940 sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power;
3943 * Consider the group unbalanced when the imbalance is larger
3944 * than the average weight of two tasks.
3946 * APZ: with cgroup the avg task weight can vary wildly and
3947 * might not be a suitable number - should we keep a
3948 * normalized nr_running number somewhere that negates
3951 if (sgs->sum_nr_running)
3952 avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
3954 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task && max_nr_running > 1)
3957 sgs->group_capacity = DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
3958 sgs->group_weight = group->group_weight;
3960 if (sgs->group_capacity > sgs->sum_nr_running)
3961 sgs->group_has_capacity = 1;
3965 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3966 * @sd: sched_domain whose statistics are to be updated.
3967 * @this_cpu: Cpu for which load balance is currently performed.
3968 * @idle: Idle status of this_cpu
3969 * @sd_idle: Idle status of the sched_domain containing group.
3970 * @cpus: Set of cpus considered for load balancing.
3971 * @balance: Should we balance.
3972 * @sds: variable to hold the statistics for this sched_domain.
3974 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3975 enum cpu_idle_type idle, int *sd_idle,
3976 const struct cpumask *cpus, int *balance,
3977 struct sd_lb_stats *sds)
3979 struct sched_domain *child = sd->child;
3980 struct sched_group *group = sd->groups;
3981 struct sg_lb_stats sgs;
3982 int load_idx, prefer_sibling = 0;
3984 if (child && child->flags & SD_PREFER_SIBLING)
3987 init_sd_power_savings_stats(sd, sds, idle);
3988 load_idx = get_sd_load_idx(sd, idle);
3993 local_group = cpumask_test_cpu(this_cpu,
3994 sched_group_cpus(group));
3995 memset(&sgs, 0, sizeof(sgs));
3996 update_sg_lb_stats(sd, group, this_cpu, idle, load_idx, sd_idle,
3997 local_group, cpus, balance, &sgs);
3999 if (local_group && balance && !(*balance))
4002 sds->total_load += sgs.group_load;
4003 sds->total_pwr += group->cpu_power;
4006 * In case the child domain prefers tasks go to siblings
4007 * first, lower the group capacity to one so that we'll try
4008 * and move all the excess tasks away. We lower the capacity
4009 * of a group only if the local group has the capacity to fit
4010 * these excess tasks, i.e. nr_running < group_capacity. The
4011 * extra check prevents the case where you always pull from the
4012 * heaviest group when it is already under-utilized (possible
4013 * with a large weight task outweighs the tasks on the system).
4015 if (prefer_sibling && !local_group && sds->this_has_capacity)
4016 sgs.group_capacity = min(sgs.group_capacity, 1UL);
4019 sds->this_load = sgs.avg_load;
4021 sds->this_nr_running = sgs.sum_nr_running;
4022 sds->this_load_per_task = sgs.sum_weighted_load;
4023 sds->this_has_capacity = sgs.group_has_capacity;
4024 sds->this_idle_cpus = sgs.idle_cpus;
4025 } else if (sgs.avg_load > sds->max_load &&
4026 (sgs.sum_nr_running > sgs.group_capacity ||
4028 sds->max_load = sgs.avg_load;
4029 sds->busiest = group;
4030 sds->busiest_nr_running = sgs.sum_nr_running;
4031 sds->busiest_idle_cpus = sgs.idle_cpus;
4032 sds->busiest_group_capacity = sgs.group_capacity;
4033 sds->busiest_group_weight = sgs.group_weight;
4034 sds->busiest_load_per_task = sgs.sum_weighted_load;
4035 sds->busiest_has_capacity = sgs.group_has_capacity;
4036 sds->group_imb = sgs.group_imb;
4039 update_sd_power_savings_stats(group, sds, local_group, &sgs);
4040 group = group->next;
4041 } while (group != sd->groups);
4045 * fix_small_imbalance - Calculate the minor imbalance that exists
4046 * amongst the groups of a sched_domain, during
4048 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
4049 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
4050 * @imbalance: Variable to store the imbalance.
4052 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
4053 int this_cpu, unsigned long *imbalance)
4055 unsigned long tmp, pwr_now = 0, pwr_move = 0;
4056 unsigned int imbn = 2;
4057 unsigned long scaled_busy_load_per_task;
4059 if (sds->this_nr_running) {
4060 sds->this_load_per_task /= sds->this_nr_running;
4061 if (sds->busiest_load_per_task >
4062 sds->this_load_per_task)
4065 sds->this_load_per_task =
4066 cpu_avg_load_per_task(this_cpu);
4068 scaled_busy_load_per_task = sds->busiest_load_per_task
4070 scaled_busy_load_per_task /= sds->busiest->cpu_power;
4072 if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
4073 (scaled_busy_load_per_task * imbn)) {
4074 *imbalance = sds->busiest_load_per_task;
4079 * OK, we don't have enough imbalance to justify moving tasks,
4080 * however we may be able to increase total CPU power used by
4084 pwr_now += sds->busiest->cpu_power *
4085 min(sds->busiest_load_per_task, sds->max_load);
4086 pwr_now += sds->this->cpu_power *
4087 min(sds->this_load_per_task, sds->this_load);
4088 pwr_now /= SCHED_LOAD_SCALE;
4090 /* Amount of load we'd subtract */
4091 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
4092 sds->busiest->cpu_power;
4093 if (sds->max_load > tmp)
4094 pwr_move += sds->busiest->cpu_power *
4095 min(sds->busiest_load_per_task, sds->max_load - tmp);
4097 /* Amount of load we'd add */
4098 if (sds->max_load * sds->busiest->cpu_power <
4099 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
4100 tmp = (sds->max_load * sds->busiest->cpu_power) /
4101 sds->this->cpu_power;
4103 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
4104 sds->this->cpu_power;
4105 pwr_move += sds->this->cpu_power *
4106 min(sds->this_load_per_task, sds->this_load + tmp);
4107 pwr_move /= SCHED_LOAD_SCALE;
4109 /* Move if we gain throughput */
4110 if (pwr_move > pwr_now)
4111 *imbalance = sds->busiest_load_per_task;
4115 * calculate_imbalance - Calculate the amount of imbalance present within the
4116 * groups of a given sched_domain during load balance.
4117 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
4118 * @this_cpu: Cpu for which currently load balance is being performed.
4119 * @imbalance: The variable to store the imbalance.
4121 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
4122 unsigned long *imbalance)
4124 unsigned long max_pull, load_above_capacity = ~0UL;
4126 sds->busiest_load_per_task /= sds->busiest_nr_running;
4127 if (sds->group_imb) {
4128 sds->busiest_load_per_task =
4129 min(sds->busiest_load_per_task, sds->avg_load);
4133 * In the presence of smp nice balancing, certain scenarios can have
4134 * max load less than avg load(as we skip the groups at or below
4135 * its cpu_power, while calculating max_load..)
4137 if (sds->max_load < sds->avg_load) {
4139 return fix_small_imbalance(sds, this_cpu, imbalance);
4142 if (!sds->group_imb) {
4144 * Don't want to pull so many tasks that a group would go idle.
4146 load_above_capacity = (sds->busiest_nr_running -
4147 sds->busiest_group_capacity);
4149 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_LOAD_SCALE);
4151 load_above_capacity /= sds->busiest->cpu_power;
4155 * We're trying to get all the cpus to the average_load, so we don't
4156 * want to push ourselves above the average load, nor do we wish to
4157 * reduce the max loaded cpu below the average load. At the same time,
4158 * we also don't want to reduce the group load below the group capacity
4159 * (so that we can implement power-savings policies etc). Thus we look
4160 * for the minimum possible imbalance.
4161 * Be careful of negative numbers as they'll appear as very large values
4162 * with unsigned longs.
4164 max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
4166 /* How much load to actually move to equalise the imbalance */
4167 *imbalance = min(max_pull * sds->busiest->cpu_power,
4168 (sds->avg_load - sds->this_load) * sds->this->cpu_power)
4172 * if *imbalance is less than the average load per runnable task
4173 * there is no gaurantee that any tasks will be moved so we'll have
4174 * a think about bumping its value to force at least one task to be
4177 if (*imbalance < sds->busiest_load_per_task)
4178 return fix_small_imbalance(sds, this_cpu, imbalance);
4182 /******* find_busiest_group() helpers end here *********************/
4185 * find_busiest_group - Returns the busiest group within the sched_domain
4186 * if there is an imbalance. If there isn't an imbalance, and
4187 * the user has opted for power-savings, it returns a group whose
4188 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
4189 * such a group exists.
4191 * Also calculates the amount of weighted load which should be moved
4192 * to restore balance.
4194 * @sd: The sched_domain whose busiest group is to be returned.
4195 * @this_cpu: The cpu for which load balancing is currently being performed.
4196 * @imbalance: Variable which stores amount of weighted load which should
4197 * be moved to restore balance/put a group to idle.
4198 * @idle: The idle status of this_cpu.
4199 * @sd_idle: The idleness of sd
4200 * @cpus: The set of CPUs under consideration for load-balancing.
4201 * @balance: Pointer to a variable indicating if this_cpu
4202 * is the appropriate cpu to perform load balancing at this_level.
4204 * Returns: - the busiest group if imbalance exists.
4205 * - If no imbalance and user has opted for power-savings balance,
4206 * return the least loaded group whose CPUs can be
4207 * put to idle by rebalancing its tasks onto our group.
4209 static struct sched_group *
4210 find_busiest_group(struct sched_domain *sd, int this_cpu,
4211 unsigned long *imbalance, enum cpu_idle_type idle,
4212 int *sd_idle, const struct cpumask *cpus, int *balance)
4214 struct sd_lb_stats sds;
4216 memset(&sds, 0, sizeof(sds));
4219 * Compute the various statistics relavent for load balancing at
4222 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
4225 /* Cases where imbalance does not exist from POV of this_cpu */
4226 /* 1) this_cpu is not the appropriate cpu to perform load balancing
4228 * 2) There is no busy sibling group to pull from.
4229 * 3) This group is the busiest group.
4230 * 4) This group is more busy than the avg busieness at this
4232 * 5) The imbalance is within the specified limit.
4234 * Note: when doing newidle balance, if the local group has excess
4235 * capacity (i.e. nr_running < group_capacity) and the busiest group
4236 * does not have any capacity, we force a load balance to pull tasks
4237 * to the local group. In this case, we skip past checks 3, 4 and 5.
4239 if (balance && !(*balance))
4242 if (!sds.busiest || sds.busiest_nr_running == 0)
4245 /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
4246 if (idle == CPU_NEWLY_IDLE && sds.this_has_capacity &&
4247 !sds.busiest_has_capacity)
4250 if (sds.this_load >= sds.max_load)
4253 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
4255 if (sds.this_load >= sds.avg_load)
4259 * In the CPU_NEWLY_IDLE, use imbalance_pct to be conservative.
4260 * And to check for busy balance use !idle_cpu instead of
4261 * CPU_NOT_IDLE. This is because HT siblings will use CPU_NOT_IDLE
4262 * even when they are idle.
4264 if (idle == CPU_NEWLY_IDLE || !idle_cpu(this_cpu)) {
4265 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
4269 * This cpu is idle. If the busiest group load doesn't
4270 * have more tasks than the number of available cpu's and
4271 * there is no imbalance between this and busiest group
4272 * wrt to idle cpu's, it is balanced.
4274 if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) &&
4275 sds.busiest_nr_running <= sds.busiest_group_weight)
4280 /* Looks like there is an imbalance. Compute it */
4281 calculate_imbalance(&sds, this_cpu, imbalance);
4286 * There is no obvious imbalance. But check if we can do some balancing
4289 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
4297 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4300 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
4301 unsigned long imbalance, const struct cpumask *cpus)
4303 struct rq *busiest = NULL, *rq;
4304 unsigned long max_load = 0;
4307 for_each_cpu(i, sched_group_cpus(group)) {
4308 unsigned long power = power_of(i);
4309 unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
4312 if (!cpumask_test_cpu(i, cpus))
4316 wl = weighted_cpuload(i);
4319 * When comparing with imbalance, use weighted_cpuload()
4320 * which is not scaled with the cpu power.
4322 if (capacity && rq->nr_running == 1 && wl > imbalance)
4326 * For the load comparisons with the other cpu's, consider
4327 * the weighted_cpuload() scaled with the cpu power, so that
4328 * the load can be moved away from the cpu that is potentially
4329 * running at a lower capacity.
4331 wl = (wl * SCHED_LOAD_SCALE) / power;
4333 if (wl > max_load) {
4343 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4344 * so long as it is large enough.
4346 #define MAX_PINNED_INTERVAL 512
4348 /* Working cpumask for load_balance and load_balance_newidle. */
4349 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4352 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4353 * tasks if there is an imbalance.
4355 static int load_balance(int this_cpu, struct rq *this_rq,
4356 struct sched_domain *sd, enum cpu_idle_type idle,
4359 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
4360 struct sched_group *group;
4361 unsigned long imbalance;
4363 unsigned long flags;
4364 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4366 cpumask_copy(cpus, cpu_active_mask);
4369 * When power savings policy is enabled for the parent domain, idle
4370 * sibling can pick up load irrespective of busy siblings. In this case,
4371 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4372 * portraying it as CPU_NOT_IDLE.
4374 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
4375 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4378 schedstat_inc(sd, lb_count[idle]);
4382 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
4389 schedstat_inc(sd, lb_nobusyg[idle]);
4393 busiest = find_busiest_queue(group, idle, imbalance, cpus);
4395 schedstat_inc(sd, lb_nobusyq[idle]);
4399 BUG_ON(busiest == this_rq);
4401 schedstat_add(sd, lb_imbalance[idle], imbalance);
4404 if (busiest->nr_running > 1) {
4406 * Attempt to move tasks. If find_busiest_group has found
4407 * an imbalance but busiest->nr_running <= 1, the group is
4408 * still unbalanced. ld_moved simply stays zero, so it is
4409 * correctly treated as an imbalance.
4411 local_irq_save(flags);
4412 double_rq_lock(this_rq, busiest);
4413 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4414 imbalance, sd, idle, &all_pinned);
4415 double_rq_unlock(this_rq, busiest);
4416 local_irq_restore(flags);
4419 * some other cpu did the load balance for us.
4421 if (ld_moved && this_cpu != smp_processor_id())
4422 resched_cpu(this_cpu);
4424 /* All tasks on this runqueue were pinned by CPU affinity */
4425 if (unlikely(all_pinned)) {
4426 cpumask_clear_cpu(cpu_of(busiest), cpus);
4427 if (!cpumask_empty(cpus))
4434 schedstat_inc(sd, lb_failed[idle]);
4436 * Increment the failure counter only on periodic balance.
4437 * We do not want newidle balance, which can be very
4438 * frequent, pollute the failure counter causing
4439 * excessive cache_hot migrations and active balances.
4441 if (idle != CPU_NEWLY_IDLE)
4442 sd->nr_balance_failed++;
4444 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
4446 spin_lock_irqsave(&busiest->lock, flags);
4448 /* don't kick the migration_thread, if the curr
4449 * task on busiest cpu can't be moved to this_cpu
4451 if (!cpumask_test_cpu(this_cpu,
4452 &busiest->curr->cpus_allowed)) {
4453 spin_unlock_irqrestore(&busiest->lock, flags);
4455 goto out_one_pinned;
4458 if (!busiest->active_balance) {
4459 busiest->active_balance = 1;
4460 busiest->push_cpu = this_cpu;
4463 spin_unlock_irqrestore(&busiest->lock, flags);
4465 wake_up_process(busiest->migration_thread);
4468 * We've kicked active balancing, reset the failure
4471 sd->nr_balance_failed = sd->cache_nice_tries+1;
4474 sd->nr_balance_failed = 0;
4476 if (likely(!active_balance)) {
4477 /* We were unbalanced, so reset the balancing interval */
4478 sd->balance_interval = sd->min_interval;
4481 * If we've begun active balancing, start to back off. This
4482 * case may not be covered by the all_pinned logic if there
4483 * is only 1 task on the busy runqueue (because we don't call
4486 if (sd->balance_interval < sd->max_interval)
4487 sd->balance_interval *= 2;
4490 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4491 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4497 schedstat_inc(sd, lb_balanced[idle]);
4499 sd->nr_balance_failed = 0;
4502 /* tune up the balancing interval */
4503 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4504 (sd->balance_interval < sd->max_interval))
4505 sd->balance_interval *= 2;
4507 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4508 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4519 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4520 * tasks if there is an imbalance.
4522 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4523 * this_rq is locked.
4526 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
4528 struct sched_group *group;
4529 struct rq *busiest = NULL;
4530 unsigned long imbalance;
4534 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4536 cpumask_copy(cpus, cpu_active_mask);
4539 * When power savings policy is enabled for the parent domain, idle
4540 * sibling can pick up load irrespective of busy siblings. In this case,
4541 * let the state of idle sibling percolate up as IDLE, instead of
4542 * portraying it as CPU_NOT_IDLE.
4544 if (sd->flags & SD_SHARE_CPUPOWER &&
4545 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4548 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
4550 update_shares_locked(this_rq, sd);
4551 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
4552 &sd_idle, cpus, NULL);
4554 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
4558 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
4560 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
4564 BUG_ON(busiest == this_rq);
4566 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
4569 if (busiest->nr_running > 1) {
4570 /* Attempt to move tasks */
4571 double_lock_balance(this_rq, busiest);
4572 /* this_rq->clock is already updated */
4573 update_rq_clock(busiest);
4574 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4575 imbalance, sd, CPU_NEWLY_IDLE,
4577 double_unlock_balance(this_rq, busiest);
4579 if (unlikely(all_pinned)) {
4580 cpumask_clear_cpu(cpu_of(busiest), cpus);
4581 if (!cpumask_empty(cpus))
4587 int active_balance = 0;
4589 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
4590 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4591 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4594 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4597 if (sd->nr_balance_failed++ < 2)
4601 * The only task running in a non-idle cpu can be moved to this
4602 * cpu in an attempt to completely freeup the other CPU
4603 * package. The same method used to move task in load_balance()
4604 * have been extended for load_balance_newidle() to speedup
4605 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4607 * The package power saving logic comes from
4608 * find_busiest_group(). If there are no imbalance, then
4609 * f_b_g() will return NULL. However when sched_mc={1,2} then
4610 * f_b_g() will select a group from which a running task may be
4611 * pulled to this cpu in order to make the other package idle.
4612 * If there is no opportunity to make a package idle and if
4613 * there are no imbalance, then f_b_g() will return NULL and no
4614 * action will be taken in load_balance_newidle().
4616 * Under normal task pull operation due to imbalance, there
4617 * will be more than one task in the source run queue and
4618 * move_tasks() will succeed. ld_moved will be true and this
4619 * active balance code will not be triggered.
4622 /* Lock busiest in correct order while this_rq is held */
4623 double_lock_balance(this_rq, busiest);
4626 * don't kick the migration_thread, if the curr
4627 * task on busiest cpu can't be moved to this_cpu
4629 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
4630 double_unlock_balance(this_rq, busiest);
4635 if (!busiest->active_balance) {
4636 busiest->active_balance = 1;
4637 busiest->push_cpu = this_cpu;
4641 double_unlock_balance(this_rq, busiest);
4643 * Should not call ttwu while holding a rq->lock
4645 spin_unlock(&this_rq->lock);
4647 wake_up_process(busiest->migration_thread);
4648 spin_lock(&this_rq->lock);
4651 sd->nr_balance_failed = 0;
4653 update_shares_locked(this_rq, sd);
4657 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
4658 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4659 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4661 sd->nr_balance_failed = 0;
4667 * idle_balance is called by schedule() if this_cpu is about to become
4668 * idle. Attempts to pull tasks from other CPUs.
4670 static void idle_balance(int this_cpu, struct rq *this_rq)
4672 struct sched_domain *sd;
4673 int pulled_task = 0;
4674 unsigned long next_balance = jiffies + HZ;
4676 this_rq->idle_stamp = this_rq->clock;
4678 if (this_rq->avg_idle < sysctl_sched_migration_cost)
4681 for_each_domain(this_cpu, sd) {
4682 unsigned long interval;
4684 if (!(sd->flags & SD_LOAD_BALANCE))
4687 if (sd->flags & SD_BALANCE_NEWIDLE)
4688 /* If we've pulled tasks over stop searching: */
4689 pulled_task = load_balance_newidle(this_cpu, this_rq,
4692 interval = msecs_to_jiffies(sd->balance_interval);
4693 if (time_after(next_balance, sd->last_balance + interval))
4694 next_balance = sd->last_balance + interval;
4696 this_rq->idle_stamp = 0;
4700 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4702 * We are going idle. next_balance may be set based on
4703 * a busy processor. So reset next_balance.
4705 this_rq->next_balance = next_balance;
4710 * active_load_balance is run by migration threads. It pushes running tasks
4711 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4712 * running on each physical CPU where possible, and avoids physical /
4713 * logical imbalances.
4715 * Called with busiest_rq locked.
4717 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
4719 int target_cpu = busiest_rq->push_cpu;
4720 struct sched_domain *sd;
4721 struct rq *target_rq;
4723 /* Is there any task to move? */
4724 if (busiest_rq->nr_running <= 1)
4727 target_rq = cpu_rq(target_cpu);
4730 * This condition is "impossible", if it occurs
4731 * we need to fix it. Originally reported by
4732 * Bjorn Helgaas on a 128-cpu setup.
4734 BUG_ON(busiest_rq == target_rq);
4736 /* move a task from busiest_rq to target_rq */
4737 double_lock_balance(busiest_rq, target_rq);
4738 update_rq_clock(busiest_rq);
4739 update_rq_clock(target_rq);
4741 /* Search for an sd spanning us and the target CPU. */
4742 for_each_domain(target_cpu, sd) {
4743 if ((sd->flags & SD_LOAD_BALANCE) &&
4744 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4749 schedstat_inc(sd, alb_count);
4751 if (move_one_task(target_rq, target_cpu, busiest_rq,
4753 schedstat_inc(sd, alb_pushed);
4755 schedstat_inc(sd, alb_failed);
4757 double_unlock_balance(busiest_rq, target_rq);
4762 atomic_t load_balancer;
4763 cpumask_var_t cpu_mask;
4764 cpumask_var_t ilb_grp_nohz_mask;
4765 } nohz ____cacheline_aligned = {
4766 .load_balancer = ATOMIC_INIT(-1),
4769 int get_nohz_load_balancer(void)
4771 return atomic_read(&nohz.load_balancer);
4774 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4776 * lowest_flag_domain - Return lowest sched_domain containing flag.
4777 * @cpu: The cpu whose lowest level of sched domain is to
4779 * @flag: The flag to check for the lowest sched_domain
4780 * for the given cpu.
4782 * Returns the lowest sched_domain of a cpu which contains the given flag.
4784 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4786 struct sched_domain *sd;
4788 for_each_domain(cpu, sd)
4789 if (sd && (sd->flags & flag))
4796 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4797 * @cpu: The cpu whose domains we're iterating over.
4798 * @sd: variable holding the value of the power_savings_sd
4800 * @flag: The flag to filter the sched_domains to be iterated.
4802 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4803 * set, starting from the lowest sched_domain to the highest.
4805 #define for_each_flag_domain(cpu, sd, flag) \
4806 for (sd = lowest_flag_domain(cpu, flag); \
4807 (sd && (sd->flags & flag)); sd = sd->parent)
4810 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4811 * @ilb_group: group to be checked for semi-idleness
4813 * Returns: 1 if the group is semi-idle. 0 otherwise.
4815 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4816 * and atleast one non-idle CPU. This helper function checks if the given
4817 * sched_group is semi-idle or not.
4819 static inline int is_semi_idle_group(struct sched_group *ilb_group)
4821 cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask,
4822 sched_group_cpus(ilb_group));
4825 * A sched_group is semi-idle when it has atleast one busy cpu
4826 * and atleast one idle cpu.
4828 if (cpumask_empty(nohz.ilb_grp_nohz_mask))
4831 if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group)))
4837 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4838 * @cpu: The cpu which is nominating a new idle_load_balancer.
4840 * Returns: Returns the id of the idle load balancer if it exists,
4841 * Else, returns >= nr_cpu_ids.
4843 * This algorithm picks the idle load balancer such that it belongs to a
4844 * semi-idle powersavings sched_domain. The idea is to try and avoid
4845 * completely idle packages/cores just for the purpose of idle load balancing
4846 * when there are other idle cpu's which are better suited for that job.
4848 static int find_new_ilb(int cpu)
4850 struct sched_domain *sd;
4851 struct sched_group *ilb_group;
4854 * Have idle load balancer selection from semi-idle packages only
4855 * when power-aware load balancing is enabled
4857 if (!(sched_smt_power_savings || sched_mc_power_savings))
4861 * Optimize for the case when we have no idle CPUs or only one
4862 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4864 if (cpumask_weight(nohz.cpu_mask) < 2)
4867 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4868 ilb_group = sd->groups;
4871 if (is_semi_idle_group(ilb_group))
4872 return cpumask_first(nohz.ilb_grp_nohz_mask);
4874 ilb_group = ilb_group->next;
4876 } while (ilb_group != sd->groups);
4880 return cpumask_first(nohz.cpu_mask);
4882 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4883 static inline int find_new_ilb(int call_cpu)
4885 return cpumask_first(nohz.cpu_mask);
4890 * This routine will try to nominate the ilb (idle load balancing)
4891 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4892 * load balancing on behalf of all those cpus. If all the cpus in the system
4893 * go into this tickless mode, then there will be no ilb owner (as there is
4894 * no need for one) and all the cpus will sleep till the next wakeup event
4897 * For the ilb owner, tick is not stopped. And this tick will be used
4898 * for idle load balancing. ilb owner will still be part of
4901 * While stopping the tick, this cpu will become the ilb owner if there
4902 * is no other owner. And will be the owner till that cpu becomes busy
4903 * or if all cpus in the system stop their ticks at which point
4904 * there is no need for ilb owner.
4906 * When the ilb owner becomes busy, it nominates another owner, during the
4907 * next busy scheduler_tick()
4909 int select_nohz_load_balancer(int stop_tick)
4911 int cpu = smp_processor_id();
4914 cpu_rq(cpu)->in_nohz_recently = 1;
4916 if (!cpu_active(cpu)) {
4917 if (atomic_read(&nohz.load_balancer) != cpu)
4921 * If we are going offline and still the leader,
4924 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4930 cpumask_set_cpu(cpu, nohz.cpu_mask);
4932 /* time for ilb owner also to sleep */
4933 if (cpumask_weight(nohz.cpu_mask) == num_active_cpus()) {
4934 if (atomic_read(&nohz.load_balancer) == cpu)
4935 atomic_set(&nohz.load_balancer, -1);
4939 if (atomic_read(&nohz.load_balancer) == -1) {
4940 /* make me the ilb owner */
4941 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4943 } else if (atomic_read(&nohz.load_balancer) == cpu) {
4946 if (!(sched_smt_power_savings ||
4947 sched_mc_power_savings))
4950 * Check to see if there is a more power-efficient
4953 new_ilb = find_new_ilb(cpu);
4954 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4955 atomic_set(&nohz.load_balancer, -1);
4956 resched_cpu(new_ilb);
4962 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4965 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4967 if (atomic_read(&nohz.load_balancer) == cpu)
4968 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4975 static DEFINE_SPINLOCK(balancing);
4978 * It checks each scheduling domain to see if it is due to be balanced,
4979 * and initiates a balancing operation if so.
4981 * Balancing parameters are set up in arch_init_sched_domains.
4983 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4986 struct rq *rq = cpu_rq(cpu);
4987 unsigned long interval;
4988 struct sched_domain *sd;
4989 /* Earliest time when we have to do rebalance again */
4990 unsigned long next_balance = jiffies + 60*HZ;
4991 int update_next_balance = 0;
4994 for_each_domain(cpu, sd) {
4995 if (!(sd->flags & SD_LOAD_BALANCE))
4998 interval = sd->balance_interval;
4999 if (idle != CPU_IDLE)
5000 interval *= sd->busy_factor;
5002 /* scale ms to jiffies */
5003 interval = msecs_to_jiffies(interval);
5004 if (unlikely(!interval))
5006 if (interval > HZ*NR_CPUS/10)
5007 interval = HZ*NR_CPUS/10;
5009 need_serialize = sd->flags & SD_SERIALIZE;
5011 if (need_serialize) {
5012 if (!spin_trylock(&balancing))
5016 if (time_after_eq(jiffies, sd->last_balance + interval)) {
5017 if (load_balance(cpu, rq, sd, idle, &balance)) {
5019 * We've pulled tasks over so either we're no
5020 * longer idle, or one of our SMT siblings is
5023 idle = CPU_NOT_IDLE;
5025 sd->last_balance = jiffies;
5028 spin_unlock(&balancing);
5030 if (time_after(next_balance, sd->last_balance + interval)) {
5031 next_balance = sd->last_balance + interval;
5032 update_next_balance = 1;
5036 * Stop the load balance at this level. There is another
5037 * CPU in our sched group which is doing load balancing more
5045 * next_balance will be updated only when there is a need.
5046 * When the cpu is attached to null domain for ex, it will not be
5049 if (likely(update_next_balance))
5050 rq->next_balance = next_balance;
5054 * run_rebalance_domains is triggered when needed from the scheduler tick.
5055 * In CONFIG_NO_HZ case, the idle load balance owner will do the
5056 * rebalancing for all the cpus for whom scheduler ticks are stopped.
5058 static void run_rebalance_domains(struct softirq_action *h)
5060 int this_cpu = smp_processor_id();
5061 struct rq *this_rq = cpu_rq(this_cpu);
5062 enum cpu_idle_type idle = this_rq->idle_at_tick ?
5063 CPU_IDLE : CPU_NOT_IDLE;
5065 rebalance_domains(this_cpu, idle);
5069 * If this cpu is the owner for idle load balancing, then do the
5070 * balancing on behalf of the other idle cpus whose ticks are
5073 if (this_rq->idle_at_tick &&
5074 atomic_read(&nohz.load_balancer) == this_cpu) {
5078 for_each_cpu(balance_cpu, nohz.cpu_mask) {
5079 if (balance_cpu == this_cpu)
5083 * If this cpu gets work to do, stop the load balancing
5084 * work being done for other cpus. Next load
5085 * balancing owner will pick it up.
5090 rebalance_domains(balance_cpu, CPU_IDLE);
5092 rq = cpu_rq(balance_cpu);
5093 if (time_after(this_rq->next_balance, rq->next_balance))
5094 this_rq->next_balance = rq->next_balance;
5100 static inline int on_null_domain(int cpu)
5102 return !rcu_dereference(cpu_rq(cpu)->sd);
5106 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
5108 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
5109 * idle load balancing owner or decide to stop the periodic load balancing,
5110 * if the whole system is idle.
5112 static inline void trigger_load_balance(struct rq *rq, int cpu)
5116 * If we were in the nohz mode recently and busy at the current
5117 * scheduler tick, then check if we need to nominate new idle
5120 if (rq->in_nohz_recently && !rq->idle_at_tick) {
5121 rq->in_nohz_recently = 0;
5123 if (atomic_read(&nohz.load_balancer) == cpu) {
5124 cpumask_clear_cpu(cpu, nohz.cpu_mask);
5125 atomic_set(&nohz.load_balancer, -1);
5128 if (atomic_read(&nohz.load_balancer) == -1) {
5129 int ilb = find_new_ilb(cpu);
5131 if (ilb < nr_cpu_ids)
5137 * If this cpu is idle and doing idle load balancing for all the
5138 * cpus with ticks stopped, is it time for that to stop?
5140 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
5141 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
5147 * If this cpu is idle and the idle load balancing is done by
5148 * someone else, then no need raise the SCHED_SOFTIRQ
5150 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
5151 cpumask_test_cpu(cpu, nohz.cpu_mask))
5154 /* Don't need to rebalance while attached to NULL domain */
5155 if (time_after_eq(jiffies, rq->next_balance) &&
5156 likely(!on_null_domain(cpu)))
5157 raise_softirq(SCHED_SOFTIRQ);
5160 #else /* CONFIG_SMP */
5163 * on UP we do not need to balance between CPUs:
5165 static inline void idle_balance(int cpu, struct rq *rq)
5171 DEFINE_PER_CPU(struct kernel_stat, kstat);
5173 EXPORT_PER_CPU_SYMBOL(kstat);
5176 * Return any ns on the sched_clock that have not yet been accounted in
5177 * @p in case that task is currently running.
5179 * Called with task_rq_lock() held on @rq.
5181 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
5185 if (task_current(rq, p)) {
5186 update_rq_clock(rq);
5187 ns = rq->clock_task - p->se.exec_start;
5195 unsigned long long task_delta_exec(struct task_struct *p)
5197 unsigned long flags;
5201 rq = task_rq_lock(p, &flags);
5202 ns = do_task_delta_exec(p, rq);
5203 task_rq_unlock(rq, &flags);
5209 * Return accounted runtime for the task.
5210 * In case the task is currently running, return the runtime plus current's
5211 * pending runtime that have not been accounted yet.
5213 unsigned long long task_sched_runtime(struct task_struct *p)
5215 unsigned long flags;
5219 rq = task_rq_lock(p, &flags);
5220 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
5221 task_rq_unlock(rq, &flags);
5227 * Return sum_exec_runtime for the thread group.
5228 * In case the task is currently running, return the sum plus current's
5229 * pending runtime that have not been accounted yet.
5231 * Note that the thread group might have other running tasks as well,
5232 * so the return value not includes other pending runtime that other
5233 * running tasks might have.
5235 unsigned long long thread_group_sched_runtime(struct task_struct *p)
5237 struct task_cputime totals;
5238 unsigned long flags;
5242 rq = task_rq_lock(p, &flags);
5243 thread_group_cputime(p, &totals);
5244 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
5245 task_rq_unlock(rq, &flags);
5251 * Account user cpu time to a process.
5252 * @p: the process that the cpu time gets accounted to
5253 * @cputime: the cpu time spent in user space since the last update
5254 * @cputime_scaled: cputime scaled by cpu frequency
5256 void account_user_time(struct task_struct *p, cputime_t cputime,
5257 cputime_t cputime_scaled)
5259 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5262 /* Add user time to process. */
5263 p->utime = cputime_add(p->utime, cputime);
5264 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5265 account_group_user_time(p, cputime);
5267 /* Add user time to cpustat. */
5268 tmp = cputime_to_cputime64(cputime);
5269 if (TASK_NICE(p) > 0)
5270 cpustat->nice = cputime64_add(cpustat->nice, tmp);
5272 cpustat->user = cputime64_add(cpustat->user, tmp);
5274 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
5275 /* Account for user time used */
5276 acct_update_integrals(p);
5280 * Account guest cpu time to a process.
5281 * @p: the process that the cpu time gets accounted to
5282 * @cputime: the cpu time spent in virtual machine since the last update
5283 * @cputime_scaled: cputime scaled by cpu frequency
5285 static void account_guest_time(struct task_struct *p, cputime_t cputime,
5286 cputime_t cputime_scaled)
5289 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5291 tmp = cputime_to_cputime64(cputime);
5293 /* Add guest time to process. */
5294 p->utime = cputime_add(p->utime, cputime);
5295 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5296 account_group_user_time(p, cputime);
5297 p->gtime = cputime_add(p->gtime, cputime);
5299 /* Add guest time to cpustat. */
5300 cpustat->user = cputime64_add(cpustat->user, tmp);
5301 cpustat->guest = cputime64_add(cpustat->guest, tmp);
5305 * Account system cpu time to a process.
5306 * @p: the process that the cpu time gets accounted to
5307 * @hardirq_offset: the offset to subtract from hardirq_count()
5308 * @cputime: the cpu time spent in kernel space since the last update
5309 * @cputime_scaled: cputime scaled by cpu frequency
5311 void account_system_time(struct task_struct *p, int hardirq_offset,
5312 cputime_t cputime, cputime_t cputime_scaled)
5314 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5317 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
5318 account_guest_time(p, cputime, cputime_scaled);
5322 /* Add system time to process. */
5323 p->stime = cputime_add(p->stime, cputime);
5324 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
5325 account_group_system_time(p, cputime);
5327 /* Add system time to cpustat. */
5328 tmp = cputime_to_cputime64(cputime);
5329 if (hardirq_count() - hardirq_offset)
5330 cpustat->irq = cputime64_add(cpustat->irq, tmp);
5331 else if (in_serving_softirq())
5332 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
5334 cpustat->system = cputime64_add(cpustat->system, tmp);
5336 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
5338 /* Account for system time used */
5339 acct_update_integrals(p);
5343 * Account for involuntary wait time.
5344 * @steal: the cpu time spent in involuntary wait
5346 void account_steal_time(cputime_t cputime)
5348 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5349 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5351 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
5355 * Account for idle time.
5356 * @cputime: the cpu time spent in idle wait
5358 void account_idle_time(cputime_t cputime)
5360 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5361 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5362 struct rq *rq = this_rq();
5364 if (atomic_read(&rq->nr_iowait) > 0)
5365 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
5367 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
5370 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
5373 * Account a single tick of cpu time.
5374 * @p: the process that the cpu time gets accounted to
5375 * @user_tick: indicates if the tick is a user or a system tick
5377 void account_process_tick(struct task_struct *p, int user_tick)
5379 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
5380 struct rq *rq = this_rq();
5383 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
5384 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
5385 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
5388 account_idle_time(cputime_one_jiffy);
5392 * Account multiple ticks of steal time.
5393 * @p: the process from which the cpu time has been stolen
5394 * @ticks: number of stolen ticks
5396 void account_steal_ticks(unsigned long ticks)
5398 account_steal_time(jiffies_to_cputime(ticks));
5402 * Account multiple ticks of idle time.
5403 * @ticks: number of stolen ticks
5405 void account_idle_ticks(unsigned long ticks)
5407 account_idle_time(jiffies_to_cputime(ticks));
5413 * Use precise platform statistics if available:
5415 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
5416 cputime_t task_utime(struct task_struct *p)
5421 cputime_t task_stime(struct task_struct *p)
5426 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5428 struct task_cputime cputime;
5430 thread_group_cputime(p, &cputime);
5432 *ut = cputime.utime;
5433 *st = cputime.stime;
5437 #ifndef nsecs_to_cputime
5438 # define nsecs_to_cputime(__nsecs) \
5439 msecs_to_cputime(div_u64((__nsecs), NSEC_PER_MSEC))
5442 cputime_t task_utime(struct task_struct *p)
5444 cputime_t utime = p->utime, total = utime + p->stime;
5448 * Use CFS's precise accounting:
5450 temp = (u64)nsecs_to_cputime(p->se.sum_exec_runtime);
5454 do_div(temp, total);
5456 utime = (cputime_t)temp;
5458 p->prev_utime = max(p->prev_utime, utime);
5459 return p->prev_utime;
5462 cputime_t task_stime(struct task_struct *p)
5467 * Use CFS's precise accounting. (we subtract utime from
5468 * the total, to make sure the total observed by userspace
5469 * grows monotonically - apps rely on that):
5471 stime = nsecs_to_cputime(p->se.sum_exec_runtime) - task_utime(p);
5474 p->prev_stime = max(p->prev_stime, stime);
5476 return p->prev_stime;
5480 * Must be called with siglock held.
5482 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5484 struct signal_struct *sig = p->signal;
5485 struct task_cputime cputime;
5486 cputime_t rtime, utime, total;
5488 thread_group_cputime(p, &cputime);
5490 total = cputime_add(cputime.utime, cputime.stime);
5491 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
5496 temp *= cputime.utime;
5497 do_div(temp, total);
5498 utime = (cputime_t)temp;
5502 sig->prev_utime = max(sig->prev_utime, utime);
5503 sig->prev_stime = max(sig->prev_stime,
5504 cputime_sub(rtime, sig->prev_utime));
5506 *ut = sig->prev_utime;
5507 *st = sig->prev_stime;
5511 inline cputime_t task_gtime(struct task_struct *p)
5517 * This function gets called by the timer code, with HZ frequency.
5518 * We call it with interrupts disabled.
5520 * It also gets called by the fork code, when changing the parent's
5523 void scheduler_tick(void)
5525 int cpu = smp_processor_id();
5526 struct rq *rq = cpu_rq(cpu);
5527 struct task_struct *curr = rq->curr;
5531 spin_lock(&rq->lock);
5532 update_rq_clock(rq);
5533 update_cpu_load(rq);
5534 curr->sched_class->task_tick(rq, curr, 0);
5535 spin_unlock(&rq->lock);
5537 perf_event_task_tick(curr, cpu);
5540 rq->idle_at_tick = idle_cpu(cpu);
5541 trigger_load_balance(rq, cpu);
5545 notrace unsigned long get_parent_ip(unsigned long addr)
5547 if (in_lock_functions(addr)) {
5548 addr = CALLER_ADDR2;
5549 if (in_lock_functions(addr))
5550 addr = CALLER_ADDR3;
5555 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5556 defined(CONFIG_PREEMPT_TRACER))
5558 void __kprobes add_preempt_count(int val)
5560 #ifdef CONFIG_DEBUG_PREEMPT
5564 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5567 preempt_count() += val;
5568 #ifdef CONFIG_DEBUG_PREEMPT
5570 * Spinlock count overflowing soon?
5572 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5575 if (preempt_count() == val)
5576 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5578 EXPORT_SYMBOL(add_preempt_count);
5580 void __kprobes sub_preempt_count(int val)
5582 #ifdef CONFIG_DEBUG_PREEMPT
5586 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5589 * Is the spinlock portion underflowing?
5591 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5592 !(preempt_count() & PREEMPT_MASK)))
5596 if (preempt_count() == val)
5597 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5598 preempt_count() -= val;
5600 EXPORT_SYMBOL(sub_preempt_count);
5605 * Print scheduling while atomic bug:
5607 static noinline void __schedule_bug(struct task_struct *prev)
5609 struct pt_regs *regs = get_irq_regs();
5611 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5612 prev->comm, prev->pid, preempt_count());
5614 debug_show_held_locks(prev);
5616 if (irqs_disabled())
5617 print_irqtrace_events(prev);
5626 * Various schedule()-time debugging checks and statistics:
5628 static inline void schedule_debug(struct task_struct *prev)
5631 * Test if we are atomic. Since do_exit() needs to call into
5632 * schedule() atomically, we ignore that path for now.
5633 * Otherwise, whine if we are scheduling when we should not be.
5635 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
5636 __schedule_bug(prev);
5638 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5640 schedstat_inc(this_rq(), sched_count);
5641 #ifdef CONFIG_SCHEDSTATS
5642 if (unlikely(prev->lock_depth >= 0)) {
5643 schedstat_inc(this_rq(), bkl_count);
5644 schedstat_inc(prev, sched_info.bkl_count);
5649 static void put_prev_task(struct rq *rq, struct task_struct *p)
5651 u64 runtime = p->se.sum_exec_runtime - p->se.prev_sum_exec_runtime;
5653 update_avg(&p->se.avg_running, runtime);
5655 if (p->state == TASK_RUNNING) {
5657 * In order to avoid avg_overlap growing stale when we are
5658 * indeed overlapping and hence not getting put to sleep, grow
5659 * the avg_overlap on preemption.
5661 * We use the average preemption runtime because that
5662 * correlates to the amount of cache footprint a task can
5665 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
5666 update_avg(&p->se.avg_overlap, runtime);
5668 update_avg(&p->se.avg_running, 0);
5670 p->sched_class->put_prev_task(rq, p);
5674 * Pick up the highest-prio task:
5676 static inline struct task_struct *
5677 pick_next_task(struct rq *rq)
5679 const struct sched_class *class;
5680 struct task_struct *p;
5683 * Optimization: we know that if all tasks are in
5684 * the fair class we can call that function directly:
5686 if (likely(rq->nr_running == rq->cfs.nr_running)) {
5687 p = fair_sched_class.pick_next_task(rq);
5692 class = sched_class_highest;
5694 p = class->pick_next_task(rq);
5698 * Will never be NULL as the idle class always
5699 * returns a non-NULL p:
5701 class = class->next;
5706 * schedule() is the main scheduler function.
5708 asmlinkage void __sched schedule(void)
5710 struct task_struct *prev, *next;
5711 unsigned long *switch_count;
5717 cpu = smp_processor_id();
5721 switch_count = &prev->nivcsw;
5723 release_kernel_lock(prev);
5724 need_resched_nonpreemptible:
5726 schedule_debug(prev);
5728 if (sched_feat(HRTICK))
5731 spin_lock_irq(&rq->lock);
5732 update_rq_clock(rq);
5733 clear_tsk_need_resched(prev);
5735 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
5736 if (unlikely(signal_pending_state(prev->state, prev)))
5737 prev->state = TASK_RUNNING;
5739 deactivate_task(rq, prev, 1);
5740 switch_count = &prev->nvcsw;
5743 pre_schedule(rq, prev);
5745 if (unlikely(!rq->nr_running))
5746 idle_balance(cpu, rq);
5748 put_prev_task(rq, prev);
5749 next = pick_next_task(rq);
5751 if (likely(prev != next)) {
5752 sched_info_switch(prev, next);
5753 perf_event_task_sched_out(prev, next, cpu);
5759 context_switch(rq, prev, next); /* unlocks the rq */
5761 * the context switch might have flipped the stack from under
5762 * us, hence refresh the local variables.
5764 cpu = smp_processor_id();
5767 spin_unlock_irq(&rq->lock);
5771 if (unlikely(reacquire_kernel_lock(current) < 0))
5772 goto need_resched_nonpreemptible;
5774 preempt_enable_no_resched();
5778 EXPORT_SYMBOL(schedule);
5782 * Look out! "owner" is an entirely speculative pointer
5783 * access and not reliable.
5785 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5790 if (!sched_feat(OWNER_SPIN))
5793 #ifdef CONFIG_DEBUG_PAGEALLOC
5795 * Need to access the cpu field knowing that
5796 * DEBUG_PAGEALLOC could have unmapped it if
5797 * the mutex owner just released it and exited.
5799 if (probe_kernel_address(&owner->cpu, cpu))
5806 * Even if the access succeeded (likely case),
5807 * the cpu field may no longer be valid.
5809 if (cpu >= nr_cpumask_bits)
5813 * We need to validate that we can do a
5814 * get_cpu() and that we have the percpu area.
5816 if (!cpu_online(cpu))
5823 * Owner changed, break to re-assess state.
5825 if (lock->owner != owner)
5829 * Is that owner really running on that cpu?
5831 if (task_thread_info(rq->curr) != owner || need_resched())
5841 #ifdef CONFIG_PREEMPT
5843 * this is the entry point to schedule() from in-kernel preemption
5844 * off of preempt_enable. Kernel preemptions off return from interrupt
5845 * occur there and call schedule directly.
5847 asmlinkage void __sched preempt_schedule(void)
5849 struct thread_info *ti = current_thread_info();
5852 * If there is a non-zero preempt_count or interrupts are disabled,
5853 * we do not want to preempt the current task. Just return..
5855 if (likely(ti->preempt_count || irqs_disabled()))
5859 add_preempt_count(PREEMPT_ACTIVE);
5861 sub_preempt_count(PREEMPT_ACTIVE);
5864 * Check again in case we missed a preemption opportunity
5865 * between schedule and now.
5868 } while (need_resched());
5870 EXPORT_SYMBOL(preempt_schedule);
5873 * this is the entry point to schedule() from kernel preemption
5874 * off of irq context.
5875 * Note, that this is called and return with irqs disabled. This will
5876 * protect us against recursive calling from irq.
5878 asmlinkage void __sched preempt_schedule_irq(void)
5880 struct thread_info *ti = current_thread_info();
5882 /* Catch callers which need to be fixed */
5883 BUG_ON(ti->preempt_count || !irqs_disabled());
5886 add_preempt_count(PREEMPT_ACTIVE);
5889 local_irq_disable();
5890 sub_preempt_count(PREEMPT_ACTIVE);
5893 * Check again in case we missed a preemption opportunity
5894 * between schedule and now.
5897 } while (need_resched());
5900 #endif /* CONFIG_PREEMPT */
5902 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
5905 return try_to_wake_up(curr->private, mode, wake_flags);
5907 EXPORT_SYMBOL(default_wake_function);
5910 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5911 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5912 * number) then we wake all the non-exclusive tasks and one exclusive task.
5914 * There are circumstances in which we can try to wake a task which has already
5915 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5916 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5918 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
5919 int nr_exclusive, int wake_flags, void *key)
5921 wait_queue_t *curr, *next;
5923 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
5924 unsigned flags = curr->flags;
5926 if (curr->func(curr, mode, wake_flags, key) &&
5927 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
5933 * __wake_up - wake up threads blocked on a waitqueue.
5935 * @mode: which threads
5936 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5937 * @key: is directly passed to the wakeup function
5939 * It may be assumed that this function implies a write memory barrier before
5940 * changing the task state if and only if any tasks are woken up.
5942 void __wake_up(wait_queue_head_t *q, unsigned int mode,
5943 int nr_exclusive, void *key)
5945 unsigned long flags;
5947 spin_lock_irqsave(&q->lock, flags);
5948 __wake_up_common(q, mode, nr_exclusive, 0, key);
5949 spin_unlock_irqrestore(&q->lock, flags);
5951 EXPORT_SYMBOL(__wake_up);
5954 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5956 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
5958 __wake_up_common(q, mode, 1, 0, NULL);
5961 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
5963 __wake_up_common(q, mode, 1, 0, key);
5967 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5969 * @mode: which threads
5970 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5971 * @key: opaque value to be passed to wakeup targets
5973 * The sync wakeup differs that the waker knows that it will schedule
5974 * away soon, so while the target thread will be woken up, it will not
5975 * be migrated to another CPU - ie. the two threads are 'synchronized'
5976 * with each other. This can prevent needless bouncing between CPUs.
5978 * On UP it can prevent extra preemption.
5980 * It may be assumed that this function implies a write memory barrier before
5981 * changing the task state if and only if any tasks are woken up.
5983 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
5984 int nr_exclusive, void *key)
5986 unsigned long flags;
5987 int wake_flags = WF_SYNC;
5992 if (unlikely(!nr_exclusive))
5995 spin_lock_irqsave(&q->lock, flags);
5996 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
5997 spin_unlock_irqrestore(&q->lock, flags);
5999 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
6002 * __wake_up_sync - see __wake_up_sync_key()
6004 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
6006 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
6008 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
6011 * complete: - signals a single thread waiting on this completion
6012 * @x: holds the state of this particular completion
6014 * This will wake up a single thread waiting on this completion. Threads will be
6015 * awakened in the same order in which they were queued.
6017 * See also complete_all(), wait_for_completion() and related routines.
6019 * It may be assumed that this function implies a write memory barrier before
6020 * changing the task state if and only if any tasks are woken up.
6022 void complete(struct completion *x)
6024 unsigned long flags;
6026 spin_lock_irqsave(&x->wait.lock, flags);
6028 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
6029 spin_unlock_irqrestore(&x->wait.lock, flags);
6031 EXPORT_SYMBOL(complete);
6034 * complete_all: - signals all threads waiting on this completion
6035 * @x: holds the state of this particular completion
6037 * This will wake up all threads waiting on this particular completion event.
6039 * It may be assumed that this function implies a write memory barrier before
6040 * changing the task state if and only if any tasks are woken up.
6042 void complete_all(struct completion *x)
6044 unsigned long flags;
6046 spin_lock_irqsave(&x->wait.lock, flags);
6047 x->done += UINT_MAX/2;
6048 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
6049 spin_unlock_irqrestore(&x->wait.lock, flags);
6051 EXPORT_SYMBOL(complete_all);
6053 static inline long __sched
6054 do_wait_for_common(struct completion *x, long timeout, int state)
6057 DECLARE_WAITQUEUE(wait, current);
6059 wait.flags |= WQ_FLAG_EXCLUSIVE;
6060 __add_wait_queue_tail(&x->wait, &wait);
6062 if (signal_pending_state(state, current)) {
6063 timeout = -ERESTARTSYS;
6066 __set_current_state(state);
6067 spin_unlock_irq(&x->wait.lock);
6068 timeout = schedule_timeout(timeout);
6069 spin_lock_irq(&x->wait.lock);
6070 } while (!x->done && timeout);
6071 __remove_wait_queue(&x->wait, &wait);
6076 return timeout ?: 1;
6080 wait_for_common(struct completion *x, long timeout, int state)
6084 spin_lock_irq(&x->wait.lock);
6085 timeout = do_wait_for_common(x, timeout, state);
6086 spin_unlock_irq(&x->wait.lock);
6091 * wait_for_completion: - waits for completion of a task
6092 * @x: holds the state of this particular completion
6094 * This waits to be signaled for completion of a specific task. It is NOT
6095 * interruptible and there is no timeout.
6097 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
6098 * and interrupt capability. Also see complete().
6100 void __sched wait_for_completion(struct completion *x)
6102 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
6104 EXPORT_SYMBOL(wait_for_completion);
6107 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
6108 * @x: holds the state of this particular completion
6109 * @timeout: timeout value in jiffies
6111 * This waits for either a completion of a specific task to be signaled or for a
6112 * specified timeout to expire. The timeout is in jiffies. It is not
6115 unsigned long __sched
6116 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
6118 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
6120 EXPORT_SYMBOL(wait_for_completion_timeout);
6123 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
6124 * @x: holds the state of this particular completion
6126 * This waits for completion of a specific task to be signaled. It is
6129 int __sched wait_for_completion_interruptible(struct completion *x)
6131 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
6132 if (t == -ERESTARTSYS)
6136 EXPORT_SYMBOL(wait_for_completion_interruptible);
6139 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
6140 * @x: holds the state of this particular completion
6141 * @timeout: timeout value in jiffies
6143 * This waits for either a completion of a specific task to be signaled or for a
6144 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
6146 unsigned long __sched
6147 wait_for_completion_interruptible_timeout(struct completion *x,
6148 unsigned long timeout)
6150 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
6152 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
6155 * wait_for_completion_killable: - waits for completion of a task (killable)
6156 * @x: holds the state of this particular completion
6158 * This waits to be signaled for completion of a specific task. It can be
6159 * interrupted by a kill signal.
6161 int __sched wait_for_completion_killable(struct completion *x)
6163 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
6164 if (t == -ERESTARTSYS)
6168 EXPORT_SYMBOL(wait_for_completion_killable);
6171 * try_wait_for_completion - try to decrement a completion without blocking
6172 * @x: completion structure
6174 * Returns: 0 if a decrement cannot be done without blocking
6175 * 1 if a decrement succeeded.
6177 * If a completion is being used as a counting completion,
6178 * attempt to decrement the counter without blocking. This
6179 * enables us to avoid waiting if the resource the completion
6180 * is protecting is not available.
6182 bool try_wait_for_completion(struct completion *x)
6184 unsigned long flags;
6187 spin_lock_irqsave(&x->wait.lock, flags);
6192 spin_unlock_irqrestore(&x->wait.lock, flags);
6195 EXPORT_SYMBOL(try_wait_for_completion);
6198 * completion_done - Test to see if a completion has any waiters
6199 * @x: completion structure
6201 * Returns: 0 if there are waiters (wait_for_completion() in progress)
6202 * 1 if there are no waiters.
6205 bool completion_done(struct completion *x)
6207 unsigned long flags;
6210 spin_lock_irqsave(&x->wait.lock, flags);
6213 spin_unlock_irqrestore(&x->wait.lock, flags);
6216 EXPORT_SYMBOL(completion_done);
6219 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
6221 unsigned long flags;
6224 init_waitqueue_entry(&wait, current);
6226 __set_current_state(state);
6228 spin_lock_irqsave(&q->lock, flags);
6229 __add_wait_queue(q, &wait);
6230 spin_unlock(&q->lock);
6231 timeout = schedule_timeout(timeout);
6232 spin_lock_irq(&q->lock);
6233 __remove_wait_queue(q, &wait);
6234 spin_unlock_irqrestore(&q->lock, flags);
6239 void __sched interruptible_sleep_on(wait_queue_head_t *q)
6241 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
6243 EXPORT_SYMBOL(interruptible_sleep_on);
6246 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
6248 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
6250 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
6252 void __sched sleep_on(wait_queue_head_t *q)
6254 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
6256 EXPORT_SYMBOL(sleep_on);
6258 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
6260 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
6262 EXPORT_SYMBOL(sleep_on_timeout);
6264 #ifdef CONFIG_RT_MUTEXES
6267 * rt_mutex_setprio - set the current priority of a task
6269 * @prio: prio value (kernel-internal form)
6271 * This function changes the 'effective' priority of a task. It does
6272 * not touch ->normal_prio like __setscheduler().
6274 * Used by the rt_mutex code to implement priority inheritance logic.
6276 void rt_mutex_setprio(struct task_struct *p, int prio)
6278 unsigned long flags;
6279 int oldprio, on_rq, running;
6281 const struct sched_class *prev_class;
6283 BUG_ON(prio < 0 || prio > MAX_PRIO);
6285 rq = task_rq_lock(p, &flags);
6286 update_rq_clock(rq);
6289 prev_class = p->sched_class;
6290 on_rq = p->se.on_rq;
6291 running = task_current(rq, p);
6293 dequeue_task(rq, p, 0);
6295 p->sched_class->put_prev_task(rq, p);
6298 p->sched_class = &rt_sched_class;
6300 p->sched_class = &fair_sched_class;
6305 p->sched_class->set_curr_task(rq);
6307 enqueue_task(rq, p, 0, oldprio < prio);
6309 check_class_changed(rq, p, prev_class, oldprio, running);
6311 task_rq_unlock(rq, &flags);
6316 void set_user_nice(struct task_struct *p, long nice)
6318 int old_prio, delta, on_rq;
6319 unsigned long flags;
6322 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
6325 * We have to be careful, if called from sys_setpriority(),
6326 * the task might be in the middle of scheduling on another CPU.
6328 rq = task_rq_lock(p, &flags);
6329 update_rq_clock(rq);
6331 * The RT priorities are set via sched_setscheduler(), but we still
6332 * allow the 'normal' nice value to be set - but as expected
6333 * it wont have any effect on scheduling until the task is
6334 * SCHED_FIFO/SCHED_RR:
6336 if (task_has_rt_policy(p)) {
6337 p->static_prio = NICE_TO_PRIO(nice);
6340 on_rq = p->se.on_rq;
6342 dequeue_task(rq, p, 0);
6344 p->static_prio = NICE_TO_PRIO(nice);
6347 p->prio = effective_prio(p);
6348 delta = p->prio - old_prio;
6351 enqueue_task(rq, p, 0, false);
6353 * If the task increased its priority or is running and
6354 * lowered its priority, then reschedule its CPU:
6356 if (delta < 0 || (delta > 0 && task_running(rq, p)))
6357 resched_task(rq->curr);
6360 task_rq_unlock(rq, &flags);
6362 EXPORT_SYMBOL(set_user_nice);
6365 * can_nice - check if a task can reduce its nice value
6369 int can_nice(const struct task_struct *p, const int nice)
6371 /* convert nice value [19,-20] to rlimit style value [1,40] */
6372 int nice_rlim = 20 - nice;
6374 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
6375 capable(CAP_SYS_NICE));
6378 #ifdef __ARCH_WANT_SYS_NICE
6381 * sys_nice - change the priority of the current process.
6382 * @increment: priority increment
6384 * sys_setpriority is a more generic, but much slower function that
6385 * does similar things.
6387 SYSCALL_DEFINE1(nice, int, increment)
6392 * Setpriority might change our priority at the same moment.
6393 * We don't have to worry. Conceptually one call occurs first
6394 * and we have a single winner.
6396 if (increment < -40)
6401 nice = TASK_NICE(current) + increment;
6407 if (increment < 0 && !can_nice(current, nice))
6410 retval = security_task_setnice(current, nice);
6414 set_user_nice(current, nice);
6421 * task_prio - return the priority value of a given task.
6422 * @p: the task in question.
6424 * This is the priority value as seen by users in /proc.
6425 * RT tasks are offset by -200. Normal tasks are centered
6426 * around 0, value goes from -16 to +15.
6428 int task_prio(const struct task_struct *p)
6430 return p->prio - MAX_RT_PRIO;
6434 * task_nice - return the nice value of a given task.
6435 * @p: the task in question.
6437 int task_nice(const struct task_struct *p)
6439 return TASK_NICE(p);
6441 EXPORT_SYMBOL(task_nice);
6444 * idle_cpu - is a given cpu idle currently?
6445 * @cpu: the processor in question.
6447 int idle_cpu(int cpu)
6449 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
6453 * idle_task - return the idle task for a given cpu.
6454 * @cpu: the processor in question.
6456 struct task_struct *idle_task(int cpu)
6458 return cpu_rq(cpu)->idle;
6462 * find_process_by_pid - find a process with a matching PID value.
6463 * @pid: the pid in question.
6465 static struct task_struct *find_process_by_pid(pid_t pid)
6467 return pid ? find_task_by_vpid(pid) : current;
6470 /* Actually do priority change: must hold rq lock. */
6472 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
6474 BUG_ON(p->se.on_rq);
6477 switch (p->policy) {
6481 p->sched_class = &fair_sched_class;
6485 p->sched_class = &rt_sched_class;
6489 p->rt_priority = prio;
6490 p->normal_prio = normal_prio(p);
6491 /* we are holding p->pi_lock already */
6492 p->prio = rt_mutex_getprio(p);
6497 * check the target process has a UID that matches the current process's
6499 static bool check_same_owner(struct task_struct *p)
6501 const struct cred *cred = current_cred(), *pcred;
6505 pcred = __task_cred(p);
6506 match = (cred->euid == pcred->euid ||
6507 cred->euid == pcred->uid);
6512 static int __sched_setscheduler(struct task_struct *p, int policy,
6513 struct sched_param *param, bool user)
6515 int retval, oldprio, oldpolicy = -1, on_rq, running;
6516 unsigned long flags;
6517 const struct sched_class *prev_class;
6521 /* may grab non-irq protected spin_locks */
6522 BUG_ON(in_interrupt());
6524 /* double check policy once rq lock held */
6526 reset_on_fork = p->sched_reset_on_fork;
6527 policy = oldpolicy = p->policy;
6529 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
6530 policy &= ~SCHED_RESET_ON_FORK;
6532 if (policy != SCHED_FIFO && policy != SCHED_RR &&
6533 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
6534 policy != SCHED_IDLE)
6539 * Valid priorities for SCHED_FIFO and SCHED_RR are
6540 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6541 * SCHED_BATCH and SCHED_IDLE is 0.
6543 if (param->sched_priority < 0 ||
6544 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
6545 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
6547 if (rt_policy(policy) != (param->sched_priority != 0))
6551 * Allow unprivileged RT tasks to decrease priority:
6553 if (user && !capable(CAP_SYS_NICE)) {
6554 if (rt_policy(policy)) {
6555 unsigned long rlim_rtprio;
6557 if (!lock_task_sighand(p, &flags))
6559 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
6560 unlock_task_sighand(p, &flags);
6562 /* can't set/change the rt policy */
6563 if (policy != p->policy && !rlim_rtprio)
6566 /* can't increase priority */
6567 if (param->sched_priority > p->rt_priority &&
6568 param->sched_priority > rlim_rtprio)
6572 * Like positive nice levels, dont allow tasks to
6573 * move out of SCHED_IDLE either:
6575 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
6578 /* can't change other user's priorities */
6579 if (!check_same_owner(p))
6582 /* Normal users shall not reset the sched_reset_on_fork flag */
6583 if (p->sched_reset_on_fork && !reset_on_fork)
6588 #ifdef CONFIG_RT_GROUP_SCHED
6590 * Do not allow realtime tasks into groups that have no runtime
6593 if (rt_bandwidth_enabled() && rt_policy(policy) &&
6594 task_group(p)->rt_bandwidth.rt_runtime == 0)
6598 retval = security_task_setscheduler(p, policy, param);
6604 * make sure no PI-waiters arrive (or leave) while we are
6605 * changing the priority of the task:
6607 spin_lock_irqsave(&p->pi_lock, flags);
6609 * To be able to change p->policy safely, the apropriate
6610 * runqueue lock must be held.
6612 rq = __task_rq_lock(p);
6613 /* recheck policy now with rq lock held */
6614 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
6615 policy = oldpolicy = -1;
6616 __task_rq_unlock(rq);
6617 spin_unlock_irqrestore(&p->pi_lock, flags);
6620 update_rq_clock(rq);
6621 on_rq = p->se.on_rq;
6622 running = task_current(rq, p);
6624 deactivate_task(rq, p, 0);
6626 p->sched_class->put_prev_task(rq, p);
6628 p->sched_reset_on_fork = reset_on_fork;
6631 prev_class = p->sched_class;
6632 __setscheduler(rq, p, policy, param->sched_priority);
6635 p->sched_class->set_curr_task(rq);
6637 activate_task(rq, p, 0);
6639 check_class_changed(rq, p, prev_class, oldprio, running);
6641 __task_rq_unlock(rq);
6642 spin_unlock_irqrestore(&p->pi_lock, flags);
6644 rt_mutex_adjust_pi(p);
6650 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6651 * @p: the task in question.
6652 * @policy: new policy.
6653 * @param: structure containing the new RT priority.
6655 * NOTE that the task may be already dead.
6657 int sched_setscheduler(struct task_struct *p, int policy,
6658 struct sched_param *param)
6660 return __sched_setscheduler(p, policy, param, true);
6662 EXPORT_SYMBOL_GPL(sched_setscheduler);
6665 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6666 * @p: the task in question.
6667 * @policy: new policy.
6668 * @param: structure containing the new RT priority.
6670 * Just like sched_setscheduler, only don't bother checking if the
6671 * current context has permission. For example, this is needed in
6672 * stop_machine(): we create temporary high priority worker threads,
6673 * but our caller might not have that capability.
6675 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
6676 struct sched_param *param)
6678 return __sched_setscheduler(p, policy, param, false);
6682 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
6684 struct sched_param lparam;
6685 struct task_struct *p;
6688 if (!param || pid < 0)
6690 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
6695 p = find_process_by_pid(pid);
6697 retval = sched_setscheduler(p, policy, &lparam);
6704 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6705 * @pid: the pid in question.
6706 * @policy: new policy.
6707 * @param: structure containing the new RT priority.
6709 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
6710 struct sched_param __user *, param)
6712 /* negative values for policy are not valid */
6716 return do_sched_setscheduler(pid, policy, param);
6720 * sys_sched_setparam - set/change the RT priority of a thread
6721 * @pid: the pid in question.
6722 * @param: structure containing the new RT priority.
6724 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
6726 return do_sched_setscheduler(pid, -1, param);
6730 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6731 * @pid: the pid in question.
6733 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
6735 struct task_struct *p;
6743 p = find_process_by_pid(pid);
6745 retval = security_task_getscheduler(p);
6748 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
6755 * sys_sched_getparam - get the RT priority of a thread
6756 * @pid: the pid in question.
6757 * @param: structure containing the RT priority.
6759 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
6761 struct sched_param lp;
6762 struct task_struct *p;
6765 if (!param || pid < 0)
6769 p = find_process_by_pid(pid);
6774 retval = security_task_getscheduler(p);
6778 lp.sched_priority = p->rt_priority;
6782 * This one might sleep, we cannot do it with a spinlock held ...
6784 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
6793 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
6795 cpumask_var_t cpus_allowed, new_mask;
6796 struct task_struct *p;
6802 p = find_process_by_pid(pid);
6809 /* Prevent p going away */
6813 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6817 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
6819 goto out_free_cpus_allowed;
6822 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
6825 retval = security_task_setscheduler(p, 0, NULL);
6829 cpuset_cpus_allowed(p, cpus_allowed);
6830 cpumask_and(new_mask, in_mask, cpus_allowed);
6832 retval = set_cpus_allowed_ptr(p, new_mask);
6835 cpuset_cpus_allowed(p, cpus_allowed);
6836 if (!cpumask_subset(new_mask, cpus_allowed)) {
6838 * We must have raced with a concurrent cpuset
6839 * update. Just reset the cpus_allowed to the
6840 * cpuset's cpus_allowed
6842 cpumask_copy(new_mask, cpus_allowed);
6847 free_cpumask_var(new_mask);
6848 out_free_cpus_allowed:
6849 free_cpumask_var(cpus_allowed);
6856 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
6857 struct cpumask *new_mask)
6859 if (len < cpumask_size())
6860 cpumask_clear(new_mask);
6861 else if (len > cpumask_size())
6862 len = cpumask_size();
6864 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6868 * sys_sched_setaffinity - set the cpu affinity of a process
6869 * @pid: pid of the process
6870 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6871 * @user_mask_ptr: user-space pointer to the new cpu mask
6873 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6874 unsigned long __user *, user_mask_ptr)
6876 cpumask_var_t new_mask;
6879 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6882 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6884 retval = sched_setaffinity(pid, new_mask);
6885 free_cpumask_var(new_mask);
6889 long sched_getaffinity(pid_t pid, struct cpumask *mask)
6891 struct task_struct *p;
6892 unsigned long flags;
6900 p = find_process_by_pid(pid);
6904 retval = security_task_getscheduler(p);
6908 rq = task_rq_lock(p, &flags);
6909 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
6910 task_rq_unlock(rq, &flags);
6920 * sys_sched_getaffinity - get the cpu affinity of a process
6921 * @pid: pid of the process
6922 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6923 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6925 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6926 unsigned long __user *, user_mask_ptr)
6931 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
6933 if (len & (sizeof(unsigned long)-1))
6936 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6939 ret = sched_getaffinity(pid, mask);
6941 size_t retlen = min_t(size_t, len, cpumask_size());
6943 if (copy_to_user(user_mask_ptr, mask, retlen))
6948 free_cpumask_var(mask);
6954 * sys_sched_yield - yield the current processor to other threads.
6956 * This function yields the current CPU to other tasks. If there are no
6957 * other threads running on this CPU then this function will return.
6959 SYSCALL_DEFINE0(sched_yield)
6961 struct rq *rq = this_rq_lock();
6963 schedstat_inc(rq, yld_count);
6964 current->sched_class->yield_task(rq);
6967 * Since we are going to call schedule() anyway, there's
6968 * no need to preempt or enable interrupts:
6970 __release(rq->lock);
6971 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
6972 _raw_spin_unlock(&rq->lock);
6973 preempt_enable_no_resched();
6980 static inline int should_resched(void)
6982 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
6985 static void __cond_resched(void)
6987 add_preempt_count(PREEMPT_ACTIVE);
6989 sub_preempt_count(PREEMPT_ACTIVE);
6992 int __sched _cond_resched(void)
6994 if (should_resched()) {
7000 EXPORT_SYMBOL(_cond_resched);
7003 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
7004 * call schedule, and on return reacquire the lock.
7006 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
7007 * operations here to prevent schedule() from being called twice (once via
7008 * spin_unlock(), once by hand).
7010 int __cond_resched_lock(spinlock_t *lock)
7012 int resched = should_resched();
7015 lockdep_assert_held(lock);
7017 if (spin_needbreak(lock) || resched) {
7028 EXPORT_SYMBOL(__cond_resched_lock);
7030 int __sched __cond_resched_softirq(void)
7032 BUG_ON(!in_softirq());
7034 if (should_resched()) {
7042 EXPORT_SYMBOL(__cond_resched_softirq);
7045 * yield - yield the current processor to other threads.
7047 * This is a shortcut for kernel-space yielding - it marks the
7048 * thread runnable and calls sys_sched_yield().
7050 void __sched yield(void)
7052 set_current_state(TASK_RUNNING);
7055 EXPORT_SYMBOL(yield);
7058 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
7059 * that process accounting knows that this is a task in IO wait state.
7061 void __sched io_schedule(void)
7063 struct rq *rq = raw_rq();
7065 delayacct_blkio_start();
7066 atomic_inc(&rq->nr_iowait);
7067 current->in_iowait = 1;
7069 current->in_iowait = 0;
7070 atomic_dec(&rq->nr_iowait);
7071 delayacct_blkio_end();
7073 EXPORT_SYMBOL(io_schedule);
7075 long __sched io_schedule_timeout(long timeout)
7077 struct rq *rq = raw_rq();
7080 delayacct_blkio_start();
7081 atomic_inc(&rq->nr_iowait);
7082 current->in_iowait = 1;
7083 ret = schedule_timeout(timeout);
7084 current->in_iowait = 0;
7085 atomic_dec(&rq->nr_iowait);
7086 delayacct_blkio_end();
7091 * sys_sched_get_priority_max - return maximum RT priority.
7092 * @policy: scheduling class.
7094 * this syscall returns the maximum rt_priority that can be used
7095 * by a given scheduling class.
7097 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
7104 ret = MAX_USER_RT_PRIO-1;
7116 * sys_sched_get_priority_min - return minimum RT priority.
7117 * @policy: scheduling class.
7119 * this syscall returns the minimum rt_priority that can be used
7120 * by a given scheduling class.
7122 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
7140 * sys_sched_rr_get_interval - return the default timeslice of a process.
7141 * @pid: pid of the process.
7142 * @interval: userspace pointer to the timeslice value.
7144 * this syscall writes the default timeslice value of a given process
7145 * into the user-space timespec buffer. A value of '0' means infinity.
7147 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
7148 struct timespec __user *, interval)
7150 struct task_struct *p;
7151 unsigned int time_slice;
7152 unsigned long flags;
7162 p = find_process_by_pid(pid);
7166 retval = security_task_getscheduler(p);
7170 rq = task_rq_lock(p, &flags);
7171 time_slice = p->sched_class->get_rr_interval(rq, p);
7172 task_rq_unlock(rq, &flags);
7175 jiffies_to_timespec(time_slice, &t);
7176 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
7184 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
7186 void sched_show_task(struct task_struct *p)
7188 unsigned long free = 0;
7191 state = p->state ? __ffs(p->state) + 1 : 0;
7192 printk(KERN_INFO "%-13.13s %c", p->comm,
7193 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
7194 #if BITS_PER_LONG == 32
7195 if (state == TASK_RUNNING)
7196 printk(KERN_CONT " running ");
7198 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
7200 if (state == TASK_RUNNING)
7201 printk(KERN_CONT " running task ");
7203 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
7205 #ifdef CONFIG_DEBUG_STACK_USAGE
7206 free = stack_not_used(p);
7208 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
7209 task_pid_nr(p), task_pid_nr(p->real_parent),
7210 (unsigned long)task_thread_info(p)->flags);
7212 show_stack(p, NULL);
7215 void show_state_filter(unsigned long state_filter)
7217 struct task_struct *g, *p;
7219 #if BITS_PER_LONG == 32
7221 " task PC stack pid father\n");
7224 " task PC stack pid father\n");
7226 read_lock(&tasklist_lock);
7227 do_each_thread(g, p) {
7229 * reset the NMI-timeout, listing all files on a slow
7230 * console might take alot of time:
7232 touch_nmi_watchdog();
7233 if (!state_filter || (p->state & state_filter))
7235 } while_each_thread(g, p);
7237 touch_all_softlockup_watchdogs();
7239 #ifdef CONFIG_SCHED_DEBUG
7240 sysrq_sched_debug_show();
7242 read_unlock(&tasklist_lock);
7244 * Only show locks if all tasks are dumped:
7246 if (state_filter == -1)
7247 debug_show_all_locks();
7250 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
7252 idle->sched_class = &idle_sched_class;
7256 * init_idle - set up an idle thread for a given CPU
7257 * @idle: task in question
7258 * @cpu: cpu the idle task belongs to
7260 * NOTE: this function does not set the idle thread's NEED_RESCHED
7261 * flag, to make booting more robust.
7263 void __cpuinit init_idle(struct task_struct *idle, int cpu)
7265 struct rq *rq = cpu_rq(cpu);
7266 unsigned long flags;
7268 spin_lock_irqsave(&rq->lock, flags);
7271 idle->state = TASK_RUNNING;
7272 idle->se.exec_start = sched_clock();
7274 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
7276 * We're having a chicken and egg problem, even though we are
7277 * holding rq->lock, the cpu isn't yet set to this cpu so the
7278 * lockdep check in task_group() will fail.
7280 * Similar case to sched_fork(). / Alternatively we could
7281 * use task_rq_lock() here and obtain the other rq->lock.
7286 __set_task_cpu(idle, cpu);
7289 rq->curr = rq->idle = idle;
7290 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
7293 spin_unlock_irqrestore(&rq->lock, flags);
7295 /* Set the preempt count _outside_ the spinlocks! */
7296 #if defined(CONFIG_PREEMPT)
7297 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
7299 task_thread_info(idle)->preempt_count = 0;
7302 * The idle tasks have their own, simple scheduling class:
7304 idle->sched_class = &idle_sched_class;
7305 ftrace_graph_init_task(idle);
7309 * In a system that switches off the HZ timer nohz_cpu_mask
7310 * indicates which cpus entered this state. This is used
7311 * in the rcu update to wait only for active cpus. For system
7312 * which do not switch off the HZ timer nohz_cpu_mask should
7313 * always be CPU_BITS_NONE.
7315 cpumask_var_t nohz_cpu_mask;
7318 * Increase the granularity value when there are more CPUs,
7319 * because with more CPUs the 'effective latency' as visible
7320 * to users decreases. But the relationship is not linear,
7321 * so pick a second-best guess by going with the log2 of the
7324 * This idea comes from the SD scheduler of Con Kolivas:
7326 static void update_sysctl(void)
7328 unsigned int cpus = min(num_online_cpus(), 8U);
7329 unsigned int factor = 1 + ilog2(cpus);
7331 #define SET_SYSCTL(name) \
7332 (sysctl_##name = (factor) * normalized_sysctl_##name)
7333 SET_SYSCTL(sched_min_granularity);
7334 SET_SYSCTL(sched_latency);
7335 SET_SYSCTL(sched_wakeup_granularity);
7336 SET_SYSCTL(sched_shares_ratelimit);
7340 static inline void sched_init_granularity(void)
7347 * This is how migration works:
7349 * 1) we queue a struct migration_req structure in the source CPU's
7350 * runqueue and wake up that CPU's migration thread.
7351 * 2) we down() the locked semaphore => thread blocks.
7352 * 3) migration thread wakes up (implicitly it forces the migrated
7353 * thread off the CPU)
7354 * 4) it gets the migration request and checks whether the migrated
7355 * task is still in the wrong runqueue.
7356 * 5) if it's in the wrong runqueue then the migration thread removes
7357 * it and puts it into the right queue.
7358 * 6) migration thread up()s the semaphore.
7359 * 7) we wake up and the migration is done.
7363 * Change a given task's CPU affinity. Migrate the thread to a
7364 * proper CPU and schedule it away if the CPU it's executing on
7365 * is removed from the allowed bitmask.
7367 * NOTE: the caller must have a valid reference to the task, the
7368 * task must not exit() & deallocate itself prematurely. The
7369 * call is not atomic; no spinlocks may be held.
7371 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
7373 struct migration_req req;
7374 unsigned long flags;
7379 * Serialize against TASK_WAKING so that ttwu() and wunt() can
7380 * drop the rq->lock and still rely on ->cpus_allowed.
7383 while (task_is_waking(p))
7385 rq = task_rq_lock(p, &flags);
7386 if (task_is_waking(p)) {
7387 task_rq_unlock(rq, &flags);
7391 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
7396 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
7397 !cpumask_equal(&p->cpus_allowed, new_mask))) {
7402 if (p->sched_class->set_cpus_allowed)
7403 p->sched_class->set_cpus_allowed(p, new_mask);
7405 cpumask_copy(&p->cpus_allowed, new_mask);
7406 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
7409 /* Can the task run on the task's current CPU? If so, we're done */
7410 if (cpumask_test_cpu(task_cpu(p), new_mask))
7413 if (migrate_task(p, cpumask_any_and(cpu_active_mask, new_mask), &req)) {
7414 /* Need help from migration thread: drop lock and wait. */
7415 struct task_struct *mt = rq->migration_thread;
7417 get_task_struct(mt);
7418 task_rq_unlock(rq, &flags);
7419 wake_up_process(mt);
7420 put_task_struct(mt);
7421 wait_for_completion(&req.done);
7422 tlb_migrate_finish(p->mm);
7426 task_rq_unlock(rq, &flags);
7430 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
7433 * Move (not current) task off this cpu, onto dest cpu. We're doing
7434 * this because either it can't run here any more (set_cpus_allowed()
7435 * away from this CPU, or CPU going down), or because we're
7436 * attempting to rebalance this task on exec (sched_exec).
7438 * So we race with normal scheduler movements, but that's OK, as long
7439 * as the task is no longer on this CPU.
7441 * Returns non-zero if task was successfully migrated.
7443 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
7445 struct rq *rq_dest, *rq_src;
7448 if (unlikely(!cpu_active(dest_cpu)))
7451 rq_src = cpu_rq(src_cpu);
7452 rq_dest = cpu_rq(dest_cpu);
7454 double_rq_lock(rq_src, rq_dest);
7455 /* Already moved. */
7456 if (task_cpu(p) != src_cpu)
7458 /* Affinity changed (again). */
7459 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7463 * If we're not on a rq, the next wake-up will ensure we're
7467 deactivate_task(rq_src, p, 0);
7468 set_task_cpu(p, dest_cpu);
7469 activate_task(rq_dest, p, 0);
7470 check_preempt_curr(rq_dest, p, 0);
7475 double_rq_unlock(rq_src, rq_dest);
7479 #define RCU_MIGRATION_IDLE 0
7480 #define RCU_MIGRATION_NEED_QS 1
7481 #define RCU_MIGRATION_GOT_QS 2
7482 #define RCU_MIGRATION_MUST_SYNC 3
7485 * migration_thread - this is a highprio system thread that performs
7486 * thread migration by bumping thread off CPU then 'pushing' onto
7489 static int migration_thread(void *data)
7492 int cpu = (long)data;
7496 BUG_ON(rq->migration_thread != current);
7498 set_current_state(TASK_INTERRUPTIBLE);
7499 while (!kthread_should_stop()) {
7500 struct migration_req *req;
7501 struct list_head *head;
7503 spin_lock_irq(&rq->lock);
7505 if (cpu_is_offline(cpu)) {
7506 spin_unlock_irq(&rq->lock);
7510 if (rq->active_balance) {
7511 active_load_balance(rq, cpu);
7512 rq->active_balance = 0;
7515 head = &rq->migration_queue;
7517 if (list_empty(head)) {
7518 spin_unlock_irq(&rq->lock);
7520 set_current_state(TASK_INTERRUPTIBLE);
7523 req = list_entry(head->next, struct migration_req, list);
7524 list_del_init(head->next);
7526 if (req->task != NULL) {
7527 spin_unlock(&rq->lock);
7528 __migrate_task(req->task, cpu, req->dest_cpu);
7529 } else if (likely(cpu == (badcpu = smp_processor_id()))) {
7530 req->dest_cpu = RCU_MIGRATION_GOT_QS;
7531 spin_unlock(&rq->lock);
7533 req->dest_cpu = RCU_MIGRATION_MUST_SYNC;
7534 spin_unlock(&rq->lock);
7535 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu, cpu);
7539 complete(&req->done);
7541 __set_current_state(TASK_RUNNING);
7546 #ifdef CONFIG_HOTPLUG_CPU
7548 * Figure out where task on dead CPU should go, use force if necessary.
7550 void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
7552 struct rq *rq = cpu_rq(dead_cpu);
7553 int needs_cpu, uninitialized_var(dest_cpu);
7554 unsigned long flags;
7556 local_irq_save(flags);
7558 spin_lock(&rq->lock);
7559 needs_cpu = (task_cpu(p) == dead_cpu) && (p->state != TASK_WAKING);
7561 dest_cpu = select_fallback_rq(dead_cpu, p);
7562 spin_unlock(&rq->lock);
7564 * It can only fail if we race with set_cpus_allowed(),
7565 * in the racer should migrate the task anyway.
7568 __migrate_task(p, dead_cpu, dest_cpu);
7569 local_irq_restore(flags);
7573 * While a dead CPU has no uninterruptible tasks queued at this point,
7574 * it might still have a nonzero ->nr_uninterruptible counter, because
7575 * for performance reasons the counter is not stricly tracking tasks to
7576 * their home CPUs. So we just add the counter to another CPU's counter,
7577 * to keep the global sum constant after CPU-down:
7579 static void migrate_nr_uninterruptible(struct rq *rq_src)
7581 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
7582 unsigned long flags;
7584 local_irq_save(flags);
7585 double_rq_lock(rq_src, rq_dest);
7586 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
7587 rq_src->nr_uninterruptible = 0;
7588 double_rq_unlock(rq_src, rq_dest);
7589 local_irq_restore(flags);
7592 /* Run through task list and migrate tasks from the dead cpu. */
7593 static void migrate_live_tasks(int src_cpu)
7595 struct task_struct *p, *t;
7597 read_lock(&tasklist_lock);
7599 do_each_thread(t, p) {
7603 if (task_cpu(p) == src_cpu)
7604 move_task_off_dead_cpu(src_cpu, p);
7605 } while_each_thread(t, p);
7607 read_unlock(&tasklist_lock);
7611 * Schedules idle task to be the next runnable task on current CPU.
7612 * It does so by boosting its priority to highest possible.
7613 * Used by CPU offline code.
7615 void sched_idle_next(void)
7617 int this_cpu = smp_processor_id();
7618 struct rq *rq = cpu_rq(this_cpu);
7619 struct task_struct *p = rq->idle;
7620 unsigned long flags;
7622 /* cpu has to be offline */
7623 BUG_ON(cpu_online(this_cpu));
7626 * Strictly not necessary since rest of the CPUs are stopped by now
7627 * and interrupts disabled on the current cpu.
7629 spin_lock_irqsave(&rq->lock, flags);
7631 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7633 update_rq_clock(rq);
7634 activate_task(rq, p, 0);
7636 spin_unlock_irqrestore(&rq->lock, flags);
7640 * Ensures that the idle task is using init_mm right before its cpu goes
7643 void idle_task_exit(void)
7645 struct mm_struct *mm = current->active_mm;
7647 BUG_ON(cpu_online(smp_processor_id()));
7650 switch_mm(mm, &init_mm, current);
7654 /* called under rq->lock with disabled interrupts */
7655 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
7657 struct rq *rq = cpu_rq(dead_cpu);
7659 /* Must be exiting, otherwise would be on tasklist. */
7660 BUG_ON(!p->exit_state);
7662 /* Cannot have done final schedule yet: would have vanished. */
7663 BUG_ON(p->state == TASK_DEAD);
7668 * Drop lock around migration; if someone else moves it,
7669 * that's OK. No task can be added to this CPU, so iteration is
7672 spin_unlock_irq(&rq->lock);
7673 move_task_off_dead_cpu(dead_cpu, p);
7674 spin_lock_irq(&rq->lock);
7679 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7680 static void migrate_dead_tasks(unsigned int dead_cpu)
7682 struct rq *rq = cpu_rq(dead_cpu);
7683 struct task_struct *next;
7686 if (!rq->nr_running)
7688 update_rq_clock(rq);
7689 next = pick_next_task(rq);
7692 next->sched_class->put_prev_task(rq, next);
7693 migrate_dead(dead_cpu, next);
7699 * remove the tasks which were accounted by rq from calc_load_tasks.
7701 static void calc_global_load_remove(struct rq *rq)
7703 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
7704 rq->calc_load_active = 0;
7706 #endif /* CONFIG_HOTPLUG_CPU */
7708 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7710 static struct ctl_table sd_ctl_dir[] = {
7712 .procname = "sched_domain",
7718 static struct ctl_table sd_ctl_root[] = {
7720 .ctl_name = CTL_KERN,
7721 .procname = "kernel",
7723 .child = sd_ctl_dir,
7728 static struct ctl_table *sd_alloc_ctl_entry(int n)
7730 struct ctl_table *entry =
7731 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
7736 static void sd_free_ctl_entry(struct ctl_table **tablep)
7738 struct ctl_table *entry;
7741 * In the intermediate directories, both the child directory and
7742 * procname are dynamically allocated and could fail but the mode
7743 * will always be set. In the lowest directory the names are
7744 * static strings and all have proc handlers.
7746 for (entry = *tablep; entry->mode; entry++) {
7748 sd_free_ctl_entry(&entry->child);
7749 if (entry->proc_handler == NULL)
7750 kfree(entry->procname);
7758 set_table_entry(struct ctl_table *entry,
7759 const char *procname, void *data, int maxlen,
7760 mode_t mode, proc_handler *proc_handler)
7762 entry->procname = procname;
7764 entry->maxlen = maxlen;
7766 entry->proc_handler = proc_handler;
7769 static struct ctl_table *
7770 sd_alloc_ctl_domain_table(struct sched_domain *sd)
7772 struct ctl_table *table = sd_alloc_ctl_entry(13);
7777 set_table_entry(&table[0], "min_interval", &sd->min_interval,
7778 sizeof(long), 0644, proc_doulongvec_minmax);
7779 set_table_entry(&table[1], "max_interval", &sd->max_interval,
7780 sizeof(long), 0644, proc_doulongvec_minmax);
7781 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
7782 sizeof(int), 0644, proc_dointvec_minmax);
7783 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
7784 sizeof(int), 0644, proc_dointvec_minmax);
7785 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
7786 sizeof(int), 0644, proc_dointvec_minmax);
7787 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
7788 sizeof(int), 0644, proc_dointvec_minmax);
7789 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
7790 sizeof(int), 0644, proc_dointvec_minmax);
7791 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
7792 sizeof(int), 0644, proc_dointvec_minmax);
7793 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
7794 sizeof(int), 0644, proc_dointvec_minmax);
7795 set_table_entry(&table[9], "cache_nice_tries",
7796 &sd->cache_nice_tries,
7797 sizeof(int), 0644, proc_dointvec_minmax);
7798 set_table_entry(&table[10], "flags", &sd->flags,
7799 sizeof(int), 0644, proc_dointvec_minmax);
7800 set_table_entry(&table[11], "name", sd->name,
7801 CORENAME_MAX_SIZE, 0444, proc_dostring);
7802 /* &table[12] is terminator */
7807 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
7809 struct ctl_table *entry, *table;
7810 struct sched_domain *sd;
7811 int domain_num = 0, i;
7814 for_each_domain(cpu, sd)
7816 entry = table = sd_alloc_ctl_entry(domain_num + 1);
7821 for_each_domain(cpu, sd) {
7822 snprintf(buf, 32, "domain%d", i);
7823 entry->procname = kstrdup(buf, GFP_KERNEL);
7825 entry->child = sd_alloc_ctl_domain_table(sd);
7832 static struct ctl_table_header *sd_sysctl_header;
7833 static void register_sched_domain_sysctl(void)
7835 int i, cpu_num = num_possible_cpus();
7836 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
7839 WARN_ON(sd_ctl_dir[0].child);
7840 sd_ctl_dir[0].child = entry;
7845 for_each_possible_cpu(i) {
7846 snprintf(buf, 32, "cpu%d", i);
7847 entry->procname = kstrdup(buf, GFP_KERNEL);
7849 entry->child = sd_alloc_ctl_cpu_table(i);
7853 WARN_ON(sd_sysctl_header);
7854 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
7857 /* may be called multiple times per register */
7858 static void unregister_sched_domain_sysctl(void)
7860 if (sd_sysctl_header)
7861 unregister_sysctl_table(sd_sysctl_header);
7862 sd_sysctl_header = NULL;
7863 if (sd_ctl_dir[0].child)
7864 sd_free_ctl_entry(&sd_ctl_dir[0].child);
7867 static void register_sched_domain_sysctl(void)
7870 static void unregister_sched_domain_sysctl(void)
7875 static void set_rq_online(struct rq *rq)
7878 const struct sched_class *class;
7880 cpumask_set_cpu(rq->cpu, rq->rd->online);
7883 for_each_class(class) {
7884 if (class->rq_online)
7885 class->rq_online(rq);
7890 static void set_rq_offline(struct rq *rq)
7893 const struct sched_class *class;
7895 for_each_class(class) {
7896 if (class->rq_offline)
7897 class->rq_offline(rq);
7900 cpumask_clear_cpu(rq->cpu, rq->rd->online);
7906 * migration_call - callback that gets triggered when a CPU is added.
7907 * Here we can start up the necessary migration thread for the new CPU.
7909 static int __cpuinit
7910 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
7912 struct task_struct *p;
7913 int cpu = (long)hcpu;
7914 unsigned long flags;
7917 switch (action & ~CPU_TASKS_FROZEN) {
7919 case CPU_UP_PREPARE:
7920 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
7923 kthread_bind(p, cpu);
7924 /* Must be high prio: stop_machine expects to yield to it. */
7925 rq = task_rq_lock(p, &flags);
7926 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7927 task_rq_unlock(rq, &flags);
7929 cpu_rq(cpu)->migration_thread = p;
7930 rq->calc_load_update = calc_load_update;
7934 /* Strictly unnecessary, as first user will wake it. */
7935 wake_up_process(cpu_rq(cpu)->migration_thread);
7937 /* Update our root-domain */
7939 spin_lock_irqsave(&rq->lock, flags);
7941 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7945 spin_unlock_irqrestore(&rq->lock, flags);
7948 #ifdef CONFIG_HOTPLUG_CPU
7949 case CPU_UP_CANCELED:
7950 if (!cpu_rq(cpu)->migration_thread)
7952 /* Unbind it from offline cpu so it can run. Fall thru. */
7953 kthread_bind(cpu_rq(cpu)->migration_thread,
7954 cpumask_any(cpu_online_mask));
7955 kthread_stop(cpu_rq(cpu)->migration_thread);
7956 put_task_struct(cpu_rq(cpu)->migration_thread);
7957 cpu_rq(cpu)->migration_thread = NULL;
7962 * Bring the migration thread down in CPU_POST_DEAD event,
7963 * since the timers should have got migrated by now and thus
7964 * we should not see a deadlock between trying to kill the
7965 * migration thread and the sched_rt_period_timer.
7968 kthread_stop(rq->migration_thread);
7969 put_task_struct(rq->migration_thread);
7970 rq->migration_thread = NULL;
7974 migrate_live_tasks(cpu);
7976 /* Idle task back to normal (off runqueue, low prio) */
7977 spin_lock_irq(&rq->lock);
7978 update_rq_clock(rq);
7979 deactivate_task(rq, rq->idle, 0);
7980 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
7981 rq->idle->sched_class = &idle_sched_class;
7982 migrate_dead_tasks(cpu);
7983 spin_unlock_irq(&rq->lock);
7984 migrate_nr_uninterruptible(rq);
7985 BUG_ON(rq->nr_running != 0);
7986 calc_global_load_remove(rq);
7988 * No need to migrate the tasks: it was best-effort if
7989 * they didn't take sched_hotcpu_mutex. Just wake up
7992 spin_lock_irq(&rq->lock);
7993 while (!list_empty(&rq->migration_queue)) {
7994 struct migration_req *req;
7996 req = list_entry(rq->migration_queue.next,
7997 struct migration_req, list);
7998 list_del_init(&req->list);
7999 spin_unlock_irq(&rq->lock);
8000 complete(&req->done);
8001 spin_lock_irq(&rq->lock);
8003 spin_unlock_irq(&rq->lock);
8007 /* Update our root-domain */
8009 spin_lock_irqsave(&rq->lock, flags);
8011 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
8014 spin_unlock_irqrestore(&rq->lock, flags);
8022 * Register at high priority so that task migration (migrate_all_tasks)
8023 * happens before everything else. This has to be lower priority than
8024 * the notifier in the perf_event subsystem, though.
8026 static struct notifier_block __cpuinitdata migration_notifier = {
8027 .notifier_call = migration_call,
8031 static int __init migration_init(void)
8033 void *cpu = (void *)(long)smp_processor_id();
8036 /* Start one for the boot CPU: */
8037 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
8038 BUG_ON(err == NOTIFY_BAD);
8039 migration_call(&migration_notifier, CPU_ONLINE, cpu);
8040 register_cpu_notifier(&migration_notifier);
8044 early_initcall(migration_init);
8049 #ifdef CONFIG_SCHED_DEBUG
8051 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
8052 struct cpumask *groupmask)
8054 struct sched_group *group = sd->groups;
8057 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
8058 cpumask_clear(groupmask);
8060 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
8062 if (!(sd->flags & SD_LOAD_BALANCE)) {
8063 printk("does not load-balance\n");
8065 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
8070 printk(KERN_CONT "span %s level %s\n", str, sd->name);
8072 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
8073 printk(KERN_ERR "ERROR: domain->span does not contain "
8076 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
8077 printk(KERN_ERR "ERROR: domain->groups does not contain"
8081 printk(KERN_DEBUG "%*s groups:", level + 1, "");
8085 printk(KERN_ERR "ERROR: group is NULL\n");
8089 if (!group->cpu_power) {
8090 printk(KERN_CONT "\n");
8091 printk(KERN_ERR "ERROR: domain->cpu_power not "
8096 if (!cpumask_weight(sched_group_cpus(group))) {
8097 printk(KERN_CONT "\n");
8098 printk(KERN_ERR "ERROR: empty group\n");
8102 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
8103 printk(KERN_CONT "\n");
8104 printk(KERN_ERR "ERROR: repeated CPUs\n");
8108 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
8110 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
8112 printk(KERN_CONT " %s", str);
8113 if (group->cpu_power != SCHED_LOAD_SCALE) {
8114 printk(KERN_CONT " (cpu_power = %d)",
8118 group = group->next;
8119 } while (group != sd->groups);
8120 printk(KERN_CONT "\n");
8122 if (!cpumask_equal(sched_domain_span(sd), groupmask))
8123 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
8126 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
8127 printk(KERN_ERR "ERROR: parent span is not a superset "
8128 "of domain->span\n");
8132 static void sched_domain_debug(struct sched_domain *sd, int cpu)
8134 cpumask_var_t groupmask;
8138 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
8142 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
8144 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
8145 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
8150 if (sched_domain_debug_one(sd, cpu, level, groupmask))
8157 free_cpumask_var(groupmask);
8159 #else /* !CONFIG_SCHED_DEBUG */
8160 # define sched_domain_debug(sd, cpu) do { } while (0)
8161 #endif /* CONFIG_SCHED_DEBUG */
8163 static int sd_degenerate(struct sched_domain *sd)
8165 if (cpumask_weight(sched_domain_span(sd)) == 1)
8168 /* Following flags need at least 2 groups */
8169 if (sd->flags & (SD_LOAD_BALANCE |
8170 SD_BALANCE_NEWIDLE |
8174 SD_SHARE_PKG_RESOURCES)) {
8175 if (sd->groups != sd->groups->next)
8179 /* Following flags don't use groups */
8180 if (sd->flags & (SD_WAKE_AFFINE))
8187 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
8189 unsigned long cflags = sd->flags, pflags = parent->flags;
8191 if (sd_degenerate(parent))
8194 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
8197 /* Flags needing groups don't count if only 1 group in parent */
8198 if (parent->groups == parent->groups->next) {
8199 pflags &= ~(SD_LOAD_BALANCE |
8200 SD_BALANCE_NEWIDLE |
8204 SD_SHARE_PKG_RESOURCES);
8205 if (nr_node_ids == 1)
8206 pflags &= ~SD_SERIALIZE;
8208 if (~cflags & pflags)
8214 static void free_rootdomain(struct root_domain *rd)
8216 synchronize_sched();
8218 cpupri_cleanup(&rd->cpupri);
8220 free_cpumask_var(rd->rto_mask);
8221 free_cpumask_var(rd->online);
8222 free_cpumask_var(rd->span);
8226 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
8228 struct root_domain *old_rd = NULL;
8229 unsigned long flags;
8231 spin_lock_irqsave(&rq->lock, flags);
8236 if (cpumask_test_cpu(rq->cpu, old_rd->online))
8239 cpumask_clear_cpu(rq->cpu, old_rd->span);
8242 * If we dont want to free the old_rt yet then
8243 * set old_rd to NULL to skip the freeing later
8246 if (!atomic_dec_and_test(&old_rd->refcount))
8250 atomic_inc(&rd->refcount);
8253 cpumask_set_cpu(rq->cpu, rd->span);
8254 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
8257 spin_unlock_irqrestore(&rq->lock, flags);
8260 free_rootdomain(old_rd);
8263 static int init_rootdomain(struct root_domain *rd, bool bootmem)
8265 gfp_t gfp = GFP_KERNEL;
8267 memset(rd, 0, sizeof(*rd));
8272 if (!alloc_cpumask_var(&rd->span, gfp))
8274 if (!alloc_cpumask_var(&rd->online, gfp))
8276 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
8279 if (cpupri_init(&rd->cpupri, bootmem) != 0)
8284 free_cpumask_var(rd->rto_mask);
8286 free_cpumask_var(rd->online);
8288 free_cpumask_var(rd->span);
8293 static void init_defrootdomain(void)
8295 init_rootdomain(&def_root_domain, true);
8297 atomic_set(&def_root_domain.refcount, 1);
8300 static struct root_domain *alloc_rootdomain(void)
8302 struct root_domain *rd;
8304 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
8308 if (init_rootdomain(rd, false) != 0) {
8317 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
8318 * hold the hotplug lock.
8321 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
8323 struct rq *rq = cpu_rq(cpu);
8324 struct sched_domain *tmp;
8326 for (tmp = sd; tmp; tmp = tmp->parent)
8327 tmp->span_weight = cpumask_weight(sched_domain_span(tmp));
8329 /* Remove the sched domains which do not contribute to scheduling. */
8330 for (tmp = sd; tmp; ) {
8331 struct sched_domain *parent = tmp->parent;
8335 if (sd_parent_degenerate(tmp, parent)) {
8336 tmp->parent = parent->parent;
8338 parent->parent->child = tmp;
8343 if (sd && sd_degenerate(sd)) {
8349 sched_domain_debug(sd, cpu);
8351 rq_attach_root(rq, rd);
8352 rcu_assign_pointer(rq->sd, sd);
8355 /* cpus with isolated domains */
8356 static cpumask_var_t cpu_isolated_map;
8358 /* Setup the mask of cpus configured for isolated domains */
8359 static int __init isolated_cpu_setup(char *str)
8361 alloc_bootmem_cpumask_var(&cpu_isolated_map);
8362 cpulist_parse(str, cpu_isolated_map);
8366 __setup("isolcpus=", isolated_cpu_setup);
8369 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
8370 * to a function which identifies what group(along with sched group) a CPU
8371 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
8372 * (due to the fact that we keep track of groups covered with a struct cpumask).
8374 * init_sched_build_groups will build a circular linked list of the groups
8375 * covered by the given span, and will set each group's ->cpumask correctly,
8376 * and ->cpu_power to 0.
8379 init_sched_build_groups(const struct cpumask *span,
8380 const struct cpumask *cpu_map,
8381 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
8382 struct sched_group **sg,
8383 struct cpumask *tmpmask),
8384 struct cpumask *covered, struct cpumask *tmpmask)
8386 struct sched_group *first = NULL, *last = NULL;
8389 cpumask_clear(covered);
8391 for_each_cpu(i, span) {
8392 struct sched_group *sg;
8393 int group = group_fn(i, cpu_map, &sg, tmpmask);
8396 if (cpumask_test_cpu(i, covered))
8399 cpumask_clear(sched_group_cpus(sg));
8402 for_each_cpu(j, span) {
8403 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
8406 cpumask_set_cpu(j, covered);
8407 cpumask_set_cpu(j, sched_group_cpus(sg));
8418 #define SD_NODES_PER_DOMAIN 16
8423 * find_next_best_node - find the next node to include in a sched_domain
8424 * @node: node whose sched_domain we're building
8425 * @used_nodes: nodes already in the sched_domain
8427 * Find the next node to include in a given scheduling domain. Simply
8428 * finds the closest node not already in the @used_nodes map.
8430 * Should use nodemask_t.
8432 static int find_next_best_node(int node, nodemask_t *used_nodes)
8434 int i, n, val, min_val, best_node = 0;
8438 for (i = 0; i < nr_node_ids; i++) {
8439 /* Start at @node */
8440 n = (node + i) % nr_node_ids;
8442 if (!nr_cpus_node(n))
8445 /* Skip already used nodes */
8446 if (node_isset(n, *used_nodes))
8449 /* Simple min distance search */
8450 val = node_distance(node, n);
8452 if (val < min_val) {
8458 node_set(best_node, *used_nodes);
8463 * sched_domain_node_span - get a cpumask for a node's sched_domain
8464 * @node: node whose cpumask we're constructing
8465 * @span: resulting cpumask
8467 * Given a node, construct a good cpumask for its sched_domain to span. It
8468 * should be one that prevents unnecessary balancing, but also spreads tasks
8471 static void sched_domain_node_span(int node, struct cpumask *span)
8473 nodemask_t used_nodes;
8476 cpumask_clear(span);
8477 nodes_clear(used_nodes);
8479 cpumask_or(span, span, cpumask_of_node(node));
8480 node_set(node, used_nodes);
8482 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
8483 int next_node = find_next_best_node(node, &used_nodes);
8485 cpumask_or(span, span, cpumask_of_node(next_node));
8488 #endif /* CONFIG_NUMA */
8490 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
8493 * The cpus mask in sched_group and sched_domain hangs off the end.
8495 * ( See the the comments in include/linux/sched.h:struct sched_group
8496 * and struct sched_domain. )
8498 struct static_sched_group {
8499 struct sched_group sg;
8500 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
8503 struct static_sched_domain {
8504 struct sched_domain sd;
8505 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
8511 cpumask_var_t domainspan;
8512 cpumask_var_t covered;
8513 cpumask_var_t notcovered;
8515 cpumask_var_t nodemask;
8516 cpumask_var_t this_sibling_map;
8517 cpumask_var_t this_core_map;
8518 cpumask_var_t send_covered;
8519 cpumask_var_t tmpmask;
8520 struct sched_group **sched_group_nodes;
8521 struct root_domain *rd;
8525 sa_sched_groups = 0,
8530 sa_this_sibling_map,
8532 sa_sched_group_nodes,
8542 * SMT sched-domains:
8544 #ifdef CONFIG_SCHED_SMT
8545 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
8546 static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
8549 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
8550 struct sched_group **sg, struct cpumask *unused)
8553 *sg = &per_cpu(sched_group_cpus, cpu).sg;
8556 #endif /* CONFIG_SCHED_SMT */
8559 * multi-core sched-domains:
8561 #ifdef CONFIG_SCHED_MC
8562 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
8563 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
8564 #endif /* CONFIG_SCHED_MC */
8566 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8568 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8569 struct sched_group **sg, struct cpumask *mask)
8573 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8574 group = cpumask_first(mask);
8576 *sg = &per_cpu(sched_group_core, group).sg;
8579 #elif defined(CONFIG_SCHED_MC)
8581 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8582 struct sched_group **sg, struct cpumask *unused)
8585 *sg = &per_cpu(sched_group_core, cpu).sg;
8590 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
8591 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
8594 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
8595 struct sched_group **sg, struct cpumask *mask)
8598 #ifdef CONFIG_SCHED_MC
8599 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
8600 group = cpumask_first(mask);
8601 #elif defined(CONFIG_SCHED_SMT)
8602 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8603 group = cpumask_first(mask);
8608 *sg = &per_cpu(sched_group_phys, group).sg;
8614 * The init_sched_build_groups can't handle what we want to do with node
8615 * groups, so roll our own. Now each node has its own list of groups which
8616 * gets dynamically allocated.
8618 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
8619 static struct sched_group ***sched_group_nodes_bycpu;
8621 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
8622 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
8624 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
8625 struct sched_group **sg,
8626 struct cpumask *nodemask)
8630 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
8631 group = cpumask_first(nodemask);
8634 *sg = &per_cpu(sched_group_allnodes, group).sg;
8638 static void init_numa_sched_groups_power(struct sched_group *group_head)
8640 struct sched_group *sg = group_head;
8646 for_each_cpu(j, sched_group_cpus(sg)) {
8647 struct sched_domain *sd;
8649 sd = &per_cpu(phys_domains, j).sd;
8650 if (j != group_first_cpu(sd->groups)) {
8652 * Only add "power" once for each
8658 sg->cpu_power += sd->groups->cpu_power;
8661 } while (sg != group_head);
8664 static int build_numa_sched_groups(struct s_data *d,
8665 const struct cpumask *cpu_map, int num)
8667 struct sched_domain *sd;
8668 struct sched_group *sg, *prev;
8671 cpumask_clear(d->covered);
8672 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
8673 if (cpumask_empty(d->nodemask)) {
8674 d->sched_group_nodes[num] = NULL;
8678 sched_domain_node_span(num, d->domainspan);
8679 cpumask_and(d->domainspan, d->domainspan, cpu_map);
8681 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8684 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
8688 d->sched_group_nodes[num] = sg;
8690 for_each_cpu(j, d->nodemask) {
8691 sd = &per_cpu(node_domains, j).sd;
8696 cpumask_copy(sched_group_cpus(sg), d->nodemask);
8698 cpumask_or(d->covered, d->covered, d->nodemask);
8701 for (j = 0; j < nr_node_ids; j++) {
8702 n = (num + j) % nr_node_ids;
8703 cpumask_complement(d->notcovered, d->covered);
8704 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
8705 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
8706 if (cpumask_empty(d->tmpmask))
8708 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
8709 if (cpumask_empty(d->tmpmask))
8711 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8715 "Can not alloc domain group for node %d\n", j);
8719 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
8720 sg->next = prev->next;
8721 cpumask_or(d->covered, d->covered, d->tmpmask);
8728 #endif /* CONFIG_NUMA */
8731 /* Free memory allocated for various sched_group structures */
8732 static void free_sched_groups(const struct cpumask *cpu_map,
8733 struct cpumask *nodemask)
8737 for_each_cpu(cpu, cpu_map) {
8738 struct sched_group **sched_group_nodes
8739 = sched_group_nodes_bycpu[cpu];
8741 if (!sched_group_nodes)
8744 for (i = 0; i < nr_node_ids; i++) {
8745 struct sched_group *oldsg, *sg = sched_group_nodes[i];
8747 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8748 if (cpumask_empty(nodemask))
8758 if (oldsg != sched_group_nodes[i])
8761 kfree(sched_group_nodes);
8762 sched_group_nodes_bycpu[cpu] = NULL;
8765 #else /* !CONFIG_NUMA */
8766 static void free_sched_groups(const struct cpumask *cpu_map,
8767 struct cpumask *nodemask)
8770 #endif /* CONFIG_NUMA */
8773 * Initialize sched groups cpu_power.
8775 * cpu_power indicates the capacity of sched group, which is used while
8776 * distributing the load between different sched groups in a sched domain.
8777 * Typically cpu_power for all the groups in a sched domain will be same unless
8778 * there are asymmetries in the topology. If there are asymmetries, group
8779 * having more cpu_power will pickup more load compared to the group having
8782 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
8784 struct sched_domain *child;
8785 struct sched_group *group;
8789 WARN_ON(!sd || !sd->groups);
8791 if (cpu != group_first_cpu(sd->groups))
8794 sd->groups->group_weight = cpumask_weight(sched_group_cpus(sd->groups));
8798 sd->groups->cpu_power = 0;
8801 power = SCHED_LOAD_SCALE;
8802 weight = cpumask_weight(sched_domain_span(sd));
8804 * SMT siblings share the power of a single core.
8805 * Usually multiple threads get a better yield out of
8806 * that one core than a single thread would have,
8807 * reflect that in sd->smt_gain.
8809 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
8810 power *= sd->smt_gain;
8812 power >>= SCHED_LOAD_SHIFT;
8814 sd->groups->cpu_power += power;
8819 * Add cpu_power of each child group to this groups cpu_power.
8821 group = child->groups;
8823 sd->groups->cpu_power += group->cpu_power;
8824 group = group->next;
8825 } while (group != child->groups);
8829 * Initializers for schedule domains
8830 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8833 #ifdef CONFIG_SCHED_DEBUG
8834 # define SD_INIT_NAME(sd, type) sd->name = #type
8836 # define SD_INIT_NAME(sd, type) do { } while (0)
8839 #define SD_INIT(sd, type) sd_init_##type(sd)
8841 #define SD_INIT_FUNC(type) \
8842 static noinline void sd_init_##type(struct sched_domain *sd) \
8844 memset(sd, 0, sizeof(*sd)); \
8845 *sd = SD_##type##_INIT; \
8846 sd->level = SD_LV_##type; \
8847 SD_INIT_NAME(sd, type); \
8852 SD_INIT_FUNC(ALLNODES)
8855 #ifdef CONFIG_SCHED_SMT
8856 SD_INIT_FUNC(SIBLING)
8858 #ifdef CONFIG_SCHED_MC
8862 static int default_relax_domain_level = -1;
8864 static int __init setup_relax_domain_level(char *str)
8868 val = simple_strtoul(str, NULL, 0);
8869 if (val < SD_LV_MAX)
8870 default_relax_domain_level = val;
8874 __setup("relax_domain_level=", setup_relax_domain_level);
8876 static void set_domain_attribute(struct sched_domain *sd,
8877 struct sched_domain_attr *attr)
8881 if (!attr || attr->relax_domain_level < 0) {
8882 if (default_relax_domain_level < 0)
8885 request = default_relax_domain_level;
8887 request = attr->relax_domain_level;
8888 if (request < sd->level) {
8889 /* turn off idle balance on this domain */
8890 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
8892 /* turn on idle balance on this domain */
8893 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
8897 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
8898 const struct cpumask *cpu_map)
8901 case sa_sched_groups:
8902 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
8903 d->sched_group_nodes = NULL;
8905 free_rootdomain(d->rd); /* fall through */
8907 free_cpumask_var(d->tmpmask); /* fall through */
8908 case sa_send_covered:
8909 free_cpumask_var(d->send_covered); /* fall through */
8910 case sa_this_core_map:
8911 free_cpumask_var(d->this_core_map); /* fall through */
8912 case sa_this_sibling_map:
8913 free_cpumask_var(d->this_sibling_map); /* fall through */
8915 free_cpumask_var(d->nodemask); /* fall through */
8916 case sa_sched_group_nodes:
8918 kfree(d->sched_group_nodes); /* fall through */
8920 free_cpumask_var(d->notcovered); /* fall through */
8922 free_cpumask_var(d->covered); /* fall through */
8924 free_cpumask_var(d->domainspan); /* fall through */
8931 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
8932 const struct cpumask *cpu_map)
8935 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
8937 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
8938 return sa_domainspan;
8939 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
8941 /* Allocate the per-node list of sched groups */
8942 d->sched_group_nodes = kcalloc(nr_node_ids,
8943 sizeof(struct sched_group *), GFP_KERNEL);
8944 if (!d->sched_group_nodes) {
8945 printk(KERN_WARNING "Can not alloc sched group node list\n");
8946 return sa_notcovered;
8948 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
8950 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
8951 return sa_sched_group_nodes;
8952 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
8954 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
8955 return sa_this_sibling_map;
8956 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
8957 return sa_this_core_map;
8958 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
8959 return sa_send_covered;
8960 d->rd = alloc_rootdomain();
8962 printk(KERN_WARNING "Cannot alloc root domain\n");
8965 return sa_rootdomain;
8968 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
8969 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
8971 struct sched_domain *sd = NULL;
8973 struct sched_domain *parent;
8976 if (cpumask_weight(cpu_map) >
8977 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
8978 sd = &per_cpu(allnodes_domains, i).sd;
8979 SD_INIT(sd, ALLNODES);
8980 set_domain_attribute(sd, attr);
8981 cpumask_copy(sched_domain_span(sd), cpu_map);
8982 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
8987 sd = &per_cpu(node_domains, i).sd;
8989 set_domain_attribute(sd, attr);
8990 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
8991 sd->parent = parent;
8994 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
8999 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
9000 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
9001 struct sched_domain *parent, int i)
9003 struct sched_domain *sd;
9004 sd = &per_cpu(phys_domains, i).sd;
9006 set_domain_attribute(sd, attr);
9007 cpumask_copy(sched_domain_span(sd), d->nodemask);
9008 sd->parent = parent;
9011 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
9015 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
9016 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
9017 struct sched_domain *parent, int i)
9019 struct sched_domain *sd = parent;
9020 #ifdef CONFIG_SCHED_MC
9021 sd = &per_cpu(core_domains, i).sd;
9023 set_domain_attribute(sd, attr);
9024 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
9025 sd->parent = parent;
9027 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
9032 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
9033 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
9034 struct sched_domain *parent, int i)
9036 struct sched_domain *sd = parent;
9037 #ifdef CONFIG_SCHED_SMT
9038 sd = &per_cpu(cpu_domains, i).sd;
9039 SD_INIT(sd, SIBLING);
9040 set_domain_attribute(sd, attr);
9041 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
9042 sd->parent = parent;
9044 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
9049 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
9050 const struct cpumask *cpu_map, int cpu)
9053 #ifdef CONFIG_SCHED_SMT
9054 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
9055 cpumask_and(d->this_sibling_map, cpu_map,
9056 topology_thread_cpumask(cpu));
9057 if (cpu == cpumask_first(d->this_sibling_map))
9058 init_sched_build_groups(d->this_sibling_map, cpu_map,
9060 d->send_covered, d->tmpmask);
9063 #ifdef CONFIG_SCHED_MC
9064 case SD_LV_MC: /* set up multi-core groups */
9065 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
9066 if (cpu == cpumask_first(d->this_core_map))
9067 init_sched_build_groups(d->this_core_map, cpu_map,
9069 d->send_covered, d->tmpmask);
9072 case SD_LV_CPU: /* set up physical groups */
9073 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
9074 if (!cpumask_empty(d->nodemask))
9075 init_sched_build_groups(d->nodemask, cpu_map,
9077 d->send_covered, d->tmpmask);
9080 case SD_LV_ALLNODES:
9081 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
9082 d->send_covered, d->tmpmask);
9091 * Build sched domains for a given set of cpus and attach the sched domains
9092 * to the individual cpus
9094 static int __build_sched_domains(const struct cpumask *cpu_map,
9095 struct sched_domain_attr *attr)
9097 enum s_alloc alloc_state = sa_none;
9099 struct sched_domain *sd;
9105 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
9106 if (alloc_state != sa_rootdomain)
9108 alloc_state = sa_sched_groups;
9111 * Set up domains for cpus specified by the cpu_map.
9113 for_each_cpu(i, cpu_map) {
9114 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
9117 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
9118 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
9119 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
9120 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
9123 for_each_cpu(i, cpu_map) {
9124 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
9125 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
9128 /* Set up physical groups */
9129 for (i = 0; i < nr_node_ids; i++)
9130 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
9133 /* Set up node groups */
9135 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
9137 for (i = 0; i < nr_node_ids; i++)
9138 if (build_numa_sched_groups(&d, cpu_map, i))
9142 /* Calculate CPU power for physical packages and nodes */
9143 #ifdef CONFIG_SCHED_SMT
9144 for_each_cpu(i, cpu_map) {
9145 sd = &per_cpu(cpu_domains, i).sd;
9146 init_sched_groups_power(i, sd);
9149 #ifdef CONFIG_SCHED_MC
9150 for_each_cpu(i, cpu_map) {
9151 sd = &per_cpu(core_domains, i).sd;
9152 init_sched_groups_power(i, sd);
9156 for_each_cpu(i, cpu_map) {
9157 sd = &per_cpu(phys_domains, i).sd;
9158 init_sched_groups_power(i, sd);
9162 for (i = 0; i < nr_node_ids; i++)
9163 init_numa_sched_groups_power(d.sched_group_nodes[i]);
9165 if (d.sd_allnodes) {
9166 struct sched_group *sg;
9168 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
9170 init_numa_sched_groups_power(sg);
9174 /* Attach the domains */
9175 for_each_cpu(i, cpu_map) {
9176 #ifdef CONFIG_SCHED_SMT
9177 sd = &per_cpu(cpu_domains, i).sd;
9178 #elif defined(CONFIG_SCHED_MC)
9179 sd = &per_cpu(core_domains, i).sd;
9181 sd = &per_cpu(phys_domains, i).sd;
9183 cpu_attach_domain(sd, d.rd, i);
9186 d.sched_group_nodes = NULL; /* don't free this we still need it */
9187 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
9191 __free_domain_allocs(&d, alloc_state, cpu_map);
9195 static int build_sched_domains(const struct cpumask *cpu_map)
9197 return __build_sched_domains(cpu_map, NULL);
9200 static struct cpumask *doms_cur; /* current sched domains */
9201 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
9202 static struct sched_domain_attr *dattr_cur;
9203 /* attribues of custom domains in 'doms_cur' */
9206 * Special case: If a kmalloc of a doms_cur partition (array of
9207 * cpumask) fails, then fallback to a single sched domain,
9208 * as determined by the single cpumask fallback_doms.
9210 static cpumask_var_t fallback_doms;
9213 * arch_update_cpu_topology lets virtualized architectures update the
9214 * cpu core maps. It is supposed to return 1 if the topology changed
9215 * or 0 if it stayed the same.
9217 int __attribute__((weak)) arch_update_cpu_topology(void)
9223 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
9224 * For now this just excludes isolated cpus, but could be used to
9225 * exclude other special cases in the future.
9227 static int arch_init_sched_domains(const struct cpumask *cpu_map)
9231 arch_update_cpu_topology();
9233 doms_cur = kmalloc(cpumask_size(), GFP_KERNEL);
9235 doms_cur = fallback_doms;
9236 cpumask_andnot(doms_cur, cpu_map, cpu_isolated_map);
9238 err = build_sched_domains(doms_cur);
9239 register_sched_domain_sysctl();
9244 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
9245 struct cpumask *tmpmask)
9247 free_sched_groups(cpu_map, tmpmask);
9251 * Detach sched domains from a group of cpus specified in cpu_map
9252 * These cpus will now be attached to the NULL domain
9254 static void detach_destroy_domains(const struct cpumask *cpu_map)
9256 /* Save because hotplug lock held. */
9257 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
9260 for_each_cpu(i, cpu_map)
9261 cpu_attach_domain(NULL, &def_root_domain, i);
9262 synchronize_sched();
9263 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
9266 /* handle null as "default" */
9267 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
9268 struct sched_domain_attr *new, int idx_new)
9270 struct sched_domain_attr tmp;
9277 return !memcmp(cur ? (cur + idx_cur) : &tmp,
9278 new ? (new + idx_new) : &tmp,
9279 sizeof(struct sched_domain_attr));
9283 * Partition sched domains as specified by the 'ndoms_new'
9284 * cpumasks in the array doms_new[] of cpumasks. This compares
9285 * doms_new[] to the current sched domain partitioning, doms_cur[].
9286 * It destroys each deleted domain and builds each new domain.
9288 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
9289 * The masks don't intersect (don't overlap.) We should setup one
9290 * sched domain for each mask. CPUs not in any of the cpumasks will
9291 * not be load balanced. If the same cpumask appears both in the
9292 * current 'doms_cur' domains and in the new 'doms_new', we can leave
9295 * The passed in 'doms_new' should be kmalloc'd. This routine takes
9296 * ownership of it and will kfree it when done with it. If the caller
9297 * failed the kmalloc call, then it can pass in doms_new == NULL &&
9298 * ndoms_new == 1, and partition_sched_domains() will fallback to
9299 * the single partition 'fallback_doms', it also forces the domains
9302 * If doms_new == NULL it will be replaced with cpu_online_mask.
9303 * ndoms_new == 0 is a special case for destroying existing domains,
9304 * and it will not create the default domain.
9306 * Call with hotplug lock held
9308 /* FIXME: Change to struct cpumask *doms_new[] */
9309 void partition_sched_domains(int ndoms_new, struct cpumask *doms_new,
9310 struct sched_domain_attr *dattr_new)
9315 mutex_lock(&sched_domains_mutex);
9317 /* always unregister in case we don't destroy any domains */
9318 unregister_sched_domain_sysctl();
9320 /* Let architecture update cpu core mappings. */
9321 new_topology = arch_update_cpu_topology();
9323 n = doms_new ? ndoms_new : 0;
9325 /* Destroy deleted domains */
9326 for (i = 0; i < ndoms_cur; i++) {
9327 for (j = 0; j < n && !new_topology; j++) {
9328 if (cpumask_equal(&doms_cur[i], &doms_new[j])
9329 && dattrs_equal(dattr_cur, i, dattr_new, j))
9332 /* no match - a current sched domain not in new doms_new[] */
9333 detach_destroy_domains(doms_cur + i);
9338 if (doms_new == NULL) {
9340 doms_new = fallback_doms;
9341 cpumask_andnot(&doms_new[0], cpu_active_mask, cpu_isolated_map);
9342 WARN_ON_ONCE(dattr_new);
9345 /* Build new domains */
9346 for (i = 0; i < ndoms_new; i++) {
9347 for (j = 0; j < ndoms_cur && !new_topology; j++) {
9348 if (cpumask_equal(&doms_new[i], &doms_cur[j])
9349 && dattrs_equal(dattr_new, i, dattr_cur, j))
9352 /* no match - add a new doms_new */
9353 __build_sched_domains(doms_new + i,
9354 dattr_new ? dattr_new + i : NULL);
9359 /* Remember the new sched domains */
9360 if (doms_cur != fallback_doms)
9362 kfree(dattr_cur); /* kfree(NULL) is safe */
9363 doms_cur = doms_new;
9364 dattr_cur = dattr_new;
9365 ndoms_cur = ndoms_new;
9367 register_sched_domain_sysctl();
9369 mutex_unlock(&sched_domains_mutex);
9372 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
9373 static void arch_reinit_sched_domains(void)
9377 /* Destroy domains first to force the rebuild */
9378 partition_sched_domains(0, NULL, NULL);
9380 rebuild_sched_domains();
9384 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
9386 unsigned int level = 0;
9388 if (sscanf(buf, "%u", &level) != 1)
9392 * level is always be positive so don't check for
9393 * level < POWERSAVINGS_BALANCE_NONE which is 0
9394 * What happens on 0 or 1 byte write,
9395 * need to check for count as well?
9398 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
9402 sched_smt_power_savings = level;
9404 sched_mc_power_savings = level;
9406 arch_reinit_sched_domains();
9411 #ifdef CONFIG_SCHED_MC
9412 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
9415 return sprintf(page, "%u\n", sched_mc_power_savings);
9417 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
9418 const char *buf, size_t count)
9420 return sched_power_savings_store(buf, count, 0);
9422 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
9423 sched_mc_power_savings_show,
9424 sched_mc_power_savings_store);
9427 #ifdef CONFIG_SCHED_SMT
9428 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
9431 return sprintf(page, "%u\n", sched_smt_power_savings);
9433 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
9434 const char *buf, size_t count)
9436 return sched_power_savings_store(buf, count, 1);
9438 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
9439 sched_smt_power_savings_show,
9440 sched_smt_power_savings_store);
9443 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
9447 #ifdef CONFIG_SCHED_SMT
9449 err = sysfs_create_file(&cls->kset.kobj,
9450 &attr_sched_smt_power_savings.attr);
9452 #ifdef CONFIG_SCHED_MC
9453 if (!err && mc_capable())
9454 err = sysfs_create_file(&cls->kset.kobj,
9455 &attr_sched_mc_power_savings.attr);
9459 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
9461 #ifndef CONFIG_CPUSETS
9463 * Add online and remove offline CPUs from the scheduler domains.
9464 * When cpusets are enabled they take over this function.
9466 static int update_sched_domains(struct notifier_block *nfb,
9467 unsigned long action, void *hcpu)
9471 case CPU_ONLINE_FROZEN:
9472 case CPU_DOWN_PREPARE:
9473 case CPU_DOWN_PREPARE_FROZEN:
9474 case CPU_DOWN_FAILED:
9475 case CPU_DOWN_FAILED_FROZEN:
9476 partition_sched_domains(1, NULL, NULL);
9485 static int update_runtime(struct notifier_block *nfb,
9486 unsigned long action, void *hcpu)
9488 int cpu = (int)(long)hcpu;
9491 case CPU_DOWN_PREPARE:
9492 case CPU_DOWN_PREPARE_FROZEN:
9493 disable_runtime(cpu_rq(cpu));
9496 case CPU_DOWN_FAILED:
9497 case CPU_DOWN_FAILED_FROZEN:
9499 case CPU_ONLINE_FROZEN:
9500 enable_runtime(cpu_rq(cpu));
9508 void __init sched_init_smp(void)
9510 cpumask_var_t non_isolated_cpus;
9512 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
9513 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
9515 #if defined(CONFIG_NUMA)
9516 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
9518 BUG_ON(sched_group_nodes_bycpu == NULL);
9521 mutex_lock(&sched_domains_mutex);
9522 arch_init_sched_domains(cpu_active_mask);
9523 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
9524 if (cpumask_empty(non_isolated_cpus))
9525 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
9526 mutex_unlock(&sched_domains_mutex);
9529 #ifndef CONFIG_CPUSETS
9530 /* XXX: Theoretical race here - CPU may be hotplugged now */
9531 hotcpu_notifier(update_sched_domains, 0);
9534 /* RT runtime code needs to handle some hotplug events */
9535 hotcpu_notifier(update_runtime, 0);
9539 /* Move init over to a non-isolated CPU */
9540 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
9542 sched_init_granularity();
9543 free_cpumask_var(non_isolated_cpus);
9545 init_sched_rt_class();
9548 void __init sched_init_smp(void)
9550 sched_init_granularity();
9552 #endif /* CONFIG_SMP */
9554 const_debug unsigned int sysctl_timer_migration = 1;
9556 int in_sched_functions(unsigned long addr)
9558 return in_lock_functions(addr) ||
9559 (addr >= (unsigned long)__sched_text_start
9560 && addr < (unsigned long)__sched_text_end);
9563 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
9565 cfs_rq->tasks_timeline = RB_ROOT;
9566 INIT_LIST_HEAD(&cfs_rq->tasks);
9567 #ifdef CONFIG_FAIR_GROUP_SCHED
9570 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9573 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
9575 struct rt_prio_array *array;
9578 array = &rt_rq->active;
9579 for (i = 0; i < MAX_RT_PRIO; i++) {
9580 INIT_LIST_HEAD(array->queue + i);
9581 __clear_bit(i, array->bitmap);
9583 /* delimiter for bitsearch: */
9584 __set_bit(MAX_RT_PRIO, array->bitmap);
9586 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9587 rt_rq->highest_prio.curr = MAX_RT_PRIO;
9589 rt_rq->highest_prio.next = MAX_RT_PRIO;
9593 rt_rq->rt_nr_migratory = 0;
9594 rt_rq->overloaded = 0;
9595 plist_head_init(&rt_rq->pushable_tasks, &rq->lock);
9599 rt_rq->rt_throttled = 0;
9600 rt_rq->rt_runtime = 0;
9601 spin_lock_init(&rt_rq->rt_runtime_lock);
9603 #ifdef CONFIG_RT_GROUP_SCHED
9604 rt_rq->rt_nr_boosted = 0;
9609 #ifdef CONFIG_FAIR_GROUP_SCHED
9610 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9611 struct sched_entity *se, int cpu, int add,
9612 struct sched_entity *parent)
9614 struct rq *rq = cpu_rq(cpu);
9615 tg->cfs_rq[cpu] = cfs_rq;
9616 init_cfs_rq(cfs_rq, rq);
9619 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
9622 /* se could be NULL for init_task_group */
9627 se->cfs_rq = &rq->cfs;
9629 se->cfs_rq = parent->my_q;
9632 se->load.weight = tg->shares;
9633 se->load.inv_weight = 0;
9634 se->parent = parent;
9638 #ifdef CONFIG_RT_GROUP_SCHED
9639 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
9640 struct sched_rt_entity *rt_se, int cpu, int add,
9641 struct sched_rt_entity *parent)
9643 struct rq *rq = cpu_rq(cpu);
9645 tg->rt_rq[cpu] = rt_rq;
9646 init_rt_rq(rt_rq, rq);
9648 rt_rq->rt_se = rt_se;
9649 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
9651 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
9653 tg->rt_se[cpu] = rt_se;
9658 rt_se->rt_rq = &rq->rt;
9660 rt_se->rt_rq = parent->my_q;
9662 rt_se->my_q = rt_rq;
9663 rt_se->parent = parent;
9664 INIT_LIST_HEAD(&rt_se->run_list);
9668 void __init sched_init(void)
9671 unsigned long alloc_size = 0, ptr;
9673 #ifdef CONFIG_FAIR_GROUP_SCHED
9674 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9676 #ifdef CONFIG_RT_GROUP_SCHED
9677 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9679 #ifdef CONFIG_CPUMASK_OFFSTACK
9680 alloc_size += num_possible_cpus() * cpumask_size();
9683 * As sched_init() is called before page_alloc is setup,
9684 * we use alloc_bootmem().
9687 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
9689 #ifdef CONFIG_FAIR_GROUP_SCHED
9690 init_task_group.se = (struct sched_entity **)ptr;
9691 ptr += nr_cpu_ids * sizeof(void **);
9693 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
9694 ptr += nr_cpu_ids * sizeof(void **);
9696 #endif /* CONFIG_FAIR_GROUP_SCHED */
9697 #ifdef CONFIG_RT_GROUP_SCHED
9698 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
9699 ptr += nr_cpu_ids * sizeof(void **);
9701 init_task_group.rt_rq = (struct rt_rq **)ptr;
9702 ptr += nr_cpu_ids * sizeof(void **);
9704 #endif /* CONFIG_RT_GROUP_SCHED */
9705 #ifdef CONFIG_CPUMASK_OFFSTACK
9706 for_each_possible_cpu(i) {
9707 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
9708 ptr += cpumask_size();
9710 #endif /* CONFIG_CPUMASK_OFFSTACK */
9714 init_defrootdomain();
9717 init_rt_bandwidth(&def_rt_bandwidth,
9718 global_rt_period(), global_rt_runtime());
9720 #ifdef CONFIG_RT_GROUP_SCHED
9721 init_rt_bandwidth(&init_task_group.rt_bandwidth,
9722 global_rt_period(), global_rt_runtime());
9723 #endif /* CONFIG_RT_GROUP_SCHED */
9725 #ifdef CONFIG_CGROUP_SCHED
9726 list_add(&init_task_group.list, &task_groups);
9727 INIT_LIST_HEAD(&init_task_group.children);
9729 #endif /* CONFIG_CGROUP_SCHED */
9731 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
9732 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
9733 __alignof__(unsigned long));
9735 for_each_possible_cpu(i) {
9739 spin_lock_init(&rq->lock);
9741 rq->calc_load_active = 0;
9742 rq->calc_load_update = jiffies + LOAD_FREQ;
9743 init_cfs_rq(&rq->cfs, rq);
9744 init_rt_rq(&rq->rt, rq);
9745 #ifdef CONFIG_FAIR_GROUP_SCHED
9746 init_task_group.shares = init_task_group_load;
9747 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9748 #ifdef CONFIG_CGROUP_SCHED
9750 * How much cpu bandwidth does init_task_group get?
9752 * In case of task-groups formed thr' the cgroup filesystem, it
9753 * gets 100% of the cpu resources in the system. This overall
9754 * system cpu resource is divided among the tasks of
9755 * init_task_group and its child task-groups in a fair manner,
9756 * based on each entity's (task or task-group's) weight
9757 * (se->load.weight).
9759 * In other words, if init_task_group has 10 tasks of weight
9760 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9761 * then A0's share of the cpu resource is:
9763 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9765 * We achieve this by letting init_task_group's tasks sit
9766 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9768 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
9770 #endif /* CONFIG_FAIR_GROUP_SCHED */
9772 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9773 #ifdef CONFIG_RT_GROUP_SCHED
9774 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
9775 #ifdef CONFIG_CGROUP_SCHED
9776 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
9777 #elif defined CONFIG_USER_SCHED
9778 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
9779 init_tg_rt_entry(&init_task_group,
9780 &per_cpu(init_rt_rq, i),
9781 &per_cpu(init_sched_rt_entity, i), i, 1,
9782 root_task_group.rt_se[i]);
9786 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
9787 rq->cpu_load[j] = 0;
9791 rq->post_schedule = 0;
9792 rq->active_balance = 0;
9793 rq->next_balance = jiffies;
9797 rq->migration_thread = NULL;
9799 rq->avg_idle = 2*sysctl_sched_migration_cost;
9800 INIT_LIST_HEAD(&rq->migration_queue);
9801 rq_attach_root(rq, &def_root_domain);
9804 atomic_set(&rq->nr_iowait, 0);
9807 set_load_weight(&init_task);
9809 #ifdef CONFIG_PREEMPT_NOTIFIERS
9810 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
9814 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9817 #ifdef CONFIG_RT_MUTEXES
9818 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
9822 * The boot idle thread does lazy MMU switching as well:
9824 atomic_inc(&init_mm.mm_count);
9825 enter_lazy_tlb(&init_mm, current);
9828 * Make us the idle thread. Technically, schedule() should not be
9829 * called from this thread, however somewhere below it might be,
9830 * but because we are the idle thread, we just pick up running again
9831 * when this runqueue becomes "idle".
9833 init_idle(current, smp_processor_id());
9835 calc_load_update = jiffies + LOAD_FREQ;
9838 * During early bootup we pretend to be a normal task:
9840 current->sched_class = &fair_sched_class;
9842 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9843 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
9846 zalloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
9847 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
9849 /* May be allocated at isolcpus cmdline parse time */
9850 if (cpu_isolated_map == NULL)
9851 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
9856 scheduler_running = 1;
9859 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9860 static inline int preempt_count_equals(int preempt_offset)
9862 int nested = preempt_count() & ~PREEMPT_ACTIVE;
9864 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
9867 void __might_sleep(char *file, int line, int preempt_offset)
9870 static unsigned long prev_jiffy; /* ratelimiting */
9872 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
9873 system_state != SYSTEM_RUNNING || oops_in_progress)
9875 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9877 prev_jiffy = jiffies;
9880 "BUG: sleeping function called from invalid context at %s:%d\n",
9883 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9884 in_atomic(), irqs_disabled(),
9885 current->pid, current->comm);
9887 debug_show_held_locks(current);
9888 if (irqs_disabled())
9889 print_irqtrace_events(current);
9893 EXPORT_SYMBOL(__might_sleep);
9896 #ifdef CONFIG_MAGIC_SYSRQ
9897 static void normalize_task(struct rq *rq, struct task_struct *p)
9901 update_rq_clock(rq);
9902 on_rq = p->se.on_rq;
9904 deactivate_task(rq, p, 0);
9905 __setscheduler(rq, p, SCHED_NORMAL, 0);
9907 activate_task(rq, p, 0);
9908 resched_task(rq->curr);
9912 void normalize_rt_tasks(void)
9914 struct task_struct *g, *p;
9915 unsigned long flags;
9918 read_lock_irqsave(&tasklist_lock, flags);
9919 do_each_thread(g, p) {
9921 * Only normalize user tasks:
9926 p->se.exec_start = 0;
9927 #ifdef CONFIG_SCHEDSTATS
9928 p->se.wait_start = 0;
9929 p->se.sleep_start = 0;
9930 p->se.block_start = 0;
9935 * Renice negative nice level userspace
9938 if (TASK_NICE(p) < 0 && p->mm)
9939 set_user_nice(p, 0);
9943 spin_lock(&p->pi_lock);
9944 rq = __task_rq_lock(p);
9946 normalize_task(rq, p);
9948 __task_rq_unlock(rq);
9949 spin_unlock(&p->pi_lock);
9950 } while_each_thread(g, p);
9952 read_unlock_irqrestore(&tasklist_lock, flags);
9955 #endif /* CONFIG_MAGIC_SYSRQ */
9959 * These functions are only useful for the IA64 MCA handling.
9961 * They can only be called when the whole system has been
9962 * stopped - every CPU needs to be quiescent, and no scheduling
9963 * activity can take place. Using them for anything else would
9964 * be a serious bug, and as a result, they aren't even visible
9965 * under any other configuration.
9969 * curr_task - return the current task for a given cpu.
9970 * @cpu: the processor in question.
9972 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9974 struct task_struct *curr_task(int cpu)
9976 return cpu_curr(cpu);
9980 * set_curr_task - set the current task for a given cpu.
9981 * @cpu: the processor in question.
9982 * @p: the task pointer to set.
9984 * Description: This function must only be used when non-maskable interrupts
9985 * are serviced on a separate stack. It allows the architecture to switch the
9986 * notion of the current task on a cpu in a non-blocking manner. This function
9987 * must be called with all CPU's synchronized, and interrupts disabled, the
9988 * and caller must save the original value of the current task (see
9989 * curr_task() above) and restore that value before reenabling interrupts and
9990 * re-starting the system.
9992 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9994 void set_curr_task(int cpu, struct task_struct *p)
10001 #ifdef CONFIG_FAIR_GROUP_SCHED
10002 static void free_fair_sched_group(struct task_group *tg)
10006 for_each_possible_cpu(i) {
10008 kfree(tg->cfs_rq[i]);
10018 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10020 struct cfs_rq *cfs_rq;
10021 struct sched_entity *se;
10025 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
10028 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
10032 tg->shares = NICE_0_LOAD;
10034 for_each_possible_cpu(i) {
10037 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
10038 GFP_KERNEL, cpu_to_node(i));
10042 se = kzalloc_node(sizeof(struct sched_entity),
10043 GFP_KERNEL, cpu_to_node(i));
10047 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
10056 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
10058 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
10059 &cpu_rq(cpu)->leaf_cfs_rq_list);
10062 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
10064 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
10066 #else /* !CONFG_FAIR_GROUP_SCHED */
10067 static inline void free_fair_sched_group(struct task_group *tg)
10072 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
10077 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
10081 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
10084 #endif /* CONFIG_FAIR_GROUP_SCHED */
10086 #ifdef CONFIG_RT_GROUP_SCHED
10087 static void free_rt_sched_group(struct task_group *tg)
10091 destroy_rt_bandwidth(&tg->rt_bandwidth);
10093 for_each_possible_cpu(i) {
10095 kfree(tg->rt_rq[i]);
10097 kfree(tg->rt_se[i]);
10105 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
10107 struct rt_rq *rt_rq;
10108 struct sched_rt_entity *rt_se;
10112 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
10115 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
10119 init_rt_bandwidth(&tg->rt_bandwidth,
10120 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
10122 for_each_possible_cpu(i) {
10125 rt_rq = kzalloc_node(sizeof(struct rt_rq),
10126 GFP_KERNEL, cpu_to_node(i));
10130 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
10131 GFP_KERNEL, cpu_to_node(i));
10135 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
10144 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
10146 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
10147 &cpu_rq(cpu)->leaf_rt_rq_list);
10150 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
10152 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
10154 #else /* !CONFIG_RT_GROUP_SCHED */
10155 static inline void free_rt_sched_group(struct task_group *tg)
10160 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
10165 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
10169 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
10172 #endif /* CONFIG_RT_GROUP_SCHED */
10174 #ifdef CONFIG_CGROUP_SCHED
10175 static void free_sched_group(struct task_group *tg)
10177 free_fair_sched_group(tg);
10178 free_rt_sched_group(tg);
10182 /* allocate runqueue etc for a new task group */
10183 struct task_group *sched_create_group(struct task_group *parent)
10185 struct task_group *tg;
10186 unsigned long flags;
10189 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
10191 return ERR_PTR(-ENOMEM);
10193 if (!alloc_fair_sched_group(tg, parent))
10196 if (!alloc_rt_sched_group(tg, parent))
10199 spin_lock_irqsave(&task_group_lock, flags);
10200 for_each_possible_cpu(i) {
10201 register_fair_sched_group(tg, i);
10202 register_rt_sched_group(tg, i);
10204 list_add_rcu(&tg->list, &task_groups);
10206 WARN_ON(!parent); /* root should already exist */
10208 tg->parent = parent;
10209 INIT_LIST_HEAD(&tg->children);
10210 list_add_rcu(&tg->siblings, &parent->children);
10211 spin_unlock_irqrestore(&task_group_lock, flags);
10216 free_sched_group(tg);
10217 return ERR_PTR(-ENOMEM);
10220 /* rcu callback to free various structures associated with a task group */
10221 static void free_sched_group_rcu(struct rcu_head *rhp)
10223 /* now it should be safe to free those cfs_rqs */
10224 free_sched_group(container_of(rhp, struct task_group, rcu));
10227 /* Destroy runqueue etc associated with a task group */
10228 void sched_destroy_group(struct task_group *tg)
10230 unsigned long flags;
10233 spin_lock_irqsave(&task_group_lock, flags);
10234 for_each_possible_cpu(i) {
10235 unregister_fair_sched_group(tg, i);
10236 unregister_rt_sched_group(tg, i);
10238 list_del_rcu(&tg->list);
10239 list_del_rcu(&tg->siblings);
10240 spin_unlock_irqrestore(&task_group_lock, flags);
10242 /* wait for possible concurrent references to cfs_rqs complete */
10243 call_rcu(&tg->rcu, free_sched_group_rcu);
10246 /* change task's runqueue when it moves between groups.
10247 * The caller of this function should have put the task in its new group
10248 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
10249 * reflect its new group.
10251 void sched_move_task(struct task_struct *tsk)
10253 int on_rq, running;
10254 unsigned long flags;
10257 rq = task_rq_lock(tsk, &flags);
10259 update_rq_clock(rq);
10261 running = task_current(rq, tsk);
10262 on_rq = tsk->se.on_rq;
10265 dequeue_task(rq, tsk, 0);
10266 if (unlikely(running))
10267 tsk->sched_class->put_prev_task(rq, tsk);
10269 #ifdef CONFIG_FAIR_GROUP_SCHED
10270 if (tsk->sched_class->task_move_group)
10271 tsk->sched_class->task_move_group(tsk, on_rq);
10274 set_task_rq(tsk, task_cpu(tsk));
10276 if (unlikely(running))
10277 tsk->sched_class->set_curr_task(rq);
10279 enqueue_task(rq, tsk, 0, false);
10281 task_rq_unlock(rq, &flags);
10283 #endif /* CONFIG_CGROUP_SCHED */
10285 #ifdef CONFIG_FAIR_GROUP_SCHED
10286 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
10288 struct cfs_rq *cfs_rq = se->cfs_rq;
10293 dequeue_entity(cfs_rq, se, 0);
10295 se->load.weight = shares;
10296 se->load.inv_weight = 0;
10299 enqueue_entity(cfs_rq, se, 0);
10302 static void set_se_shares(struct sched_entity *se, unsigned long shares)
10304 struct cfs_rq *cfs_rq = se->cfs_rq;
10305 struct rq *rq = cfs_rq->rq;
10306 unsigned long flags;
10308 spin_lock_irqsave(&rq->lock, flags);
10309 __set_se_shares(se, shares);
10310 spin_unlock_irqrestore(&rq->lock, flags);
10313 static DEFINE_MUTEX(shares_mutex);
10315 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10318 unsigned long flags;
10321 * We can't change the weight of the root cgroup.
10326 if (shares < MIN_SHARES)
10327 shares = MIN_SHARES;
10328 else if (shares > MAX_SHARES)
10329 shares = MAX_SHARES;
10331 mutex_lock(&shares_mutex);
10332 if (tg->shares == shares)
10335 spin_lock_irqsave(&task_group_lock, flags);
10336 for_each_possible_cpu(i)
10337 unregister_fair_sched_group(tg, i);
10338 list_del_rcu(&tg->siblings);
10339 spin_unlock_irqrestore(&task_group_lock, flags);
10341 /* wait for any ongoing reference to this group to finish */
10342 synchronize_sched();
10345 * Now we are free to modify the group's share on each cpu
10346 * w/o tripping rebalance_share or load_balance_fair.
10348 tg->shares = shares;
10349 for_each_possible_cpu(i) {
10351 * force a rebalance
10353 cfs_rq_set_shares(tg->cfs_rq[i], 0);
10354 set_se_shares(tg->se[i], shares);
10358 * Enable load balance activity on this group, by inserting it back on
10359 * each cpu's rq->leaf_cfs_rq_list.
10361 spin_lock_irqsave(&task_group_lock, flags);
10362 for_each_possible_cpu(i)
10363 register_fair_sched_group(tg, i);
10364 list_add_rcu(&tg->siblings, &tg->parent->children);
10365 spin_unlock_irqrestore(&task_group_lock, flags);
10367 mutex_unlock(&shares_mutex);
10371 unsigned long sched_group_shares(struct task_group *tg)
10377 #ifdef CONFIG_RT_GROUP_SCHED
10379 * Ensure that the real time constraints are schedulable.
10381 static DEFINE_MUTEX(rt_constraints_mutex);
10383 static unsigned long to_ratio(u64 period, u64 runtime)
10385 if (runtime == RUNTIME_INF)
10388 return div64_u64(runtime << 20, period);
10391 /* Must be called with tasklist_lock held */
10392 static inline int tg_has_rt_tasks(struct task_group *tg)
10394 struct task_struct *g, *p;
10396 do_each_thread(g, p) {
10397 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
10399 } while_each_thread(g, p);
10404 struct rt_schedulable_data {
10405 struct task_group *tg;
10410 static int tg_schedulable(struct task_group *tg, void *data)
10412 struct rt_schedulable_data *d = data;
10413 struct task_group *child;
10414 unsigned long total, sum = 0;
10415 u64 period, runtime;
10417 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10418 runtime = tg->rt_bandwidth.rt_runtime;
10421 period = d->rt_period;
10422 runtime = d->rt_runtime;
10426 * Cannot have more runtime than the period.
10428 if (runtime > period && runtime != RUNTIME_INF)
10432 * Ensure we don't starve existing RT tasks.
10434 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
10437 total = to_ratio(period, runtime);
10440 * Nobody can have more than the global setting allows.
10442 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
10446 * The sum of our children's runtime should not exceed our own.
10448 list_for_each_entry_rcu(child, &tg->children, siblings) {
10449 period = ktime_to_ns(child->rt_bandwidth.rt_period);
10450 runtime = child->rt_bandwidth.rt_runtime;
10452 if (child == d->tg) {
10453 period = d->rt_period;
10454 runtime = d->rt_runtime;
10457 sum += to_ratio(period, runtime);
10466 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
10468 struct rt_schedulable_data data = {
10470 .rt_period = period,
10471 .rt_runtime = runtime,
10474 return walk_tg_tree(tg_schedulable, tg_nop, &data);
10477 static int tg_set_bandwidth(struct task_group *tg,
10478 u64 rt_period, u64 rt_runtime)
10482 mutex_lock(&rt_constraints_mutex);
10483 read_lock(&tasklist_lock);
10484 err = __rt_schedulable(tg, rt_period, rt_runtime);
10488 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10489 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
10490 tg->rt_bandwidth.rt_runtime = rt_runtime;
10492 for_each_possible_cpu(i) {
10493 struct rt_rq *rt_rq = tg->rt_rq[i];
10495 spin_lock(&rt_rq->rt_runtime_lock);
10496 rt_rq->rt_runtime = rt_runtime;
10497 spin_unlock(&rt_rq->rt_runtime_lock);
10499 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10501 read_unlock(&tasklist_lock);
10502 mutex_unlock(&rt_constraints_mutex);
10507 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
10509 u64 rt_runtime, rt_period;
10511 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10512 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
10513 if (rt_runtime_us < 0)
10514 rt_runtime = RUNTIME_INF;
10516 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10519 long sched_group_rt_runtime(struct task_group *tg)
10523 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
10526 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
10527 do_div(rt_runtime_us, NSEC_PER_USEC);
10528 return rt_runtime_us;
10531 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
10533 u64 rt_runtime, rt_period;
10535 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
10536 rt_runtime = tg->rt_bandwidth.rt_runtime;
10538 if (rt_period == 0)
10541 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10544 long sched_group_rt_period(struct task_group *tg)
10548 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
10549 do_div(rt_period_us, NSEC_PER_USEC);
10550 return rt_period_us;
10553 static int sched_rt_global_constraints(void)
10555 u64 runtime, period;
10558 if (sysctl_sched_rt_period <= 0)
10561 runtime = global_rt_runtime();
10562 period = global_rt_period();
10565 * Sanity check on the sysctl variables.
10567 if (runtime > period && runtime != RUNTIME_INF)
10570 mutex_lock(&rt_constraints_mutex);
10571 read_lock(&tasklist_lock);
10572 ret = __rt_schedulable(NULL, 0, 0);
10573 read_unlock(&tasklist_lock);
10574 mutex_unlock(&rt_constraints_mutex);
10579 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
10581 /* Don't accept realtime tasks when there is no way for them to run */
10582 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
10588 #else /* !CONFIG_RT_GROUP_SCHED */
10589 static int sched_rt_global_constraints(void)
10591 unsigned long flags;
10594 if (sysctl_sched_rt_period <= 0)
10598 * There's always some RT tasks in the root group
10599 * -- migration, kstopmachine etc..
10601 if (sysctl_sched_rt_runtime == 0)
10604 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
10605 for_each_possible_cpu(i) {
10606 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
10608 spin_lock(&rt_rq->rt_runtime_lock);
10609 rt_rq->rt_runtime = global_rt_runtime();
10610 spin_unlock(&rt_rq->rt_runtime_lock);
10612 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
10616 #endif /* CONFIG_RT_GROUP_SCHED */
10618 int sched_rt_handler(struct ctl_table *table, int write,
10619 void __user *buffer, size_t *lenp,
10623 int old_period, old_runtime;
10624 static DEFINE_MUTEX(mutex);
10626 mutex_lock(&mutex);
10627 old_period = sysctl_sched_rt_period;
10628 old_runtime = sysctl_sched_rt_runtime;
10630 ret = proc_dointvec(table, write, buffer, lenp, ppos);
10632 if (!ret && write) {
10633 ret = sched_rt_global_constraints();
10635 sysctl_sched_rt_period = old_period;
10636 sysctl_sched_rt_runtime = old_runtime;
10638 def_rt_bandwidth.rt_runtime = global_rt_runtime();
10639 def_rt_bandwidth.rt_period =
10640 ns_to_ktime(global_rt_period());
10643 mutex_unlock(&mutex);
10648 #ifdef CONFIG_CGROUP_SCHED
10650 /* return corresponding task_group object of a cgroup */
10651 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
10653 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
10654 struct task_group, css);
10657 static struct cgroup_subsys_state *
10658 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
10660 struct task_group *tg, *parent;
10662 if (!cgrp->parent) {
10663 /* This is early initialization for the top cgroup */
10664 return &init_task_group.css;
10667 parent = cgroup_tg(cgrp->parent);
10668 tg = sched_create_group(parent);
10670 return ERR_PTR(-ENOMEM);
10676 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10678 struct task_group *tg = cgroup_tg(cgrp);
10680 sched_destroy_group(tg);
10684 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
10686 #ifdef CONFIG_RT_GROUP_SCHED
10687 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
10690 /* We don't support RT-tasks being in separate groups */
10691 if (tsk->sched_class != &fair_sched_class)
10698 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10699 struct task_struct *tsk, bool threadgroup)
10701 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
10705 struct task_struct *c;
10707 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
10708 retval = cpu_cgroup_can_attach_task(cgrp, c);
10720 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10721 struct cgroup *old_cont, struct task_struct *tsk,
10724 sched_move_task(tsk);
10726 struct task_struct *c;
10728 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
10729 sched_move_task(c);
10735 #ifdef CONFIG_FAIR_GROUP_SCHED
10736 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
10739 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
10742 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
10744 struct task_group *tg = cgroup_tg(cgrp);
10746 return (u64) tg->shares;
10748 #endif /* CONFIG_FAIR_GROUP_SCHED */
10750 #ifdef CONFIG_RT_GROUP_SCHED
10751 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
10754 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
10757 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
10759 return sched_group_rt_runtime(cgroup_tg(cgrp));
10762 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
10765 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
10768 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
10770 return sched_group_rt_period(cgroup_tg(cgrp));
10772 #endif /* CONFIG_RT_GROUP_SCHED */
10774 static struct cftype cpu_files[] = {
10775 #ifdef CONFIG_FAIR_GROUP_SCHED
10778 .read_u64 = cpu_shares_read_u64,
10779 .write_u64 = cpu_shares_write_u64,
10782 #ifdef CONFIG_RT_GROUP_SCHED
10784 .name = "rt_runtime_us",
10785 .read_s64 = cpu_rt_runtime_read,
10786 .write_s64 = cpu_rt_runtime_write,
10789 .name = "rt_period_us",
10790 .read_u64 = cpu_rt_period_read_uint,
10791 .write_u64 = cpu_rt_period_write_uint,
10796 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
10798 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
10801 struct cgroup_subsys cpu_cgroup_subsys = {
10803 .create = cpu_cgroup_create,
10804 .destroy = cpu_cgroup_destroy,
10805 .can_attach = cpu_cgroup_can_attach,
10806 .attach = cpu_cgroup_attach,
10807 .populate = cpu_cgroup_populate,
10808 .subsys_id = cpu_cgroup_subsys_id,
10812 #endif /* CONFIG_CGROUP_SCHED */
10814 #ifdef CONFIG_CGROUP_CPUACCT
10817 * CPU accounting code for task groups.
10819 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10820 * (balbir@in.ibm.com).
10823 /* track cpu usage of a group of tasks and its child groups */
10825 struct cgroup_subsys_state css;
10826 /* cpuusage holds pointer to a u64-type object on every cpu */
10828 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
10829 struct cpuacct *parent;
10832 struct cgroup_subsys cpuacct_subsys;
10834 /* return cpu accounting group corresponding to this container */
10835 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
10837 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
10838 struct cpuacct, css);
10841 /* return cpu accounting group to which this task belongs */
10842 static inline struct cpuacct *task_ca(struct task_struct *tsk)
10844 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
10845 struct cpuacct, css);
10848 /* create a new cpu accounting group */
10849 static struct cgroup_subsys_state *cpuacct_create(
10850 struct cgroup_subsys *ss, struct cgroup *cgrp)
10852 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
10858 ca->cpuusage = alloc_percpu(u64);
10862 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10863 if (percpu_counter_init(&ca->cpustat[i], 0))
10864 goto out_free_counters;
10867 ca->parent = cgroup_ca(cgrp->parent);
10873 percpu_counter_destroy(&ca->cpustat[i]);
10874 free_percpu(ca->cpuusage);
10878 return ERR_PTR(-ENOMEM);
10881 /* destroy an existing cpu accounting group */
10883 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10885 struct cpuacct *ca = cgroup_ca(cgrp);
10888 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10889 percpu_counter_destroy(&ca->cpustat[i]);
10890 free_percpu(ca->cpuusage);
10894 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
10896 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10899 #ifndef CONFIG_64BIT
10901 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10903 spin_lock_irq(&cpu_rq(cpu)->lock);
10905 spin_unlock_irq(&cpu_rq(cpu)->lock);
10913 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
10915 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10917 #ifndef CONFIG_64BIT
10919 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10921 spin_lock_irq(&cpu_rq(cpu)->lock);
10923 spin_unlock_irq(&cpu_rq(cpu)->lock);
10929 /* return total cpu usage (in nanoseconds) of a group */
10930 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
10932 struct cpuacct *ca = cgroup_ca(cgrp);
10933 u64 totalcpuusage = 0;
10936 for_each_present_cpu(i)
10937 totalcpuusage += cpuacct_cpuusage_read(ca, i);
10939 return totalcpuusage;
10942 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
10945 struct cpuacct *ca = cgroup_ca(cgrp);
10954 for_each_present_cpu(i)
10955 cpuacct_cpuusage_write(ca, i, 0);
10961 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
10962 struct seq_file *m)
10964 struct cpuacct *ca = cgroup_ca(cgroup);
10968 for_each_present_cpu(i) {
10969 percpu = cpuacct_cpuusage_read(ca, i);
10970 seq_printf(m, "%llu ", (unsigned long long) percpu);
10972 seq_printf(m, "\n");
10976 static const char *cpuacct_stat_desc[] = {
10977 [CPUACCT_STAT_USER] = "user",
10978 [CPUACCT_STAT_SYSTEM] = "system",
10981 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
10982 struct cgroup_map_cb *cb)
10984 struct cpuacct *ca = cgroup_ca(cgrp);
10987 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
10988 s64 val = percpu_counter_read(&ca->cpustat[i]);
10989 val = cputime64_to_clock_t(val);
10990 cb->fill(cb, cpuacct_stat_desc[i], val);
10995 static struct cftype files[] = {
10998 .read_u64 = cpuusage_read,
10999 .write_u64 = cpuusage_write,
11002 .name = "usage_percpu",
11003 .read_seq_string = cpuacct_percpu_seq_read,
11007 .read_map = cpuacct_stats_show,
11011 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
11013 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
11017 * charge this task's execution time to its accounting group.
11019 * called with rq->lock held.
11021 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
11023 struct cpuacct *ca;
11026 if (unlikely(!cpuacct_subsys.active))
11029 cpu = task_cpu(tsk);
11035 for (; ca; ca = ca->parent) {
11036 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
11037 *cpuusage += cputime;
11044 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
11045 * in cputime_t units. As a result, cpuacct_update_stats calls
11046 * percpu_counter_add with values large enough to always overflow the
11047 * per cpu batch limit causing bad SMP scalability.
11049 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
11050 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
11051 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
11054 #define CPUACCT_BATCH \
11055 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
11057 #define CPUACCT_BATCH 0
11061 * Charge the system/user time to the task's accounting group.
11063 static void cpuacct_update_stats(struct task_struct *tsk,
11064 enum cpuacct_stat_index idx, cputime_t val)
11066 struct cpuacct *ca;
11067 int batch = CPUACCT_BATCH;
11069 if (unlikely(!cpuacct_subsys.active))
11076 __percpu_counter_add(&ca->cpustat[idx], val, batch);
11082 struct cgroup_subsys cpuacct_subsys = {
11084 .create = cpuacct_create,
11085 .destroy = cpuacct_destroy,
11086 .populate = cpuacct_populate,
11087 .subsys_id = cpuacct_subsys_id,
11089 #endif /* CONFIG_CGROUP_CPUACCT */
11093 int rcu_expedited_torture_stats(char *page)
11097 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
11099 void synchronize_sched_expedited(void)
11102 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
11104 #else /* #ifndef CONFIG_SMP */
11106 static DEFINE_PER_CPU(struct migration_req, rcu_migration_req);
11107 static DEFINE_MUTEX(rcu_sched_expedited_mutex);
11109 #define RCU_EXPEDITED_STATE_POST -2
11110 #define RCU_EXPEDITED_STATE_IDLE -1
11112 static int rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
11114 int rcu_expedited_torture_stats(char *page)
11119 cnt += sprintf(&page[cnt], "state: %d /", rcu_expedited_state);
11120 for_each_online_cpu(cpu) {
11121 cnt += sprintf(&page[cnt], " %d:%d",
11122 cpu, per_cpu(rcu_migration_req, cpu).dest_cpu);
11124 cnt += sprintf(&page[cnt], "\n");
11127 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
11129 static long synchronize_sched_expedited_count;
11132 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
11133 * approach to force grace period to end quickly. This consumes
11134 * significant time on all CPUs, and is thus not recommended for
11135 * any sort of common-case code.
11137 * Note that it is illegal to call this function while holding any
11138 * lock that is acquired by a CPU-hotplug notifier. Failing to
11139 * observe this restriction will result in deadlock.
11141 void synchronize_sched_expedited(void)
11144 unsigned long flags;
11145 bool need_full_sync = 0;
11147 struct migration_req *req;
11151 smp_mb(); /* ensure prior mod happens before capturing snap. */
11152 snap = ACCESS_ONCE(synchronize_sched_expedited_count) + 1;
11154 while (!mutex_trylock(&rcu_sched_expedited_mutex)) {
11156 if (trycount++ < 10)
11157 udelay(trycount * num_online_cpus());
11159 synchronize_sched();
11162 if (ACCESS_ONCE(synchronize_sched_expedited_count) - snap > 0) {
11163 smp_mb(); /* ensure test happens before caller kfree */
11168 rcu_expedited_state = RCU_EXPEDITED_STATE_POST;
11169 for_each_online_cpu(cpu) {
11171 req = &per_cpu(rcu_migration_req, cpu);
11172 init_completion(&req->done);
11174 req->dest_cpu = RCU_MIGRATION_NEED_QS;
11175 spin_lock_irqsave(&rq->lock, flags);
11176 list_add(&req->list, &rq->migration_queue);
11177 spin_unlock_irqrestore(&rq->lock, flags);
11178 wake_up_process(rq->migration_thread);
11180 for_each_online_cpu(cpu) {
11181 rcu_expedited_state = cpu;
11182 req = &per_cpu(rcu_migration_req, cpu);
11184 wait_for_completion(&req->done);
11185 spin_lock_irqsave(&rq->lock, flags);
11186 if (unlikely(req->dest_cpu == RCU_MIGRATION_MUST_SYNC))
11187 need_full_sync = 1;
11188 req->dest_cpu = RCU_MIGRATION_IDLE;
11189 spin_unlock_irqrestore(&rq->lock, flags);
11191 rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
11192 mutex_unlock(&rcu_sched_expedited_mutex);
11194 if (need_full_sync)
11195 synchronize_sched();
11197 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
11199 #endif /* #else #ifndef CONFIG_SMP */