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/proc_fs.h>
59 #include <linux/seq_file.h>
60 #include <linux/stop_machine.h>
61 #include <linux/sysctl.h>
62 #include <linux/syscalls.h>
63 #include <linux/times.h>
64 #include <linux/tsacct_kern.h>
65 #include <linux/kprobes.h>
66 #include <linux/delayacct.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
73 #include <linux/ftrace.h>
74 #include <linux/slab.h>
77 #include <asm/irq_regs.h>
78 #include <asm/mutex.h>
80 #include "sched_cpupri.h"
81 #include "workqueue_sched.h"
82 #include "sched_autogroup.h"
84 #define CREATE_TRACE_POINTS
85 #include <trace/events/sched.h>
88 * Convert user-nice values [ -20 ... 0 ... 19 ]
89 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
92 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
93 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
94 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
97 * 'User priority' is the nice value converted to something we
98 * can work with better when scaling various scheduler parameters,
99 * it's a [ 0 ... 39 ] range.
101 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
102 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
103 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
106 * Helpers for converting nanosecond timing to jiffy resolution
108 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
110 #define NICE_0_LOAD SCHED_LOAD_SCALE
111 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
114 * These are the 'tuning knobs' of the scheduler:
116 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
117 * Timeslices get refilled after they expire.
119 #define DEF_TIMESLICE (100 * HZ / 1000)
122 * single value that denotes runtime == period, ie unlimited time.
124 #define RUNTIME_INF ((u64)~0ULL)
126 static inline int rt_policy(int policy)
128 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
133 static inline int task_has_rt_policy(struct task_struct *p)
135 return rt_policy(p->policy);
139 * This is the priority-queue data structure of the RT scheduling class:
141 struct rt_prio_array {
142 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
143 struct list_head queue[MAX_RT_PRIO];
146 struct rt_bandwidth {
147 /* nests inside the rq lock: */
148 raw_spinlock_t rt_runtime_lock;
151 struct hrtimer rt_period_timer;
154 static struct rt_bandwidth def_rt_bandwidth;
156 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
158 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
160 struct rt_bandwidth *rt_b =
161 container_of(timer, struct rt_bandwidth, rt_period_timer);
167 now = hrtimer_cb_get_time(timer);
168 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
173 idle = do_sched_rt_period_timer(rt_b, overrun);
176 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
180 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
182 rt_b->rt_period = ns_to_ktime(period);
183 rt_b->rt_runtime = runtime;
185 raw_spin_lock_init(&rt_b->rt_runtime_lock);
187 hrtimer_init(&rt_b->rt_period_timer,
188 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
189 rt_b->rt_period_timer.function = sched_rt_period_timer;
192 static inline int rt_bandwidth_enabled(void)
194 return sysctl_sched_rt_runtime >= 0;
197 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
201 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
204 if (hrtimer_active(&rt_b->rt_period_timer))
207 raw_spin_lock(&rt_b->rt_runtime_lock);
212 if (hrtimer_active(&rt_b->rt_period_timer))
215 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
216 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
218 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
219 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
220 delta = ktime_to_ns(ktime_sub(hard, soft));
221 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
222 HRTIMER_MODE_ABS_PINNED, 0);
224 raw_spin_unlock(&rt_b->rt_runtime_lock);
227 #ifdef CONFIG_RT_GROUP_SCHED
228 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
230 hrtimer_cancel(&rt_b->rt_period_timer);
235 * sched_domains_mutex serializes calls to arch_init_sched_domains,
236 * detach_destroy_domains and partition_sched_domains.
238 static DEFINE_MUTEX(sched_domains_mutex);
240 #ifdef CONFIG_CGROUP_SCHED
242 #include <linux/cgroup.h>
246 static LIST_HEAD(task_groups);
248 /* task group related information */
250 struct cgroup_subsys_state css;
252 #ifdef CONFIG_FAIR_GROUP_SCHED
253 /* schedulable entities of this group on each cpu */
254 struct sched_entity **se;
255 /* runqueue "owned" by this group on each cpu */
256 struct cfs_rq **cfs_rq;
257 unsigned long shares;
259 atomic_t load_weight;
262 #ifdef CONFIG_RT_GROUP_SCHED
263 struct sched_rt_entity **rt_se;
264 struct rt_rq **rt_rq;
266 struct rt_bandwidth rt_bandwidth;
270 struct list_head list;
272 struct task_group *parent;
273 struct list_head siblings;
274 struct list_head children;
276 #ifdef CONFIG_SCHED_AUTOGROUP
277 struct autogroup *autogroup;
281 #define root_task_group init_task_group
283 /* task_group_lock serializes the addition/removal of task groups */
284 static DEFINE_SPINLOCK(task_group_lock);
286 #ifdef CONFIG_FAIR_GROUP_SCHED
288 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
291 * A weight of 0 or 1 can cause arithmetics problems.
292 * A weight of a cfs_rq is the sum of weights of which entities
293 * are queued on this cfs_rq, so a weight of a entity should not be
294 * too large, so as the shares value of a task group.
295 * (The default weight is 1024 - so there's no practical
296 * limitation from this.)
299 #define MAX_SHARES (1UL << 18)
301 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
304 /* Default task group.
305 * Every task in system belong to this group at bootup.
307 struct task_group init_task_group;
309 #endif /* CONFIG_CGROUP_SCHED */
311 /* CFS-related fields in a runqueue */
313 struct load_weight load;
314 unsigned long nr_running;
319 struct rb_root tasks_timeline;
320 struct rb_node *rb_leftmost;
322 struct list_head tasks;
323 struct list_head *balance_iterator;
326 * 'curr' points to currently running entity on this cfs_rq.
327 * It is set to NULL otherwise (i.e when none are currently running).
329 struct sched_entity *curr, *next, *last;
331 unsigned int nr_spread_over;
333 #ifdef CONFIG_FAIR_GROUP_SCHED
334 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
337 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
338 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
339 * (like users, containers etc.)
341 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
342 * list is used during load balance.
345 struct list_head leaf_cfs_rq_list;
346 struct task_group *tg; /* group that "owns" this runqueue */
350 * the part of load.weight contributed by tasks
352 unsigned long task_weight;
355 * h_load = weight * f(tg)
357 * Where f(tg) is the recursive weight fraction assigned to
360 unsigned long h_load;
363 * Maintaining per-cpu shares distribution for group scheduling
365 * load_stamp is the last time we updated the load average
366 * load_last is the last time we updated the load average and saw load
367 * load_unacc_exec_time is currently unaccounted execution time
371 u64 load_stamp, load_last, load_unacc_exec_time;
373 unsigned long load_contribution;
378 /* Real-Time classes' related field in a runqueue: */
380 struct rt_prio_array active;
381 unsigned long rt_nr_running;
382 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
384 int curr; /* highest queued rt task prio */
386 int next; /* next highest */
391 unsigned long rt_nr_migratory;
392 unsigned long rt_nr_total;
394 struct plist_head pushable_tasks;
399 /* Nests inside the rq lock: */
400 raw_spinlock_t rt_runtime_lock;
402 #ifdef CONFIG_RT_GROUP_SCHED
403 unsigned long rt_nr_boosted;
406 struct list_head leaf_rt_rq_list;
407 struct task_group *tg;
414 * We add the notion of a root-domain which will be used to define per-domain
415 * variables. Each exclusive cpuset essentially defines an island domain by
416 * fully partitioning the member cpus from any other cpuset. Whenever a new
417 * exclusive cpuset is created, we also create and attach a new root-domain
424 cpumask_var_t online;
427 * The "RT overload" flag: it gets set if a CPU has more than
428 * one runnable RT task.
430 cpumask_var_t rto_mask;
432 struct cpupri cpupri;
436 * By default the system creates a single root-domain with all cpus as
437 * members (mimicking the global state we have today).
439 static struct root_domain def_root_domain;
441 #endif /* CONFIG_SMP */
444 * This is the main, per-CPU runqueue data structure.
446 * Locking rule: those places that want to lock multiple runqueues
447 * (such as the load balancing or the thread migration code), lock
448 * acquire operations must be ordered by ascending &runqueue.
455 * nr_running and cpu_load should be in the same cacheline because
456 * remote CPUs use both these fields when doing load calculation.
458 unsigned long nr_running;
459 #define CPU_LOAD_IDX_MAX 5
460 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
461 unsigned long last_load_update_tick;
464 unsigned char nohz_balance_kick;
466 unsigned int skip_clock_update;
468 /* capture load from *all* tasks on this cpu: */
469 struct load_weight load;
470 unsigned long nr_load_updates;
476 #ifdef CONFIG_FAIR_GROUP_SCHED
477 /* list of leaf cfs_rq on this cpu: */
478 struct list_head leaf_cfs_rq_list;
480 #ifdef CONFIG_RT_GROUP_SCHED
481 struct list_head leaf_rt_rq_list;
485 * This is part of a global counter where only the total sum
486 * over all CPUs matters. A task can increase this counter on
487 * one CPU and if it got migrated afterwards it may decrease
488 * it on another CPU. Always updated under the runqueue lock:
490 unsigned long nr_uninterruptible;
492 struct task_struct *curr, *idle, *stop;
493 unsigned long next_balance;
494 struct mm_struct *prev_mm;
502 struct root_domain *rd;
503 struct sched_domain *sd;
505 unsigned long cpu_power;
507 unsigned char idle_at_tick;
508 /* For active balancing */
512 struct cpu_stop_work active_balance_work;
513 /* cpu of this runqueue: */
517 unsigned long avg_load_per_task;
525 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
529 /* calc_load related fields */
530 unsigned long calc_load_update;
531 long calc_load_active;
533 #ifdef CONFIG_SCHED_HRTICK
535 int hrtick_csd_pending;
536 struct call_single_data hrtick_csd;
538 struct hrtimer hrtick_timer;
541 #ifdef CONFIG_SCHEDSTATS
543 struct sched_info rq_sched_info;
544 unsigned long long rq_cpu_time;
545 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
547 /* sys_sched_yield() stats */
548 unsigned int yld_count;
550 /* schedule() stats */
551 unsigned int sched_switch;
552 unsigned int sched_count;
553 unsigned int sched_goidle;
555 /* try_to_wake_up() stats */
556 unsigned int ttwu_count;
557 unsigned int ttwu_local;
560 unsigned int bkl_count;
564 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
567 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags);
569 static inline int cpu_of(struct rq *rq)
578 #define rcu_dereference_check_sched_domain(p) \
579 rcu_dereference_check((p), \
580 rcu_read_lock_sched_held() || \
581 lockdep_is_held(&sched_domains_mutex))
584 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
585 * See detach_destroy_domains: synchronize_sched for details.
587 * The domain tree of any CPU may only be accessed from within
588 * preempt-disabled sections.
590 #define for_each_domain(cpu, __sd) \
591 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
593 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
594 #define this_rq() (&__get_cpu_var(runqueues))
595 #define task_rq(p) cpu_rq(task_cpu(p))
596 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
597 #define raw_rq() (&__raw_get_cpu_var(runqueues))
599 #ifdef CONFIG_CGROUP_SCHED
602 * Return the group to which this tasks belongs.
604 * We use task_subsys_state_check() and extend the RCU verification
605 * with lockdep_is_held(&task_rq(p)->lock) because cpu_cgroup_attach()
606 * holds that lock for each task it moves into the cgroup. Therefore
607 * by holding that lock, we pin the task to the current cgroup.
609 static inline struct task_group *task_group(struct task_struct *p)
611 struct task_group *tg;
612 struct cgroup_subsys_state *css;
614 css = task_subsys_state_check(p, cpu_cgroup_subsys_id,
615 lockdep_is_held(&task_rq(p)->lock));
616 tg = container_of(css, struct task_group, css);
618 return autogroup_task_group(p, tg);
621 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
622 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
624 #ifdef CONFIG_FAIR_GROUP_SCHED
625 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
626 p->se.parent = task_group(p)->se[cpu];
629 #ifdef CONFIG_RT_GROUP_SCHED
630 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
631 p->rt.parent = task_group(p)->rt_se[cpu];
635 #else /* CONFIG_CGROUP_SCHED */
637 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
638 static inline struct task_group *task_group(struct task_struct *p)
643 #endif /* CONFIG_CGROUP_SCHED */
645 static u64 irq_time_cpu(int cpu);
646 static void sched_irq_time_avg_update(struct rq *rq, u64 irq_time);
648 inline void update_rq_clock(struct rq *rq)
650 if (!rq->skip_clock_update) {
651 int cpu = cpu_of(rq);
654 rq->clock = sched_clock_cpu(cpu);
655 irq_time = irq_time_cpu(cpu);
656 if (rq->clock - irq_time > rq->clock_task)
657 rq->clock_task = rq->clock - irq_time;
659 sched_irq_time_avg_update(rq, irq_time);
664 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
666 #ifdef CONFIG_SCHED_DEBUG
667 # define const_debug __read_mostly
669 # define const_debug static const
674 * @cpu: the processor in question.
676 * Returns true if the current cpu runqueue is locked.
677 * This interface allows printk to be called with the runqueue lock
678 * held and know whether or not it is OK to wake up the klogd.
680 int runqueue_is_locked(int cpu)
682 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
686 * Debugging: various feature bits
689 #define SCHED_FEAT(name, enabled) \
690 __SCHED_FEAT_##name ,
693 #include "sched_features.h"
698 #define SCHED_FEAT(name, enabled) \
699 (1UL << __SCHED_FEAT_##name) * enabled |
701 const_debug unsigned int sysctl_sched_features =
702 #include "sched_features.h"
707 #ifdef CONFIG_SCHED_DEBUG
708 #define SCHED_FEAT(name, enabled) \
711 static __read_mostly char *sched_feat_names[] = {
712 #include "sched_features.h"
718 static int sched_feat_show(struct seq_file *m, void *v)
722 for (i = 0; sched_feat_names[i]; i++) {
723 if (!(sysctl_sched_features & (1UL << i)))
725 seq_printf(m, "%s ", sched_feat_names[i]);
733 sched_feat_write(struct file *filp, const char __user *ubuf,
734 size_t cnt, loff_t *ppos)
744 if (copy_from_user(&buf, ubuf, cnt))
750 if (strncmp(buf, "NO_", 3) == 0) {
755 for (i = 0; sched_feat_names[i]; i++) {
756 if (strcmp(cmp, sched_feat_names[i]) == 0) {
758 sysctl_sched_features &= ~(1UL << i);
760 sysctl_sched_features |= (1UL << i);
765 if (!sched_feat_names[i])
773 static int sched_feat_open(struct inode *inode, struct file *filp)
775 return single_open(filp, sched_feat_show, NULL);
778 static const struct file_operations sched_feat_fops = {
779 .open = sched_feat_open,
780 .write = sched_feat_write,
783 .release = single_release,
786 static __init int sched_init_debug(void)
788 debugfs_create_file("sched_features", 0644, NULL, NULL,
793 late_initcall(sched_init_debug);
797 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
800 * Number of tasks to iterate in a single balance run.
801 * Limited because this is done with IRQs disabled.
803 const_debug unsigned int sysctl_sched_nr_migrate = 32;
806 * period over which we average the RT time consumption, measured
811 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
814 * period over which we measure -rt task cpu usage in us.
817 unsigned int sysctl_sched_rt_period = 1000000;
819 static __read_mostly int scheduler_running;
822 * part of the period that we allow rt tasks to run in us.
825 int sysctl_sched_rt_runtime = 950000;
827 static inline u64 global_rt_period(void)
829 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
832 static inline u64 global_rt_runtime(void)
834 if (sysctl_sched_rt_runtime < 0)
837 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
840 #ifndef prepare_arch_switch
841 # define prepare_arch_switch(next) do { } while (0)
843 #ifndef finish_arch_switch
844 # define finish_arch_switch(prev) do { } while (0)
847 static inline int task_current(struct rq *rq, struct task_struct *p)
849 return rq->curr == p;
852 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
853 static inline int task_running(struct rq *rq, struct task_struct *p)
855 return task_current(rq, p);
858 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
862 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
864 #ifdef CONFIG_DEBUG_SPINLOCK
865 /* this is a valid case when another task releases the spinlock */
866 rq->lock.owner = current;
869 * If we are tracking spinlock dependencies then we have to
870 * fix up the runqueue lock - which gets 'carried over' from
873 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
875 raw_spin_unlock_irq(&rq->lock);
878 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
879 static inline int task_running(struct rq *rq, struct task_struct *p)
884 return task_current(rq, p);
888 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
892 * We can optimise this out completely for !SMP, because the
893 * SMP rebalancing from interrupt is the only thing that cares
898 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
899 raw_spin_unlock_irq(&rq->lock);
901 raw_spin_unlock(&rq->lock);
905 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
909 * After ->oncpu is cleared, the task can be moved to a different CPU.
910 * We must ensure this doesn't happen until the switch is completely
916 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
920 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
923 * Check whether the task is waking, we use this to synchronize ->cpus_allowed
926 static inline int task_is_waking(struct task_struct *p)
928 return unlikely(p->state == TASK_WAKING);
932 * __task_rq_lock - lock the runqueue a given task resides on.
933 * Must be called interrupts disabled.
935 static inline struct rq *__task_rq_lock(struct task_struct *p)
942 raw_spin_lock(&rq->lock);
943 if (likely(rq == task_rq(p)))
945 raw_spin_unlock(&rq->lock);
950 * task_rq_lock - lock the runqueue a given task resides on and disable
951 * interrupts. Note the ordering: we can safely lookup the task_rq without
952 * explicitly disabling preemption.
954 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
960 local_irq_save(*flags);
962 raw_spin_lock(&rq->lock);
963 if (likely(rq == task_rq(p)))
965 raw_spin_unlock_irqrestore(&rq->lock, *flags);
969 static void __task_rq_unlock(struct rq *rq)
972 raw_spin_unlock(&rq->lock);
975 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
978 raw_spin_unlock_irqrestore(&rq->lock, *flags);
982 * this_rq_lock - lock this runqueue and disable interrupts.
984 static struct rq *this_rq_lock(void)
991 raw_spin_lock(&rq->lock);
996 #ifdef CONFIG_SCHED_HRTICK
998 * Use HR-timers to deliver accurate preemption points.
1000 * Its all a bit involved since we cannot program an hrt while holding the
1001 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1004 * When we get rescheduled we reprogram the hrtick_timer outside of the
1010 * - enabled by features
1011 * - hrtimer is actually high res
1013 static inline int hrtick_enabled(struct rq *rq)
1015 if (!sched_feat(HRTICK))
1017 if (!cpu_active(cpu_of(rq)))
1019 return hrtimer_is_hres_active(&rq->hrtick_timer);
1022 static void hrtick_clear(struct rq *rq)
1024 if (hrtimer_active(&rq->hrtick_timer))
1025 hrtimer_cancel(&rq->hrtick_timer);
1029 * High-resolution timer tick.
1030 * Runs from hardirq context with interrupts disabled.
1032 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1034 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1036 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1038 raw_spin_lock(&rq->lock);
1039 update_rq_clock(rq);
1040 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1041 raw_spin_unlock(&rq->lock);
1043 return HRTIMER_NORESTART;
1048 * called from hardirq (IPI) context
1050 static void __hrtick_start(void *arg)
1052 struct rq *rq = arg;
1054 raw_spin_lock(&rq->lock);
1055 hrtimer_restart(&rq->hrtick_timer);
1056 rq->hrtick_csd_pending = 0;
1057 raw_spin_unlock(&rq->lock);
1061 * Called to set the hrtick timer state.
1063 * called with rq->lock held and irqs disabled
1065 static void hrtick_start(struct rq *rq, u64 delay)
1067 struct hrtimer *timer = &rq->hrtick_timer;
1068 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1070 hrtimer_set_expires(timer, time);
1072 if (rq == this_rq()) {
1073 hrtimer_restart(timer);
1074 } else if (!rq->hrtick_csd_pending) {
1075 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1076 rq->hrtick_csd_pending = 1;
1081 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1083 int cpu = (int)(long)hcpu;
1086 case CPU_UP_CANCELED:
1087 case CPU_UP_CANCELED_FROZEN:
1088 case CPU_DOWN_PREPARE:
1089 case CPU_DOWN_PREPARE_FROZEN:
1091 case CPU_DEAD_FROZEN:
1092 hrtick_clear(cpu_rq(cpu));
1099 static __init void init_hrtick(void)
1101 hotcpu_notifier(hotplug_hrtick, 0);
1105 * Called to set the hrtick timer state.
1107 * called with rq->lock held and irqs disabled
1109 static void hrtick_start(struct rq *rq, u64 delay)
1111 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1112 HRTIMER_MODE_REL_PINNED, 0);
1115 static inline void init_hrtick(void)
1118 #endif /* CONFIG_SMP */
1120 static void init_rq_hrtick(struct rq *rq)
1123 rq->hrtick_csd_pending = 0;
1125 rq->hrtick_csd.flags = 0;
1126 rq->hrtick_csd.func = __hrtick_start;
1127 rq->hrtick_csd.info = rq;
1130 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1131 rq->hrtick_timer.function = hrtick;
1133 #else /* CONFIG_SCHED_HRTICK */
1134 static inline void hrtick_clear(struct rq *rq)
1138 static inline void init_rq_hrtick(struct rq *rq)
1142 static inline void init_hrtick(void)
1145 #endif /* CONFIG_SCHED_HRTICK */
1148 * resched_task - mark a task 'to be rescheduled now'.
1150 * On UP this means the setting of the need_resched flag, on SMP it
1151 * might also involve a cross-CPU call to trigger the scheduler on
1156 #ifndef tsk_is_polling
1157 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1160 static void resched_task(struct task_struct *p)
1164 assert_raw_spin_locked(&task_rq(p)->lock);
1166 if (test_tsk_need_resched(p))
1169 set_tsk_need_resched(p);
1172 if (cpu == smp_processor_id())
1175 /* NEED_RESCHED must be visible before we test polling */
1177 if (!tsk_is_polling(p))
1178 smp_send_reschedule(cpu);
1181 static void resched_cpu(int cpu)
1183 struct rq *rq = cpu_rq(cpu);
1184 unsigned long flags;
1186 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1188 resched_task(cpu_curr(cpu));
1189 raw_spin_unlock_irqrestore(&rq->lock, flags);
1194 * In the semi idle case, use the nearest busy cpu for migrating timers
1195 * from an idle cpu. This is good for power-savings.
1197 * We don't do similar optimization for completely idle system, as
1198 * selecting an idle cpu will add more delays to the timers than intended
1199 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1201 int get_nohz_timer_target(void)
1203 int cpu = smp_processor_id();
1205 struct sched_domain *sd;
1207 for_each_domain(cpu, sd) {
1208 for_each_cpu(i, sched_domain_span(sd))
1215 * When add_timer_on() enqueues a timer into the timer wheel of an
1216 * idle CPU then this timer might expire before the next timer event
1217 * which is scheduled to wake up that CPU. In case of a completely
1218 * idle system the next event might even be infinite time into the
1219 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1220 * leaves the inner idle loop so the newly added timer is taken into
1221 * account when the CPU goes back to idle and evaluates the timer
1222 * wheel for the next timer event.
1224 void wake_up_idle_cpu(int cpu)
1226 struct rq *rq = cpu_rq(cpu);
1228 if (cpu == smp_processor_id())
1232 * This is safe, as this function is called with the timer
1233 * wheel base lock of (cpu) held. When the CPU is on the way
1234 * to idle and has not yet set rq->curr to idle then it will
1235 * be serialized on the timer wheel base lock and take the new
1236 * timer into account automatically.
1238 if (rq->curr != rq->idle)
1242 * We can set TIF_RESCHED on the idle task of the other CPU
1243 * lockless. The worst case is that the other CPU runs the
1244 * idle task through an additional NOOP schedule()
1246 set_tsk_need_resched(rq->idle);
1248 /* NEED_RESCHED must be visible before we test polling */
1250 if (!tsk_is_polling(rq->idle))
1251 smp_send_reschedule(cpu);
1254 #endif /* CONFIG_NO_HZ */
1256 static u64 sched_avg_period(void)
1258 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1261 static void sched_avg_update(struct rq *rq)
1263 s64 period = sched_avg_period();
1265 while ((s64)(rq->clock - rq->age_stamp) > period) {
1267 * Inline assembly required to prevent the compiler
1268 * optimising this loop into a divmod call.
1269 * See __iter_div_u64_rem() for another example of this.
1271 asm("" : "+rm" (rq->age_stamp));
1272 rq->age_stamp += period;
1277 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1279 rq->rt_avg += rt_delta;
1280 sched_avg_update(rq);
1283 #else /* !CONFIG_SMP */
1284 static void resched_task(struct task_struct *p)
1286 assert_raw_spin_locked(&task_rq(p)->lock);
1287 set_tsk_need_resched(p);
1290 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1294 static void sched_avg_update(struct rq *rq)
1297 #endif /* CONFIG_SMP */
1299 #if BITS_PER_LONG == 32
1300 # define WMULT_CONST (~0UL)
1302 # define WMULT_CONST (1UL << 32)
1305 #define WMULT_SHIFT 32
1308 * Shift right and round:
1310 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1313 * delta *= weight / lw
1315 static unsigned long
1316 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1317 struct load_weight *lw)
1321 if (!lw->inv_weight) {
1322 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1325 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1329 tmp = (u64)delta_exec * weight;
1331 * Check whether we'd overflow the 64-bit multiplication:
1333 if (unlikely(tmp > WMULT_CONST))
1334 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1337 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1339 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1342 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1348 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1354 static inline void update_load_set(struct load_weight *lw, unsigned long w)
1361 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1362 * of tasks with abnormal "nice" values across CPUs the contribution that
1363 * each task makes to its run queue's load is weighted according to its
1364 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1365 * scaled version of the new time slice allocation that they receive on time
1369 #define WEIGHT_IDLEPRIO 3
1370 #define WMULT_IDLEPRIO 1431655765
1373 * Nice levels are multiplicative, with a gentle 10% change for every
1374 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1375 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1376 * that remained on nice 0.
1378 * The "10% effect" is relative and cumulative: from _any_ nice level,
1379 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1380 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1381 * If a task goes up by ~10% and another task goes down by ~10% then
1382 * the relative distance between them is ~25%.)
1384 static const int prio_to_weight[40] = {
1385 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1386 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1387 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1388 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1389 /* 0 */ 1024, 820, 655, 526, 423,
1390 /* 5 */ 335, 272, 215, 172, 137,
1391 /* 10 */ 110, 87, 70, 56, 45,
1392 /* 15 */ 36, 29, 23, 18, 15,
1396 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1398 * In cases where the weight does not change often, we can use the
1399 * precalculated inverse to speed up arithmetics by turning divisions
1400 * into multiplications:
1402 static const u32 prio_to_wmult[40] = {
1403 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1404 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1405 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1406 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1407 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1408 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1409 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1410 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1413 /* Time spent by the tasks of the cpu accounting group executing in ... */
1414 enum cpuacct_stat_index {
1415 CPUACCT_STAT_USER, /* ... user mode */
1416 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1418 CPUACCT_STAT_NSTATS,
1421 #ifdef CONFIG_CGROUP_CPUACCT
1422 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1423 static void cpuacct_update_stats(struct task_struct *tsk,
1424 enum cpuacct_stat_index idx, cputime_t val);
1426 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1427 static inline void cpuacct_update_stats(struct task_struct *tsk,
1428 enum cpuacct_stat_index idx, cputime_t val) {}
1431 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1433 update_load_add(&rq->load, load);
1436 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1438 update_load_sub(&rq->load, load);
1441 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1442 typedef int (*tg_visitor)(struct task_group *, void *);
1445 * Iterate the full tree, calling @down when first entering a node and @up when
1446 * leaving it for the final time.
1448 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1450 struct task_group *parent, *child;
1454 parent = &root_task_group;
1456 ret = (*down)(parent, data);
1459 list_for_each_entry_rcu(child, &parent->children, siblings) {
1466 ret = (*up)(parent, data);
1471 parent = parent->parent;
1480 static int tg_nop(struct task_group *tg, void *data)
1487 /* Used instead of source_load when we know the type == 0 */
1488 static unsigned long weighted_cpuload(const int cpu)
1490 return cpu_rq(cpu)->load.weight;
1494 * Return a low guess at the load of a migration-source cpu weighted
1495 * according to the scheduling class and "nice" value.
1497 * We want to under-estimate the load of migration sources, to
1498 * balance conservatively.
1500 static unsigned long source_load(int cpu, int type)
1502 struct rq *rq = cpu_rq(cpu);
1503 unsigned long total = weighted_cpuload(cpu);
1505 if (type == 0 || !sched_feat(LB_BIAS))
1508 return min(rq->cpu_load[type-1], total);
1512 * Return a high guess at the load of a migration-target cpu weighted
1513 * according to the scheduling class and "nice" value.
1515 static unsigned long target_load(int cpu, int type)
1517 struct rq *rq = cpu_rq(cpu);
1518 unsigned long total = weighted_cpuload(cpu);
1520 if (type == 0 || !sched_feat(LB_BIAS))
1523 return max(rq->cpu_load[type-1], total);
1526 static unsigned long power_of(int cpu)
1528 return cpu_rq(cpu)->cpu_power;
1531 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1533 static unsigned long cpu_avg_load_per_task(int cpu)
1535 struct rq *rq = cpu_rq(cpu);
1536 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1539 rq->avg_load_per_task = rq->load.weight / nr_running;
1541 rq->avg_load_per_task = 0;
1543 return rq->avg_load_per_task;
1546 #ifdef CONFIG_FAIR_GROUP_SCHED
1549 * Compute the cpu's hierarchical load factor for each task group.
1550 * This needs to be done in a top-down fashion because the load of a child
1551 * group is a fraction of its parents load.
1553 static int tg_load_down(struct task_group *tg, void *data)
1556 long cpu = (long)data;
1559 load = cpu_rq(cpu)->load.weight;
1561 load = tg->parent->cfs_rq[cpu]->h_load;
1562 load *= tg->se[cpu]->load.weight;
1563 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1566 tg->cfs_rq[cpu]->h_load = load;
1571 static void update_h_load(long cpu)
1573 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1578 #ifdef CONFIG_PREEMPT
1580 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1583 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1584 * way at the expense of forcing extra atomic operations in all
1585 * invocations. This assures that the double_lock is acquired using the
1586 * same underlying policy as the spinlock_t on this architecture, which
1587 * reduces latency compared to the unfair variant below. However, it
1588 * also adds more overhead and therefore may reduce throughput.
1590 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1591 __releases(this_rq->lock)
1592 __acquires(busiest->lock)
1593 __acquires(this_rq->lock)
1595 raw_spin_unlock(&this_rq->lock);
1596 double_rq_lock(this_rq, busiest);
1603 * Unfair double_lock_balance: Optimizes throughput at the expense of
1604 * latency by eliminating extra atomic operations when the locks are
1605 * already in proper order on entry. This favors lower cpu-ids and will
1606 * grant the double lock to lower cpus over higher ids under contention,
1607 * regardless of entry order into the function.
1609 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1610 __releases(this_rq->lock)
1611 __acquires(busiest->lock)
1612 __acquires(this_rq->lock)
1616 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1617 if (busiest < this_rq) {
1618 raw_spin_unlock(&this_rq->lock);
1619 raw_spin_lock(&busiest->lock);
1620 raw_spin_lock_nested(&this_rq->lock,
1621 SINGLE_DEPTH_NESTING);
1624 raw_spin_lock_nested(&busiest->lock,
1625 SINGLE_DEPTH_NESTING);
1630 #endif /* CONFIG_PREEMPT */
1633 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1635 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1637 if (unlikely(!irqs_disabled())) {
1638 /* printk() doesn't work good under rq->lock */
1639 raw_spin_unlock(&this_rq->lock);
1643 return _double_lock_balance(this_rq, busiest);
1646 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1647 __releases(busiest->lock)
1649 raw_spin_unlock(&busiest->lock);
1650 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1654 * double_rq_lock - safely lock two runqueues
1656 * Note this does not disable interrupts like task_rq_lock,
1657 * you need to do so manually before calling.
1659 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1660 __acquires(rq1->lock)
1661 __acquires(rq2->lock)
1663 BUG_ON(!irqs_disabled());
1665 raw_spin_lock(&rq1->lock);
1666 __acquire(rq2->lock); /* Fake it out ;) */
1669 raw_spin_lock(&rq1->lock);
1670 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1672 raw_spin_lock(&rq2->lock);
1673 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1679 * double_rq_unlock - safely unlock two runqueues
1681 * Note this does not restore interrupts like task_rq_unlock,
1682 * you need to do so manually after calling.
1684 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1685 __releases(rq1->lock)
1686 __releases(rq2->lock)
1688 raw_spin_unlock(&rq1->lock);
1690 raw_spin_unlock(&rq2->lock);
1692 __release(rq2->lock);
1697 static void calc_load_account_idle(struct rq *this_rq);
1698 static void update_sysctl(void);
1699 static int get_update_sysctl_factor(void);
1700 static void update_cpu_load(struct rq *this_rq);
1702 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1704 set_task_rq(p, cpu);
1707 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1708 * successfuly executed on another CPU. We must ensure that updates of
1709 * per-task data have been completed by this moment.
1712 task_thread_info(p)->cpu = cpu;
1716 static const struct sched_class rt_sched_class;
1718 #define sched_class_highest (&stop_sched_class)
1719 #define for_each_class(class) \
1720 for (class = sched_class_highest; class; class = class->next)
1722 #include "sched_stats.h"
1724 static void inc_nr_running(struct rq *rq)
1729 static void dec_nr_running(struct rq *rq)
1734 static void set_load_weight(struct task_struct *p)
1737 * SCHED_IDLE tasks get minimal weight:
1739 if (p->policy == SCHED_IDLE) {
1740 p->se.load.weight = WEIGHT_IDLEPRIO;
1741 p->se.load.inv_weight = WMULT_IDLEPRIO;
1745 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1746 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1749 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1751 update_rq_clock(rq);
1752 sched_info_queued(p);
1753 p->sched_class->enqueue_task(rq, p, flags);
1757 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1759 update_rq_clock(rq);
1760 sched_info_dequeued(p);
1761 p->sched_class->dequeue_task(rq, p, flags);
1766 * activate_task - move a task to the runqueue.
1768 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1770 if (task_contributes_to_load(p))
1771 rq->nr_uninterruptible--;
1773 enqueue_task(rq, p, flags);
1778 * deactivate_task - remove a task from the runqueue.
1780 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1782 if (task_contributes_to_load(p))
1783 rq->nr_uninterruptible++;
1785 dequeue_task(rq, p, flags);
1789 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1792 * There are no locks covering percpu hardirq/softirq time.
1793 * They are only modified in account_system_vtime, on corresponding CPU
1794 * with interrupts disabled. So, writes are safe.
1795 * They are read and saved off onto struct rq in update_rq_clock().
1796 * This may result in other CPU reading this CPU's irq time and can
1797 * race with irq/account_system_vtime on this CPU. We would either get old
1798 * or new value (or semi updated value on 32 bit) with a side effect of
1799 * accounting a slice of irq time to wrong task when irq is in progress
1800 * while we read rq->clock. That is a worthy compromise in place of having
1801 * locks on each irq in account_system_time.
1803 static DEFINE_PER_CPU(u64, cpu_hardirq_time);
1804 static DEFINE_PER_CPU(u64, cpu_softirq_time);
1806 static DEFINE_PER_CPU(u64, irq_start_time);
1807 static int sched_clock_irqtime;
1809 void enable_sched_clock_irqtime(void)
1811 sched_clock_irqtime = 1;
1814 void disable_sched_clock_irqtime(void)
1816 sched_clock_irqtime = 0;
1819 static u64 irq_time_cpu(int cpu)
1821 if (!sched_clock_irqtime)
1824 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
1827 void account_system_vtime(struct task_struct *curr)
1829 unsigned long flags;
1833 if (!sched_clock_irqtime)
1836 local_irq_save(flags);
1838 cpu = smp_processor_id();
1839 now = sched_clock_cpu(cpu);
1840 delta = now - per_cpu(irq_start_time, cpu);
1841 per_cpu(irq_start_time, cpu) = now;
1843 * We do not account for softirq time from ksoftirqd here.
1844 * We want to continue accounting softirq time to ksoftirqd thread
1845 * in that case, so as not to confuse scheduler with a special task
1846 * that do not consume any time, but still wants to run.
1848 if (hardirq_count())
1849 per_cpu(cpu_hardirq_time, cpu) += delta;
1850 else if (in_serving_softirq() && !(curr->flags & PF_KSOFTIRQD))
1851 per_cpu(cpu_softirq_time, cpu) += delta;
1853 local_irq_restore(flags);
1855 EXPORT_SYMBOL_GPL(account_system_vtime);
1857 static void sched_irq_time_avg_update(struct rq *rq, u64 curr_irq_time)
1859 if (sched_clock_irqtime && sched_feat(NONIRQ_POWER)) {
1860 u64 delta_irq = curr_irq_time - rq->prev_irq_time;
1861 rq->prev_irq_time = curr_irq_time;
1862 sched_rt_avg_update(rq, delta_irq);
1868 static u64 irq_time_cpu(int cpu)
1873 static void sched_irq_time_avg_update(struct rq *rq, u64 curr_irq_time) { }
1877 #include "sched_idletask.c"
1878 #include "sched_fair.c"
1879 #include "sched_rt.c"
1880 #include "sched_autogroup.c"
1881 #include "sched_stoptask.c"
1882 #ifdef CONFIG_SCHED_DEBUG
1883 # include "sched_debug.c"
1886 void sched_set_stop_task(int cpu, struct task_struct *stop)
1888 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
1889 struct task_struct *old_stop = cpu_rq(cpu)->stop;
1893 * Make it appear like a SCHED_FIFO task, its something
1894 * userspace knows about and won't get confused about.
1896 * Also, it will make PI more or less work without too
1897 * much confusion -- but then, stop work should not
1898 * rely on PI working anyway.
1900 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
1902 stop->sched_class = &stop_sched_class;
1905 cpu_rq(cpu)->stop = stop;
1909 * Reset it back to a normal scheduling class so that
1910 * it can die in pieces.
1912 old_stop->sched_class = &rt_sched_class;
1917 * __normal_prio - return the priority that is based on the static prio
1919 static inline int __normal_prio(struct task_struct *p)
1921 return p->static_prio;
1925 * Calculate the expected normal priority: i.e. priority
1926 * without taking RT-inheritance into account. Might be
1927 * boosted by interactivity modifiers. Changes upon fork,
1928 * setprio syscalls, and whenever the interactivity
1929 * estimator recalculates.
1931 static inline int normal_prio(struct task_struct *p)
1935 if (task_has_rt_policy(p))
1936 prio = MAX_RT_PRIO-1 - p->rt_priority;
1938 prio = __normal_prio(p);
1943 * Calculate the current priority, i.e. the priority
1944 * taken into account by the scheduler. This value might
1945 * be boosted by RT tasks, or might be boosted by
1946 * interactivity modifiers. Will be RT if the task got
1947 * RT-boosted. If not then it returns p->normal_prio.
1949 static int effective_prio(struct task_struct *p)
1951 p->normal_prio = normal_prio(p);
1953 * If we are RT tasks or we were boosted to RT priority,
1954 * keep the priority unchanged. Otherwise, update priority
1955 * to the normal priority:
1957 if (!rt_prio(p->prio))
1958 return p->normal_prio;
1963 * task_curr - is this task currently executing on a CPU?
1964 * @p: the task in question.
1966 inline int task_curr(const struct task_struct *p)
1968 return cpu_curr(task_cpu(p)) == p;
1971 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1972 const struct sched_class *prev_class,
1973 int oldprio, int running)
1975 if (prev_class != p->sched_class) {
1976 if (prev_class->switched_from)
1977 prev_class->switched_from(rq, p, running);
1978 p->sched_class->switched_to(rq, p, running);
1980 p->sched_class->prio_changed(rq, p, oldprio, running);
1983 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1985 const struct sched_class *class;
1987 if (p->sched_class == rq->curr->sched_class) {
1988 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1990 for_each_class(class) {
1991 if (class == rq->curr->sched_class)
1993 if (class == p->sched_class) {
1994 resched_task(rq->curr);
2001 * A queue event has occurred, and we're going to schedule. In
2002 * this case, we can save a useless back to back clock update.
2004 if (test_tsk_need_resched(rq->curr))
2005 rq->skip_clock_update = 1;
2010 * Is this task likely cache-hot:
2013 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2017 if (p->sched_class != &fair_sched_class)
2020 if (unlikely(p->policy == SCHED_IDLE))
2024 * Buddy candidates are cache hot:
2026 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2027 (&p->se == cfs_rq_of(&p->se)->next ||
2028 &p->se == cfs_rq_of(&p->se)->last))
2031 if (sysctl_sched_migration_cost == -1)
2033 if (sysctl_sched_migration_cost == 0)
2036 delta = now - p->se.exec_start;
2038 return delta < (s64)sysctl_sched_migration_cost;
2041 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2043 #ifdef CONFIG_SCHED_DEBUG
2045 * We should never call set_task_cpu() on a blocked task,
2046 * ttwu() will sort out the placement.
2048 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2049 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2052 trace_sched_migrate_task(p, new_cpu);
2054 if (task_cpu(p) != new_cpu) {
2055 p->se.nr_migrations++;
2056 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2059 __set_task_cpu(p, new_cpu);
2062 struct migration_arg {
2063 struct task_struct *task;
2067 static int migration_cpu_stop(void *data);
2070 * The task's runqueue lock must be held.
2071 * Returns true if you have to wait for migration thread.
2073 static bool migrate_task(struct task_struct *p, struct rq *rq)
2076 * If the task is not on a runqueue (and not running), then
2077 * the next wake-up will properly place the task.
2079 return p->se.on_rq || task_running(rq, p);
2083 * wait_task_inactive - wait for a thread to unschedule.
2085 * If @match_state is nonzero, it's the @p->state value just checked and
2086 * not expected to change. If it changes, i.e. @p might have woken up,
2087 * then return zero. When we succeed in waiting for @p to be off its CPU,
2088 * we return a positive number (its total switch count). If a second call
2089 * a short while later returns the same number, the caller can be sure that
2090 * @p has remained unscheduled the whole time.
2092 * The caller must ensure that the task *will* unschedule sometime soon,
2093 * else this function might spin for a *long* time. This function can't
2094 * be called with interrupts off, or it may introduce deadlock with
2095 * smp_call_function() if an IPI is sent by the same process we are
2096 * waiting to become inactive.
2098 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2100 unsigned long flags;
2107 * We do the initial early heuristics without holding
2108 * any task-queue locks at all. We'll only try to get
2109 * the runqueue lock when things look like they will
2115 * If the task is actively running on another CPU
2116 * still, just relax and busy-wait without holding
2119 * NOTE! Since we don't hold any locks, it's not
2120 * even sure that "rq" stays as the right runqueue!
2121 * But we don't care, since "task_running()" will
2122 * return false if the runqueue has changed and p
2123 * is actually now running somewhere else!
2125 while (task_running(rq, p)) {
2126 if (match_state && unlikely(p->state != match_state))
2132 * Ok, time to look more closely! We need the rq
2133 * lock now, to be *sure*. If we're wrong, we'll
2134 * just go back and repeat.
2136 rq = task_rq_lock(p, &flags);
2137 trace_sched_wait_task(p);
2138 running = task_running(rq, p);
2139 on_rq = p->se.on_rq;
2141 if (!match_state || p->state == match_state)
2142 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2143 task_rq_unlock(rq, &flags);
2146 * If it changed from the expected state, bail out now.
2148 if (unlikely(!ncsw))
2152 * Was it really running after all now that we
2153 * checked with the proper locks actually held?
2155 * Oops. Go back and try again..
2157 if (unlikely(running)) {
2163 * It's not enough that it's not actively running,
2164 * it must be off the runqueue _entirely_, and not
2167 * So if it was still runnable (but just not actively
2168 * running right now), it's preempted, and we should
2169 * yield - it could be a while.
2171 if (unlikely(on_rq)) {
2172 schedule_timeout_uninterruptible(1);
2177 * Ahh, all good. It wasn't running, and it wasn't
2178 * runnable, which means that it will never become
2179 * running in the future either. We're all done!
2188 * kick_process - kick a running thread to enter/exit the kernel
2189 * @p: the to-be-kicked thread
2191 * Cause a process which is running on another CPU to enter
2192 * kernel-mode, without any delay. (to get signals handled.)
2194 * NOTE: this function doesnt have to take the runqueue lock,
2195 * because all it wants to ensure is that the remote task enters
2196 * the kernel. If the IPI races and the task has been migrated
2197 * to another CPU then no harm is done and the purpose has been
2200 void kick_process(struct task_struct *p)
2206 if ((cpu != smp_processor_id()) && task_curr(p))
2207 smp_send_reschedule(cpu);
2210 EXPORT_SYMBOL_GPL(kick_process);
2211 #endif /* CONFIG_SMP */
2214 * task_oncpu_function_call - call a function on the cpu on which a task runs
2215 * @p: the task to evaluate
2216 * @func: the function to be called
2217 * @info: the function call argument
2219 * Calls the function @func when the task is currently running. This might
2220 * be on the current CPU, which just calls the function directly
2222 void task_oncpu_function_call(struct task_struct *p,
2223 void (*func) (void *info), void *info)
2230 smp_call_function_single(cpu, func, info, 1);
2236 * ->cpus_allowed is protected by either TASK_WAKING or rq->lock held.
2238 static int select_fallback_rq(int cpu, struct task_struct *p)
2241 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2243 /* Look for allowed, online CPU in same node. */
2244 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2245 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2248 /* Any allowed, online CPU? */
2249 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2250 if (dest_cpu < nr_cpu_ids)
2253 /* No more Mr. Nice Guy. */
2254 dest_cpu = cpuset_cpus_allowed_fallback(p);
2256 * Don't tell them about moving exiting tasks or
2257 * kernel threads (both mm NULL), since they never
2260 if (p->mm && printk_ratelimit()) {
2261 printk(KERN_INFO "process %d (%s) no longer affine to cpu%d\n",
2262 task_pid_nr(p), p->comm, cpu);
2269 * The caller (fork, wakeup) owns TASK_WAKING, ->cpus_allowed is stable.
2272 int select_task_rq(struct rq *rq, struct task_struct *p, int sd_flags, int wake_flags)
2274 int cpu = p->sched_class->select_task_rq(rq, p, sd_flags, wake_flags);
2277 * In order not to call set_task_cpu() on a blocking task we need
2278 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2281 * Since this is common to all placement strategies, this lives here.
2283 * [ this allows ->select_task() to simply return task_cpu(p) and
2284 * not worry about this generic constraint ]
2286 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2288 cpu = select_fallback_rq(task_cpu(p), p);
2293 static void update_avg(u64 *avg, u64 sample)
2295 s64 diff = sample - *avg;
2300 static inline void ttwu_activate(struct task_struct *p, struct rq *rq,
2301 bool is_sync, bool is_migrate, bool is_local,
2302 unsigned long en_flags)
2304 schedstat_inc(p, se.statistics.nr_wakeups);
2306 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2308 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2310 schedstat_inc(p, se.statistics.nr_wakeups_local);
2312 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2314 activate_task(rq, p, en_flags);
2317 static inline void ttwu_post_activation(struct task_struct *p, struct rq *rq,
2318 int wake_flags, bool success)
2320 trace_sched_wakeup(p, success);
2321 check_preempt_curr(rq, p, wake_flags);
2323 p->state = TASK_RUNNING;
2325 if (p->sched_class->task_woken)
2326 p->sched_class->task_woken(rq, p);
2328 if (unlikely(rq->idle_stamp)) {
2329 u64 delta = rq->clock - rq->idle_stamp;
2330 u64 max = 2*sysctl_sched_migration_cost;
2335 update_avg(&rq->avg_idle, delta);
2339 /* if a worker is waking up, notify workqueue */
2340 if ((p->flags & PF_WQ_WORKER) && success)
2341 wq_worker_waking_up(p, cpu_of(rq));
2345 * try_to_wake_up - wake up a thread
2346 * @p: the thread to be awakened
2347 * @state: the mask of task states that can be woken
2348 * @wake_flags: wake modifier flags (WF_*)
2350 * Put it on the run-queue if it's not already there. The "current"
2351 * thread is always on the run-queue (except when the actual
2352 * re-schedule is in progress), and as such you're allowed to do
2353 * the simpler "current->state = TASK_RUNNING" to mark yourself
2354 * runnable without the overhead of this.
2356 * Returns %true if @p was woken up, %false if it was already running
2357 * or @state didn't match @p's state.
2359 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2362 int cpu, orig_cpu, this_cpu, success = 0;
2363 unsigned long flags;
2364 unsigned long en_flags = ENQUEUE_WAKEUP;
2367 this_cpu = get_cpu();
2370 rq = task_rq_lock(p, &flags);
2371 if (!(p->state & state))
2381 if (unlikely(task_running(rq, p)))
2385 * In order to handle concurrent wakeups and release the rq->lock
2386 * we put the task in TASK_WAKING state.
2388 * First fix up the nr_uninterruptible count:
2390 if (task_contributes_to_load(p)) {
2391 if (likely(cpu_online(orig_cpu)))
2392 rq->nr_uninterruptible--;
2394 this_rq()->nr_uninterruptible--;
2396 p->state = TASK_WAKING;
2398 if (p->sched_class->task_waking) {
2399 p->sched_class->task_waking(rq, p);
2400 en_flags |= ENQUEUE_WAKING;
2403 cpu = select_task_rq(rq, p, SD_BALANCE_WAKE, wake_flags);
2404 if (cpu != orig_cpu)
2405 set_task_cpu(p, cpu);
2406 __task_rq_unlock(rq);
2409 raw_spin_lock(&rq->lock);
2412 * We migrated the task without holding either rq->lock, however
2413 * since the task is not on the task list itself, nobody else
2414 * will try and migrate the task, hence the rq should match the
2415 * cpu we just moved it to.
2417 WARN_ON(task_cpu(p) != cpu);
2418 WARN_ON(p->state != TASK_WAKING);
2420 #ifdef CONFIG_SCHEDSTATS
2421 schedstat_inc(rq, ttwu_count);
2422 if (cpu == this_cpu)
2423 schedstat_inc(rq, ttwu_local);
2425 struct sched_domain *sd;
2426 for_each_domain(this_cpu, sd) {
2427 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2428 schedstat_inc(sd, ttwu_wake_remote);
2433 #endif /* CONFIG_SCHEDSTATS */
2436 #endif /* CONFIG_SMP */
2437 ttwu_activate(p, rq, wake_flags & WF_SYNC, orig_cpu != cpu,
2438 cpu == this_cpu, en_flags);
2441 ttwu_post_activation(p, rq, wake_flags, success);
2443 task_rq_unlock(rq, &flags);
2450 * try_to_wake_up_local - try to wake up a local task with rq lock held
2451 * @p: the thread to be awakened
2453 * Put @p on the run-queue if it's not alredy there. The caller must
2454 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2455 * the current task. this_rq() stays locked over invocation.
2457 static void try_to_wake_up_local(struct task_struct *p)
2459 struct rq *rq = task_rq(p);
2460 bool success = false;
2462 BUG_ON(rq != this_rq());
2463 BUG_ON(p == current);
2464 lockdep_assert_held(&rq->lock);
2466 if (!(p->state & TASK_NORMAL))
2470 if (likely(!task_running(rq, p))) {
2471 schedstat_inc(rq, ttwu_count);
2472 schedstat_inc(rq, ttwu_local);
2474 ttwu_activate(p, rq, false, false, true, ENQUEUE_WAKEUP);
2477 ttwu_post_activation(p, rq, 0, success);
2481 * wake_up_process - Wake up a specific process
2482 * @p: The process to be woken up.
2484 * Attempt to wake up the nominated process and move it to the set of runnable
2485 * processes. Returns 1 if the process was woken up, 0 if it was already
2488 * It may be assumed that this function implies a write memory barrier before
2489 * changing the task state if and only if any tasks are woken up.
2491 int wake_up_process(struct task_struct *p)
2493 return try_to_wake_up(p, TASK_ALL, 0);
2495 EXPORT_SYMBOL(wake_up_process);
2497 int wake_up_state(struct task_struct *p, unsigned int state)
2499 return try_to_wake_up(p, state, 0);
2503 * Perform scheduler related setup for a newly forked process p.
2504 * p is forked by current.
2506 * __sched_fork() is basic setup used by init_idle() too:
2508 static void __sched_fork(struct task_struct *p)
2510 p->se.exec_start = 0;
2511 p->se.sum_exec_runtime = 0;
2512 p->se.prev_sum_exec_runtime = 0;
2513 p->se.nr_migrations = 0;
2515 #ifdef CONFIG_SCHEDSTATS
2516 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2519 INIT_LIST_HEAD(&p->rt.run_list);
2521 INIT_LIST_HEAD(&p->se.group_node);
2523 #ifdef CONFIG_PREEMPT_NOTIFIERS
2524 INIT_HLIST_HEAD(&p->preempt_notifiers);
2529 * fork()/clone()-time setup:
2531 void sched_fork(struct task_struct *p, int clone_flags)
2533 int cpu = get_cpu();
2537 * We mark the process as running here. This guarantees that
2538 * nobody will actually run it, and a signal or other external
2539 * event cannot wake it up and insert it on the runqueue either.
2541 p->state = TASK_RUNNING;
2544 * Revert to default priority/policy on fork if requested.
2546 if (unlikely(p->sched_reset_on_fork)) {
2547 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2548 p->policy = SCHED_NORMAL;
2549 p->normal_prio = p->static_prio;
2552 if (PRIO_TO_NICE(p->static_prio) < 0) {
2553 p->static_prio = NICE_TO_PRIO(0);
2554 p->normal_prio = p->static_prio;
2559 * We don't need the reset flag anymore after the fork. It has
2560 * fulfilled its duty:
2562 p->sched_reset_on_fork = 0;
2566 * Make sure we do not leak PI boosting priority to the child.
2568 p->prio = current->normal_prio;
2570 if (!rt_prio(p->prio))
2571 p->sched_class = &fair_sched_class;
2573 if (p->sched_class->task_fork)
2574 p->sched_class->task_fork(p);
2577 * The child is not yet in the pid-hash so no cgroup attach races,
2578 * and the cgroup is pinned to this child due to cgroup_fork()
2579 * is ran before sched_fork().
2581 * Silence PROVE_RCU.
2584 set_task_cpu(p, cpu);
2587 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2588 if (likely(sched_info_on()))
2589 memset(&p->sched_info, 0, sizeof(p->sched_info));
2591 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2594 #ifdef CONFIG_PREEMPT
2595 /* Want to start with kernel preemption disabled. */
2596 task_thread_info(p)->preempt_count = 1;
2598 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2604 * wake_up_new_task - wake up a newly created task for the first time.
2606 * This function will do some initial scheduler statistics housekeeping
2607 * that must be done for every newly created context, then puts the task
2608 * on the runqueue and wakes it.
2610 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2612 unsigned long flags;
2614 int cpu __maybe_unused = get_cpu();
2617 rq = task_rq_lock(p, &flags);
2618 p->state = TASK_WAKING;
2621 * Fork balancing, do it here and not earlier because:
2622 * - cpus_allowed can change in the fork path
2623 * - any previously selected cpu might disappear through hotplug
2625 * We set TASK_WAKING so that select_task_rq() can drop rq->lock
2626 * without people poking at ->cpus_allowed.
2628 cpu = select_task_rq(rq, p, SD_BALANCE_FORK, 0);
2629 set_task_cpu(p, cpu);
2631 p->state = TASK_RUNNING;
2632 task_rq_unlock(rq, &flags);
2635 rq = task_rq_lock(p, &flags);
2636 activate_task(rq, p, 0);
2637 trace_sched_wakeup_new(p, 1);
2638 check_preempt_curr(rq, p, WF_FORK);
2640 if (p->sched_class->task_woken)
2641 p->sched_class->task_woken(rq, p);
2643 task_rq_unlock(rq, &flags);
2647 #ifdef CONFIG_PREEMPT_NOTIFIERS
2650 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2651 * @notifier: notifier struct to register
2653 void preempt_notifier_register(struct preempt_notifier *notifier)
2655 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2657 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2660 * preempt_notifier_unregister - no longer interested in preemption notifications
2661 * @notifier: notifier struct to unregister
2663 * This is safe to call from within a preemption notifier.
2665 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2667 hlist_del(¬ifier->link);
2669 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2671 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2673 struct preempt_notifier *notifier;
2674 struct hlist_node *node;
2676 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2677 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2681 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2682 struct task_struct *next)
2684 struct preempt_notifier *notifier;
2685 struct hlist_node *node;
2687 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2688 notifier->ops->sched_out(notifier, next);
2691 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2693 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2698 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2699 struct task_struct *next)
2703 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2706 * prepare_task_switch - prepare to switch tasks
2707 * @rq: the runqueue preparing to switch
2708 * @prev: the current task that is being switched out
2709 * @next: the task we are going to switch to.
2711 * This is called with the rq lock held and interrupts off. It must
2712 * be paired with a subsequent finish_task_switch after the context
2715 * prepare_task_switch sets up locking and calls architecture specific
2719 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2720 struct task_struct *next)
2722 fire_sched_out_preempt_notifiers(prev, next);
2723 prepare_lock_switch(rq, next);
2724 prepare_arch_switch(next);
2728 * finish_task_switch - clean up after a task-switch
2729 * @rq: runqueue associated with task-switch
2730 * @prev: the thread we just switched away from.
2732 * finish_task_switch must be called after the context switch, paired
2733 * with a prepare_task_switch call before the context switch.
2734 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2735 * and do any other architecture-specific cleanup actions.
2737 * Note that we may have delayed dropping an mm in context_switch(). If
2738 * so, we finish that here outside of the runqueue lock. (Doing it
2739 * with the lock held can cause deadlocks; see schedule() for
2742 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2743 __releases(rq->lock)
2745 struct mm_struct *mm = rq->prev_mm;
2751 * A task struct has one reference for the use as "current".
2752 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2753 * schedule one last time. The schedule call will never return, and
2754 * the scheduled task must drop that reference.
2755 * The test for TASK_DEAD must occur while the runqueue locks are
2756 * still held, otherwise prev could be scheduled on another cpu, die
2757 * there before we look at prev->state, and then the reference would
2759 * Manfred Spraul <manfred@colorfullife.com>
2761 prev_state = prev->state;
2762 finish_arch_switch(prev);
2763 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2764 local_irq_disable();
2765 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2766 perf_event_task_sched_in(current);
2767 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2769 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2770 finish_lock_switch(rq, prev);
2772 fire_sched_in_preempt_notifiers(current);
2775 if (unlikely(prev_state == TASK_DEAD)) {
2777 * Remove function-return probe instances associated with this
2778 * task and put them back on the free list.
2780 kprobe_flush_task(prev);
2781 put_task_struct(prev);
2787 /* assumes rq->lock is held */
2788 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2790 if (prev->sched_class->pre_schedule)
2791 prev->sched_class->pre_schedule(rq, prev);
2794 /* rq->lock is NOT held, but preemption is disabled */
2795 static inline void post_schedule(struct rq *rq)
2797 if (rq->post_schedule) {
2798 unsigned long flags;
2800 raw_spin_lock_irqsave(&rq->lock, flags);
2801 if (rq->curr->sched_class->post_schedule)
2802 rq->curr->sched_class->post_schedule(rq);
2803 raw_spin_unlock_irqrestore(&rq->lock, flags);
2805 rq->post_schedule = 0;
2811 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2815 static inline void post_schedule(struct rq *rq)
2822 * schedule_tail - first thing a freshly forked thread must call.
2823 * @prev: the thread we just switched away from.
2825 asmlinkage void schedule_tail(struct task_struct *prev)
2826 __releases(rq->lock)
2828 struct rq *rq = this_rq();
2830 finish_task_switch(rq, prev);
2833 * FIXME: do we need to worry about rq being invalidated by the
2838 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2839 /* In this case, finish_task_switch does not reenable preemption */
2842 if (current->set_child_tid)
2843 put_user(task_pid_vnr(current), current->set_child_tid);
2847 * context_switch - switch to the new MM and the new
2848 * thread's register state.
2851 context_switch(struct rq *rq, struct task_struct *prev,
2852 struct task_struct *next)
2854 struct mm_struct *mm, *oldmm;
2856 prepare_task_switch(rq, prev, next);
2857 trace_sched_switch(prev, next);
2859 oldmm = prev->active_mm;
2861 * For paravirt, this is coupled with an exit in switch_to to
2862 * combine the page table reload and the switch backend into
2865 arch_start_context_switch(prev);
2868 next->active_mm = oldmm;
2869 atomic_inc(&oldmm->mm_count);
2870 enter_lazy_tlb(oldmm, next);
2872 switch_mm(oldmm, mm, next);
2875 prev->active_mm = NULL;
2876 rq->prev_mm = oldmm;
2879 * Since the runqueue lock will be released by the next
2880 * task (which is an invalid locking op but in the case
2881 * of the scheduler it's an obvious special-case), so we
2882 * do an early lockdep release here:
2884 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2885 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2888 /* Here we just switch the register state and the stack. */
2889 switch_to(prev, next, prev);
2893 * this_rq must be evaluated again because prev may have moved
2894 * CPUs since it called schedule(), thus the 'rq' on its stack
2895 * frame will be invalid.
2897 finish_task_switch(this_rq(), prev);
2901 * nr_running, nr_uninterruptible and nr_context_switches:
2903 * externally visible scheduler statistics: current number of runnable
2904 * threads, current number of uninterruptible-sleeping threads, total
2905 * number of context switches performed since bootup.
2907 unsigned long nr_running(void)
2909 unsigned long i, sum = 0;
2911 for_each_online_cpu(i)
2912 sum += cpu_rq(i)->nr_running;
2917 unsigned long nr_uninterruptible(void)
2919 unsigned long i, sum = 0;
2921 for_each_possible_cpu(i)
2922 sum += cpu_rq(i)->nr_uninterruptible;
2925 * Since we read the counters lockless, it might be slightly
2926 * inaccurate. Do not allow it to go below zero though:
2928 if (unlikely((long)sum < 0))
2934 unsigned long long nr_context_switches(void)
2937 unsigned long long sum = 0;
2939 for_each_possible_cpu(i)
2940 sum += cpu_rq(i)->nr_switches;
2945 unsigned long nr_iowait(void)
2947 unsigned long i, sum = 0;
2949 for_each_possible_cpu(i)
2950 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2955 unsigned long nr_iowait_cpu(int cpu)
2957 struct rq *this = cpu_rq(cpu);
2958 return atomic_read(&this->nr_iowait);
2961 unsigned long this_cpu_load(void)
2963 struct rq *this = this_rq();
2964 return this->cpu_load[0];
2968 /* Variables and functions for calc_load */
2969 static atomic_long_t calc_load_tasks;
2970 static unsigned long calc_load_update;
2971 unsigned long avenrun[3];
2972 EXPORT_SYMBOL(avenrun);
2974 static long calc_load_fold_active(struct rq *this_rq)
2976 long nr_active, delta = 0;
2978 nr_active = this_rq->nr_running;
2979 nr_active += (long) this_rq->nr_uninterruptible;
2981 if (nr_active != this_rq->calc_load_active) {
2982 delta = nr_active - this_rq->calc_load_active;
2983 this_rq->calc_load_active = nr_active;
2991 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
2993 * When making the ILB scale, we should try to pull this in as well.
2995 static atomic_long_t calc_load_tasks_idle;
2997 static void calc_load_account_idle(struct rq *this_rq)
3001 delta = calc_load_fold_active(this_rq);
3003 atomic_long_add(delta, &calc_load_tasks_idle);
3006 static long calc_load_fold_idle(void)
3011 * Its got a race, we don't care...
3013 if (atomic_long_read(&calc_load_tasks_idle))
3014 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
3019 static void calc_load_account_idle(struct rq *this_rq)
3023 static inline long calc_load_fold_idle(void)
3030 * get_avenrun - get the load average array
3031 * @loads: pointer to dest load array
3032 * @offset: offset to add
3033 * @shift: shift count to shift the result left
3035 * These values are estimates at best, so no need for locking.
3037 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3039 loads[0] = (avenrun[0] + offset) << shift;
3040 loads[1] = (avenrun[1] + offset) << shift;
3041 loads[2] = (avenrun[2] + offset) << shift;
3044 static unsigned long
3045 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3048 load += active * (FIXED_1 - exp);
3049 return load >> FSHIFT;
3053 * calc_load - update the avenrun load estimates 10 ticks after the
3054 * CPUs have updated calc_load_tasks.
3056 void calc_global_load(void)
3058 unsigned long upd = calc_load_update + 10;
3061 if (time_before(jiffies, upd))
3064 active = atomic_long_read(&calc_load_tasks);
3065 active = active > 0 ? active * FIXED_1 : 0;
3067 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3068 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3069 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3071 calc_load_update += LOAD_FREQ;
3075 * Called from update_cpu_load() to periodically update this CPU's
3078 static void calc_load_account_active(struct rq *this_rq)
3082 if (time_before(jiffies, this_rq->calc_load_update))
3085 delta = calc_load_fold_active(this_rq);
3086 delta += calc_load_fold_idle();
3088 atomic_long_add(delta, &calc_load_tasks);
3090 this_rq->calc_load_update += LOAD_FREQ;
3094 * The exact cpuload at various idx values, calculated at every tick would be
3095 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3097 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3098 * on nth tick when cpu may be busy, then we have:
3099 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3100 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3102 * decay_load_missed() below does efficient calculation of
3103 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3104 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3106 * The calculation is approximated on a 128 point scale.
3107 * degrade_zero_ticks is the number of ticks after which load at any
3108 * particular idx is approximated to be zero.
3109 * degrade_factor is a precomputed table, a row for each load idx.
3110 * Each column corresponds to degradation factor for a power of two ticks,
3111 * based on 128 point scale.
3113 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3114 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3116 * With this power of 2 load factors, we can degrade the load n times
3117 * by looking at 1 bits in n and doing as many mult/shift instead of
3118 * n mult/shifts needed by the exact degradation.
3120 #define DEGRADE_SHIFT 7
3121 static const unsigned char
3122 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3123 static const unsigned char
3124 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3125 {0, 0, 0, 0, 0, 0, 0, 0},
3126 {64, 32, 8, 0, 0, 0, 0, 0},
3127 {96, 72, 40, 12, 1, 0, 0},
3128 {112, 98, 75, 43, 15, 1, 0},
3129 {120, 112, 98, 76, 45, 16, 2} };
3132 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3133 * would be when CPU is idle and so we just decay the old load without
3134 * adding any new load.
3136 static unsigned long
3137 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3141 if (!missed_updates)
3144 if (missed_updates >= degrade_zero_ticks[idx])
3148 return load >> missed_updates;
3150 while (missed_updates) {
3151 if (missed_updates % 2)
3152 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3154 missed_updates >>= 1;
3161 * Update rq->cpu_load[] statistics. This function is usually called every
3162 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3163 * every tick. We fix it up based on jiffies.
3165 static void update_cpu_load(struct rq *this_rq)
3167 unsigned long this_load = this_rq->load.weight;
3168 unsigned long curr_jiffies = jiffies;
3169 unsigned long pending_updates;
3172 this_rq->nr_load_updates++;
3174 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3175 if (curr_jiffies == this_rq->last_load_update_tick)
3178 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3179 this_rq->last_load_update_tick = curr_jiffies;
3181 /* Update our load: */
3182 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3183 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3184 unsigned long old_load, new_load;
3186 /* scale is effectively 1 << i now, and >> i divides by scale */
3188 old_load = this_rq->cpu_load[i];
3189 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3190 new_load = this_load;
3192 * Round up the averaging division if load is increasing. This
3193 * prevents us from getting stuck on 9 if the load is 10, for
3196 if (new_load > old_load)
3197 new_load += scale - 1;
3199 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3202 sched_avg_update(this_rq);
3205 static void update_cpu_load_active(struct rq *this_rq)
3207 update_cpu_load(this_rq);
3209 calc_load_account_active(this_rq);
3215 * sched_exec - execve() is a valuable balancing opportunity, because at
3216 * this point the task has the smallest effective memory and cache footprint.
3218 void sched_exec(void)
3220 struct task_struct *p = current;
3221 unsigned long flags;
3225 rq = task_rq_lock(p, &flags);
3226 dest_cpu = p->sched_class->select_task_rq(rq, p, SD_BALANCE_EXEC, 0);
3227 if (dest_cpu == smp_processor_id())
3231 * select_task_rq() can race against ->cpus_allowed
3233 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed) &&
3234 likely(cpu_active(dest_cpu)) && migrate_task(p, rq)) {
3235 struct migration_arg arg = { p, dest_cpu };
3237 task_rq_unlock(rq, &flags);
3238 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
3242 task_rq_unlock(rq, &flags);
3247 DEFINE_PER_CPU(struct kernel_stat, kstat);
3249 EXPORT_PER_CPU_SYMBOL(kstat);
3252 * Return any ns on the sched_clock that have not yet been accounted in
3253 * @p in case that task is currently running.
3255 * Called with task_rq_lock() held on @rq.
3257 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3261 if (task_current(rq, p)) {
3262 update_rq_clock(rq);
3263 ns = rq->clock_task - p->se.exec_start;
3271 unsigned long long task_delta_exec(struct task_struct *p)
3273 unsigned long flags;
3277 rq = task_rq_lock(p, &flags);
3278 ns = do_task_delta_exec(p, rq);
3279 task_rq_unlock(rq, &flags);
3285 * Return accounted runtime for the task.
3286 * In case the task is currently running, return the runtime plus current's
3287 * pending runtime that have not been accounted yet.
3289 unsigned long long task_sched_runtime(struct task_struct *p)
3291 unsigned long flags;
3295 rq = task_rq_lock(p, &flags);
3296 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3297 task_rq_unlock(rq, &flags);
3303 * Return sum_exec_runtime for the thread group.
3304 * In case the task is currently running, return the sum plus current's
3305 * pending runtime that have not been accounted yet.
3307 * Note that the thread group might have other running tasks as well,
3308 * so the return value not includes other pending runtime that other
3309 * running tasks might have.
3311 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3313 struct task_cputime totals;
3314 unsigned long flags;
3318 rq = task_rq_lock(p, &flags);
3319 thread_group_cputime(p, &totals);
3320 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3321 task_rq_unlock(rq, &flags);
3327 * Account user cpu time to a process.
3328 * @p: the process that the cpu time gets accounted to
3329 * @cputime: the cpu time spent in user space since the last update
3330 * @cputime_scaled: cputime scaled by cpu frequency
3332 void account_user_time(struct task_struct *p, cputime_t cputime,
3333 cputime_t cputime_scaled)
3335 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3338 /* Add user time to process. */
3339 p->utime = cputime_add(p->utime, cputime);
3340 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3341 account_group_user_time(p, cputime);
3343 /* Add user time to cpustat. */
3344 tmp = cputime_to_cputime64(cputime);
3345 if (TASK_NICE(p) > 0)
3346 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3348 cpustat->user = cputime64_add(cpustat->user, tmp);
3350 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3351 /* Account for user time used */
3352 acct_update_integrals(p);
3356 * Account guest cpu time to a process.
3357 * @p: the process that the cpu time gets accounted to
3358 * @cputime: the cpu time spent in virtual machine since the last update
3359 * @cputime_scaled: cputime scaled by cpu frequency
3361 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3362 cputime_t cputime_scaled)
3365 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3367 tmp = cputime_to_cputime64(cputime);
3369 /* Add guest time to process. */
3370 p->utime = cputime_add(p->utime, cputime);
3371 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3372 account_group_user_time(p, cputime);
3373 p->gtime = cputime_add(p->gtime, cputime);
3375 /* Add guest time to cpustat. */
3376 if (TASK_NICE(p) > 0) {
3377 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3378 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3380 cpustat->user = cputime64_add(cpustat->user, tmp);
3381 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3386 * Account system cpu time to a process.
3387 * @p: the process that the cpu time gets accounted to
3388 * @hardirq_offset: the offset to subtract from hardirq_count()
3389 * @cputime: the cpu time spent in kernel space since the last update
3390 * @cputime_scaled: cputime scaled by cpu frequency
3392 void account_system_time(struct task_struct *p, int hardirq_offset,
3393 cputime_t cputime, cputime_t cputime_scaled)
3395 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3398 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3399 account_guest_time(p, cputime, cputime_scaled);
3403 /* Add system time to process. */
3404 p->stime = cputime_add(p->stime, cputime);
3405 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3406 account_group_system_time(p, cputime);
3408 /* Add system time to cpustat. */
3409 tmp = cputime_to_cputime64(cputime);
3410 if (hardirq_count() - hardirq_offset)
3411 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3412 else if (in_serving_softirq())
3413 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3415 cpustat->system = cputime64_add(cpustat->system, tmp);
3417 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3419 /* Account for system time used */
3420 acct_update_integrals(p);
3424 * Account for involuntary wait time.
3425 * @steal: the cpu time spent in involuntary wait
3427 void account_steal_time(cputime_t cputime)
3429 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3430 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3432 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3436 * Account for idle time.
3437 * @cputime: the cpu time spent in idle wait
3439 void account_idle_time(cputime_t cputime)
3441 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3442 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3443 struct rq *rq = this_rq();
3445 if (atomic_read(&rq->nr_iowait) > 0)
3446 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3448 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3451 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3454 * Account a single tick of cpu time.
3455 * @p: the process that the cpu time gets accounted to
3456 * @user_tick: indicates if the tick is a user or a system tick
3458 void account_process_tick(struct task_struct *p, int user_tick)
3460 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3461 struct rq *rq = this_rq();
3464 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3465 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3466 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3469 account_idle_time(cputime_one_jiffy);
3473 * Account multiple ticks of steal time.
3474 * @p: the process from which the cpu time has been stolen
3475 * @ticks: number of stolen ticks
3477 void account_steal_ticks(unsigned long ticks)
3479 account_steal_time(jiffies_to_cputime(ticks));
3483 * Account multiple ticks of idle time.
3484 * @ticks: number of stolen ticks
3486 void account_idle_ticks(unsigned long ticks)
3488 account_idle_time(jiffies_to_cputime(ticks));
3494 * Use precise platform statistics if available:
3496 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3497 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3503 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3505 struct task_cputime cputime;
3507 thread_group_cputime(p, &cputime);
3509 *ut = cputime.utime;
3510 *st = cputime.stime;
3514 #ifndef nsecs_to_cputime
3515 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3518 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3520 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
3523 * Use CFS's precise accounting:
3525 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3531 do_div(temp, total);
3532 utime = (cputime_t)temp;
3537 * Compare with previous values, to keep monotonicity:
3539 p->prev_utime = max(p->prev_utime, utime);
3540 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
3542 *ut = p->prev_utime;
3543 *st = p->prev_stime;
3547 * Must be called with siglock held.
3549 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3551 struct signal_struct *sig = p->signal;
3552 struct task_cputime cputime;
3553 cputime_t rtime, utime, total;
3555 thread_group_cputime(p, &cputime);
3557 total = cputime_add(cputime.utime, cputime.stime);
3558 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3563 temp *= cputime.utime;
3564 do_div(temp, total);
3565 utime = (cputime_t)temp;
3569 sig->prev_utime = max(sig->prev_utime, utime);
3570 sig->prev_stime = max(sig->prev_stime,
3571 cputime_sub(rtime, sig->prev_utime));
3573 *ut = sig->prev_utime;
3574 *st = sig->prev_stime;
3579 * This function gets called by the timer code, with HZ frequency.
3580 * We call it with interrupts disabled.
3582 * It also gets called by the fork code, when changing the parent's
3585 void scheduler_tick(void)
3587 int cpu = smp_processor_id();
3588 struct rq *rq = cpu_rq(cpu);
3589 struct task_struct *curr = rq->curr;
3593 raw_spin_lock(&rq->lock);
3594 update_rq_clock(rq);
3595 update_cpu_load_active(rq);
3596 curr->sched_class->task_tick(rq, curr, 0);
3597 raw_spin_unlock(&rq->lock);
3599 perf_event_task_tick();
3602 rq->idle_at_tick = idle_cpu(cpu);
3603 trigger_load_balance(rq, cpu);
3607 notrace unsigned long get_parent_ip(unsigned long addr)
3609 if (in_lock_functions(addr)) {
3610 addr = CALLER_ADDR2;
3611 if (in_lock_functions(addr))
3612 addr = CALLER_ADDR3;
3617 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3618 defined(CONFIG_PREEMPT_TRACER))
3620 void __kprobes add_preempt_count(int val)
3622 #ifdef CONFIG_DEBUG_PREEMPT
3626 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3629 preempt_count() += val;
3630 #ifdef CONFIG_DEBUG_PREEMPT
3632 * Spinlock count overflowing soon?
3634 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3637 if (preempt_count() == val)
3638 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3640 EXPORT_SYMBOL(add_preempt_count);
3642 void __kprobes sub_preempt_count(int val)
3644 #ifdef CONFIG_DEBUG_PREEMPT
3648 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3651 * Is the spinlock portion underflowing?
3653 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3654 !(preempt_count() & PREEMPT_MASK)))
3658 if (preempt_count() == val)
3659 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3660 preempt_count() -= val;
3662 EXPORT_SYMBOL(sub_preempt_count);
3667 * Print scheduling while atomic bug:
3669 static noinline void __schedule_bug(struct task_struct *prev)
3671 struct pt_regs *regs = get_irq_regs();
3673 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3674 prev->comm, prev->pid, preempt_count());
3676 debug_show_held_locks(prev);
3678 if (irqs_disabled())
3679 print_irqtrace_events(prev);
3688 * Various schedule()-time debugging checks and statistics:
3690 static inline void schedule_debug(struct task_struct *prev)
3693 * Test if we are atomic. Since do_exit() needs to call into
3694 * schedule() atomically, we ignore that path for now.
3695 * Otherwise, whine if we are scheduling when we should not be.
3697 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3698 __schedule_bug(prev);
3700 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3702 schedstat_inc(this_rq(), sched_count);
3703 #ifdef CONFIG_SCHEDSTATS
3704 if (unlikely(prev->lock_depth >= 0)) {
3705 schedstat_inc(this_rq(), bkl_count);
3706 schedstat_inc(prev, sched_info.bkl_count);
3711 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3714 update_rq_clock(rq);
3715 rq->skip_clock_update = 0;
3716 prev->sched_class->put_prev_task(rq, prev);
3720 * Pick up the highest-prio task:
3722 static inline struct task_struct *
3723 pick_next_task(struct rq *rq)
3725 const struct sched_class *class;
3726 struct task_struct *p;
3729 * Optimization: we know that if all tasks are in
3730 * the fair class we can call that function directly:
3732 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3733 p = fair_sched_class.pick_next_task(rq);
3738 for_each_class(class) {
3739 p = class->pick_next_task(rq);
3744 BUG(); /* the idle class will always have a runnable task */
3748 * schedule() is the main scheduler function.
3750 asmlinkage void __sched schedule(void)
3752 struct task_struct *prev, *next;
3753 unsigned long *switch_count;
3759 cpu = smp_processor_id();
3761 rcu_note_context_switch(cpu);
3764 release_kernel_lock(prev);
3765 need_resched_nonpreemptible:
3767 schedule_debug(prev);
3769 if (sched_feat(HRTICK))
3772 raw_spin_lock_irq(&rq->lock);
3773 clear_tsk_need_resched(prev);
3775 switch_count = &prev->nivcsw;
3776 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3777 if (unlikely(signal_pending_state(prev->state, prev))) {
3778 prev->state = TASK_RUNNING;
3781 * If a worker is going to sleep, notify and
3782 * ask workqueue whether it wants to wake up a
3783 * task to maintain concurrency. If so, wake
3786 if (prev->flags & PF_WQ_WORKER) {
3787 struct task_struct *to_wakeup;
3789 to_wakeup = wq_worker_sleeping(prev, cpu);
3791 try_to_wake_up_local(to_wakeup);
3793 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3795 switch_count = &prev->nvcsw;
3798 pre_schedule(rq, prev);
3800 if (unlikely(!rq->nr_running))
3801 idle_balance(cpu, rq);
3803 put_prev_task(rq, prev);
3804 next = pick_next_task(rq);
3806 if (likely(prev != next)) {
3807 sched_info_switch(prev, next);
3808 perf_event_task_sched_out(prev, next);
3814 context_switch(rq, prev, next); /* unlocks the rq */
3816 * The context switch have flipped the stack from under us
3817 * and restored the local variables which were saved when
3818 * this task called schedule() in the past. prev == current
3819 * is still correct, but it can be moved to another cpu/rq.
3821 cpu = smp_processor_id();
3824 raw_spin_unlock_irq(&rq->lock);
3828 if (unlikely(reacquire_kernel_lock(prev)))
3829 goto need_resched_nonpreemptible;
3831 preempt_enable_no_resched();
3835 EXPORT_SYMBOL(schedule);
3837 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3839 * Look out! "owner" is an entirely speculative pointer
3840 * access and not reliable.
3842 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
3847 if (!sched_feat(OWNER_SPIN))
3850 #ifdef CONFIG_DEBUG_PAGEALLOC
3852 * Need to access the cpu field knowing that
3853 * DEBUG_PAGEALLOC could have unmapped it if
3854 * the mutex owner just released it and exited.
3856 if (probe_kernel_address(&owner->cpu, cpu))
3863 * Even if the access succeeded (likely case),
3864 * the cpu field may no longer be valid.
3866 if (cpu >= nr_cpumask_bits)
3870 * We need to validate that we can do a
3871 * get_cpu() and that we have the percpu area.
3873 if (!cpu_online(cpu))
3880 * Owner changed, break to re-assess state.
3882 if (lock->owner != owner) {
3884 * If the lock has switched to a different owner,
3885 * we likely have heavy contention. Return 0 to quit
3886 * optimistic spinning and not contend further:
3894 * Is that owner really running on that cpu?
3896 if (task_thread_info(rq->curr) != owner || need_resched())
3899 arch_mutex_cpu_relax();
3906 #ifdef CONFIG_PREEMPT
3908 * this is the entry point to schedule() from in-kernel preemption
3909 * off of preempt_enable. Kernel preemptions off return from interrupt
3910 * occur there and call schedule directly.
3912 asmlinkage void __sched notrace preempt_schedule(void)
3914 struct thread_info *ti = current_thread_info();
3917 * If there is a non-zero preempt_count or interrupts are disabled,
3918 * we do not want to preempt the current task. Just return..
3920 if (likely(ti->preempt_count || irqs_disabled()))
3924 add_preempt_count_notrace(PREEMPT_ACTIVE);
3926 sub_preempt_count_notrace(PREEMPT_ACTIVE);
3929 * Check again in case we missed a preemption opportunity
3930 * between schedule and now.
3933 } while (need_resched());
3935 EXPORT_SYMBOL(preempt_schedule);
3938 * this is the entry point to schedule() from kernel preemption
3939 * off of irq context.
3940 * Note, that this is called and return with irqs disabled. This will
3941 * protect us against recursive calling from irq.
3943 asmlinkage void __sched preempt_schedule_irq(void)
3945 struct thread_info *ti = current_thread_info();
3947 /* Catch callers which need to be fixed */
3948 BUG_ON(ti->preempt_count || !irqs_disabled());
3951 add_preempt_count(PREEMPT_ACTIVE);
3954 local_irq_disable();
3955 sub_preempt_count(PREEMPT_ACTIVE);
3958 * Check again in case we missed a preemption opportunity
3959 * between schedule and now.
3962 } while (need_resched());
3965 #endif /* CONFIG_PREEMPT */
3967 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3970 return try_to_wake_up(curr->private, mode, wake_flags);
3972 EXPORT_SYMBOL(default_wake_function);
3975 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3976 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3977 * number) then we wake all the non-exclusive tasks and one exclusive task.
3979 * There are circumstances in which we can try to wake a task which has already
3980 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3981 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3983 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3984 int nr_exclusive, int wake_flags, void *key)
3986 wait_queue_t *curr, *next;
3988 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3989 unsigned flags = curr->flags;
3991 if (curr->func(curr, mode, wake_flags, key) &&
3992 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3998 * __wake_up - wake up threads blocked on a waitqueue.
4000 * @mode: which threads
4001 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4002 * @key: is directly passed to the wakeup function
4004 * It may be assumed that this function implies a write memory barrier before
4005 * changing the task state if and only if any tasks are woken up.
4007 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4008 int nr_exclusive, void *key)
4010 unsigned long flags;
4012 spin_lock_irqsave(&q->lock, flags);
4013 __wake_up_common(q, mode, nr_exclusive, 0, key);
4014 spin_unlock_irqrestore(&q->lock, flags);
4016 EXPORT_SYMBOL(__wake_up);
4019 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4021 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4023 __wake_up_common(q, mode, 1, 0, NULL);
4025 EXPORT_SYMBOL_GPL(__wake_up_locked);
4027 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
4029 __wake_up_common(q, mode, 1, 0, key);
4033 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4035 * @mode: which threads
4036 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4037 * @key: opaque value to be passed to wakeup targets
4039 * The sync wakeup differs that the waker knows that it will schedule
4040 * away soon, so while the target thread will be woken up, it will not
4041 * be migrated to another CPU - ie. the two threads are 'synchronized'
4042 * with each other. This can prevent needless bouncing between CPUs.
4044 * On UP it can prevent extra preemption.
4046 * It may be assumed that this function implies a write memory barrier before
4047 * changing the task state if and only if any tasks are woken up.
4049 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4050 int nr_exclusive, void *key)
4052 unsigned long flags;
4053 int wake_flags = WF_SYNC;
4058 if (unlikely(!nr_exclusive))
4061 spin_lock_irqsave(&q->lock, flags);
4062 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4063 spin_unlock_irqrestore(&q->lock, flags);
4065 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4068 * __wake_up_sync - see __wake_up_sync_key()
4070 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4072 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4074 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4077 * complete: - signals a single thread waiting on this completion
4078 * @x: holds the state of this particular completion
4080 * This will wake up a single thread waiting on this completion. Threads will be
4081 * awakened in the same order in which they were queued.
4083 * See also complete_all(), wait_for_completion() and related routines.
4085 * It may be assumed that this function implies a write memory barrier before
4086 * changing the task state if and only if any tasks are woken up.
4088 void complete(struct completion *x)
4090 unsigned long flags;
4092 spin_lock_irqsave(&x->wait.lock, flags);
4094 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4095 spin_unlock_irqrestore(&x->wait.lock, flags);
4097 EXPORT_SYMBOL(complete);
4100 * complete_all: - signals all threads waiting on this completion
4101 * @x: holds the state of this particular completion
4103 * This will wake up all threads waiting on this particular completion event.
4105 * It may be assumed that this function implies a write memory barrier before
4106 * changing the task state if and only if any tasks are woken up.
4108 void complete_all(struct completion *x)
4110 unsigned long flags;
4112 spin_lock_irqsave(&x->wait.lock, flags);
4113 x->done += UINT_MAX/2;
4114 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4115 spin_unlock_irqrestore(&x->wait.lock, flags);
4117 EXPORT_SYMBOL(complete_all);
4119 static inline long __sched
4120 do_wait_for_common(struct completion *x, long timeout, int state)
4123 DECLARE_WAITQUEUE(wait, current);
4125 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4127 if (signal_pending_state(state, current)) {
4128 timeout = -ERESTARTSYS;
4131 __set_current_state(state);
4132 spin_unlock_irq(&x->wait.lock);
4133 timeout = schedule_timeout(timeout);
4134 spin_lock_irq(&x->wait.lock);
4135 } while (!x->done && timeout);
4136 __remove_wait_queue(&x->wait, &wait);
4141 return timeout ?: 1;
4145 wait_for_common(struct completion *x, long timeout, int state)
4149 spin_lock_irq(&x->wait.lock);
4150 timeout = do_wait_for_common(x, timeout, state);
4151 spin_unlock_irq(&x->wait.lock);
4156 * wait_for_completion: - waits for completion of a task
4157 * @x: holds the state of this particular completion
4159 * This waits to be signaled for completion of a specific task. It is NOT
4160 * interruptible and there is no timeout.
4162 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4163 * and interrupt capability. Also see complete().
4165 void __sched wait_for_completion(struct completion *x)
4167 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4169 EXPORT_SYMBOL(wait_for_completion);
4172 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4173 * @x: holds the state of this particular completion
4174 * @timeout: timeout value in jiffies
4176 * This waits for either a completion of a specific task to be signaled or for a
4177 * specified timeout to expire. The timeout is in jiffies. It is not
4180 unsigned long __sched
4181 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4183 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4185 EXPORT_SYMBOL(wait_for_completion_timeout);
4188 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4189 * @x: holds the state of this particular completion
4191 * This waits for completion of a specific task to be signaled. It is
4194 int __sched wait_for_completion_interruptible(struct completion *x)
4196 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4197 if (t == -ERESTARTSYS)
4201 EXPORT_SYMBOL(wait_for_completion_interruptible);
4204 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4205 * @x: holds the state of this particular completion
4206 * @timeout: timeout value in jiffies
4208 * This waits for either a completion of a specific task to be signaled or for a
4209 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4211 unsigned long __sched
4212 wait_for_completion_interruptible_timeout(struct completion *x,
4213 unsigned long timeout)
4215 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4217 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4220 * wait_for_completion_killable: - waits for completion of a task (killable)
4221 * @x: holds the state of this particular completion
4223 * This waits to be signaled for completion of a specific task. It can be
4224 * interrupted by a kill signal.
4226 int __sched wait_for_completion_killable(struct completion *x)
4228 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4229 if (t == -ERESTARTSYS)
4233 EXPORT_SYMBOL(wait_for_completion_killable);
4236 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4237 * @x: holds the state of this particular completion
4238 * @timeout: timeout value in jiffies
4240 * This waits for either a completion of a specific task to be
4241 * signaled or for a specified timeout to expire. It can be
4242 * interrupted by a kill signal. The timeout is in jiffies.
4244 unsigned long __sched
4245 wait_for_completion_killable_timeout(struct completion *x,
4246 unsigned long timeout)
4248 return wait_for_common(x, timeout, TASK_KILLABLE);
4250 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
4253 * try_wait_for_completion - try to decrement a completion without blocking
4254 * @x: completion structure
4256 * Returns: 0 if a decrement cannot be done without blocking
4257 * 1 if a decrement succeeded.
4259 * If a completion is being used as a counting completion,
4260 * attempt to decrement the counter without blocking. This
4261 * enables us to avoid waiting if the resource the completion
4262 * is protecting is not available.
4264 bool try_wait_for_completion(struct completion *x)
4266 unsigned long flags;
4269 spin_lock_irqsave(&x->wait.lock, flags);
4274 spin_unlock_irqrestore(&x->wait.lock, flags);
4277 EXPORT_SYMBOL(try_wait_for_completion);
4280 * completion_done - Test to see if a completion has any waiters
4281 * @x: completion structure
4283 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4284 * 1 if there are no waiters.
4287 bool completion_done(struct completion *x)
4289 unsigned long flags;
4292 spin_lock_irqsave(&x->wait.lock, flags);
4295 spin_unlock_irqrestore(&x->wait.lock, flags);
4298 EXPORT_SYMBOL(completion_done);
4301 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4303 unsigned long flags;
4306 init_waitqueue_entry(&wait, current);
4308 __set_current_state(state);
4310 spin_lock_irqsave(&q->lock, flags);
4311 __add_wait_queue(q, &wait);
4312 spin_unlock(&q->lock);
4313 timeout = schedule_timeout(timeout);
4314 spin_lock_irq(&q->lock);
4315 __remove_wait_queue(q, &wait);
4316 spin_unlock_irqrestore(&q->lock, flags);
4321 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4323 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4325 EXPORT_SYMBOL(interruptible_sleep_on);
4328 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4330 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4332 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4334 void __sched sleep_on(wait_queue_head_t *q)
4336 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4338 EXPORT_SYMBOL(sleep_on);
4340 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4342 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4344 EXPORT_SYMBOL(sleep_on_timeout);
4346 #ifdef CONFIG_RT_MUTEXES
4349 * rt_mutex_setprio - set the current priority of a task
4351 * @prio: prio value (kernel-internal form)
4353 * This function changes the 'effective' priority of a task. It does
4354 * not touch ->normal_prio like __setscheduler().
4356 * Used by the rt_mutex code to implement priority inheritance logic.
4358 void rt_mutex_setprio(struct task_struct *p, int prio)
4360 unsigned long flags;
4361 int oldprio, on_rq, running;
4363 const struct sched_class *prev_class;
4365 BUG_ON(prio < 0 || prio > MAX_PRIO);
4367 rq = task_rq_lock(p, &flags);
4369 trace_sched_pi_setprio(p, prio);
4371 prev_class = p->sched_class;
4372 on_rq = p->se.on_rq;
4373 running = task_current(rq, p);
4375 dequeue_task(rq, p, 0);
4377 p->sched_class->put_prev_task(rq, p);
4380 p->sched_class = &rt_sched_class;
4382 p->sched_class = &fair_sched_class;
4387 p->sched_class->set_curr_task(rq);
4389 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4391 check_class_changed(rq, p, prev_class, oldprio, running);
4393 task_rq_unlock(rq, &flags);
4398 void set_user_nice(struct task_struct *p, long nice)
4400 int old_prio, delta, on_rq;
4401 unsigned long flags;
4404 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4407 * We have to be careful, if called from sys_setpriority(),
4408 * the task might be in the middle of scheduling on another CPU.
4410 rq = task_rq_lock(p, &flags);
4412 * The RT priorities are set via sched_setscheduler(), but we still
4413 * allow the 'normal' nice value to be set - but as expected
4414 * it wont have any effect on scheduling until the task is
4415 * SCHED_FIFO/SCHED_RR:
4417 if (task_has_rt_policy(p)) {
4418 p->static_prio = NICE_TO_PRIO(nice);
4421 on_rq = p->se.on_rq;
4423 dequeue_task(rq, p, 0);
4425 p->static_prio = NICE_TO_PRIO(nice);
4428 p->prio = effective_prio(p);
4429 delta = p->prio - old_prio;
4432 enqueue_task(rq, p, 0);
4434 * If the task increased its priority or is running and
4435 * lowered its priority, then reschedule its CPU:
4437 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4438 resched_task(rq->curr);
4441 task_rq_unlock(rq, &flags);
4443 EXPORT_SYMBOL(set_user_nice);
4446 * can_nice - check if a task can reduce its nice value
4450 int can_nice(const struct task_struct *p, const int nice)
4452 /* convert nice value [19,-20] to rlimit style value [1,40] */
4453 int nice_rlim = 20 - nice;
4455 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4456 capable(CAP_SYS_NICE));
4459 #ifdef __ARCH_WANT_SYS_NICE
4462 * sys_nice - change the priority of the current process.
4463 * @increment: priority increment
4465 * sys_setpriority is a more generic, but much slower function that
4466 * does similar things.
4468 SYSCALL_DEFINE1(nice, int, increment)
4473 * Setpriority might change our priority at the same moment.
4474 * We don't have to worry. Conceptually one call occurs first
4475 * and we have a single winner.
4477 if (increment < -40)
4482 nice = TASK_NICE(current) + increment;
4488 if (increment < 0 && !can_nice(current, nice))
4491 retval = security_task_setnice(current, nice);
4495 set_user_nice(current, nice);
4502 * task_prio - return the priority value of a given task.
4503 * @p: the task in question.
4505 * This is the priority value as seen by users in /proc.
4506 * RT tasks are offset by -200. Normal tasks are centered
4507 * around 0, value goes from -16 to +15.
4509 int task_prio(const struct task_struct *p)
4511 return p->prio - MAX_RT_PRIO;
4515 * task_nice - return the nice value of a given task.
4516 * @p: the task in question.
4518 int task_nice(const struct task_struct *p)
4520 return TASK_NICE(p);
4522 EXPORT_SYMBOL(task_nice);
4525 * idle_cpu - is a given cpu idle currently?
4526 * @cpu: the processor in question.
4528 int idle_cpu(int cpu)
4530 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4534 * idle_task - return the idle task for a given cpu.
4535 * @cpu: the processor in question.
4537 struct task_struct *idle_task(int cpu)
4539 return cpu_rq(cpu)->idle;
4543 * find_process_by_pid - find a process with a matching PID value.
4544 * @pid: the pid in question.
4546 static struct task_struct *find_process_by_pid(pid_t pid)
4548 return pid ? find_task_by_vpid(pid) : current;
4551 /* Actually do priority change: must hold rq lock. */
4553 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4555 BUG_ON(p->se.on_rq);
4558 p->rt_priority = prio;
4559 p->normal_prio = normal_prio(p);
4560 /* we are holding p->pi_lock already */
4561 p->prio = rt_mutex_getprio(p);
4562 if (rt_prio(p->prio))
4563 p->sched_class = &rt_sched_class;
4565 p->sched_class = &fair_sched_class;
4570 * check the target process has a UID that matches the current process's
4572 static bool check_same_owner(struct task_struct *p)
4574 const struct cred *cred = current_cred(), *pcred;
4578 pcred = __task_cred(p);
4579 match = (cred->euid == pcred->euid ||
4580 cred->euid == pcred->uid);
4585 static int __sched_setscheduler(struct task_struct *p, int policy,
4586 const struct sched_param *param, bool user)
4588 int retval, oldprio, oldpolicy = -1, on_rq, running;
4589 unsigned long flags;
4590 const struct sched_class *prev_class;
4594 /* may grab non-irq protected spin_locks */
4595 BUG_ON(in_interrupt());
4597 /* double check policy once rq lock held */
4599 reset_on_fork = p->sched_reset_on_fork;
4600 policy = oldpolicy = p->policy;
4602 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4603 policy &= ~SCHED_RESET_ON_FORK;
4605 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4606 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4607 policy != SCHED_IDLE)
4612 * Valid priorities for SCHED_FIFO and SCHED_RR are
4613 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4614 * SCHED_BATCH and SCHED_IDLE is 0.
4616 if (param->sched_priority < 0 ||
4617 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4618 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4620 if (rt_policy(policy) != (param->sched_priority != 0))
4624 * Allow unprivileged RT tasks to decrease priority:
4626 if (user && !capable(CAP_SYS_NICE)) {
4627 if (rt_policy(policy)) {
4628 unsigned long rlim_rtprio =
4629 task_rlimit(p, RLIMIT_RTPRIO);
4631 /* can't set/change the rt policy */
4632 if (policy != p->policy && !rlim_rtprio)
4635 /* can't increase priority */
4636 if (param->sched_priority > p->rt_priority &&
4637 param->sched_priority > rlim_rtprio)
4641 * Like positive nice levels, dont allow tasks to
4642 * move out of SCHED_IDLE either:
4644 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4647 /* can't change other user's priorities */
4648 if (!check_same_owner(p))
4651 /* Normal users shall not reset the sched_reset_on_fork flag */
4652 if (p->sched_reset_on_fork && !reset_on_fork)
4657 retval = security_task_setscheduler(p);
4663 * make sure no PI-waiters arrive (or leave) while we are
4664 * changing the priority of the task:
4666 raw_spin_lock_irqsave(&p->pi_lock, flags);
4668 * To be able to change p->policy safely, the apropriate
4669 * runqueue lock must be held.
4671 rq = __task_rq_lock(p);
4674 * Changing the policy of the stop threads its a very bad idea
4676 if (p == rq->stop) {
4677 __task_rq_unlock(rq);
4678 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4682 #ifdef CONFIG_RT_GROUP_SCHED
4685 * Do not allow realtime tasks into groups that have no runtime
4688 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4689 task_group(p)->rt_bandwidth.rt_runtime == 0) {
4690 __task_rq_unlock(rq);
4691 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4697 /* recheck policy now with rq lock held */
4698 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4699 policy = oldpolicy = -1;
4700 __task_rq_unlock(rq);
4701 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4704 on_rq = p->se.on_rq;
4705 running = task_current(rq, p);
4707 deactivate_task(rq, p, 0);
4709 p->sched_class->put_prev_task(rq, p);
4711 p->sched_reset_on_fork = reset_on_fork;
4714 prev_class = p->sched_class;
4715 __setscheduler(rq, p, policy, param->sched_priority);
4718 p->sched_class->set_curr_task(rq);
4720 activate_task(rq, p, 0);
4722 check_class_changed(rq, p, prev_class, oldprio, running);
4724 __task_rq_unlock(rq);
4725 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4727 rt_mutex_adjust_pi(p);
4733 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4734 * @p: the task in question.
4735 * @policy: new policy.
4736 * @param: structure containing the new RT priority.
4738 * NOTE that the task may be already dead.
4740 int sched_setscheduler(struct task_struct *p, int policy,
4741 const struct sched_param *param)
4743 return __sched_setscheduler(p, policy, param, true);
4745 EXPORT_SYMBOL_GPL(sched_setscheduler);
4748 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4749 * @p: the task in question.
4750 * @policy: new policy.
4751 * @param: structure containing the new RT priority.
4753 * Just like sched_setscheduler, only don't bother checking if the
4754 * current context has permission. For example, this is needed in
4755 * stop_machine(): we create temporary high priority worker threads,
4756 * but our caller might not have that capability.
4758 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4759 const struct sched_param *param)
4761 return __sched_setscheduler(p, policy, param, false);
4765 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4767 struct sched_param lparam;
4768 struct task_struct *p;
4771 if (!param || pid < 0)
4773 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4778 p = find_process_by_pid(pid);
4780 retval = sched_setscheduler(p, policy, &lparam);
4787 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4788 * @pid: the pid in question.
4789 * @policy: new policy.
4790 * @param: structure containing the new RT priority.
4792 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4793 struct sched_param __user *, param)
4795 /* negative values for policy are not valid */
4799 return do_sched_setscheduler(pid, policy, param);
4803 * sys_sched_setparam - set/change the RT priority of a thread
4804 * @pid: the pid in question.
4805 * @param: structure containing the new RT priority.
4807 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4809 return do_sched_setscheduler(pid, -1, param);
4813 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4814 * @pid: the pid in question.
4816 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4818 struct task_struct *p;
4826 p = find_process_by_pid(pid);
4828 retval = security_task_getscheduler(p);
4831 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4838 * sys_sched_getparam - get the RT priority of a thread
4839 * @pid: the pid in question.
4840 * @param: structure containing the RT priority.
4842 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4844 struct sched_param lp;
4845 struct task_struct *p;
4848 if (!param || pid < 0)
4852 p = find_process_by_pid(pid);
4857 retval = security_task_getscheduler(p);
4861 lp.sched_priority = p->rt_priority;
4865 * This one might sleep, we cannot do it with a spinlock held ...
4867 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4876 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4878 cpumask_var_t cpus_allowed, new_mask;
4879 struct task_struct *p;
4885 p = find_process_by_pid(pid);
4892 /* Prevent p going away */
4896 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4900 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4902 goto out_free_cpus_allowed;
4905 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
4908 retval = security_task_setscheduler(p);
4912 cpuset_cpus_allowed(p, cpus_allowed);
4913 cpumask_and(new_mask, in_mask, cpus_allowed);
4915 retval = set_cpus_allowed_ptr(p, new_mask);
4918 cpuset_cpus_allowed(p, cpus_allowed);
4919 if (!cpumask_subset(new_mask, cpus_allowed)) {
4921 * We must have raced with a concurrent cpuset
4922 * update. Just reset the cpus_allowed to the
4923 * cpuset's cpus_allowed
4925 cpumask_copy(new_mask, cpus_allowed);
4930 free_cpumask_var(new_mask);
4931 out_free_cpus_allowed:
4932 free_cpumask_var(cpus_allowed);
4939 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4940 struct cpumask *new_mask)
4942 if (len < cpumask_size())
4943 cpumask_clear(new_mask);
4944 else if (len > cpumask_size())
4945 len = cpumask_size();
4947 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4951 * sys_sched_setaffinity - set the cpu affinity of a process
4952 * @pid: pid of the process
4953 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4954 * @user_mask_ptr: user-space pointer to the new cpu mask
4956 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4957 unsigned long __user *, user_mask_ptr)
4959 cpumask_var_t new_mask;
4962 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4965 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4967 retval = sched_setaffinity(pid, new_mask);
4968 free_cpumask_var(new_mask);
4972 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4974 struct task_struct *p;
4975 unsigned long flags;
4983 p = find_process_by_pid(pid);
4987 retval = security_task_getscheduler(p);
4991 rq = task_rq_lock(p, &flags);
4992 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4993 task_rq_unlock(rq, &flags);
5003 * sys_sched_getaffinity - get the cpu affinity of a process
5004 * @pid: pid of the process
5005 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5006 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5008 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5009 unsigned long __user *, user_mask_ptr)
5014 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5016 if (len & (sizeof(unsigned long)-1))
5019 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5022 ret = sched_getaffinity(pid, mask);
5024 size_t retlen = min_t(size_t, len, cpumask_size());
5026 if (copy_to_user(user_mask_ptr, mask, retlen))
5031 free_cpumask_var(mask);
5037 * sys_sched_yield - yield the current processor to other threads.
5039 * This function yields the current CPU to other tasks. If there are no
5040 * other threads running on this CPU then this function will return.
5042 SYSCALL_DEFINE0(sched_yield)
5044 struct rq *rq = this_rq_lock();
5046 schedstat_inc(rq, yld_count);
5047 current->sched_class->yield_task(rq);
5050 * Since we are going to call schedule() anyway, there's
5051 * no need to preempt or enable interrupts:
5053 __release(rq->lock);
5054 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5055 do_raw_spin_unlock(&rq->lock);
5056 preempt_enable_no_resched();
5063 static inline int should_resched(void)
5065 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5068 static void __cond_resched(void)
5070 add_preempt_count(PREEMPT_ACTIVE);
5072 sub_preempt_count(PREEMPT_ACTIVE);
5075 int __sched _cond_resched(void)
5077 if (should_resched()) {
5083 EXPORT_SYMBOL(_cond_resched);
5086 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5087 * call schedule, and on return reacquire the lock.
5089 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5090 * operations here to prevent schedule() from being called twice (once via
5091 * spin_unlock(), once by hand).
5093 int __cond_resched_lock(spinlock_t *lock)
5095 int resched = should_resched();
5098 lockdep_assert_held(lock);
5100 if (spin_needbreak(lock) || resched) {
5111 EXPORT_SYMBOL(__cond_resched_lock);
5113 int __sched __cond_resched_softirq(void)
5115 BUG_ON(!in_softirq());
5117 if (should_resched()) {
5125 EXPORT_SYMBOL(__cond_resched_softirq);
5128 * yield - yield the current processor to other threads.
5130 * This is a shortcut for kernel-space yielding - it marks the
5131 * thread runnable and calls sys_sched_yield().
5133 void __sched yield(void)
5135 set_current_state(TASK_RUNNING);
5138 EXPORT_SYMBOL(yield);
5141 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5142 * that process accounting knows that this is a task in IO wait state.
5144 void __sched io_schedule(void)
5146 struct rq *rq = raw_rq();
5148 delayacct_blkio_start();
5149 atomic_inc(&rq->nr_iowait);
5150 current->in_iowait = 1;
5152 current->in_iowait = 0;
5153 atomic_dec(&rq->nr_iowait);
5154 delayacct_blkio_end();
5156 EXPORT_SYMBOL(io_schedule);
5158 long __sched io_schedule_timeout(long timeout)
5160 struct rq *rq = raw_rq();
5163 delayacct_blkio_start();
5164 atomic_inc(&rq->nr_iowait);
5165 current->in_iowait = 1;
5166 ret = schedule_timeout(timeout);
5167 current->in_iowait = 0;
5168 atomic_dec(&rq->nr_iowait);
5169 delayacct_blkio_end();
5174 * sys_sched_get_priority_max - return maximum RT priority.
5175 * @policy: scheduling class.
5177 * this syscall returns the maximum rt_priority that can be used
5178 * by a given scheduling class.
5180 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5187 ret = MAX_USER_RT_PRIO-1;
5199 * sys_sched_get_priority_min - return minimum RT priority.
5200 * @policy: scheduling class.
5202 * this syscall returns the minimum rt_priority that can be used
5203 * by a given scheduling class.
5205 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5223 * sys_sched_rr_get_interval - return the default timeslice of a process.
5224 * @pid: pid of the process.
5225 * @interval: userspace pointer to the timeslice value.
5227 * this syscall writes the default timeslice value of a given process
5228 * into the user-space timespec buffer. A value of '0' means infinity.
5230 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5231 struct timespec __user *, interval)
5233 struct task_struct *p;
5234 unsigned int time_slice;
5235 unsigned long flags;
5245 p = find_process_by_pid(pid);
5249 retval = security_task_getscheduler(p);
5253 rq = task_rq_lock(p, &flags);
5254 time_slice = p->sched_class->get_rr_interval(rq, p);
5255 task_rq_unlock(rq, &flags);
5258 jiffies_to_timespec(time_slice, &t);
5259 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5267 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5269 void sched_show_task(struct task_struct *p)
5271 unsigned long free = 0;
5274 state = p->state ? __ffs(p->state) + 1 : 0;
5275 printk(KERN_INFO "%-15.15s %c", p->comm,
5276 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5277 #if BITS_PER_LONG == 32
5278 if (state == TASK_RUNNING)
5279 printk(KERN_CONT " running ");
5281 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5283 if (state == TASK_RUNNING)
5284 printk(KERN_CONT " running task ");
5286 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5288 #ifdef CONFIG_DEBUG_STACK_USAGE
5289 free = stack_not_used(p);
5291 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5292 task_pid_nr(p), task_pid_nr(p->real_parent),
5293 (unsigned long)task_thread_info(p)->flags);
5295 show_stack(p, NULL);
5298 void show_state_filter(unsigned long state_filter)
5300 struct task_struct *g, *p;
5302 #if BITS_PER_LONG == 32
5304 " task PC stack pid father\n");
5307 " task PC stack pid father\n");
5309 read_lock(&tasklist_lock);
5310 do_each_thread(g, p) {
5312 * reset the NMI-timeout, listing all files on a slow
5313 * console might take alot of time:
5315 touch_nmi_watchdog();
5316 if (!state_filter || (p->state & state_filter))
5318 } while_each_thread(g, p);
5320 touch_all_softlockup_watchdogs();
5322 #ifdef CONFIG_SCHED_DEBUG
5323 sysrq_sched_debug_show();
5325 read_unlock(&tasklist_lock);
5327 * Only show locks if all tasks are dumped:
5330 debug_show_all_locks();
5333 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5335 idle->sched_class = &idle_sched_class;
5339 * init_idle - set up an idle thread for a given CPU
5340 * @idle: task in question
5341 * @cpu: cpu the idle task belongs to
5343 * NOTE: this function does not set the idle thread's NEED_RESCHED
5344 * flag, to make booting more robust.
5346 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5348 struct rq *rq = cpu_rq(cpu);
5349 unsigned long flags;
5351 raw_spin_lock_irqsave(&rq->lock, flags);
5354 idle->state = TASK_RUNNING;
5355 idle->se.exec_start = sched_clock();
5357 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
5359 * We're having a chicken and egg problem, even though we are
5360 * holding rq->lock, the cpu isn't yet set to this cpu so the
5361 * lockdep check in task_group() will fail.
5363 * Similar case to sched_fork(). / Alternatively we could
5364 * use task_rq_lock() here and obtain the other rq->lock.
5369 __set_task_cpu(idle, cpu);
5372 rq->curr = rq->idle = idle;
5373 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5376 raw_spin_unlock_irqrestore(&rq->lock, flags);
5378 /* Set the preempt count _outside_ the spinlocks! */
5379 #if defined(CONFIG_PREEMPT)
5380 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5382 task_thread_info(idle)->preempt_count = 0;
5385 * The idle tasks have their own, simple scheduling class:
5387 idle->sched_class = &idle_sched_class;
5388 ftrace_graph_init_task(idle);
5392 * In a system that switches off the HZ timer nohz_cpu_mask
5393 * indicates which cpus entered this state. This is used
5394 * in the rcu update to wait only for active cpus. For system
5395 * which do not switch off the HZ timer nohz_cpu_mask should
5396 * always be CPU_BITS_NONE.
5398 cpumask_var_t nohz_cpu_mask;
5401 * Increase the granularity value when there are more CPUs,
5402 * because with more CPUs the 'effective latency' as visible
5403 * to users decreases. But the relationship is not linear,
5404 * so pick a second-best guess by going with the log2 of the
5407 * This idea comes from the SD scheduler of Con Kolivas:
5409 static int get_update_sysctl_factor(void)
5411 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5412 unsigned int factor;
5414 switch (sysctl_sched_tunable_scaling) {
5415 case SCHED_TUNABLESCALING_NONE:
5418 case SCHED_TUNABLESCALING_LINEAR:
5421 case SCHED_TUNABLESCALING_LOG:
5423 factor = 1 + ilog2(cpus);
5430 static void update_sysctl(void)
5432 unsigned int factor = get_update_sysctl_factor();
5434 #define SET_SYSCTL(name) \
5435 (sysctl_##name = (factor) * normalized_sysctl_##name)
5436 SET_SYSCTL(sched_min_granularity);
5437 SET_SYSCTL(sched_latency);
5438 SET_SYSCTL(sched_wakeup_granularity);
5442 static inline void sched_init_granularity(void)
5449 * This is how migration works:
5451 * 1) we invoke migration_cpu_stop() on the target CPU using
5453 * 2) stopper starts to run (implicitly forcing the migrated thread
5455 * 3) it checks whether the migrated task is still in the wrong runqueue.
5456 * 4) if it's in the wrong runqueue then the migration thread removes
5457 * it and puts it into the right queue.
5458 * 5) stopper completes and stop_one_cpu() returns and the migration
5463 * Change a given task's CPU affinity. Migrate the thread to a
5464 * proper CPU and schedule it away if the CPU it's executing on
5465 * is removed from the allowed bitmask.
5467 * NOTE: the caller must have a valid reference to the task, the
5468 * task must not exit() & deallocate itself prematurely. The
5469 * call is not atomic; no spinlocks may be held.
5471 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5473 unsigned long flags;
5475 unsigned int dest_cpu;
5479 * Serialize against TASK_WAKING so that ttwu() and wunt() can
5480 * drop the rq->lock and still rely on ->cpus_allowed.
5483 while (task_is_waking(p))
5485 rq = task_rq_lock(p, &flags);
5486 if (task_is_waking(p)) {
5487 task_rq_unlock(rq, &flags);
5491 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5496 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5497 !cpumask_equal(&p->cpus_allowed, new_mask))) {
5502 if (p->sched_class->set_cpus_allowed)
5503 p->sched_class->set_cpus_allowed(p, new_mask);
5505 cpumask_copy(&p->cpus_allowed, new_mask);
5506 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5509 /* Can the task run on the task's current CPU? If so, we're done */
5510 if (cpumask_test_cpu(task_cpu(p), new_mask))
5513 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
5514 if (migrate_task(p, rq)) {
5515 struct migration_arg arg = { p, dest_cpu };
5516 /* Need help from migration thread: drop lock and wait. */
5517 task_rq_unlock(rq, &flags);
5518 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
5519 tlb_migrate_finish(p->mm);
5523 task_rq_unlock(rq, &flags);
5527 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5530 * Move (not current) task off this cpu, onto dest cpu. We're doing
5531 * this because either it can't run here any more (set_cpus_allowed()
5532 * away from this CPU, or CPU going down), or because we're
5533 * attempting to rebalance this task on exec (sched_exec).
5535 * So we race with normal scheduler movements, but that's OK, as long
5536 * as the task is no longer on this CPU.
5538 * Returns non-zero if task was successfully migrated.
5540 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5542 struct rq *rq_dest, *rq_src;
5545 if (unlikely(!cpu_active(dest_cpu)))
5548 rq_src = cpu_rq(src_cpu);
5549 rq_dest = cpu_rq(dest_cpu);
5551 double_rq_lock(rq_src, rq_dest);
5552 /* Already moved. */
5553 if (task_cpu(p) != src_cpu)
5555 /* Affinity changed (again). */
5556 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
5560 * If we're not on a rq, the next wake-up will ensure we're
5564 deactivate_task(rq_src, p, 0);
5565 set_task_cpu(p, dest_cpu);
5566 activate_task(rq_dest, p, 0);
5567 check_preempt_curr(rq_dest, p, 0);
5572 double_rq_unlock(rq_src, rq_dest);
5577 * migration_cpu_stop - this will be executed by a highprio stopper thread
5578 * and performs thread migration by bumping thread off CPU then
5579 * 'pushing' onto another runqueue.
5581 static int migration_cpu_stop(void *data)
5583 struct migration_arg *arg = data;
5586 * The original target cpu might have gone down and we might
5587 * be on another cpu but it doesn't matter.
5589 local_irq_disable();
5590 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
5595 #ifdef CONFIG_HOTPLUG_CPU
5598 * Ensures that the idle task is using init_mm right before its cpu goes
5601 void idle_task_exit(void)
5603 struct mm_struct *mm = current->active_mm;
5605 BUG_ON(cpu_online(smp_processor_id()));
5608 switch_mm(mm, &init_mm, current);
5613 * While a dead CPU has no uninterruptible tasks queued at this point,
5614 * it might still have a nonzero ->nr_uninterruptible counter, because
5615 * for performance reasons the counter is not stricly tracking tasks to
5616 * their home CPUs. So we just add the counter to another CPU's counter,
5617 * to keep the global sum constant after CPU-down:
5619 static void migrate_nr_uninterruptible(struct rq *rq_src)
5621 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5623 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5624 rq_src->nr_uninterruptible = 0;
5628 * remove the tasks which were accounted by rq from calc_load_tasks.
5630 static void calc_global_load_remove(struct rq *rq)
5632 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5633 rq->calc_load_active = 0;
5637 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5638 * try_to_wake_up()->select_task_rq().
5640 * Called with rq->lock held even though we'er in stop_machine() and
5641 * there's no concurrency possible, we hold the required locks anyway
5642 * because of lock validation efforts.
5644 static void migrate_tasks(unsigned int dead_cpu)
5646 struct rq *rq = cpu_rq(dead_cpu);
5647 struct task_struct *next, *stop = rq->stop;
5651 * Fudge the rq selection such that the below task selection loop
5652 * doesn't get stuck on the currently eligible stop task.
5654 * We're currently inside stop_machine() and the rq is either stuck
5655 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5656 * either way we should never end up calling schedule() until we're
5663 * There's this thread running, bail when that's the only
5666 if (rq->nr_running == 1)
5669 next = pick_next_task(rq);
5671 next->sched_class->put_prev_task(rq, next);
5673 /* Find suitable destination for @next, with force if needed. */
5674 dest_cpu = select_fallback_rq(dead_cpu, next);
5675 raw_spin_unlock(&rq->lock);
5677 __migrate_task(next, dead_cpu, dest_cpu);
5679 raw_spin_lock(&rq->lock);
5685 #endif /* CONFIG_HOTPLUG_CPU */
5687 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5689 static struct ctl_table sd_ctl_dir[] = {
5691 .procname = "sched_domain",
5697 static struct ctl_table sd_ctl_root[] = {
5699 .procname = "kernel",
5701 .child = sd_ctl_dir,
5706 static struct ctl_table *sd_alloc_ctl_entry(int n)
5708 struct ctl_table *entry =
5709 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5714 static void sd_free_ctl_entry(struct ctl_table **tablep)
5716 struct ctl_table *entry;
5719 * In the intermediate directories, both the child directory and
5720 * procname are dynamically allocated and could fail but the mode
5721 * will always be set. In the lowest directory the names are
5722 * static strings and all have proc handlers.
5724 for (entry = *tablep; entry->mode; entry++) {
5726 sd_free_ctl_entry(&entry->child);
5727 if (entry->proc_handler == NULL)
5728 kfree(entry->procname);
5736 set_table_entry(struct ctl_table *entry,
5737 const char *procname, void *data, int maxlen,
5738 mode_t mode, proc_handler *proc_handler)
5740 entry->procname = procname;
5742 entry->maxlen = maxlen;
5744 entry->proc_handler = proc_handler;
5747 static struct ctl_table *
5748 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5750 struct ctl_table *table = sd_alloc_ctl_entry(13);
5755 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5756 sizeof(long), 0644, proc_doulongvec_minmax);
5757 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5758 sizeof(long), 0644, proc_doulongvec_minmax);
5759 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5760 sizeof(int), 0644, proc_dointvec_minmax);
5761 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5762 sizeof(int), 0644, proc_dointvec_minmax);
5763 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5764 sizeof(int), 0644, proc_dointvec_minmax);
5765 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5766 sizeof(int), 0644, proc_dointvec_minmax);
5767 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5768 sizeof(int), 0644, proc_dointvec_minmax);
5769 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5770 sizeof(int), 0644, proc_dointvec_minmax);
5771 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5772 sizeof(int), 0644, proc_dointvec_minmax);
5773 set_table_entry(&table[9], "cache_nice_tries",
5774 &sd->cache_nice_tries,
5775 sizeof(int), 0644, proc_dointvec_minmax);
5776 set_table_entry(&table[10], "flags", &sd->flags,
5777 sizeof(int), 0644, proc_dointvec_minmax);
5778 set_table_entry(&table[11], "name", sd->name,
5779 CORENAME_MAX_SIZE, 0444, proc_dostring);
5780 /* &table[12] is terminator */
5785 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5787 struct ctl_table *entry, *table;
5788 struct sched_domain *sd;
5789 int domain_num = 0, i;
5792 for_each_domain(cpu, sd)
5794 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5799 for_each_domain(cpu, sd) {
5800 snprintf(buf, 32, "domain%d", i);
5801 entry->procname = kstrdup(buf, GFP_KERNEL);
5803 entry->child = sd_alloc_ctl_domain_table(sd);
5810 static struct ctl_table_header *sd_sysctl_header;
5811 static void register_sched_domain_sysctl(void)
5813 int i, cpu_num = num_possible_cpus();
5814 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5817 WARN_ON(sd_ctl_dir[0].child);
5818 sd_ctl_dir[0].child = entry;
5823 for_each_possible_cpu(i) {
5824 snprintf(buf, 32, "cpu%d", i);
5825 entry->procname = kstrdup(buf, GFP_KERNEL);
5827 entry->child = sd_alloc_ctl_cpu_table(i);
5831 WARN_ON(sd_sysctl_header);
5832 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5835 /* may be called multiple times per register */
5836 static void unregister_sched_domain_sysctl(void)
5838 if (sd_sysctl_header)
5839 unregister_sysctl_table(sd_sysctl_header);
5840 sd_sysctl_header = NULL;
5841 if (sd_ctl_dir[0].child)
5842 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5845 static void register_sched_domain_sysctl(void)
5848 static void unregister_sched_domain_sysctl(void)
5853 static void set_rq_online(struct rq *rq)
5856 const struct sched_class *class;
5858 cpumask_set_cpu(rq->cpu, rq->rd->online);
5861 for_each_class(class) {
5862 if (class->rq_online)
5863 class->rq_online(rq);
5868 static void set_rq_offline(struct rq *rq)
5871 const struct sched_class *class;
5873 for_each_class(class) {
5874 if (class->rq_offline)
5875 class->rq_offline(rq);
5878 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5884 * migration_call - callback that gets triggered when a CPU is added.
5885 * Here we can start up the necessary migration thread for the new CPU.
5887 static int __cpuinit
5888 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5890 int cpu = (long)hcpu;
5891 unsigned long flags;
5892 struct rq *rq = cpu_rq(cpu);
5894 switch (action & ~CPU_TASKS_FROZEN) {
5896 case CPU_UP_PREPARE:
5897 rq->calc_load_update = calc_load_update;
5901 /* Update our root-domain */
5902 raw_spin_lock_irqsave(&rq->lock, flags);
5904 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5908 raw_spin_unlock_irqrestore(&rq->lock, flags);
5911 #ifdef CONFIG_HOTPLUG_CPU
5913 /* Update our root-domain */
5914 raw_spin_lock_irqsave(&rq->lock, flags);
5916 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5920 BUG_ON(rq->nr_running != 1); /* the migration thread */
5921 raw_spin_unlock_irqrestore(&rq->lock, flags);
5923 migrate_nr_uninterruptible(rq);
5924 calc_global_load_remove(rq);
5932 * Register at high priority so that task migration (migrate_all_tasks)
5933 * happens before everything else. This has to be lower priority than
5934 * the notifier in the perf_event subsystem, though.
5936 static struct notifier_block __cpuinitdata migration_notifier = {
5937 .notifier_call = migration_call,
5938 .priority = CPU_PRI_MIGRATION,
5941 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
5942 unsigned long action, void *hcpu)
5944 switch (action & ~CPU_TASKS_FROZEN) {
5946 case CPU_DOWN_FAILED:
5947 set_cpu_active((long)hcpu, true);
5954 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
5955 unsigned long action, void *hcpu)
5957 switch (action & ~CPU_TASKS_FROZEN) {
5958 case CPU_DOWN_PREPARE:
5959 set_cpu_active((long)hcpu, false);
5966 static int __init migration_init(void)
5968 void *cpu = (void *)(long)smp_processor_id();
5971 /* Initialize migration for the boot CPU */
5972 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5973 BUG_ON(err == NOTIFY_BAD);
5974 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5975 register_cpu_notifier(&migration_notifier);
5977 /* Register cpu active notifiers */
5978 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5979 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5983 early_initcall(migration_init);
5988 #ifdef CONFIG_SCHED_DEBUG
5990 static __read_mostly int sched_domain_debug_enabled;
5992 static int __init sched_domain_debug_setup(char *str)
5994 sched_domain_debug_enabled = 1;
5998 early_param("sched_debug", sched_domain_debug_setup);
6000 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6001 struct cpumask *groupmask)
6003 struct sched_group *group = sd->groups;
6006 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6007 cpumask_clear(groupmask);
6009 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6011 if (!(sd->flags & SD_LOAD_BALANCE)) {
6012 printk("does not load-balance\n");
6014 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6019 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6021 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6022 printk(KERN_ERR "ERROR: domain->span does not contain "
6025 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6026 printk(KERN_ERR "ERROR: domain->groups does not contain"
6030 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6034 printk(KERN_ERR "ERROR: group is NULL\n");
6038 if (!group->cpu_power) {
6039 printk(KERN_CONT "\n");
6040 printk(KERN_ERR "ERROR: domain->cpu_power not "
6045 if (!cpumask_weight(sched_group_cpus(group))) {
6046 printk(KERN_CONT "\n");
6047 printk(KERN_ERR "ERROR: empty group\n");
6051 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6052 printk(KERN_CONT "\n");
6053 printk(KERN_ERR "ERROR: repeated CPUs\n");
6057 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6059 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6061 printk(KERN_CONT " %s", str);
6062 if (group->cpu_power != SCHED_LOAD_SCALE) {
6063 printk(KERN_CONT " (cpu_power = %d)",
6067 group = group->next;
6068 } while (group != sd->groups);
6069 printk(KERN_CONT "\n");
6071 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6072 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6075 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6076 printk(KERN_ERR "ERROR: parent span is not a superset "
6077 "of domain->span\n");
6081 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6083 cpumask_var_t groupmask;
6086 if (!sched_domain_debug_enabled)
6090 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6094 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6096 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6097 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6102 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6109 free_cpumask_var(groupmask);
6111 #else /* !CONFIG_SCHED_DEBUG */
6112 # define sched_domain_debug(sd, cpu) do { } while (0)
6113 #endif /* CONFIG_SCHED_DEBUG */
6115 static int sd_degenerate(struct sched_domain *sd)
6117 if (cpumask_weight(sched_domain_span(sd)) == 1)
6120 /* Following flags need at least 2 groups */
6121 if (sd->flags & (SD_LOAD_BALANCE |
6122 SD_BALANCE_NEWIDLE |
6126 SD_SHARE_PKG_RESOURCES)) {
6127 if (sd->groups != sd->groups->next)
6131 /* Following flags don't use groups */
6132 if (sd->flags & (SD_WAKE_AFFINE))
6139 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6141 unsigned long cflags = sd->flags, pflags = parent->flags;
6143 if (sd_degenerate(parent))
6146 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6149 /* Flags needing groups don't count if only 1 group in parent */
6150 if (parent->groups == parent->groups->next) {
6151 pflags &= ~(SD_LOAD_BALANCE |
6152 SD_BALANCE_NEWIDLE |
6156 SD_SHARE_PKG_RESOURCES);
6157 if (nr_node_ids == 1)
6158 pflags &= ~SD_SERIALIZE;
6160 if (~cflags & pflags)
6166 static void free_rootdomain(struct root_domain *rd)
6168 synchronize_sched();
6170 cpupri_cleanup(&rd->cpupri);
6172 free_cpumask_var(rd->rto_mask);
6173 free_cpumask_var(rd->online);
6174 free_cpumask_var(rd->span);
6178 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6180 struct root_domain *old_rd = NULL;
6181 unsigned long flags;
6183 raw_spin_lock_irqsave(&rq->lock, flags);
6188 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6191 cpumask_clear_cpu(rq->cpu, old_rd->span);
6194 * If we dont want to free the old_rt yet then
6195 * set old_rd to NULL to skip the freeing later
6198 if (!atomic_dec_and_test(&old_rd->refcount))
6202 atomic_inc(&rd->refcount);
6205 cpumask_set_cpu(rq->cpu, rd->span);
6206 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6209 raw_spin_unlock_irqrestore(&rq->lock, flags);
6212 free_rootdomain(old_rd);
6215 static int init_rootdomain(struct root_domain *rd)
6217 memset(rd, 0, sizeof(*rd));
6219 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
6221 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
6223 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6226 if (cpupri_init(&rd->cpupri) != 0)
6231 free_cpumask_var(rd->rto_mask);
6233 free_cpumask_var(rd->online);
6235 free_cpumask_var(rd->span);
6240 static void init_defrootdomain(void)
6242 init_rootdomain(&def_root_domain);
6244 atomic_set(&def_root_domain.refcount, 1);
6247 static struct root_domain *alloc_rootdomain(void)
6249 struct root_domain *rd;
6251 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6255 if (init_rootdomain(rd) != 0) {
6264 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6265 * hold the hotplug lock.
6268 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6270 struct rq *rq = cpu_rq(cpu);
6271 struct sched_domain *tmp;
6273 for (tmp = sd; tmp; tmp = tmp->parent)
6274 tmp->span_weight = cpumask_weight(sched_domain_span(tmp));
6276 /* Remove the sched domains which do not contribute to scheduling. */
6277 for (tmp = sd; tmp; ) {
6278 struct sched_domain *parent = tmp->parent;
6282 if (sd_parent_degenerate(tmp, parent)) {
6283 tmp->parent = parent->parent;
6285 parent->parent->child = tmp;
6290 if (sd && sd_degenerate(sd)) {
6296 sched_domain_debug(sd, cpu);
6298 rq_attach_root(rq, rd);
6299 rcu_assign_pointer(rq->sd, sd);
6302 /* cpus with isolated domains */
6303 static cpumask_var_t cpu_isolated_map;
6305 /* Setup the mask of cpus configured for isolated domains */
6306 static int __init isolated_cpu_setup(char *str)
6308 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6309 cpulist_parse(str, cpu_isolated_map);
6313 __setup("isolcpus=", isolated_cpu_setup);
6316 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6317 * to a function which identifies what group(along with sched group) a CPU
6318 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6319 * (due to the fact that we keep track of groups covered with a struct cpumask).
6321 * init_sched_build_groups will build a circular linked list of the groups
6322 * covered by the given span, and will set each group's ->cpumask correctly,
6323 * and ->cpu_power to 0.
6326 init_sched_build_groups(const struct cpumask *span,
6327 const struct cpumask *cpu_map,
6328 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
6329 struct sched_group **sg,
6330 struct cpumask *tmpmask),
6331 struct cpumask *covered, struct cpumask *tmpmask)
6333 struct sched_group *first = NULL, *last = NULL;
6336 cpumask_clear(covered);
6338 for_each_cpu(i, span) {
6339 struct sched_group *sg;
6340 int group = group_fn(i, cpu_map, &sg, tmpmask);
6343 if (cpumask_test_cpu(i, covered))
6346 cpumask_clear(sched_group_cpus(sg));
6349 for_each_cpu(j, span) {
6350 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6353 cpumask_set_cpu(j, covered);
6354 cpumask_set_cpu(j, sched_group_cpus(sg));
6365 #define SD_NODES_PER_DOMAIN 16
6370 * find_next_best_node - find the next node to include in a sched_domain
6371 * @node: node whose sched_domain we're building
6372 * @used_nodes: nodes already in the sched_domain
6374 * Find the next node to include in a given scheduling domain. Simply
6375 * finds the closest node not already in the @used_nodes map.
6377 * Should use nodemask_t.
6379 static int find_next_best_node(int node, nodemask_t *used_nodes)
6381 int i, n, val, min_val, best_node = 0;
6385 for (i = 0; i < nr_node_ids; i++) {
6386 /* Start at @node */
6387 n = (node + i) % nr_node_ids;
6389 if (!nr_cpus_node(n))
6392 /* Skip already used nodes */
6393 if (node_isset(n, *used_nodes))
6396 /* Simple min distance search */
6397 val = node_distance(node, n);
6399 if (val < min_val) {
6405 node_set(best_node, *used_nodes);
6410 * sched_domain_node_span - get a cpumask for a node's sched_domain
6411 * @node: node whose cpumask we're constructing
6412 * @span: resulting cpumask
6414 * Given a node, construct a good cpumask for its sched_domain to span. It
6415 * should be one that prevents unnecessary balancing, but also spreads tasks
6418 static void sched_domain_node_span(int node, struct cpumask *span)
6420 nodemask_t used_nodes;
6423 cpumask_clear(span);
6424 nodes_clear(used_nodes);
6426 cpumask_or(span, span, cpumask_of_node(node));
6427 node_set(node, used_nodes);
6429 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6430 int next_node = find_next_best_node(node, &used_nodes);
6432 cpumask_or(span, span, cpumask_of_node(next_node));
6435 #endif /* CONFIG_NUMA */
6437 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6440 * The cpus mask in sched_group and sched_domain hangs off the end.
6442 * ( See the the comments in include/linux/sched.h:struct sched_group
6443 * and struct sched_domain. )
6445 struct static_sched_group {
6446 struct sched_group sg;
6447 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
6450 struct static_sched_domain {
6451 struct sched_domain sd;
6452 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
6458 cpumask_var_t domainspan;
6459 cpumask_var_t covered;
6460 cpumask_var_t notcovered;
6462 cpumask_var_t nodemask;
6463 cpumask_var_t this_sibling_map;
6464 cpumask_var_t this_core_map;
6465 cpumask_var_t this_book_map;
6466 cpumask_var_t send_covered;
6467 cpumask_var_t tmpmask;
6468 struct sched_group **sched_group_nodes;
6469 struct root_domain *rd;
6473 sa_sched_groups = 0,
6479 sa_this_sibling_map,
6481 sa_sched_group_nodes,
6491 * SMT sched-domains:
6493 #ifdef CONFIG_SCHED_SMT
6494 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
6495 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
6498 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
6499 struct sched_group **sg, struct cpumask *unused)
6502 *sg = &per_cpu(sched_groups, cpu).sg;
6505 #endif /* CONFIG_SCHED_SMT */
6508 * multi-core sched-domains:
6510 #ifdef CONFIG_SCHED_MC
6511 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
6512 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
6515 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6516 struct sched_group **sg, struct cpumask *mask)
6519 #ifdef CONFIG_SCHED_SMT
6520 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6521 group = cpumask_first(mask);
6526 *sg = &per_cpu(sched_group_core, group).sg;
6529 #endif /* CONFIG_SCHED_MC */
6532 * book sched-domains:
6534 #ifdef CONFIG_SCHED_BOOK
6535 static DEFINE_PER_CPU(struct static_sched_domain, book_domains);
6536 static DEFINE_PER_CPU(struct static_sched_group, sched_group_book);
6539 cpu_to_book_group(int cpu, const struct cpumask *cpu_map,
6540 struct sched_group **sg, struct cpumask *mask)
6543 #ifdef CONFIG_SCHED_MC
6544 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6545 group = cpumask_first(mask);
6546 #elif defined(CONFIG_SCHED_SMT)
6547 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6548 group = cpumask_first(mask);
6551 *sg = &per_cpu(sched_group_book, group).sg;
6554 #endif /* CONFIG_SCHED_BOOK */
6556 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
6557 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
6560 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
6561 struct sched_group **sg, struct cpumask *mask)
6564 #ifdef CONFIG_SCHED_BOOK
6565 cpumask_and(mask, cpu_book_mask(cpu), cpu_map);
6566 group = cpumask_first(mask);
6567 #elif defined(CONFIG_SCHED_MC)
6568 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6569 group = cpumask_first(mask);
6570 #elif defined(CONFIG_SCHED_SMT)
6571 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6572 group = cpumask_first(mask);
6577 *sg = &per_cpu(sched_group_phys, group).sg;
6583 * The init_sched_build_groups can't handle what we want to do with node
6584 * groups, so roll our own. Now each node has its own list of groups which
6585 * gets dynamically allocated.
6587 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
6588 static struct sched_group ***sched_group_nodes_bycpu;
6590 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
6591 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
6593 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
6594 struct sched_group **sg,
6595 struct cpumask *nodemask)
6599 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
6600 group = cpumask_first(nodemask);
6603 *sg = &per_cpu(sched_group_allnodes, group).sg;
6607 static void init_numa_sched_groups_power(struct sched_group *group_head)
6609 struct sched_group *sg = group_head;
6615 for_each_cpu(j, sched_group_cpus(sg)) {
6616 struct sched_domain *sd;
6618 sd = &per_cpu(phys_domains, j).sd;
6619 if (j != group_first_cpu(sd->groups)) {
6621 * Only add "power" once for each
6627 sg->cpu_power += sd->groups->cpu_power;
6630 } while (sg != group_head);
6633 static int build_numa_sched_groups(struct s_data *d,
6634 const struct cpumask *cpu_map, int num)
6636 struct sched_domain *sd;
6637 struct sched_group *sg, *prev;
6640 cpumask_clear(d->covered);
6641 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
6642 if (cpumask_empty(d->nodemask)) {
6643 d->sched_group_nodes[num] = NULL;
6647 sched_domain_node_span(num, d->domainspan);
6648 cpumask_and(d->domainspan, d->domainspan, cpu_map);
6650 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6653 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
6657 d->sched_group_nodes[num] = sg;
6659 for_each_cpu(j, d->nodemask) {
6660 sd = &per_cpu(node_domains, j).sd;
6665 cpumask_copy(sched_group_cpus(sg), d->nodemask);
6667 cpumask_or(d->covered, d->covered, d->nodemask);
6670 for (j = 0; j < nr_node_ids; j++) {
6671 n = (num + j) % nr_node_ids;
6672 cpumask_complement(d->notcovered, d->covered);
6673 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
6674 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
6675 if (cpumask_empty(d->tmpmask))
6677 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
6678 if (cpumask_empty(d->tmpmask))
6680 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6684 "Can not alloc domain group for node %d\n", j);
6688 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
6689 sg->next = prev->next;
6690 cpumask_or(d->covered, d->covered, d->tmpmask);
6697 #endif /* CONFIG_NUMA */
6700 /* Free memory allocated for various sched_group structures */
6701 static void free_sched_groups(const struct cpumask *cpu_map,
6702 struct cpumask *nodemask)
6706 for_each_cpu(cpu, cpu_map) {
6707 struct sched_group **sched_group_nodes
6708 = sched_group_nodes_bycpu[cpu];
6710 if (!sched_group_nodes)
6713 for (i = 0; i < nr_node_ids; i++) {
6714 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6716 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
6717 if (cpumask_empty(nodemask))
6727 if (oldsg != sched_group_nodes[i])
6730 kfree(sched_group_nodes);
6731 sched_group_nodes_bycpu[cpu] = NULL;
6734 #else /* !CONFIG_NUMA */
6735 static void free_sched_groups(const struct cpumask *cpu_map,
6736 struct cpumask *nodemask)
6739 #endif /* CONFIG_NUMA */
6742 * Initialize sched groups cpu_power.
6744 * cpu_power indicates the capacity of sched group, which is used while
6745 * distributing the load between different sched groups in a sched domain.
6746 * Typically cpu_power for all the groups in a sched domain will be same unless
6747 * there are asymmetries in the topology. If there are asymmetries, group
6748 * having more cpu_power will pickup more load compared to the group having
6751 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6753 struct sched_domain *child;
6754 struct sched_group *group;
6758 WARN_ON(!sd || !sd->groups);
6760 if (cpu != group_first_cpu(sd->groups))
6763 sd->groups->group_weight = cpumask_weight(sched_group_cpus(sd->groups));
6767 sd->groups->cpu_power = 0;
6770 power = SCHED_LOAD_SCALE;
6771 weight = cpumask_weight(sched_domain_span(sd));
6773 * SMT siblings share the power of a single core.
6774 * Usually multiple threads get a better yield out of
6775 * that one core than a single thread would have,
6776 * reflect that in sd->smt_gain.
6778 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
6779 power *= sd->smt_gain;
6781 power >>= SCHED_LOAD_SHIFT;
6783 sd->groups->cpu_power += power;
6788 * Add cpu_power of each child group to this groups cpu_power.
6790 group = child->groups;
6792 sd->groups->cpu_power += group->cpu_power;
6793 group = group->next;
6794 } while (group != child->groups);
6798 * Initializers for schedule domains
6799 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6802 #ifdef CONFIG_SCHED_DEBUG
6803 # define SD_INIT_NAME(sd, type) sd->name = #type
6805 # define SD_INIT_NAME(sd, type) do { } while (0)
6808 #define SD_INIT(sd, type) sd_init_##type(sd)
6810 #define SD_INIT_FUNC(type) \
6811 static noinline void sd_init_##type(struct sched_domain *sd) \
6813 memset(sd, 0, sizeof(*sd)); \
6814 *sd = SD_##type##_INIT; \
6815 sd->level = SD_LV_##type; \
6816 SD_INIT_NAME(sd, type); \
6821 SD_INIT_FUNC(ALLNODES)
6824 #ifdef CONFIG_SCHED_SMT
6825 SD_INIT_FUNC(SIBLING)
6827 #ifdef CONFIG_SCHED_MC
6830 #ifdef CONFIG_SCHED_BOOK
6834 static int default_relax_domain_level = -1;
6836 static int __init setup_relax_domain_level(char *str)
6840 val = simple_strtoul(str, NULL, 0);
6841 if (val < SD_LV_MAX)
6842 default_relax_domain_level = val;
6846 __setup("relax_domain_level=", setup_relax_domain_level);
6848 static void set_domain_attribute(struct sched_domain *sd,
6849 struct sched_domain_attr *attr)
6853 if (!attr || attr->relax_domain_level < 0) {
6854 if (default_relax_domain_level < 0)
6857 request = default_relax_domain_level;
6859 request = attr->relax_domain_level;
6860 if (request < sd->level) {
6861 /* turn off idle balance on this domain */
6862 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6864 /* turn on idle balance on this domain */
6865 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6869 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6870 const struct cpumask *cpu_map)
6873 case sa_sched_groups:
6874 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
6875 d->sched_group_nodes = NULL;
6877 free_rootdomain(d->rd); /* fall through */
6879 free_cpumask_var(d->tmpmask); /* fall through */
6880 case sa_send_covered:
6881 free_cpumask_var(d->send_covered); /* fall through */
6882 case sa_this_book_map:
6883 free_cpumask_var(d->this_book_map); /* fall through */
6884 case sa_this_core_map:
6885 free_cpumask_var(d->this_core_map); /* fall through */
6886 case sa_this_sibling_map:
6887 free_cpumask_var(d->this_sibling_map); /* fall through */
6889 free_cpumask_var(d->nodemask); /* fall through */
6890 case sa_sched_group_nodes:
6892 kfree(d->sched_group_nodes); /* fall through */
6894 free_cpumask_var(d->notcovered); /* fall through */
6896 free_cpumask_var(d->covered); /* fall through */
6898 free_cpumask_var(d->domainspan); /* fall through */
6905 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6906 const struct cpumask *cpu_map)
6909 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
6911 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
6912 return sa_domainspan;
6913 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
6915 /* Allocate the per-node list of sched groups */
6916 d->sched_group_nodes = kcalloc(nr_node_ids,
6917 sizeof(struct sched_group *), GFP_KERNEL);
6918 if (!d->sched_group_nodes) {
6919 printk(KERN_WARNING "Can not alloc sched group node list\n");
6920 return sa_notcovered;
6922 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
6924 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
6925 return sa_sched_group_nodes;
6926 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
6928 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
6929 return sa_this_sibling_map;
6930 if (!alloc_cpumask_var(&d->this_book_map, GFP_KERNEL))
6931 return sa_this_core_map;
6932 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
6933 return sa_this_book_map;
6934 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
6935 return sa_send_covered;
6936 d->rd = alloc_rootdomain();
6938 printk(KERN_WARNING "Cannot alloc root domain\n");
6941 return sa_rootdomain;
6944 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
6945 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
6947 struct sched_domain *sd = NULL;
6949 struct sched_domain *parent;
6952 if (cpumask_weight(cpu_map) >
6953 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
6954 sd = &per_cpu(allnodes_domains, i).sd;
6955 SD_INIT(sd, ALLNODES);
6956 set_domain_attribute(sd, attr);
6957 cpumask_copy(sched_domain_span(sd), cpu_map);
6958 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
6963 sd = &per_cpu(node_domains, i).sd;
6965 set_domain_attribute(sd, attr);
6966 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
6967 sd->parent = parent;
6970 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
6975 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
6976 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6977 struct sched_domain *parent, int i)
6979 struct sched_domain *sd;
6980 sd = &per_cpu(phys_domains, i).sd;
6982 set_domain_attribute(sd, attr);
6983 cpumask_copy(sched_domain_span(sd), d->nodemask);
6984 sd->parent = parent;
6987 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
6991 static struct sched_domain *__build_book_sched_domain(struct s_data *d,
6992 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6993 struct sched_domain *parent, int i)
6995 struct sched_domain *sd = parent;
6996 #ifdef CONFIG_SCHED_BOOK
6997 sd = &per_cpu(book_domains, i).sd;
6999 set_domain_attribute(sd, attr);
7000 cpumask_and(sched_domain_span(sd), cpu_map, cpu_book_mask(i));
7001 sd->parent = parent;
7003 cpu_to_book_group(i, cpu_map, &sd->groups, d->tmpmask);
7008 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
7009 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7010 struct sched_domain *parent, int i)
7012 struct sched_domain *sd = parent;
7013 #ifdef CONFIG_SCHED_MC
7014 sd = &per_cpu(core_domains, i).sd;
7016 set_domain_attribute(sd, attr);
7017 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
7018 sd->parent = parent;
7020 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
7025 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
7026 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7027 struct sched_domain *parent, int i)
7029 struct sched_domain *sd = parent;
7030 #ifdef CONFIG_SCHED_SMT
7031 sd = &per_cpu(cpu_domains, i).sd;
7032 SD_INIT(sd, SIBLING);
7033 set_domain_attribute(sd, attr);
7034 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
7035 sd->parent = parent;
7037 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
7042 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
7043 const struct cpumask *cpu_map, int cpu)
7046 #ifdef CONFIG_SCHED_SMT
7047 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
7048 cpumask_and(d->this_sibling_map, cpu_map,
7049 topology_thread_cpumask(cpu));
7050 if (cpu == cpumask_first(d->this_sibling_map))
7051 init_sched_build_groups(d->this_sibling_map, cpu_map,
7053 d->send_covered, d->tmpmask);
7056 #ifdef CONFIG_SCHED_MC
7057 case SD_LV_MC: /* set up multi-core groups */
7058 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
7059 if (cpu == cpumask_first(d->this_core_map))
7060 init_sched_build_groups(d->this_core_map, cpu_map,
7062 d->send_covered, d->tmpmask);
7065 #ifdef CONFIG_SCHED_BOOK
7066 case SD_LV_BOOK: /* set up book groups */
7067 cpumask_and(d->this_book_map, cpu_map, cpu_book_mask(cpu));
7068 if (cpu == cpumask_first(d->this_book_map))
7069 init_sched_build_groups(d->this_book_map, cpu_map,
7071 d->send_covered, d->tmpmask);
7074 case SD_LV_CPU: /* set up physical groups */
7075 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
7076 if (!cpumask_empty(d->nodemask))
7077 init_sched_build_groups(d->nodemask, cpu_map,
7079 d->send_covered, d->tmpmask);
7082 case SD_LV_ALLNODES:
7083 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
7084 d->send_covered, d->tmpmask);
7093 * Build sched domains for a given set of cpus and attach the sched domains
7094 * to the individual cpus
7096 static int __build_sched_domains(const struct cpumask *cpu_map,
7097 struct sched_domain_attr *attr)
7099 enum s_alloc alloc_state = sa_none;
7101 struct sched_domain *sd;
7107 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7108 if (alloc_state != sa_rootdomain)
7110 alloc_state = sa_sched_groups;
7113 * Set up domains for cpus specified by the cpu_map.
7115 for_each_cpu(i, cpu_map) {
7116 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
7119 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
7120 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
7121 sd = __build_book_sched_domain(&d, cpu_map, attr, sd, i);
7122 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
7123 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
7126 for_each_cpu(i, cpu_map) {
7127 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
7128 build_sched_groups(&d, SD_LV_BOOK, cpu_map, i);
7129 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
7132 /* Set up physical groups */
7133 for (i = 0; i < nr_node_ids; i++)
7134 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
7137 /* Set up node groups */
7139 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
7141 for (i = 0; i < nr_node_ids; i++)
7142 if (build_numa_sched_groups(&d, cpu_map, i))
7146 /* Calculate CPU power for physical packages and nodes */
7147 #ifdef CONFIG_SCHED_SMT
7148 for_each_cpu(i, cpu_map) {
7149 sd = &per_cpu(cpu_domains, i).sd;
7150 init_sched_groups_power(i, sd);
7153 #ifdef CONFIG_SCHED_MC
7154 for_each_cpu(i, cpu_map) {
7155 sd = &per_cpu(core_domains, i).sd;
7156 init_sched_groups_power(i, sd);
7159 #ifdef CONFIG_SCHED_BOOK
7160 for_each_cpu(i, cpu_map) {
7161 sd = &per_cpu(book_domains, i).sd;
7162 init_sched_groups_power(i, sd);
7166 for_each_cpu(i, cpu_map) {
7167 sd = &per_cpu(phys_domains, i).sd;
7168 init_sched_groups_power(i, sd);
7172 for (i = 0; i < nr_node_ids; i++)
7173 init_numa_sched_groups_power(d.sched_group_nodes[i]);
7175 if (d.sd_allnodes) {
7176 struct sched_group *sg;
7178 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7180 init_numa_sched_groups_power(sg);
7184 /* Attach the domains */
7185 for_each_cpu(i, cpu_map) {
7186 #ifdef CONFIG_SCHED_SMT
7187 sd = &per_cpu(cpu_domains, i).sd;
7188 #elif defined(CONFIG_SCHED_MC)
7189 sd = &per_cpu(core_domains, i).sd;
7190 #elif defined(CONFIG_SCHED_BOOK)
7191 sd = &per_cpu(book_domains, i).sd;
7193 sd = &per_cpu(phys_domains, i).sd;
7195 cpu_attach_domain(sd, d.rd, i);
7198 d.sched_group_nodes = NULL; /* don't free this we still need it */
7199 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
7203 __free_domain_allocs(&d, alloc_state, cpu_map);
7207 static int build_sched_domains(const struct cpumask *cpu_map)
7209 return __build_sched_domains(cpu_map, NULL);
7212 static cpumask_var_t *doms_cur; /* current sched domains */
7213 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7214 static struct sched_domain_attr *dattr_cur;
7215 /* attribues of custom domains in 'doms_cur' */
7218 * Special case: If a kmalloc of a doms_cur partition (array of
7219 * cpumask) fails, then fallback to a single sched domain,
7220 * as determined by the single cpumask fallback_doms.
7222 static cpumask_var_t fallback_doms;
7225 * arch_update_cpu_topology lets virtualized architectures update the
7226 * cpu core maps. It is supposed to return 1 if the topology changed
7227 * or 0 if it stayed the same.
7229 int __attribute__((weak)) arch_update_cpu_topology(void)
7234 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7237 cpumask_var_t *doms;
7239 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7242 for (i = 0; i < ndoms; i++) {
7243 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7244 free_sched_domains(doms, i);
7251 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7254 for (i = 0; i < ndoms; i++)
7255 free_cpumask_var(doms[i]);
7260 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7261 * For now this just excludes isolated cpus, but could be used to
7262 * exclude other special cases in the future.
7264 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7268 arch_update_cpu_topology();
7270 doms_cur = alloc_sched_domains(ndoms_cur);
7272 doms_cur = &fallback_doms;
7273 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7275 err = build_sched_domains(doms_cur[0]);
7276 register_sched_domain_sysctl();
7281 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7282 struct cpumask *tmpmask)
7284 free_sched_groups(cpu_map, tmpmask);
7288 * Detach sched domains from a group of cpus specified in cpu_map
7289 * These cpus will now be attached to the NULL domain
7291 static void detach_destroy_domains(const struct cpumask *cpu_map)
7293 /* Save because hotplug lock held. */
7294 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7297 for_each_cpu(i, cpu_map)
7298 cpu_attach_domain(NULL, &def_root_domain, i);
7299 synchronize_sched();
7300 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7303 /* handle null as "default" */
7304 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7305 struct sched_domain_attr *new, int idx_new)
7307 struct sched_domain_attr tmp;
7314 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7315 new ? (new + idx_new) : &tmp,
7316 sizeof(struct sched_domain_attr));
7320 * Partition sched domains as specified by the 'ndoms_new'
7321 * cpumasks in the array doms_new[] of cpumasks. This compares
7322 * doms_new[] to the current sched domain partitioning, doms_cur[].
7323 * It destroys each deleted domain and builds each new domain.
7325 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7326 * The masks don't intersect (don't overlap.) We should setup one
7327 * sched domain for each mask. CPUs not in any of the cpumasks will
7328 * not be load balanced. If the same cpumask appears both in the
7329 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7332 * The passed in 'doms_new' should be allocated using
7333 * alloc_sched_domains. This routine takes ownership of it and will
7334 * free_sched_domains it when done with it. If the caller failed the
7335 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7336 * and partition_sched_domains() will fallback to the single partition
7337 * 'fallback_doms', it also forces the domains to be rebuilt.
7339 * If doms_new == NULL it will be replaced with cpu_online_mask.
7340 * ndoms_new == 0 is a special case for destroying existing domains,
7341 * and it will not create the default domain.
7343 * Call with hotplug lock held
7345 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7346 struct sched_domain_attr *dattr_new)
7351 mutex_lock(&sched_domains_mutex);
7353 /* always unregister in case we don't destroy any domains */
7354 unregister_sched_domain_sysctl();
7356 /* Let architecture update cpu core mappings. */
7357 new_topology = arch_update_cpu_topology();
7359 n = doms_new ? ndoms_new : 0;
7361 /* Destroy deleted domains */
7362 for (i = 0; i < ndoms_cur; i++) {
7363 for (j = 0; j < n && !new_topology; j++) {
7364 if (cpumask_equal(doms_cur[i], doms_new[j])
7365 && dattrs_equal(dattr_cur, i, dattr_new, j))
7368 /* no match - a current sched domain not in new doms_new[] */
7369 detach_destroy_domains(doms_cur[i]);
7374 if (doms_new == NULL) {
7376 doms_new = &fallback_doms;
7377 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7378 WARN_ON_ONCE(dattr_new);
7381 /* Build new domains */
7382 for (i = 0; i < ndoms_new; i++) {
7383 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7384 if (cpumask_equal(doms_new[i], doms_cur[j])
7385 && dattrs_equal(dattr_new, i, dattr_cur, j))
7388 /* no match - add a new doms_new */
7389 __build_sched_domains(doms_new[i],
7390 dattr_new ? dattr_new + i : NULL);
7395 /* Remember the new sched domains */
7396 if (doms_cur != &fallback_doms)
7397 free_sched_domains(doms_cur, ndoms_cur);
7398 kfree(dattr_cur); /* kfree(NULL) is safe */
7399 doms_cur = doms_new;
7400 dattr_cur = dattr_new;
7401 ndoms_cur = ndoms_new;
7403 register_sched_domain_sysctl();
7405 mutex_unlock(&sched_domains_mutex);
7408 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7409 static void arch_reinit_sched_domains(void)
7413 /* Destroy domains first to force the rebuild */
7414 partition_sched_domains(0, NULL, NULL);
7416 rebuild_sched_domains();
7420 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7422 unsigned int level = 0;
7424 if (sscanf(buf, "%u", &level) != 1)
7428 * level is always be positive so don't check for
7429 * level < POWERSAVINGS_BALANCE_NONE which is 0
7430 * What happens on 0 or 1 byte write,
7431 * need to check for count as well?
7434 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7438 sched_smt_power_savings = level;
7440 sched_mc_power_savings = level;
7442 arch_reinit_sched_domains();
7447 #ifdef CONFIG_SCHED_MC
7448 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7449 struct sysdev_class_attribute *attr,
7452 return sprintf(page, "%u\n", sched_mc_power_savings);
7454 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7455 struct sysdev_class_attribute *attr,
7456 const char *buf, size_t count)
7458 return sched_power_savings_store(buf, count, 0);
7460 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7461 sched_mc_power_savings_show,
7462 sched_mc_power_savings_store);
7465 #ifdef CONFIG_SCHED_SMT
7466 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7467 struct sysdev_class_attribute *attr,
7470 return sprintf(page, "%u\n", sched_smt_power_savings);
7472 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7473 struct sysdev_class_attribute *attr,
7474 const char *buf, size_t count)
7476 return sched_power_savings_store(buf, count, 1);
7478 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7479 sched_smt_power_savings_show,
7480 sched_smt_power_savings_store);
7483 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7487 #ifdef CONFIG_SCHED_SMT
7489 err = sysfs_create_file(&cls->kset.kobj,
7490 &attr_sched_smt_power_savings.attr);
7492 #ifdef CONFIG_SCHED_MC
7493 if (!err && mc_capable())
7494 err = sysfs_create_file(&cls->kset.kobj,
7495 &attr_sched_mc_power_savings.attr);
7499 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7502 * Update cpusets according to cpu_active mask. If cpusets are
7503 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7504 * around partition_sched_domains().
7506 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7509 switch (action & ~CPU_TASKS_FROZEN) {
7511 case CPU_DOWN_FAILED:
7512 cpuset_update_active_cpus();
7519 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7522 switch (action & ~CPU_TASKS_FROZEN) {
7523 case CPU_DOWN_PREPARE:
7524 cpuset_update_active_cpus();
7531 static int update_runtime(struct notifier_block *nfb,
7532 unsigned long action, void *hcpu)
7534 int cpu = (int)(long)hcpu;
7537 case CPU_DOWN_PREPARE:
7538 case CPU_DOWN_PREPARE_FROZEN:
7539 disable_runtime(cpu_rq(cpu));
7542 case CPU_DOWN_FAILED:
7543 case CPU_DOWN_FAILED_FROZEN:
7545 case CPU_ONLINE_FROZEN:
7546 enable_runtime(cpu_rq(cpu));
7554 void __init sched_init_smp(void)
7556 cpumask_var_t non_isolated_cpus;
7558 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7559 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7561 #if defined(CONFIG_NUMA)
7562 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7564 BUG_ON(sched_group_nodes_bycpu == NULL);
7567 mutex_lock(&sched_domains_mutex);
7568 arch_init_sched_domains(cpu_active_mask);
7569 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7570 if (cpumask_empty(non_isolated_cpus))
7571 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7572 mutex_unlock(&sched_domains_mutex);
7575 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7576 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7578 /* RT runtime code needs to handle some hotplug events */
7579 hotcpu_notifier(update_runtime, 0);
7583 /* Move init over to a non-isolated CPU */
7584 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7586 sched_init_granularity();
7587 free_cpumask_var(non_isolated_cpus);
7589 init_sched_rt_class();
7592 void __init sched_init_smp(void)
7594 sched_init_granularity();
7596 #endif /* CONFIG_SMP */
7598 const_debug unsigned int sysctl_timer_migration = 1;
7600 int in_sched_functions(unsigned long addr)
7602 return in_lock_functions(addr) ||
7603 (addr >= (unsigned long)__sched_text_start
7604 && addr < (unsigned long)__sched_text_end);
7607 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7609 cfs_rq->tasks_timeline = RB_ROOT;
7610 INIT_LIST_HEAD(&cfs_rq->tasks);
7611 #ifdef CONFIG_FAIR_GROUP_SCHED
7614 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7617 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7619 struct rt_prio_array *array;
7622 array = &rt_rq->active;
7623 for (i = 0; i < MAX_RT_PRIO; i++) {
7624 INIT_LIST_HEAD(array->queue + i);
7625 __clear_bit(i, array->bitmap);
7627 /* delimiter for bitsearch: */
7628 __set_bit(MAX_RT_PRIO, array->bitmap);
7630 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7631 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7633 rt_rq->highest_prio.next = MAX_RT_PRIO;
7637 rt_rq->rt_nr_migratory = 0;
7638 rt_rq->overloaded = 0;
7639 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
7643 rt_rq->rt_throttled = 0;
7644 rt_rq->rt_runtime = 0;
7645 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7647 #ifdef CONFIG_RT_GROUP_SCHED
7648 rt_rq->rt_nr_boosted = 0;
7653 #ifdef CONFIG_FAIR_GROUP_SCHED
7654 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7655 struct sched_entity *se, int cpu,
7656 struct sched_entity *parent)
7658 struct rq *rq = cpu_rq(cpu);
7659 tg->cfs_rq[cpu] = cfs_rq;
7660 init_cfs_rq(cfs_rq, rq);
7664 /* se could be NULL for init_task_group */
7669 se->cfs_rq = &rq->cfs;
7671 se->cfs_rq = parent->my_q;
7674 update_load_set(&se->load, 0);
7675 se->parent = parent;
7679 #ifdef CONFIG_RT_GROUP_SCHED
7680 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7681 struct sched_rt_entity *rt_se, int cpu,
7682 struct sched_rt_entity *parent)
7684 struct rq *rq = cpu_rq(cpu);
7686 tg->rt_rq[cpu] = rt_rq;
7687 init_rt_rq(rt_rq, rq);
7689 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7691 tg->rt_se[cpu] = rt_se;
7696 rt_se->rt_rq = &rq->rt;
7698 rt_se->rt_rq = parent->my_q;
7700 rt_se->my_q = rt_rq;
7701 rt_se->parent = parent;
7702 INIT_LIST_HEAD(&rt_se->run_list);
7706 void __init sched_init(void)
7709 unsigned long alloc_size = 0, ptr;
7711 #ifdef CONFIG_FAIR_GROUP_SCHED
7712 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7714 #ifdef CONFIG_RT_GROUP_SCHED
7715 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7717 #ifdef CONFIG_CPUMASK_OFFSTACK
7718 alloc_size += num_possible_cpus() * cpumask_size();
7721 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7723 #ifdef CONFIG_FAIR_GROUP_SCHED
7724 init_task_group.se = (struct sched_entity **)ptr;
7725 ptr += nr_cpu_ids * sizeof(void **);
7727 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7728 ptr += nr_cpu_ids * sizeof(void **);
7730 #endif /* CONFIG_FAIR_GROUP_SCHED */
7731 #ifdef CONFIG_RT_GROUP_SCHED
7732 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7733 ptr += nr_cpu_ids * sizeof(void **);
7735 init_task_group.rt_rq = (struct rt_rq **)ptr;
7736 ptr += nr_cpu_ids * sizeof(void **);
7738 #endif /* CONFIG_RT_GROUP_SCHED */
7739 #ifdef CONFIG_CPUMASK_OFFSTACK
7740 for_each_possible_cpu(i) {
7741 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
7742 ptr += cpumask_size();
7744 #endif /* CONFIG_CPUMASK_OFFSTACK */
7748 init_defrootdomain();
7751 init_rt_bandwidth(&def_rt_bandwidth,
7752 global_rt_period(), global_rt_runtime());
7754 #ifdef CONFIG_RT_GROUP_SCHED
7755 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7756 global_rt_period(), global_rt_runtime());
7757 #endif /* CONFIG_RT_GROUP_SCHED */
7759 #ifdef CONFIG_CGROUP_SCHED
7760 list_add(&init_task_group.list, &task_groups);
7761 INIT_LIST_HEAD(&init_task_group.children);
7762 autogroup_init(&init_task);
7763 #endif /* CONFIG_CGROUP_SCHED */
7765 for_each_possible_cpu(i) {
7769 raw_spin_lock_init(&rq->lock);
7771 rq->calc_load_active = 0;
7772 rq->calc_load_update = jiffies + LOAD_FREQ;
7773 init_cfs_rq(&rq->cfs, rq);
7774 init_rt_rq(&rq->rt, rq);
7775 #ifdef CONFIG_FAIR_GROUP_SCHED
7776 init_task_group.shares = init_task_group_load;
7777 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7778 #ifdef CONFIG_CGROUP_SCHED
7780 * How much cpu bandwidth does init_task_group get?
7782 * In case of task-groups formed thr' the cgroup filesystem, it
7783 * gets 100% of the cpu resources in the system. This overall
7784 * system cpu resource is divided among the tasks of
7785 * init_task_group and its child task-groups in a fair manner,
7786 * based on each entity's (task or task-group's) weight
7787 * (se->load.weight).
7789 * In other words, if init_task_group has 10 tasks of weight
7790 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7791 * then A0's share of the cpu resource is:
7793 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7795 * We achieve this by letting init_task_group's tasks sit
7796 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7798 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, NULL);
7800 #endif /* CONFIG_FAIR_GROUP_SCHED */
7802 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7803 #ifdef CONFIG_RT_GROUP_SCHED
7804 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7805 #ifdef CONFIG_CGROUP_SCHED
7806 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, NULL);
7810 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7811 rq->cpu_load[j] = 0;
7813 rq->last_load_update_tick = jiffies;
7818 rq->cpu_power = SCHED_LOAD_SCALE;
7819 rq->post_schedule = 0;
7820 rq->active_balance = 0;
7821 rq->next_balance = jiffies;
7826 rq->avg_idle = 2*sysctl_sched_migration_cost;
7827 rq_attach_root(rq, &def_root_domain);
7829 rq->nohz_balance_kick = 0;
7830 init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb, i));
7834 atomic_set(&rq->nr_iowait, 0);
7837 set_load_weight(&init_task);
7839 #ifdef CONFIG_PREEMPT_NOTIFIERS
7840 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7844 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7847 #ifdef CONFIG_RT_MUTEXES
7848 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
7852 * The boot idle thread does lazy MMU switching as well:
7854 atomic_inc(&init_mm.mm_count);
7855 enter_lazy_tlb(&init_mm, current);
7858 * Make us the idle thread. Technically, schedule() should not be
7859 * called from this thread, however somewhere below it might be,
7860 * but because we are the idle thread, we just pick up running again
7861 * when this runqueue becomes "idle".
7863 init_idle(current, smp_processor_id());
7865 calc_load_update = jiffies + LOAD_FREQ;
7868 * During early bootup we pretend to be a normal task:
7870 current->sched_class = &fair_sched_class;
7872 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
7873 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
7876 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
7877 alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT);
7878 atomic_set(&nohz.load_balancer, nr_cpu_ids);
7879 atomic_set(&nohz.first_pick_cpu, nr_cpu_ids);
7880 atomic_set(&nohz.second_pick_cpu, nr_cpu_ids);
7882 /* May be allocated at isolcpus cmdline parse time */
7883 if (cpu_isolated_map == NULL)
7884 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7889 scheduler_running = 1;
7892 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7893 static inline int preempt_count_equals(int preempt_offset)
7895 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7897 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
7900 void __might_sleep(const char *file, int line, int preempt_offset)
7903 static unsigned long prev_jiffy; /* ratelimiting */
7905 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7906 system_state != SYSTEM_RUNNING || oops_in_progress)
7908 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7910 prev_jiffy = jiffies;
7913 "BUG: sleeping function called from invalid context at %s:%d\n",
7916 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7917 in_atomic(), irqs_disabled(),
7918 current->pid, current->comm);
7920 debug_show_held_locks(current);
7921 if (irqs_disabled())
7922 print_irqtrace_events(current);
7926 EXPORT_SYMBOL(__might_sleep);
7929 #ifdef CONFIG_MAGIC_SYSRQ
7930 static void normalize_task(struct rq *rq, struct task_struct *p)
7934 on_rq = p->se.on_rq;
7936 deactivate_task(rq, p, 0);
7937 __setscheduler(rq, p, SCHED_NORMAL, 0);
7939 activate_task(rq, p, 0);
7940 resched_task(rq->curr);
7944 void normalize_rt_tasks(void)
7946 struct task_struct *g, *p;
7947 unsigned long flags;
7950 read_lock_irqsave(&tasklist_lock, flags);
7951 do_each_thread(g, p) {
7953 * Only normalize user tasks:
7958 p->se.exec_start = 0;
7959 #ifdef CONFIG_SCHEDSTATS
7960 p->se.statistics.wait_start = 0;
7961 p->se.statistics.sleep_start = 0;
7962 p->se.statistics.block_start = 0;
7967 * Renice negative nice level userspace
7970 if (TASK_NICE(p) < 0 && p->mm)
7971 set_user_nice(p, 0);
7975 raw_spin_lock(&p->pi_lock);
7976 rq = __task_rq_lock(p);
7978 normalize_task(rq, p);
7980 __task_rq_unlock(rq);
7981 raw_spin_unlock(&p->pi_lock);
7982 } while_each_thread(g, p);
7984 read_unlock_irqrestore(&tasklist_lock, flags);
7987 #endif /* CONFIG_MAGIC_SYSRQ */
7989 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7991 * These functions are only useful for the IA64 MCA handling, or kdb.
7993 * They can only be called when the whole system has been
7994 * stopped - every CPU needs to be quiescent, and no scheduling
7995 * activity can take place. Using them for anything else would
7996 * be a serious bug, and as a result, they aren't even visible
7997 * under any other configuration.
8001 * curr_task - return the current task for a given cpu.
8002 * @cpu: the processor in question.
8004 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8006 struct task_struct *curr_task(int cpu)
8008 return cpu_curr(cpu);
8011 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8015 * set_curr_task - set the current task for a given cpu.
8016 * @cpu: the processor in question.
8017 * @p: the task pointer to set.
8019 * Description: This function must only be used when non-maskable interrupts
8020 * are serviced on a separate stack. It allows the architecture to switch the
8021 * notion of the current task on a cpu in a non-blocking manner. This function
8022 * must be called with all CPU's synchronized, and interrupts disabled, the
8023 * and caller must save the original value of the current task (see
8024 * curr_task() above) and restore that value before reenabling interrupts and
8025 * re-starting the system.
8027 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8029 void set_curr_task(int cpu, struct task_struct *p)
8036 #ifdef CONFIG_FAIR_GROUP_SCHED
8037 static void free_fair_sched_group(struct task_group *tg)
8041 for_each_possible_cpu(i) {
8043 kfree(tg->cfs_rq[i]);
8053 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8055 struct cfs_rq *cfs_rq;
8056 struct sched_entity *se;
8060 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8063 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8067 tg->shares = NICE_0_LOAD;
8069 for_each_possible_cpu(i) {
8072 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8073 GFP_KERNEL, cpu_to_node(i));
8077 se = kzalloc_node(sizeof(struct sched_entity),
8078 GFP_KERNEL, cpu_to_node(i));
8082 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8093 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8095 struct rq *rq = cpu_rq(cpu);
8096 unsigned long flags;
8099 * Only empty task groups can be destroyed; so we can speculatively
8100 * check on_list without danger of it being re-added.
8102 if (!tg->cfs_rq[cpu]->on_list)
8105 raw_spin_lock_irqsave(&rq->lock, flags);
8106 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8107 raw_spin_unlock_irqrestore(&rq->lock, flags);
8109 #else /* !CONFG_FAIR_GROUP_SCHED */
8110 static inline void free_fair_sched_group(struct task_group *tg)
8115 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8120 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8123 #endif /* CONFIG_FAIR_GROUP_SCHED */
8125 #ifdef CONFIG_RT_GROUP_SCHED
8126 static void free_rt_sched_group(struct task_group *tg)
8130 destroy_rt_bandwidth(&tg->rt_bandwidth);
8132 for_each_possible_cpu(i) {
8134 kfree(tg->rt_rq[i]);
8136 kfree(tg->rt_se[i]);
8144 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8146 struct rt_rq *rt_rq;
8147 struct sched_rt_entity *rt_se;
8151 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8154 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8158 init_rt_bandwidth(&tg->rt_bandwidth,
8159 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8161 for_each_possible_cpu(i) {
8164 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8165 GFP_KERNEL, cpu_to_node(i));
8169 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8170 GFP_KERNEL, cpu_to_node(i));
8174 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
8184 #else /* !CONFIG_RT_GROUP_SCHED */
8185 static inline void free_rt_sched_group(struct task_group *tg)
8190 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8194 #endif /* CONFIG_RT_GROUP_SCHED */
8196 #ifdef CONFIG_CGROUP_SCHED
8197 static void free_sched_group(struct task_group *tg)
8199 free_fair_sched_group(tg);
8200 free_rt_sched_group(tg);
8204 /* allocate runqueue etc for a new task group */
8205 struct task_group *sched_create_group(struct task_group *parent)
8207 struct task_group *tg;
8208 unsigned long flags;
8210 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8212 return ERR_PTR(-ENOMEM);
8214 if (!alloc_fair_sched_group(tg, parent))
8217 if (!alloc_rt_sched_group(tg, parent))
8220 spin_lock_irqsave(&task_group_lock, flags);
8221 list_add_rcu(&tg->list, &task_groups);
8223 WARN_ON(!parent); /* root should already exist */
8225 tg->parent = parent;
8226 INIT_LIST_HEAD(&tg->children);
8227 list_add_rcu(&tg->siblings, &parent->children);
8228 spin_unlock_irqrestore(&task_group_lock, flags);
8233 free_sched_group(tg);
8234 return ERR_PTR(-ENOMEM);
8237 /* rcu callback to free various structures associated with a task group */
8238 static void free_sched_group_rcu(struct rcu_head *rhp)
8240 /* now it should be safe to free those cfs_rqs */
8241 free_sched_group(container_of(rhp, struct task_group, rcu));
8244 /* Destroy runqueue etc associated with a task group */
8245 void sched_destroy_group(struct task_group *tg)
8247 unsigned long flags;
8250 /* end participation in shares distribution */
8251 for_each_possible_cpu(i)
8252 unregister_fair_sched_group(tg, i);
8254 spin_lock_irqsave(&task_group_lock, flags);
8255 list_del_rcu(&tg->list);
8256 list_del_rcu(&tg->siblings);
8257 spin_unlock_irqrestore(&task_group_lock, flags);
8259 /* wait for possible concurrent references to cfs_rqs complete */
8260 call_rcu(&tg->rcu, free_sched_group_rcu);
8263 /* change task's runqueue when it moves between groups.
8264 * The caller of this function should have put the task in its new group
8265 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8266 * reflect its new group.
8268 void sched_move_task(struct task_struct *tsk)
8271 unsigned long flags;
8274 rq = task_rq_lock(tsk, &flags);
8276 running = task_current(rq, tsk);
8277 on_rq = tsk->se.on_rq;
8280 dequeue_task(rq, tsk, 0);
8281 if (unlikely(running))
8282 tsk->sched_class->put_prev_task(rq, tsk);
8284 #ifdef CONFIG_FAIR_GROUP_SCHED
8285 if (tsk->sched_class->task_move_group)
8286 tsk->sched_class->task_move_group(tsk, on_rq);
8289 set_task_rq(tsk, task_cpu(tsk));
8291 if (unlikely(running))
8292 tsk->sched_class->set_curr_task(rq);
8294 enqueue_task(rq, tsk, 0);
8296 task_rq_unlock(rq, &flags);
8298 #endif /* CONFIG_CGROUP_SCHED */
8300 #ifdef CONFIG_FAIR_GROUP_SCHED
8301 static DEFINE_MUTEX(shares_mutex);
8303 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8306 unsigned long flags;
8309 * We can't change the weight of the root cgroup.
8314 if (shares < MIN_SHARES)
8315 shares = MIN_SHARES;
8316 else if (shares > MAX_SHARES)
8317 shares = MAX_SHARES;
8319 mutex_lock(&shares_mutex);
8320 if (tg->shares == shares)
8323 tg->shares = shares;
8324 for_each_possible_cpu(i) {
8325 struct rq *rq = cpu_rq(i);
8326 struct sched_entity *se;
8329 /* Propagate contribution to hierarchy */
8330 raw_spin_lock_irqsave(&rq->lock, flags);
8331 for_each_sched_entity(se)
8332 update_cfs_shares(group_cfs_rq(se), 0);
8333 raw_spin_unlock_irqrestore(&rq->lock, flags);
8337 mutex_unlock(&shares_mutex);
8341 unsigned long sched_group_shares(struct task_group *tg)
8347 #ifdef CONFIG_RT_GROUP_SCHED
8349 * Ensure that the real time constraints are schedulable.
8351 static DEFINE_MUTEX(rt_constraints_mutex);
8353 static unsigned long to_ratio(u64 period, u64 runtime)
8355 if (runtime == RUNTIME_INF)
8358 return div64_u64(runtime << 20, period);
8361 /* Must be called with tasklist_lock held */
8362 static inline int tg_has_rt_tasks(struct task_group *tg)
8364 struct task_struct *g, *p;
8366 do_each_thread(g, p) {
8367 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8369 } while_each_thread(g, p);
8374 struct rt_schedulable_data {
8375 struct task_group *tg;
8380 static int tg_schedulable(struct task_group *tg, void *data)
8382 struct rt_schedulable_data *d = data;
8383 struct task_group *child;
8384 unsigned long total, sum = 0;
8385 u64 period, runtime;
8387 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8388 runtime = tg->rt_bandwidth.rt_runtime;
8391 period = d->rt_period;
8392 runtime = d->rt_runtime;
8396 * Cannot have more runtime than the period.
8398 if (runtime > period && runtime != RUNTIME_INF)
8402 * Ensure we don't starve existing RT tasks.
8404 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8407 total = to_ratio(period, runtime);
8410 * Nobody can have more than the global setting allows.
8412 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8416 * The sum of our children's runtime should not exceed our own.
8418 list_for_each_entry_rcu(child, &tg->children, siblings) {
8419 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8420 runtime = child->rt_bandwidth.rt_runtime;
8422 if (child == d->tg) {
8423 period = d->rt_period;
8424 runtime = d->rt_runtime;
8427 sum += to_ratio(period, runtime);
8436 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8438 struct rt_schedulable_data data = {
8440 .rt_period = period,
8441 .rt_runtime = runtime,
8444 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8447 static int tg_set_bandwidth(struct task_group *tg,
8448 u64 rt_period, u64 rt_runtime)
8452 mutex_lock(&rt_constraints_mutex);
8453 read_lock(&tasklist_lock);
8454 err = __rt_schedulable(tg, rt_period, rt_runtime);
8458 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8459 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8460 tg->rt_bandwidth.rt_runtime = rt_runtime;
8462 for_each_possible_cpu(i) {
8463 struct rt_rq *rt_rq = tg->rt_rq[i];
8465 raw_spin_lock(&rt_rq->rt_runtime_lock);
8466 rt_rq->rt_runtime = rt_runtime;
8467 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8469 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8471 read_unlock(&tasklist_lock);
8472 mutex_unlock(&rt_constraints_mutex);
8477 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8479 u64 rt_runtime, rt_period;
8481 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8482 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8483 if (rt_runtime_us < 0)
8484 rt_runtime = RUNTIME_INF;
8486 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8489 long sched_group_rt_runtime(struct task_group *tg)
8493 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8496 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8497 do_div(rt_runtime_us, NSEC_PER_USEC);
8498 return rt_runtime_us;
8501 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8503 u64 rt_runtime, rt_period;
8505 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8506 rt_runtime = tg->rt_bandwidth.rt_runtime;
8511 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8514 long sched_group_rt_period(struct task_group *tg)
8518 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8519 do_div(rt_period_us, NSEC_PER_USEC);
8520 return rt_period_us;
8523 static int sched_rt_global_constraints(void)
8525 u64 runtime, period;
8528 if (sysctl_sched_rt_period <= 0)
8531 runtime = global_rt_runtime();
8532 period = global_rt_period();
8535 * Sanity check on the sysctl variables.
8537 if (runtime > period && runtime != RUNTIME_INF)
8540 mutex_lock(&rt_constraints_mutex);
8541 read_lock(&tasklist_lock);
8542 ret = __rt_schedulable(NULL, 0, 0);
8543 read_unlock(&tasklist_lock);
8544 mutex_unlock(&rt_constraints_mutex);
8549 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8551 /* Don't accept realtime tasks when there is no way for them to run */
8552 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8558 #else /* !CONFIG_RT_GROUP_SCHED */
8559 static int sched_rt_global_constraints(void)
8561 unsigned long flags;
8564 if (sysctl_sched_rt_period <= 0)
8568 * There's always some RT tasks in the root group
8569 * -- migration, kstopmachine etc..
8571 if (sysctl_sched_rt_runtime == 0)
8574 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8575 for_each_possible_cpu(i) {
8576 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8578 raw_spin_lock(&rt_rq->rt_runtime_lock);
8579 rt_rq->rt_runtime = global_rt_runtime();
8580 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8582 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8586 #endif /* CONFIG_RT_GROUP_SCHED */
8588 int sched_rt_handler(struct ctl_table *table, int write,
8589 void __user *buffer, size_t *lenp,
8593 int old_period, old_runtime;
8594 static DEFINE_MUTEX(mutex);
8597 old_period = sysctl_sched_rt_period;
8598 old_runtime = sysctl_sched_rt_runtime;
8600 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8602 if (!ret && write) {
8603 ret = sched_rt_global_constraints();
8605 sysctl_sched_rt_period = old_period;
8606 sysctl_sched_rt_runtime = old_runtime;
8608 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8609 def_rt_bandwidth.rt_period =
8610 ns_to_ktime(global_rt_period());
8613 mutex_unlock(&mutex);
8618 #ifdef CONFIG_CGROUP_SCHED
8620 /* return corresponding task_group object of a cgroup */
8621 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8623 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8624 struct task_group, css);
8627 static struct cgroup_subsys_state *
8628 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8630 struct task_group *tg, *parent;
8632 if (!cgrp->parent) {
8633 /* This is early initialization for the top cgroup */
8634 return &init_task_group.css;
8637 parent = cgroup_tg(cgrp->parent);
8638 tg = sched_create_group(parent);
8640 return ERR_PTR(-ENOMEM);
8646 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8648 struct task_group *tg = cgroup_tg(cgrp);
8650 sched_destroy_group(tg);
8654 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8656 #ifdef CONFIG_RT_GROUP_SCHED
8657 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
8660 /* We don't support RT-tasks being in separate groups */
8661 if (tsk->sched_class != &fair_sched_class)
8668 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8669 struct task_struct *tsk, bool threadgroup)
8671 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
8675 struct task_struct *c;
8677 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8678 retval = cpu_cgroup_can_attach_task(cgrp, c);
8690 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8691 struct cgroup *old_cont, struct task_struct *tsk,
8694 sched_move_task(tsk);
8696 struct task_struct *c;
8698 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8705 #ifdef CONFIG_FAIR_GROUP_SCHED
8706 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8709 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8712 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8714 struct task_group *tg = cgroup_tg(cgrp);
8716 return (u64) tg->shares;
8718 #endif /* CONFIG_FAIR_GROUP_SCHED */
8720 #ifdef CONFIG_RT_GROUP_SCHED
8721 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8724 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8727 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8729 return sched_group_rt_runtime(cgroup_tg(cgrp));
8732 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8735 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8738 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8740 return sched_group_rt_period(cgroup_tg(cgrp));
8742 #endif /* CONFIG_RT_GROUP_SCHED */
8744 static struct cftype cpu_files[] = {
8745 #ifdef CONFIG_FAIR_GROUP_SCHED
8748 .read_u64 = cpu_shares_read_u64,
8749 .write_u64 = cpu_shares_write_u64,
8752 #ifdef CONFIG_RT_GROUP_SCHED
8754 .name = "rt_runtime_us",
8755 .read_s64 = cpu_rt_runtime_read,
8756 .write_s64 = cpu_rt_runtime_write,
8759 .name = "rt_period_us",
8760 .read_u64 = cpu_rt_period_read_uint,
8761 .write_u64 = cpu_rt_period_write_uint,
8766 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8768 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8771 struct cgroup_subsys cpu_cgroup_subsys = {
8773 .create = cpu_cgroup_create,
8774 .destroy = cpu_cgroup_destroy,
8775 .can_attach = cpu_cgroup_can_attach,
8776 .attach = cpu_cgroup_attach,
8777 .populate = cpu_cgroup_populate,
8778 .subsys_id = cpu_cgroup_subsys_id,
8782 #endif /* CONFIG_CGROUP_SCHED */
8784 #ifdef CONFIG_CGROUP_CPUACCT
8787 * CPU accounting code for task groups.
8789 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8790 * (balbir@in.ibm.com).
8793 /* track cpu usage of a group of tasks and its child groups */
8795 struct cgroup_subsys_state css;
8796 /* cpuusage holds pointer to a u64-type object on every cpu */
8797 u64 __percpu *cpuusage;
8798 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
8799 struct cpuacct *parent;
8802 struct cgroup_subsys cpuacct_subsys;
8804 /* return cpu accounting group corresponding to this container */
8805 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8807 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8808 struct cpuacct, css);
8811 /* return cpu accounting group to which this task belongs */
8812 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8814 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8815 struct cpuacct, css);
8818 /* create a new cpu accounting group */
8819 static struct cgroup_subsys_state *cpuacct_create(
8820 struct cgroup_subsys *ss, struct cgroup *cgrp)
8822 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8828 ca->cpuusage = alloc_percpu(u64);
8832 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8833 if (percpu_counter_init(&ca->cpustat[i], 0))
8834 goto out_free_counters;
8837 ca->parent = cgroup_ca(cgrp->parent);
8843 percpu_counter_destroy(&ca->cpustat[i]);
8844 free_percpu(ca->cpuusage);
8848 return ERR_PTR(-ENOMEM);
8851 /* destroy an existing cpu accounting group */
8853 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8855 struct cpuacct *ca = cgroup_ca(cgrp);
8858 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8859 percpu_counter_destroy(&ca->cpustat[i]);
8860 free_percpu(ca->cpuusage);
8864 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
8866 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8869 #ifndef CONFIG_64BIT
8871 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8873 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8875 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8883 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
8885 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8887 #ifndef CONFIG_64BIT
8889 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8891 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8893 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8899 /* return total cpu usage (in nanoseconds) of a group */
8900 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8902 struct cpuacct *ca = cgroup_ca(cgrp);
8903 u64 totalcpuusage = 0;
8906 for_each_present_cpu(i)
8907 totalcpuusage += cpuacct_cpuusage_read(ca, i);
8909 return totalcpuusage;
8912 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8915 struct cpuacct *ca = cgroup_ca(cgrp);
8924 for_each_present_cpu(i)
8925 cpuacct_cpuusage_write(ca, i, 0);
8931 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
8934 struct cpuacct *ca = cgroup_ca(cgroup);
8938 for_each_present_cpu(i) {
8939 percpu = cpuacct_cpuusage_read(ca, i);
8940 seq_printf(m, "%llu ", (unsigned long long) percpu);
8942 seq_printf(m, "\n");
8946 static const char *cpuacct_stat_desc[] = {
8947 [CPUACCT_STAT_USER] = "user",
8948 [CPUACCT_STAT_SYSTEM] = "system",
8951 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
8952 struct cgroup_map_cb *cb)
8954 struct cpuacct *ca = cgroup_ca(cgrp);
8957 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
8958 s64 val = percpu_counter_read(&ca->cpustat[i]);
8959 val = cputime64_to_clock_t(val);
8960 cb->fill(cb, cpuacct_stat_desc[i], val);
8965 static struct cftype files[] = {
8968 .read_u64 = cpuusage_read,
8969 .write_u64 = cpuusage_write,
8972 .name = "usage_percpu",
8973 .read_seq_string = cpuacct_percpu_seq_read,
8977 .read_map = cpuacct_stats_show,
8981 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
8983 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
8987 * charge this task's execution time to its accounting group.
8989 * called with rq->lock held.
8991 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8996 if (unlikely(!cpuacct_subsys.active))
8999 cpu = task_cpu(tsk);
9005 for (; ca; ca = ca->parent) {
9006 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9007 *cpuusage += cputime;
9014 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9015 * in cputime_t units. As a result, cpuacct_update_stats calls
9016 * percpu_counter_add with values large enough to always overflow the
9017 * per cpu batch limit causing bad SMP scalability.
9019 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9020 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9021 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9024 #define CPUACCT_BATCH \
9025 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9027 #define CPUACCT_BATCH 0
9031 * Charge the system/user time to the task's accounting group.
9033 static void cpuacct_update_stats(struct task_struct *tsk,
9034 enum cpuacct_stat_index idx, cputime_t val)
9037 int batch = CPUACCT_BATCH;
9039 if (unlikely(!cpuacct_subsys.active))
9046 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9052 struct cgroup_subsys cpuacct_subsys = {
9054 .create = cpuacct_create,
9055 .destroy = cpuacct_destroy,
9056 .populate = cpuacct_populate,
9057 .subsys_id = cpuacct_subsys_id,
9059 #endif /* CONFIG_CGROUP_CPUACCT */
9063 void synchronize_sched_expedited(void)
9067 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9069 #else /* #ifndef CONFIG_SMP */
9071 static atomic_t synchronize_sched_expedited_count = ATOMIC_INIT(0);
9073 static int synchronize_sched_expedited_cpu_stop(void *data)
9076 * There must be a full memory barrier on each affected CPU
9077 * between the time that try_stop_cpus() is called and the
9078 * time that it returns.
9080 * In the current initial implementation of cpu_stop, the
9081 * above condition is already met when the control reaches
9082 * this point and the following smp_mb() is not strictly
9083 * necessary. Do smp_mb() anyway for documentation and
9084 * robustness against future implementation changes.
9086 smp_mb(); /* See above comment block. */
9091 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
9092 * approach to force grace period to end quickly. This consumes
9093 * significant time on all CPUs, and is thus not recommended for
9094 * any sort of common-case code.
9096 * Note that it is illegal to call this function while holding any
9097 * lock that is acquired by a CPU-hotplug notifier. Failing to
9098 * observe this restriction will result in deadlock.
9100 void synchronize_sched_expedited(void)
9102 int snap, trycount = 0;
9104 smp_mb(); /* ensure prior mod happens before capturing snap. */
9105 snap = atomic_read(&synchronize_sched_expedited_count) + 1;
9107 while (try_stop_cpus(cpu_online_mask,
9108 synchronize_sched_expedited_cpu_stop,
9111 if (trycount++ < 10)
9112 udelay(trycount * num_online_cpus());
9114 synchronize_sched();
9117 if (atomic_read(&synchronize_sched_expedited_count) - snap > 0) {
9118 smp_mb(); /* ensure test happens before caller kfree */
9123 atomic_inc(&synchronize_sched_expedited_count);
9124 smp_mb__after_atomic_inc(); /* ensure post-GP actions seen after GP. */
9127 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9129 #endif /* #else #ifndef CONFIG_SMP */