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;
2599 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2606 * wake_up_new_task - wake up a newly created task for the first time.
2608 * This function will do some initial scheduler statistics housekeeping
2609 * that must be done for every newly created context, then puts the task
2610 * on the runqueue and wakes it.
2612 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2614 unsigned long flags;
2616 int cpu __maybe_unused = get_cpu();
2619 rq = task_rq_lock(p, &flags);
2620 p->state = TASK_WAKING;
2623 * Fork balancing, do it here and not earlier because:
2624 * - cpus_allowed can change in the fork path
2625 * - any previously selected cpu might disappear through hotplug
2627 * We set TASK_WAKING so that select_task_rq() can drop rq->lock
2628 * without people poking at ->cpus_allowed.
2630 cpu = select_task_rq(rq, p, SD_BALANCE_FORK, 0);
2631 set_task_cpu(p, cpu);
2633 p->state = TASK_RUNNING;
2634 task_rq_unlock(rq, &flags);
2637 rq = task_rq_lock(p, &flags);
2638 activate_task(rq, p, 0);
2639 trace_sched_wakeup_new(p, 1);
2640 check_preempt_curr(rq, p, WF_FORK);
2642 if (p->sched_class->task_woken)
2643 p->sched_class->task_woken(rq, p);
2645 task_rq_unlock(rq, &flags);
2649 #ifdef CONFIG_PREEMPT_NOTIFIERS
2652 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2653 * @notifier: notifier struct to register
2655 void preempt_notifier_register(struct preempt_notifier *notifier)
2657 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2659 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2662 * preempt_notifier_unregister - no longer interested in preemption notifications
2663 * @notifier: notifier struct to unregister
2665 * This is safe to call from within a preemption notifier.
2667 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2669 hlist_del(¬ifier->link);
2671 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2673 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2675 struct preempt_notifier *notifier;
2676 struct hlist_node *node;
2678 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2679 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2683 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2684 struct task_struct *next)
2686 struct preempt_notifier *notifier;
2687 struct hlist_node *node;
2689 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2690 notifier->ops->sched_out(notifier, next);
2693 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2695 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2700 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2701 struct task_struct *next)
2705 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2708 * prepare_task_switch - prepare to switch tasks
2709 * @rq: the runqueue preparing to switch
2710 * @prev: the current task that is being switched out
2711 * @next: the task we are going to switch to.
2713 * This is called with the rq lock held and interrupts off. It must
2714 * be paired with a subsequent finish_task_switch after the context
2717 * prepare_task_switch sets up locking and calls architecture specific
2721 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2722 struct task_struct *next)
2724 fire_sched_out_preempt_notifiers(prev, next);
2725 prepare_lock_switch(rq, next);
2726 prepare_arch_switch(next);
2730 * finish_task_switch - clean up after a task-switch
2731 * @rq: runqueue associated with task-switch
2732 * @prev: the thread we just switched away from.
2734 * finish_task_switch must be called after the context switch, paired
2735 * with a prepare_task_switch call before the context switch.
2736 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2737 * and do any other architecture-specific cleanup actions.
2739 * Note that we may have delayed dropping an mm in context_switch(). If
2740 * so, we finish that here outside of the runqueue lock. (Doing it
2741 * with the lock held can cause deadlocks; see schedule() for
2744 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2745 __releases(rq->lock)
2747 struct mm_struct *mm = rq->prev_mm;
2753 * A task struct has one reference for the use as "current".
2754 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2755 * schedule one last time. The schedule call will never return, and
2756 * the scheduled task must drop that reference.
2757 * The test for TASK_DEAD must occur while the runqueue locks are
2758 * still held, otherwise prev could be scheduled on another cpu, die
2759 * there before we look at prev->state, and then the reference would
2761 * Manfred Spraul <manfred@colorfullife.com>
2763 prev_state = prev->state;
2764 finish_arch_switch(prev);
2765 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2766 local_irq_disable();
2767 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2768 perf_event_task_sched_in(current);
2769 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2771 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2772 finish_lock_switch(rq, prev);
2774 fire_sched_in_preempt_notifiers(current);
2777 if (unlikely(prev_state == TASK_DEAD)) {
2779 * Remove function-return probe instances associated with this
2780 * task and put them back on the free list.
2782 kprobe_flush_task(prev);
2783 put_task_struct(prev);
2789 /* assumes rq->lock is held */
2790 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2792 if (prev->sched_class->pre_schedule)
2793 prev->sched_class->pre_schedule(rq, prev);
2796 /* rq->lock is NOT held, but preemption is disabled */
2797 static inline void post_schedule(struct rq *rq)
2799 if (rq->post_schedule) {
2800 unsigned long flags;
2802 raw_spin_lock_irqsave(&rq->lock, flags);
2803 if (rq->curr->sched_class->post_schedule)
2804 rq->curr->sched_class->post_schedule(rq);
2805 raw_spin_unlock_irqrestore(&rq->lock, flags);
2807 rq->post_schedule = 0;
2813 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2817 static inline void post_schedule(struct rq *rq)
2824 * schedule_tail - first thing a freshly forked thread must call.
2825 * @prev: the thread we just switched away from.
2827 asmlinkage void schedule_tail(struct task_struct *prev)
2828 __releases(rq->lock)
2830 struct rq *rq = this_rq();
2832 finish_task_switch(rq, prev);
2835 * FIXME: do we need to worry about rq being invalidated by the
2840 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2841 /* In this case, finish_task_switch does not reenable preemption */
2844 if (current->set_child_tid)
2845 put_user(task_pid_vnr(current), current->set_child_tid);
2849 * context_switch - switch to the new MM and the new
2850 * thread's register state.
2853 context_switch(struct rq *rq, struct task_struct *prev,
2854 struct task_struct *next)
2856 struct mm_struct *mm, *oldmm;
2858 prepare_task_switch(rq, prev, next);
2859 trace_sched_switch(prev, next);
2861 oldmm = prev->active_mm;
2863 * For paravirt, this is coupled with an exit in switch_to to
2864 * combine the page table reload and the switch backend into
2867 arch_start_context_switch(prev);
2870 next->active_mm = oldmm;
2871 atomic_inc(&oldmm->mm_count);
2872 enter_lazy_tlb(oldmm, next);
2874 switch_mm(oldmm, mm, next);
2877 prev->active_mm = NULL;
2878 rq->prev_mm = oldmm;
2881 * Since the runqueue lock will be released by the next
2882 * task (which is an invalid locking op but in the case
2883 * of the scheduler it's an obvious special-case), so we
2884 * do an early lockdep release here:
2886 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2887 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2890 /* Here we just switch the register state and the stack. */
2891 switch_to(prev, next, prev);
2895 * this_rq must be evaluated again because prev may have moved
2896 * CPUs since it called schedule(), thus the 'rq' on its stack
2897 * frame will be invalid.
2899 finish_task_switch(this_rq(), prev);
2903 * nr_running, nr_uninterruptible and nr_context_switches:
2905 * externally visible scheduler statistics: current number of runnable
2906 * threads, current number of uninterruptible-sleeping threads, total
2907 * number of context switches performed since bootup.
2909 unsigned long nr_running(void)
2911 unsigned long i, sum = 0;
2913 for_each_online_cpu(i)
2914 sum += cpu_rq(i)->nr_running;
2919 unsigned long nr_uninterruptible(void)
2921 unsigned long i, sum = 0;
2923 for_each_possible_cpu(i)
2924 sum += cpu_rq(i)->nr_uninterruptible;
2927 * Since we read the counters lockless, it might be slightly
2928 * inaccurate. Do not allow it to go below zero though:
2930 if (unlikely((long)sum < 0))
2936 unsigned long long nr_context_switches(void)
2939 unsigned long long sum = 0;
2941 for_each_possible_cpu(i)
2942 sum += cpu_rq(i)->nr_switches;
2947 unsigned long nr_iowait(void)
2949 unsigned long i, sum = 0;
2951 for_each_possible_cpu(i)
2952 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2957 unsigned long nr_iowait_cpu(int cpu)
2959 struct rq *this = cpu_rq(cpu);
2960 return atomic_read(&this->nr_iowait);
2963 unsigned long this_cpu_load(void)
2965 struct rq *this = this_rq();
2966 return this->cpu_load[0];
2970 /* Variables and functions for calc_load */
2971 static atomic_long_t calc_load_tasks;
2972 static unsigned long calc_load_update;
2973 unsigned long avenrun[3];
2974 EXPORT_SYMBOL(avenrun);
2976 static long calc_load_fold_active(struct rq *this_rq)
2978 long nr_active, delta = 0;
2980 nr_active = this_rq->nr_running;
2981 nr_active += (long) this_rq->nr_uninterruptible;
2983 if (nr_active != this_rq->calc_load_active) {
2984 delta = nr_active - this_rq->calc_load_active;
2985 this_rq->calc_load_active = nr_active;
2993 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
2995 * When making the ILB scale, we should try to pull this in as well.
2997 static atomic_long_t calc_load_tasks_idle;
2999 static void calc_load_account_idle(struct rq *this_rq)
3003 delta = calc_load_fold_active(this_rq);
3005 atomic_long_add(delta, &calc_load_tasks_idle);
3008 static long calc_load_fold_idle(void)
3013 * Its got a race, we don't care...
3015 if (atomic_long_read(&calc_load_tasks_idle))
3016 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
3021 static void calc_load_account_idle(struct rq *this_rq)
3025 static inline long calc_load_fold_idle(void)
3032 * get_avenrun - get the load average array
3033 * @loads: pointer to dest load array
3034 * @offset: offset to add
3035 * @shift: shift count to shift the result left
3037 * These values are estimates at best, so no need for locking.
3039 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3041 loads[0] = (avenrun[0] + offset) << shift;
3042 loads[1] = (avenrun[1] + offset) << shift;
3043 loads[2] = (avenrun[2] + offset) << shift;
3046 static unsigned long
3047 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3050 load += active * (FIXED_1 - exp);
3051 return load >> FSHIFT;
3055 * calc_load - update the avenrun load estimates 10 ticks after the
3056 * CPUs have updated calc_load_tasks.
3058 void calc_global_load(void)
3060 unsigned long upd = calc_load_update + 10;
3063 if (time_before(jiffies, upd))
3066 active = atomic_long_read(&calc_load_tasks);
3067 active = active > 0 ? active * FIXED_1 : 0;
3069 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3070 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3071 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3073 calc_load_update += LOAD_FREQ;
3077 * Called from update_cpu_load() to periodically update this CPU's
3080 static void calc_load_account_active(struct rq *this_rq)
3084 if (time_before(jiffies, this_rq->calc_load_update))
3087 delta = calc_load_fold_active(this_rq);
3088 delta += calc_load_fold_idle();
3090 atomic_long_add(delta, &calc_load_tasks);
3092 this_rq->calc_load_update += LOAD_FREQ;
3096 * The exact cpuload at various idx values, calculated at every tick would be
3097 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3099 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3100 * on nth tick when cpu may be busy, then we have:
3101 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3102 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3104 * decay_load_missed() below does efficient calculation of
3105 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3106 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3108 * The calculation is approximated on a 128 point scale.
3109 * degrade_zero_ticks is the number of ticks after which load at any
3110 * particular idx is approximated to be zero.
3111 * degrade_factor is a precomputed table, a row for each load idx.
3112 * Each column corresponds to degradation factor for a power of two ticks,
3113 * based on 128 point scale.
3115 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3116 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3118 * With this power of 2 load factors, we can degrade the load n times
3119 * by looking at 1 bits in n and doing as many mult/shift instead of
3120 * n mult/shifts needed by the exact degradation.
3122 #define DEGRADE_SHIFT 7
3123 static const unsigned char
3124 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3125 static const unsigned char
3126 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3127 {0, 0, 0, 0, 0, 0, 0, 0},
3128 {64, 32, 8, 0, 0, 0, 0, 0},
3129 {96, 72, 40, 12, 1, 0, 0},
3130 {112, 98, 75, 43, 15, 1, 0},
3131 {120, 112, 98, 76, 45, 16, 2} };
3134 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3135 * would be when CPU is idle and so we just decay the old load without
3136 * adding any new load.
3138 static unsigned long
3139 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3143 if (!missed_updates)
3146 if (missed_updates >= degrade_zero_ticks[idx])
3150 return load >> missed_updates;
3152 while (missed_updates) {
3153 if (missed_updates % 2)
3154 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3156 missed_updates >>= 1;
3163 * Update rq->cpu_load[] statistics. This function is usually called every
3164 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3165 * every tick. We fix it up based on jiffies.
3167 static void update_cpu_load(struct rq *this_rq)
3169 unsigned long this_load = this_rq->load.weight;
3170 unsigned long curr_jiffies = jiffies;
3171 unsigned long pending_updates;
3174 this_rq->nr_load_updates++;
3176 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3177 if (curr_jiffies == this_rq->last_load_update_tick)
3180 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3181 this_rq->last_load_update_tick = curr_jiffies;
3183 /* Update our load: */
3184 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3185 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3186 unsigned long old_load, new_load;
3188 /* scale is effectively 1 << i now, and >> i divides by scale */
3190 old_load = this_rq->cpu_load[i];
3191 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3192 new_load = this_load;
3194 * Round up the averaging division if load is increasing. This
3195 * prevents us from getting stuck on 9 if the load is 10, for
3198 if (new_load > old_load)
3199 new_load += scale - 1;
3201 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3204 sched_avg_update(this_rq);
3207 static void update_cpu_load_active(struct rq *this_rq)
3209 update_cpu_load(this_rq);
3211 calc_load_account_active(this_rq);
3217 * sched_exec - execve() is a valuable balancing opportunity, because at
3218 * this point the task has the smallest effective memory and cache footprint.
3220 void sched_exec(void)
3222 struct task_struct *p = current;
3223 unsigned long flags;
3227 rq = task_rq_lock(p, &flags);
3228 dest_cpu = p->sched_class->select_task_rq(rq, p, SD_BALANCE_EXEC, 0);
3229 if (dest_cpu == smp_processor_id())
3233 * select_task_rq() can race against ->cpus_allowed
3235 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed) &&
3236 likely(cpu_active(dest_cpu)) && migrate_task(p, rq)) {
3237 struct migration_arg arg = { p, dest_cpu };
3239 task_rq_unlock(rq, &flags);
3240 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
3244 task_rq_unlock(rq, &flags);
3249 DEFINE_PER_CPU(struct kernel_stat, kstat);
3251 EXPORT_PER_CPU_SYMBOL(kstat);
3254 * Return any ns on the sched_clock that have not yet been accounted in
3255 * @p in case that task is currently running.
3257 * Called with task_rq_lock() held on @rq.
3259 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3263 if (task_current(rq, p)) {
3264 update_rq_clock(rq);
3265 ns = rq->clock_task - p->se.exec_start;
3273 unsigned long long task_delta_exec(struct task_struct *p)
3275 unsigned long flags;
3279 rq = task_rq_lock(p, &flags);
3280 ns = do_task_delta_exec(p, rq);
3281 task_rq_unlock(rq, &flags);
3287 * Return accounted runtime for the task.
3288 * In case the task is currently running, return the runtime plus current's
3289 * pending runtime that have not been accounted yet.
3291 unsigned long long task_sched_runtime(struct task_struct *p)
3293 unsigned long flags;
3297 rq = task_rq_lock(p, &flags);
3298 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3299 task_rq_unlock(rq, &flags);
3305 * Return sum_exec_runtime for the thread group.
3306 * In case the task is currently running, return the sum plus current's
3307 * pending runtime that have not been accounted yet.
3309 * Note that the thread group might have other running tasks as well,
3310 * so the return value not includes other pending runtime that other
3311 * running tasks might have.
3313 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3315 struct task_cputime totals;
3316 unsigned long flags;
3320 rq = task_rq_lock(p, &flags);
3321 thread_group_cputime(p, &totals);
3322 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3323 task_rq_unlock(rq, &flags);
3329 * Account user cpu time to a process.
3330 * @p: the process that the cpu time gets accounted to
3331 * @cputime: the cpu time spent in user space since the last update
3332 * @cputime_scaled: cputime scaled by cpu frequency
3334 void account_user_time(struct task_struct *p, cputime_t cputime,
3335 cputime_t cputime_scaled)
3337 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3340 /* Add user time to process. */
3341 p->utime = cputime_add(p->utime, cputime);
3342 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3343 account_group_user_time(p, cputime);
3345 /* Add user time to cpustat. */
3346 tmp = cputime_to_cputime64(cputime);
3347 if (TASK_NICE(p) > 0)
3348 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3350 cpustat->user = cputime64_add(cpustat->user, tmp);
3352 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3353 /* Account for user time used */
3354 acct_update_integrals(p);
3358 * Account guest cpu time to a process.
3359 * @p: the process that the cpu time gets accounted to
3360 * @cputime: the cpu time spent in virtual machine since the last update
3361 * @cputime_scaled: cputime scaled by cpu frequency
3363 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3364 cputime_t cputime_scaled)
3367 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3369 tmp = cputime_to_cputime64(cputime);
3371 /* Add guest time to process. */
3372 p->utime = cputime_add(p->utime, cputime);
3373 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3374 account_group_user_time(p, cputime);
3375 p->gtime = cputime_add(p->gtime, cputime);
3377 /* Add guest time to cpustat. */
3378 if (TASK_NICE(p) > 0) {
3379 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3380 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3382 cpustat->user = cputime64_add(cpustat->user, tmp);
3383 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3388 * Account system cpu time to a process.
3389 * @p: the process that the cpu time gets accounted to
3390 * @hardirq_offset: the offset to subtract from hardirq_count()
3391 * @cputime: the cpu time spent in kernel space since the last update
3392 * @cputime_scaled: cputime scaled by cpu frequency
3394 void account_system_time(struct task_struct *p, int hardirq_offset,
3395 cputime_t cputime, cputime_t cputime_scaled)
3397 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3400 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3401 account_guest_time(p, cputime, cputime_scaled);
3405 /* Add system time to process. */
3406 p->stime = cputime_add(p->stime, cputime);
3407 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3408 account_group_system_time(p, cputime);
3410 /* Add system time to cpustat. */
3411 tmp = cputime_to_cputime64(cputime);
3412 if (hardirq_count() - hardirq_offset)
3413 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3414 else if (in_serving_softirq())
3415 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3417 cpustat->system = cputime64_add(cpustat->system, tmp);
3419 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3421 /* Account for system time used */
3422 acct_update_integrals(p);
3426 * Account for involuntary wait time.
3427 * @steal: the cpu time spent in involuntary wait
3429 void account_steal_time(cputime_t cputime)
3431 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3432 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3434 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3438 * Account for idle time.
3439 * @cputime: the cpu time spent in idle wait
3441 void account_idle_time(cputime_t cputime)
3443 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3444 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3445 struct rq *rq = this_rq();
3447 if (atomic_read(&rq->nr_iowait) > 0)
3448 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3450 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3453 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3456 * Account a single tick of cpu time.
3457 * @p: the process that the cpu time gets accounted to
3458 * @user_tick: indicates if the tick is a user or a system tick
3460 void account_process_tick(struct task_struct *p, int user_tick)
3462 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3463 struct rq *rq = this_rq();
3466 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3467 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3468 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3471 account_idle_time(cputime_one_jiffy);
3475 * Account multiple ticks of steal time.
3476 * @p: the process from which the cpu time has been stolen
3477 * @ticks: number of stolen ticks
3479 void account_steal_ticks(unsigned long ticks)
3481 account_steal_time(jiffies_to_cputime(ticks));
3485 * Account multiple ticks of idle time.
3486 * @ticks: number of stolen ticks
3488 void account_idle_ticks(unsigned long ticks)
3490 account_idle_time(jiffies_to_cputime(ticks));
3496 * Use precise platform statistics if available:
3498 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3499 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3505 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3507 struct task_cputime cputime;
3509 thread_group_cputime(p, &cputime);
3511 *ut = cputime.utime;
3512 *st = cputime.stime;
3516 #ifndef nsecs_to_cputime
3517 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3520 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3522 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
3525 * Use CFS's precise accounting:
3527 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3533 do_div(temp, total);
3534 utime = (cputime_t)temp;
3539 * Compare with previous values, to keep monotonicity:
3541 p->prev_utime = max(p->prev_utime, utime);
3542 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
3544 *ut = p->prev_utime;
3545 *st = p->prev_stime;
3549 * Must be called with siglock held.
3551 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3553 struct signal_struct *sig = p->signal;
3554 struct task_cputime cputime;
3555 cputime_t rtime, utime, total;
3557 thread_group_cputime(p, &cputime);
3559 total = cputime_add(cputime.utime, cputime.stime);
3560 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3565 temp *= cputime.utime;
3566 do_div(temp, total);
3567 utime = (cputime_t)temp;
3571 sig->prev_utime = max(sig->prev_utime, utime);
3572 sig->prev_stime = max(sig->prev_stime,
3573 cputime_sub(rtime, sig->prev_utime));
3575 *ut = sig->prev_utime;
3576 *st = sig->prev_stime;
3581 * This function gets called by the timer code, with HZ frequency.
3582 * We call it with interrupts disabled.
3584 * It also gets called by the fork code, when changing the parent's
3587 void scheduler_tick(void)
3589 int cpu = smp_processor_id();
3590 struct rq *rq = cpu_rq(cpu);
3591 struct task_struct *curr = rq->curr;
3595 raw_spin_lock(&rq->lock);
3596 update_rq_clock(rq);
3597 update_cpu_load_active(rq);
3598 curr->sched_class->task_tick(rq, curr, 0);
3599 raw_spin_unlock(&rq->lock);
3601 perf_event_task_tick();
3604 rq->idle_at_tick = idle_cpu(cpu);
3605 trigger_load_balance(rq, cpu);
3609 notrace unsigned long get_parent_ip(unsigned long addr)
3611 if (in_lock_functions(addr)) {
3612 addr = CALLER_ADDR2;
3613 if (in_lock_functions(addr))
3614 addr = CALLER_ADDR3;
3619 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3620 defined(CONFIG_PREEMPT_TRACER))
3622 void __kprobes add_preempt_count(int val)
3624 #ifdef CONFIG_DEBUG_PREEMPT
3628 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3631 preempt_count() += val;
3632 #ifdef CONFIG_DEBUG_PREEMPT
3634 * Spinlock count overflowing soon?
3636 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3639 if (preempt_count() == val)
3640 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3642 EXPORT_SYMBOL(add_preempt_count);
3644 void __kprobes sub_preempt_count(int val)
3646 #ifdef CONFIG_DEBUG_PREEMPT
3650 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3653 * Is the spinlock portion underflowing?
3655 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3656 !(preempt_count() & PREEMPT_MASK)))
3660 if (preempt_count() == val)
3661 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3662 preempt_count() -= val;
3664 EXPORT_SYMBOL(sub_preempt_count);
3669 * Print scheduling while atomic bug:
3671 static noinline void __schedule_bug(struct task_struct *prev)
3673 struct pt_regs *regs = get_irq_regs();
3675 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3676 prev->comm, prev->pid, preempt_count());
3678 debug_show_held_locks(prev);
3680 if (irqs_disabled())
3681 print_irqtrace_events(prev);
3690 * Various schedule()-time debugging checks and statistics:
3692 static inline void schedule_debug(struct task_struct *prev)
3695 * Test if we are atomic. Since do_exit() needs to call into
3696 * schedule() atomically, we ignore that path for now.
3697 * Otherwise, whine if we are scheduling when we should not be.
3699 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3700 __schedule_bug(prev);
3702 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3704 schedstat_inc(this_rq(), sched_count);
3705 #ifdef CONFIG_SCHEDSTATS
3706 if (unlikely(prev->lock_depth >= 0)) {
3707 schedstat_inc(this_rq(), bkl_count);
3708 schedstat_inc(prev, sched_info.bkl_count);
3713 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3716 update_rq_clock(rq);
3717 rq->skip_clock_update = 0;
3718 prev->sched_class->put_prev_task(rq, prev);
3722 * Pick up the highest-prio task:
3724 static inline struct task_struct *
3725 pick_next_task(struct rq *rq)
3727 const struct sched_class *class;
3728 struct task_struct *p;
3731 * Optimization: we know that if all tasks are in
3732 * the fair class we can call that function directly:
3734 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3735 p = fair_sched_class.pick_next_task(rq);
3740 for_each_class(class) {
3741 p = class->pick_next_task(rq);
3746 BUG(); /* the idle class will always have a runnable task */
3750 * schedule() is the main scheduler function.
3752 asmlinkage void __sched schedule(void)
3754 struct task_struct *prev, *next;
3755 unsigned long *switch_count;
3761 cpu = smp_processor_id();
3763 rcu_note_context_switch(cpu);
3766 release_kernel_lock(prev);
3767 need_resched_nonpreemptible:
3769 schedule_debug(prev);
3771 if (sched_feat(HRTICK))
3774 raw_spin_lock_irq(&rq->lock);
3775 clear_tsk_need_resched(prev);
3777 switch_count = &prev->nivcsw;
3778 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3779 if (unlikely(signal_pending_state(prev->state, prev))) {
3780 prev->state = TASK_RUNNING;
3783 * If a worker is going to sleep, notify and
3784 * ask workqueue whether it wants to wake up a
3785 * task to maintain concurrency. If so, wake
3788 if (prev->flags & PF_WQ_WORKER) {
3789 struct task_struct *to_wakeup;
3791 to_wakeup = wq_worker_sleeping(prev, cpu);
3793 try_to_wake_up_local(to_wakeup);
3795 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3797 switch_count = &prev->nvcsw;
3800 pre_schedule(rq, prev);
3802 if (unlikely(!rq->nr_running))
3803 idle_balance(cpu, rq);
3805 put_prev_task(rq, prev);
3806 next = pick_next_task(rq);
3808 if (likely(prev != next)) {
3809 sched_info_switch(prev, next);
3810 perf_event_task_sched_out(prev, next);
3816 context_switch(rq, prev, next); /* unlocks the rq */
3818 * The context switch have flipped the stack from under us
3819 * and restored the local variables which were saved when
3820 * this task called schedule() in the past. prev == current
3821 * is still correct, but it can be moved to another cpu/rq.
3823 cpu = smp_processor_id();
3826 raw_spin_unlock_irq(&rq->lock);
3830 if (unlikely(reacquire_kernel_lock(prev)))
3831 goto need_resched_nonpreemptible;
3833 preempt_enable_no_resched();
3837 EXPORT_SYMBOL(schedule);
3839 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3841 * Look out! "owner" is an entirely speculative pointer
3842 * access and not reliable.
3844 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
3849 if (!sched_feat(OWNER_SPIN))
3852 #ifdef CONFIG_DEBUG_PAGEALLOC
3854 * Need to access the cpu field knowing that
3855 * DEBUG_PAGEALLOC could have unmapped it if
3856 * the mutex owner just released it and exited.
3858 if (probe_kernel_address(&owner->cpu, cpu))
3865 * Even if the access succeeded (likely case),
3866 * the cpu field may no longer be valid.
3868 if (cpu >= nr_cpumask_bits)
3872 * We need to validate that we can do a
3873 * get_cpu() and that we have the percpu area.
3875 if (!cpu_online(cpu))
3882 * Owner changed, break to re-assess state.
3884 if (lock->owner != owner) {
3886 * If the lock has switched to a different owner,
3887 * we likely have heavy contention. Return 0 to quit
3888 * optimistic spinning and not contend further:
3896 * Is that owner really running on that cpu?
3898 if (task_thread_info(rq->curr) != owner || need_resched())
3901 arch_mutex_cpu_relax();
3908 #ifdef CONFIG_PREEMPT
3910 * this is the entry point to schedule() from in-kernel preemption
3911 * off of preempt_enable. Kernel preemptions off return from interrupt
3912 * occur there and call schedule directly.
3914 asmlinkage void __sched notrace preempt_schedule(void)
3916 struct thread_info *ti = current_thread_info();
3919 * If there is a non-zero preempt_count or interrupts are disabled,
3920 * we do not want to preempt the current task. Just return..
3922 if (likely(ti->preempt_count || irqs_disabled()))
3926 add_preempt_count_notrace(PREEMPT_ACTIVE);
3928 sub_preempt_count_notrace(PREEMPT_ACTIVE);
3931 * Check again in case we missed a preemption opportunity
3932 * between schedule and now.
3935 } while (need_resched());
3937 EXPORT_SYMBOL(preempt_schedule);
3940 * this is the entry point to schedule() from kernel preemption
3941 * off of irq context.
3942 * Note, that this is called and return with irqs disabled. This will
3943 * protect us against recursive calling from irq.
3945 asmlinkage void __sched preempt_schedule_irq(void)
3947 struct thread_info *ti = current_thread_info();
3949 /* Catch callers which need to be fixed */
3950 BUG_ON(ti->preempt_count || !irqs_disabled());
3953 add_preempt_count(PREEMPT_ACTIVE);
3956 local_irq_disable();
3957 sub_preempt_count(PREEMPT_ACTIVE);
3960 * Check again in case we missed a preemption opportunity
3961 * between schedule and now.
3964 } while (need_resched());
3967 #endif /* CONFIG_PREEMPT */
3969 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3972 return try_to_wake_up(curr->private, mode, wake_flags);
3974 EXPORT_SYMBOL(default_wake_function);
3977 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3978 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3979 * number) then we wake all the non-exclusive tasks and one exclusive task.
3981 * There are circumstances in which we can try to wake a task which has already
3982 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3983 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3985 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3986 int nr_exclusive, int wake_flags, void *key)
3988 wait_queue_t *curr, *next;
3990 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3991 unsigned flags = curr->flags;
3993 if (curr->func(curr, mode, wake_flags, key) &&
3994 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4000 * __wake_up - wake up threads blocked on a waitqueue.
4002 * @mode: which threads
4003 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4004 * @key: is directly passed to the wakeup function
4006 * It may be assumed that this function implies a write memory barrier before
4007 * changing the task state if and only if any tasks are woken up.
4009 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4010 int nr_exclusive, void *key)
4012 unsigned long flags;
4014 spin_lock_irqsave(&q->lock, flags);
4015 __wake_up_common(q, mode, nr_exclusive, 0, key);
4016 spin_unlock_irqrestore(&q->lock, flags);
4018 EXPORT_SYMBOL(__wake_up);
4021 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4023 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4025 __wake_up_common(q, mode, 1, 0, NULL);
4027 EXPORT_SYMBOL_GPL(__wake_up_locked);
4029 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
4031 __wake_up_common(q, mode, 1, 0, key);
4035 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4037 * @mode: which threads
4038 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4039 * @key: opaque value to be passed to wakeup targets
4041 * The sync wakeup differs that the waker knows that it will schedule
4042 * away soon, so while the target thread will be woken up, it will not
4043 * be migrated to another CPU - ie. the two threads are 'synchronized'
4044 * with each other. This can prevent needless bouncing between CPUs.
4046 * On UP it can prevent extra preemption.
4048 * It may be assumed that this function implies a write memory barrier before
4049 * changing the task state if and only if any tasks are woken up.
4051 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4052 int nr_exclusive, void *key)
4054 unsigned long flags;
4055 int wake_flags = WF_SYNC;
4060 if (unlikely(!nr_exclusive))
4063 spin_lock_irqsave(&q->lock, flags);
4064 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4065 spin_unlock_irqrestore(&q->lock, flags);
4067 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4070 * __wake_up_sync - see __wake_up_sync_key()
4072 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4074 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4076 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4079 * complete: - signals a single thread waiting on this completion
4080 * @x: holds the state of this particular completion
4082 * This will wake up a single thread waiting on this completion. Threads will be
4083 * awakened in the same order in which they were queued.
4085 * See also complete_all(), wait_for_completion() and related routines.
4087 * It may be assumed that this function implies a write memory barrier before
4088 * changing the task state if and only if any tasks are woken up.
4090 void complete(struct completion *x)
4092 unsigned long flags;
4094 spin_lock_irqsave(&x->wait.lock, flags);
4096 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4097 spin_unlock_irqrestore(&x->wait.lock, flags);
4099 EXPORT_SYMBOL(complete);
4102 * complete_all: - signals all threads waiting on this completion
4103 * @x: holds the state of this particular completion
4105 * This will wake up all threads waiting on this particular completion event.
4107 * It may be assumed that this function implies a write memory barrier before
4108 * changing the task state if and only if any tasks are woken up.
4110 void complete_all(struct completion *x)
4112 unsigned long flags;
4114 spin_lock_irqsave(&x->wait.lock, flags);
4115 x->done += UINT_MAX/2;
4116 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4117 spin_unlock_irqrestore(&x->wait.lock, flags);
4119 EXPORT_SYMBOL(complete_all);
4121 static inline long __sched
4122 do_wait_for_common(struct completion *x, long timeout, int state)
4125 DECLARE_WAITQUEUE(wait, current);
4127 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4129 if (signal_pending_state(state, current)) {
4130 timeout = -ERESTARTSYS;
4133 __set_current_state(state);
4134 spin_unlock_irq(&x->wait.lock);
4135 timeout = schedule_timeout(timeout);
4136 spin_lock_irq(&x->wait.lock);
4137 } while (!x->done && timeout);
4138 __remove_wait_queue(&x->wait, &wait);
4143 return timeout ?: 1;
4147 wait_for_common(struct completion *x, long timeout, int state)
4151 spin_lock_irq(&x->wait.lock);
4152 timeout = do_wait_for_common(x, timeout, state);
4153 spin_unlock_irq(&x->wait.lock);
4158 * wait_for_completion: - waits for completion of a task
4159 * @x: holds the state of this particular completion
4161 * This waits to be signaled for completion of a specific task. It is NOT
4162 * interruptible and there is no timeout.
4164 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4165 * and interrupt capability. Also see complete().
4167 void __sched wait_for_completion(struct completion *x)
4169 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4171 EXPORT_SYMBOL(wait_for_completion);
4174 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4175 * @x: holds the state of this particular completion
4176 * @timeout: timeout value in jiffies
4178 * This waits for either a completion of a specific task to be signaled or for a
4179 * specified timeout to expire. The timeout is in jiffies. It is not
4182 unsigned long __sched
4183 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4185 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4187 EXPORT_SYMBOL(wait_for_completion_timeout);
4190 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4191 * @x: holds the state of this particular completion
4193 * This waits for completion of a specific task to be signaled. It is
4196 int __sched wait_for_completion_interruptible(struct completion *x)
4198 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4199 if (t == -ERESTARTSYS)
4203 EXPORT_SYMBOL(wait_for_completion_interruptible);
4206 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4207 * @x: holds the state of this particular completion
4208 * @timeout: timeout value in jiffies
4210 * This waits for either a completion of a specific task to be signaled or for a
4211 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4213 unsigned long __sched
4214 wait_for_completion_interruptible_timeout(struct completion *x,
4215 unsigned long timeout)
4217 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4219 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4222 * wait_for_completion_killable: - waits for completion of a task (killable)
4223 * @x: holds the state of this particular completion
4225 * This waits to be signaled for completion of a specific task. It can be
4226 * interrupted by a kill signal.
4228 int __sched wait_for_completion_killable(struct completion *x)
4230 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4231 if (t == -ERESTARTSYS)
4235 EXPORT_SYMBOL(wait_for_completion_killable);
4238 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4239 * @x: holds the state of this particular completion
4240 * @timeout: timeout value in jiffies
4242 * This waits for either a completion of a specific task to be
4243 * signaled or for a specified timeout to expire. It can be
4244 * interrupted by a kill signal. The timeout is in jiffies.
4246 unsigned long __sched
4247 wait_for_completion_killable_timeout(struct completion *x,
4248 unsigned long timeout)
4250 return wait_for_common(x, timeout, TASK_KILLABLE);
4252 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
4255 * try_wait_for_completion - try to decrement a completion without blocking
4256 * @x: completion structure
4258 * Returns: 0 if a decrement cannot be done without blocking
4259 * 1 if a decrement succeeded.
4261 * If a completion is being used as a counting completion,
4262 * attempt to decrement the counter without blocking. This
4263 * enables us to avoid waiting if the resource the completion
4264 * is protecting is not available.
4266 bool try_wait_for_completion(struct completion *x)
4268 unsigned long flags;
4271 spin_lock_irqsave(&x->wait.lock, flags);
4276 spin_unlock_irqrestore(&x->wait.lock, flags);
4279 EXPORT_SYMBOL(try_wait_for_completion);
4282 * completion_done - Test to see if a completion has any waiters
4283 * @x: completion structure
4285 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4286 * 1 if there are no waiters.
4289 bool completion_done(struct completion *x)
4291 unsigned long flags;
4294 spin_lock_irqsave(&x->wait.lock, flags);
4297 spin_unlock_irqrestore(&x->wait.lock, flags);
4300 EXPORT_SYMBOL(completion_done);
4303 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4305 unsigned long flags;
4308 init_waitqueue_entry(&wait, current);
4310 __set_current_state(state);
4312 spin_lock_irqsave(&q->lock, flags);
4313 __add_wait_queue(q, &wait);
4314 spin_unlock(&q->lock);
4315 timeout = schedule_timeout(timeout);
4316 spin_lock_irq(&q->lock);
4317 __remove_wait_queue(q, &wait);
4318 spin_unlock_irqrestore(&q->lock, flags);
4323 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4325 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4327 EXPORT_SYMBOL(interruptible_sleep_on);
4330 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4332 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4334 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4336 void __sched sleep_on(wait_queue_head_t *q)
4338 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4340 EXPORT_SYMBOL(sleep_on);
4342 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4344 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4346 EXPORT_SYMBOL(sleep_on_timeout);
4348 #ifdef CONFIG_RT_MUTEXES
4351 * rt_mutex_setprio - set the current priority of a task
4353 * @prio: prio value (kernel-internal form)
4355 * This function changes the 'effective' priority of a task. It does
4356 * not touch ->normal_prio like __setscheduler().
4358 * Used by the rt_mutex code to implement priority inheritance logic.
4360 void rt_mutex_setprio(struct task_struct *p, int prio)
4362 unsigned long flags;
4363 int oldprio, on_rq, running;
4365 const struct sched_class *prev_class;
4367 BUG_ON(prio < 0 || prio > MAX_PRIO);
4369 rq = task_rq_lock(p, &flags);
4371 trace_sched_pi_setprio(p, prio);
4373 prev_class = p->sched_class;
4374 on_rq = p->se.on_rq;
4375 running = task_current(rq, p);
4377 dequeue_task(rq, p, 0);
4379 p->sched_class->put_prev_task(rq, p);
4382 p->sched_class = &rt_sched_class;
4384 p->sched_class = &fair_sched_class;
4389 p->sched_class->set_curr_task(rq);
4391 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4393 check_class_changed(rq, p, prev_class, oldprio, running);
4395 task_rq_unlock(rq, &flags);
4400 void set_user_nice(struct task_struct *p, long nice)
4402 int old_prio, delta, on_rq;
4403 unsigned long flags;
4406 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4409 * We have to be careful, if called from sys_setpriority(),
4410 * the task might be in the middle of scheduling on another CPU.
4412 rq = task_rq_lock(p, &flags);
4414 * The RT priorities are set via sched_setscheduler(), but we still
4415 * allow the 'normal' nice value to be set - but as expected
4416 * it wont have any effect on scheduling until the task is
4417 * SCHED_FIFO/SCHED_RR:
4419 if (task_has_rt_policy(p)) {
4420 p->static_prio = NICE_TO_PRIO(nice);
4423 on_rq = p->se.on_rq;
4425 dequeue_task(rq, p, 0);
4427 p->static_prio = NICE_TO_PRIO(nice);
4430 p->prio = effective_prio(p);
4431 delta = p->prio - old_prio;
4434 enqueue_task(rq, p, 0);
4436 * If the task increased its priority or is running and
4437 * lowered its priority, then reschedule its CPU:
4439 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4440 resched_task(rq->curr);
4443 task_rq_unlock(rq, &flags);
4445 EXPORT_SYMBOL(set_user_nice);
4448 * can_nice - check if a task can reduce its nice value
4452 int can_nice(const struct task_struct *p, const int nice)
4454 /* convert nice value [19,-20] to rlimit style value [1,40] */
4455 int nice_rlim = 20 - nice;
4457 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4458 capable(CAP_SYS_NICE));
4461 #ifdef __ARCH_WANT_SYS_NICE
4464 * sys_nice - change the priority of the current process.
4465 * @increment: priority increment
4467 * sys_setpriority is a more generic, but much slower function that
4468 * does similar things.
4470 SYSCALL_DEFINE1(nice, int, increment)
4475 * Setpriority might change our priority at the same moment.
4476 * We don't have to worry. Conceptually one call occurs first
4477 * and we have a single winner.
4479 if (increment < -40)
4484 nice = TASK_NICE(current) + increment;
4490 if (increment < 0 && !can_nice(current, nice))
4493 retval = security_task_setnice(current, nice);
4497 set_user_nice(current, nice);
4504 * task_prio - return the priority value of a given task.
4505 * @p: the task in question.
4507 * This is the priority value as seen by users in /proc.
4508 * RT tasks are offset by -200. Normal tasks are centered
4509 * around 0, value goes from -16 to +15.
4511 int task_prio(const struct task_struct *p)
4513 return p->prio - MAX_RT_PRIO;
4517 * task_nice - return the nice value of a given task.
4518 * @p: the task in question.
4520 int task_nice(const struct task_struct *p)
4522 return TASK_NICE(p);
4524 EXPORT_SYMBOL(task_nice);
4527 * idle_cpu - is a given cpu idle currently?
4528 * @cpu: the processor in question.
4530 int idle_cpu(int cpu)
4532 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4536 * idle_task - return the idle task for a given cpu.
4537 * @cpu: the processor in question.
4539 struct task_struct *idle_task(int cpu)
4541 return cpu_rq(cpu)->idle;
4545 * find_process_by_pid - find a process with a matching PID value.
4546 * @pid: the pid in question.
4548 static struct task_struct *find_process_by_pid(pid_t pid)
4550 return pid ? find_task_by_vpid(pid) : current;
4553 /* Actually do priority change: must hold rq lock. */
4555 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4557 BUG_ON(p->se.on_rq);
4560 p->rt_priority = prio;
4561 p->normal_prio = normal_prio(p);
4562 /* we are holding p->pi_lock already */
4563 p->prio = rt_mutex_getprio(p);
4564 if (rt_prio(p->prio))
4565 p->sched_class = &rt_sched_class;
4567 p->sched_class = &fair_sched_class;
4572 * check the target process has a UID that matches the current process's
4574 static bool check_same_owner(struct task_struct *p)
4576 const struct cred *cred = current_cred(), *pcred;
4580 pcred = __task_cred(p);
4581 match = (cred->euid == pcred->euid ||
4582 cred->euid == pcred->uid);
4587 static int __sched_setscheduler(struct task_struct *p, int policy,
4588 const struct sched_param *param, bool user)
4590 int retval, oldprio, oldpolicy = -1, on_rq, running;
4591 unsigned long flags;
4592 const struct sched_class *prev_class;
4596 /* may grab non-irq protected spin_locks */
4597 BUG_ON(in_interrupt());
4599 /* double check policy once rq lock held */
4601 reset_on_fork = p->sched_reset_on_fork;
4602 policy = oldpolicy = p->policy;
4604 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4605 policy &= ~SCHED_RESET_ON_FORK;
4607 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4608 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4609 policy != SCHED_IDLE)
4614 * Valid priorities for SCHED_FIFO and SCHED_RR are
4615 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4616 * SCHED_BATCH and SCHED_IDLE is 0.
4618 if (param->sched_priority < 0 ||
4619 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4620 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4622 if (rt_policy(policy) != (param->sched_priority != 0))
4626 * Allow unprivileged RT tasks to decrease priority:
4628 if (user && !capable(CAP_SYS_NICE)) {
4629 if (rt_policy(policy)) {
4630 unsigned long rlim_rtprio =
4631 task_rlimit(p, RLIMIT_RTPRIO);
4633 /* can't set/change the rt policy */
4634 if (policy != p->policy && !rlim_rtprio)
4637 /* can't increase priority */
4638 if (param->sched_priority > p->rt_priority &&
4639 param->sched_priority > rlim_rtprio)
4643 * Like positive nice levels, dont allow tasks to
4644 * move out of SCHED_IDLE either:
4646 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4649 /* can't change other user's priorities */
4650 if (!check_same_owner(p))
4653 /* Normal users shall not reset the sched_reset_on_fork flag */
4654 if (p->sched_reset_on_fork && !reset_on_fork)
4659 retval = security_task_setscheduler(p);
4665 * make sure no PI-waiters arrive (or leave) while we are
4666 * changing the priority of the task:
4668 raw_spin_lock_irqsave(&p->pi_lock, flags);
4670 * To be able to change p->policy safely, the apropriate
4671 * runqueue lock must be held.
4673 rq = __task_rq_lock(p);
4676 * Changing the policy of the stop threads its a very bad idea
4678 if (p == rq->stop) {
4679 __task_rq_unlock(rq);
4680 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4684 #ifdef CONFIG_RT_GROUP_SCHED
4687 * Do not allow realtime tasks into groups that have no runtime
4690 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4691 task_group(p)->rt_bandwidth.rt_runtime == 0) {
4692 __task_rq_unlock(rq);
4693 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4699 /* recheck policy now with rq lock held */
4700 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4701 policy = oldpolicy = -1;
4702 __task_rq_unlock(rq);
4703 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4706 on_rq = p->se.on_rq;
4707 running = task_current(rq, p);
4709 deactivate_task(rq, p, 0);
4711 p->sched_class->put_prev_task(rq, p);
4713 p->sched_reset_on_fork = reset_on_fork;
4716 prev_class = p->sched_class;
4717 __setscheduler(rq, p, policy, param->sched_priority);
4720 p->sched_class->set_curr_task(rq);
4722 activate_task(rq, p, 0);
4724 check_class_changed(rq, p, prev_class, oldprio, running);
4726 __task_rq_unlock(rq);
4727 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4729 rt_mutex_adjust_pi(p);
4735 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4736 * @p: the task in question.
4737 * @policy: new policy.
4738 * @param: structure containing the new RT priority.
4740 * NOTE that the task may be already dead.
4742 int sched_setscheduler(struct task_struct *p, int policy,
4743 const struct sched_param *param)
4745 return __sched_setscheduler(p, policy, param, true);
4747 EXPORT_SYMBOL_GPL(sched_setscheduler);
4750 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4751 * @p: the task in question.
4752 * @policy: new policy.
4753 * @param: structure containing the new RT priority.
4755 * Just like sched_setscheduler, only don't bother checking if the
4756 * current context has permission. For example, this is needed in
4757 * stop_machine(): we create temporary high priority worker threads,
4758 * but our caller might not have that capability.
4760 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4761 const struct sched_param *param)
4763 return __sched_setscheduler(p, policy, param, false);
4767 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4769 struct sched_param lparam;
4770 struct task_struct *p;
4773 if (!param || pid < 0)
4775 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4780 p = find_process_by_pid(pid);
4782 retval = sched_setscheduler(p, policy, &lparam);
4789 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4790 * @pid: the pid in question.
4791 * @policy: new policy.
4792 * @param: structure containing the new RT priority.
4794 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4795 struct sched_param __user *, param)
4797 /* negative values for policy are not valid */
4801 return do_sched_setscheduler(pid, policy, param);
4805 * sys_sched_setparam - set/change the RT priority of a thread
4806 * @pid: the pid in question.
4807 * @param: structure containing the new RT priority.
4809 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4811 return do_sched_setscheduler(pid, -1, param);
4815 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4816 * @pid: the pid in question.
4818 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4820 struct task_struct *p;
4828 p = find_process_by_pid(pid);
4830 retval = security_task_getscheduler(p);
4833 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4840 * sys_sched_getparam - get the RT priority of a thread
4841 * @pid: the pid in question.
4842 * @param: structure containing the RT priority.
4844 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4846 struct sched_param lp;
4847 struct task_struct *p;
4850 if (!param || pid < 0)
4854 p = find_process_by_pid(pid);
4859 retval = security_task_getscheduler(p);
4863 lp.sched_priority = p->rt_priority;
4867 * This one might sleep, we cannot do it with a spinlock held ...
4869 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4878 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4880 cpumask_var_t cpus_allowed, new_mask;
4881 struct task_struct *p;
4887 p = find_process_by_pid(pid);
4894 /* Prevent p going away */
4898 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4902 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4904 goto out_free_cpus_allowed;
4907 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
4910 retval = security_task_setscheduler(p);
4914 cpuset_cpus_allowed(p, cpus_allowed);
4915 cpumask_and(new_mask, in_mask, cpus_allowed);
4917 retval = set_cpus_allowed_ptr(p, new_mask);
4920 cpuset_cpus_allowed(p, cpus_allowed);
4921 if (!cpumask_subset(new_mask, cpus_allowed)) {
4923 * We must have raced with a concurrent cpuset
4924 * update. Just reset the cpus_allowed to the
4925 * cpuset's cpus_allowed
4927 cpumask_copy(new_mask, cpus_allowed);
4932 free_cpumask_var(new_mask);
4933 out_free_cpus_allowed:
4934 free_cpumask_var(cpus_allowed);
4941 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4942 struct cpumask *new_mask)
4944 if (len < cpumask_size())
4945 cpumask_clear(new_mask);
4946 else if (len > cpumask_size())
4947 len = cpumask_size();
4949 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4953 * sys_sched_setaffinity - set the cpu affinity of a process
4954 * @pid: pid of the process
4955 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4956 * @user_mask_ptr: user-space pointer to the new cpu mask
4958 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4959 unsigned long __user *, user_mask_ptr)
4961 cpumask_var_t new_mask;
4964 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4967 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4969 retval = sched_setaffinity(pid, new_mask);
4970 free_cpumask_var(new_mask);
4974 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4976 struct task_struct *p;
4977 unsigned long flags;
4985 p = find_process_by_pid(pid);
4989 retval = security_task_getscheduler(p);
4993 rq = task_rq_lock(p, &flags);
4994 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4995 task_rq_unlock(rq, &flags);
5005 * sys_sched_getaffinity - get the cpu affinity of a process
5006 * @pid: pid of the process
5007 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5008 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5010 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5011 unsigned long __user *, user_mask_ptr)
5016 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5018 if (len & (sizeof(unsigned long)-1))
5021 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5024 ret = sched_getaffinity(pid, mask);
5026 size_t retlen = min_t(size_t, len, cpumask_size());
5028 if (copy_to_user(user_mask_ptr, mask, retlen))
5033 free_cpumask_var(mask);
5039 * sys_sched_yield - yield the current processor to other threads.
5041 * This function yields the current CPU to other tasks. If there are no
5042 * other threads running on this CPU then this function will return.
5044 SYSCALL_DEFINE0(sched_yield)
5046 struct rq *rq = this_rq_lock();
5048 schedstat_inc(rq, yld_count);
5049 current->sched_class->yield_task(rq);
5052 * Since we are going to call schedule() anyway, there's
5053 * no need to preempt or enable interrupts:
5055 __release(rq->lock);
5056 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5057 do_raw_spin_unlock(&rq->lock);
5058 preempt_enable_no_resched();
5065 static inline int should_resched(void)
5067 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5070 static void __cond_resched(void)
5072 add_preempt_count(PREEMPT_ACTIVE);
5074 sub_preempt_count(PREEMPT_ACTIVE);
5077 int __sched _cond_resched(void)
5079 if (should_resched()) {
5085 EXPORT_SYMBOL(_cond_resched);
5088 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5089 * call schedule, and on return reacquire the lock.
5091 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5092 * operations here to prevent schedule() from being called twice (once via
5093 * spin_unlock(), once by hand).
5095 int __cond_resched_lock(spinlock_t *lock)
5097 int resched = should_resched();
5100 lockdep_assert_held(lock);
5102 if (spin_needbreak(lock) || resched) {
5113 EXPORT_SYMBOL(__cond_resched_lock);
5115 int __sched __cond_resched_softirq(void)
5117 BUG_ON(!in_softirq());
5119 if (should_resched()) {
5127 EXPORT_SYMBOL(__cond_resched_softirq);
5130 * yield - yield the current processor to other threads.
5132 * This is a shortcut for kernel-space yielding - it marks the
5133 * thread runnable and calls sys_sched_yield().
5135 void __sched yield(void)
5137 set_current_state(TASK_RUNNING);
5140 EXPORT_SYMBOL(yield);
5143 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5144 * that process accounting knows that this is a task in IO wait state.
5146 void __sched io_schedule(void)
5148 struct rq *rq = raw_rq();
5150 delayacct_blkio_start();
5151 atomic_inc(&rq->nr_iowait);
5152 current->in_iowait = 1;
5154 current->in_iowait = 0;
5155 atomic_dec(&rq->nr_iowait);
5156 delayacct_blkio_end();
5158 EXPORT_SYMBOL(io_schedule);
5160 long __sched io_schedule_timeout(long timeout)
5162 struct rq *rq = raw_rq();
5165 delayacct_blkio_start();
5166 atomic_inc(&rq->nr_iowait);
5167 current->in_iowait = 1;
5168 ret = schedule_timeout(timeout);
5169 current->in_iowait = 0;
5170 atomic_dec(&rq->nr_iowait);
5171 delayacct_blkio_end();
5176 * sys_sched_get_priority_max - return maximum RT priority.
5177 * @policy: scheduling class.
5179 * this syscall returns the maximum rt_priority that can be used
5180 * by a given scheduling class.
5182 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5189 ret = MAX_USER_RT_PRIO-1;
5201 * sys_sched_get_priority_min - return minimum RT priority.
5202 * @policy: scheduling class.
5204 * this syscall returns the minimum rt_priority that can be used
5205 * by a given scheduling class.
5207 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5225 * sys_sched_rr_get_interval - return the default timeslice of a process.
5226 * @pid: pid of the process.
5227 * @interval: userspace pointer to the timeslice value.
5229 * this syscall writes the default timeslice value of a given process
5230 * into the user-space timespec buffer. A value of '0' means infinity.
5232 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5233 struct timespec __user *, interval)
5235 struct task_struct *p;
5236 unsigned int time_slice;
5237 unsigned long flags;
5247 p = find_process_by_pid(pid);
5251 retval = security_task_getscheduler(p);
5255 rq = task_rq_lock(p, &flags);
5256 time_slice = p->sched_class->get_rr_interval(rq, p);
5257 task_rq_unlock(rq, &flags);
5260 jiffies_to_timespec(time_slice, &t);
5261 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5269 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5271 void sched_show_task(struct task_struct *p)
5273 unsigned long free = 0;
5276 state = p->state ? __ffs(p->state) + 1 : 0;
5277 printk(KERN_INFO "%-15.15s %c", p->comm,
5278 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5279 #if BITS_PER_LONG == 32
5280 if (state == TASK_RUNNING)
5281 printk(KERN_CONT " running ");
5283 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5285 if (state == TASK_RUNNING)
5286 printk(KERN_CONT " running task ");
5288 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5290 #ifdef CONFIG_DEBUG_STACK_USAGE
5291 free = stack_not_used(p);
5293 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5294 task_pid_nr(p), task_pid_nr(p->real_parent),
5295 (unsigned long)task_thread_info(p)->flags);
5297 show_stack(p, NULL);
5300 void show_state_filter(unsigned long state_filter)
5302 struct task_struct *g, *p;
5304 #if BITS_PER_LONG == 32
5306 " task PC stack pid father\n");
5309 " task PC stack pid father\n");
5311 read_lock(&tasklist_lock);
5312 do_each_thread(g, p) {
5314 * reset the NMI-timeout, listing all files on a slow
5315 * console might take alot of time:
5317 touch_nmi_watchdog();
5318 if (!state_filter || (p->state & state_filter))
5320 } while_each_thread(g, p);
5322 touch_all_softlockup_watchdogs();
5324 #ifdef CONFIG_SCHED_DEBUG
5325 sysrq_sched_debug_show();
5327 read_unlock(&tasklist_lock);
5329 * Only show locks if all tasks are dumped:
5332 debug_show_all_locks();
5335 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5337 idle->sched_class = &idle_sched_class;
5341 * init_idle - set up an idle thread for a given CPU
5342 * @idle: task in question
5343 * @cpu: cpu the idle task belongs to
5345 * NOTE: this function does not set the idle thread's NEED_RESCHED
5346 * flag, to make booting more robust.
5348 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5350 struct rq *rq = cpu_rq(cpu);
5351 unsigned long flags;
5353 raw_spin_lock_irqsave(&rq->lock, flags);
5356 idle->state = TASK_RUNNING;
5357 idle->se.exec_start = sched_clock();
5359 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
5361 * We're having a chicken and egg problem, even though we are
5362 * holding rq->lock, the cpu isn't yet set to this cpu so the
5363 * lockdep check in task_group() will fail.
5365 * Similar case to sched_fork(). / Alternatively we could
5366 * use task_rq_lock() here and obtain the other rq->lock.
5371 __set_task_cpu(idle, cpu);
5374 rq->curr = rq->idle = idle;
5375 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5378 raw_spin_unlock_irqrestore(&rq->lock, flags);
5380 /* Set the preempt count _outside_ the spinlocks! */
5381 #if defined(CONFIG_PREEMPT)
5382 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5384 task_thread_info(idle)->preempt_count = 0;
5387 * The idle tasks have their own, simple scheduling class:
5389 idle->sched_class = &idle_sched_class;
5390 ftrace_graph_init_task(idle);
5394 * In a system that switches off the HZ timer nohz_cpu_mask
5395 * indicates which cpus entered this state. This is used
5396 * in the rcu update to wait only for active cpus. For system
5397 * which do not switch off the HZ timer nohz_cpu_mask should
5398 * always be CPU_BITS_NONE.
5400 cpumask_var_t nohz_cpu_mask;
5403 * Increase the granularity value when there are more CPUs,
5404 * because with more CPUs the 'effective latency' as visible
5405 * to users decreases. But the relationship is not linear,
5406 * so pick a second-best guess by going with the log2 of the
5409 * This idea comes from the SD scheduler of Con Kolivas:
5411 static int get_update_sysctl_factor(void)
5413 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5414 unsigned int factor;
5416 switch (sysctl_sched_tunable_scaling) {
5417 case SCHED_TUNABLESCALING_NONE:
5420 case SCHED_TUNABLESCALING_LINEAR:
5423 case SCHED_TUNABLESCALING_LOG:
5425 factor = 1 + ilog2(cpus);
5432 static void update_sysctl(void)
5434 unsigned int factor = get_update_sysctl_factor();
5436 #define SET_SYSCTL(name) \
5437 (sysctl_##name = (factor) * normalized_sysctl_##name)
5438 SET_SYSCTL(sched_min_granularity);
5439 SET_SYSCTL(sched_latency);
5440 SET_SYSCTL(sched_wakeup_granularity);
5444 static inline void sched_init_granularity(void)
5451 * This is how migration works:
5453 * 1) we invoke migration_cpu_stop() on the target CPU using
5455 * 2) stopper starts to run (implicitly forcing the migrated thread
5457 * 3) it checks whether the migrated task is still in the wrong runqueue.
5458 * 4) if it's in the wrong runqueue then the migration thread removes
5459 * it and puts it into the right queue.
5460 * 5) stopper completes and stop_one_cpu() returns and the migration
5465 * Change a given task's CPU affinity. Migrate the thread to a
5466 * proper CPU and schedule it away if the CPU it's executing on
5467 * is removed from the allowed bitmask.
5469 * NOTE: the caller must have a valid reference to the task, the
5470 * task must not exit() & deallocate itself prematurely. The
5471 * call is not atomic; no spinlocks may be held.
5473 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5475 unsigned long flags;
5477 unsigned int dest_cpu;
5481 * Serialize against TASK_WAKING so that ttwu() and wunt() can
5482 * drop the rq->lock and still rely on ->cpus_allowed.
5485 while (task_is_waking(p))
5487 rq = task_rq_lock(p, &flags);
5488 if (task_is_waking(p)) {
5489 task_rq_unlock(rq, &flags);
5493 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5498 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5499 !cpumask_equal(&p->cpus_allowed, new_mask))) {
5504 if (p->sched_class->set_cpus_allowed)
5505 p->sched_class->set_cpus_allowed(p, new_mask);
5507 cpumask_copy(&p->cpus_allowed, new_mask);
5508 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5511 /* Can the task run on the task's current CPU? If so, we're done */
5512 if (cpumask_test_cpu(task_cpu(p), new_mask))
5515 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
5516 if (migrate_task(p, rq)) {
5517 struct migration_arg arg = { p, dest_cpu };
5518 /* Need help from migration thread: drop lock and wait. */
5519 task_rq_unlock(rq, &flags);
5520 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
5521 tlb_migrate_finish(p->mm);
5525 task_rq_unlock(rq, &flags);
5529 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5532 * Move (not current) task off this cpu, onto dest cpu. We're doing
5533 * this because either it can't run here any more (set_cpus_allowed()
5534 * away from this CPU, or CPU going down), or because we're
5535 * attempting to rebalance this task on exec (sched_exec).
5537 * So we race with normal scheduler movements, but that's OK, as long
5538 * as the task is no longer on this CPU.
5540 * Returns non-zero if task was successfully migrated.
5542 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5544 struct rq *rq_dest, *rq_src;
5547 if (unlikely(!cpu_active(dest_cpu)))
5550 rq_src = cpu_rq(src_cpu);
5551 rq_dest = cpu_rq(dest_cpu);
5553 double_rq_lock(rq_src, rq_dest);
5554 /* Already moved. */
5555 if (task_cpu(p) != src_cpu)
5557 /* Affinity changed (again). */
5558 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
5562 * If we're not on a rq, the next wake-up will ensure we're
5566 deactivate_task(rq_src, p, 0);
5567 set_task_cpu(p, dest_cpu);
5568 activate_task(rq_dest, p, 0);
5569 check_preempt_curr(rq_dest, p, 0);
5574 double_rq_unlock(rq_src, rq_dest);
5579 * migration_cpu_stop - this will be executed by a highprio stopper thread
5580 * and performs thread migration by bumping thread off CPU then
5581 * 'pushing' onto another runqueue.
5583 static int migration_cpu_stop(void *data)
5585 struct migration_arg *arg = data;
5588 * The original target cpu might have gone down and we might
5589 * be on another cpu but it doesn't matter.
5591 local_irq_disable();
5592 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
5597 #ifdef CONFIG_HOTPLUG_CPU
5600 * Ensures that the idle task is using init_mm right before its cpu goes
5603 void idle_task_exit(void)
5605 struct mm_struct *mm = current->active_mm;
5607 BUG_ON(cpu_online(smp_processor_id()));
5610 switch_mm(mm, &init_mm, current);
5615 * While a dead CPU has no uninterruptible tasks queued at this point,
5616 * it might still have a nonzero ->nr_uninterruptible counter, because
5617 * for performance reasons the counter is not stricly tracking tasks to
5618 * their home CPUs. So we just add the counter to another CPU's counter,
5619 * to keep the global sum constant after CPU-down:
5621 static void migrate_nr_uninterruptible(struct rq *rq_src)
5623 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5625 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5626 rq_src->nr_uninterruptible = 0;
5630 * remove the tasks which were accounted by rq from calc_load_tasks.
5632 static void calc_global_load_remove(struct rq *rq)
5634 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5635 rq->calc_load_active = 0;
5639 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5640 * try_to_wake_up()->select_task_rq().
5642 * Called with rq->lock held even though we'er in stop_machine() and
5643 * there's no concurrency possible, we hold the required locks anyway
5644 * because of lock validation efforts.
5646 static void migrate_tasks(unsigned int dead_cpu)
5648 struct rq *rq = cpu_rq(dead_cpu);
5649 struct task_struct *next, *stop = rq->stop;
5653 * Fudge the rq selection such that the below task selection loop
5654 * doesn't get stuck on the currently eligible stop task.
5656 * We're currently inside stop_machine() and the rq is either stuck
5657 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5658 * either way we should never end up calling schedule() until we're
5665 * There's this thread running, bail when that's the only
5668 if (rq->nr_running == 1)
5671 next = pick_next_task(rq);
5673 next->sched_class->put_prev_task(rq, next);
5675 /* Find suitable destination for @next, with force if needed. */
5676 dest_cpu = select_fallback_rq(dead_cpu, next);
5677 raw_spin_unlock(&rq->lock);
5679 __migrate_task(next, dead_cpu, dest_cpu);
5681 raw_spin_lock(&rq->lock);
5687 #endif /* CONFIG_HOTPLUG_CPU */
5689 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5691 static struct ctl_table sd_ctl_dir[] = {
5693 .procname = "sched_domain",
5699 static struct ctl_table sd_ctl_root[] = {
5701 .procname = "kernel",
5703 .child = sd_ctl_dir,
5708 static struct ctl_table *sd_alloc_ctl_entry(int n)
5710 struct ctl_table *entry =
5711 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5716 static void sd_free_ctl_entry(struct ctl_table **tablep)
5718 struct ctl_table *entry;
5721 * In the intermediate directories, both the child directory and
5722 * procname are dynamically allocated and could fail but the mode
5723 * will always be set. In the lowest directory the names are
5724 * static strings and all have proc handlers.
5726 for (entry = *tablep; entry->mode; entry++) {
5728 sd_free_ctl_entry(&entry->child);
5729 if (entry->proc_handler == NULL)
5730 kfree(entry->procname);
5738 set_table_entry(struct ctl_table *entry,
5739 const char *procname, void *data, int maxlen,
5740 mode_t mode, proc_handler *proc_handler)
5742 entry->procname = procname;
5744 entry->maxlen = maxlen;
5746 entry->proc_handler = proc_handler;
5749 static struct ctl_table *
5750 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5752 struct ctl_table *table = sd_alloc_ctl_entry(13);
5757 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5758 sizeof(long), 0644, proc_doulongvec_minmax);
5759 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5760 sizeof(long), 0644, proc_doulongvec_minmax);
5761 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5762 sizeof(int), 0644, proc_dointvec_minmax);
5763 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5764 sizeof(int), 0644, proc_dointvec_minmax);
5765 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5766 sizeof(int), 0644, proc_dointvec_minmax);
5767 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5768 sizeof(int), 0644, proc_dointvec_minmax);
5769 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5770 sizeof(int), 0644, proc_dointvec_minmax);
5771 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5772 sizeof(int), 0644, proc_dointvec_minmax);
5773 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5774 sizeof(int), 0644, proc_dointvec_minmax);
5775 set_table_entry(&table[9], "cache_nice_tries",
5776 &sd->cache_nice_tries,
5777 sizeof(int), 0644, proc_dointvec_minmax);
5778 set_table_entry(&table[10], "flags", &sd->flags,
5779 sizeof(int), 0644, proc_dointvec_minmax);
5780 set_table_entry(&table[11], "name", sd->name,
5781 CORENAME_MAX_SIZE, 0444, proc_dostring);
5782 /* &table[12] is terminator */
5787 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5789 struct ctl_table *entry, *table;
5790 struct sched_domain *sd;
5791 int domain_num = 0, i;
5794 for_each_domain(cpu, sd)
5796 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5801 for_each_domain(cpu, sd) {
5802 snprintf(buf, 32, "domain%d", i);
5803 entry->procname = kstrdup(buf, GFP_KERNEL);
5805 entry->child = sd_alloc_ctl_domain_table(sd);
5812 static struct ctl_table_header *sd_sysctl_header;
5813 static void register_sched_domain_sysctl(void)
5815 int i, cpu_num = num_possible_cpus();
5816 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5819 WARN_ON(sd_ctl_dir[0].child);
5820 sd_ctl_dir[0].child = entry;
5825 for_each_possible_cpu(i) {
5826 snprintf(buf, 32, "cpu%d", i);
5827 entry->procname = kstrdup(buf, GFP_KERNEL);
5829 entry->child = sd_alloc_ctl_cpu_table(i);
5833 WARN_ON(sd_sysctl_header);
5834 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5837 /* may be called multiple times per register */
5838 static void unregister_sched_domain_sysctl(void)
5840 if (sd_sysctl_header)
5841 unregister_sysctl_table(sd_sysctl_header);
5842 sd_sysctl_header = NULL;
5843 if (sd_ctl_dir[0].child)
5844 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5847 static void register_sched_domain_sysctl(void)
5850 static void unregister_sched_domain_sysctl(void)
5855 static void set_rq_online(struct rq *rq)
5858 const struct sched_class *class;
5860 cpumask_set_cpu(rq->cpu, rq->rd->online);
5863 for_each_class(class) {
5864 if (class->rq_online)
5865 class->rq_online(rq);
5870 static void set_rq_offline(struct rq *rq)
5873 const struct sched_class *class;
5875 for_each_class(class) {
5876 if (class->rq_offline)
5877 class->rq_offline(rq);
5880 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5886 * migration_call - callback that gets triggered when a CPU is added.
5887 * Here we can start up the necessary migration thread for the new CPU.
5889 static int __cpuinit
5890 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5892 int cpu = (long)hcpu;
5893 unsigned long flags;
5894 struct rq *rq = cpu_rq(cpu);
5896 switch (action & ~CPU_TASKS_FROZEN) {
5898 case CPU_UP_PREPARE:
5899 rq->calc_load_update = calc_load_update;
5903 /* Update our root-domain */
5904 raw_spin_lock_irqsave(&rq->lock, flags);
5906 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5910 raw_spin_unlock_irqrestore(&rq->lock, flags);
5913 #ifdef CONFIG_HOTPLUG_CPU
5915 /* Update our root-domain */
5916 raw_spin_lock_irqsave(&rq->lock, flags);
5918 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5922 BUG_ON(rq->nr_running != 1); /* the migration thread */
5923 raw_spin_unlock_irqrestore(&rq->lock, flags);
5925 migrate_nr_uninterruptible(rq);
5926 calc_global_load_remove(rq);
5934 * Register at high priority so that task migration (migrate_all_tasks)
5935 * happens before everything else. This has to be lower priority than
5936 * the notifier in the perf_event subsystem, though.
5938 static struct notifier_block __cpuinitdata migration_notifier = {
5939 .notifier_call = migration_call,
5940 .priority = CPU_PRI_MIGRATION,
5943 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
5944 unsigned long action, void *hcpu)
5946 switch (action & ~CPU_TASKS_FROZEN) {
5948 case CPU_DOWN_FAILED:
5949 set_cpu_active((long)hcpu, true);
5956 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
5957 unsigned long action, void *hcpu)
5959 switch (action & ~CPU_TASKS_FROZEN) {
5960 case CPU_DOWN_PREPARE:
5961 set_cpu_active((long)hcpu, false);
5968 static int __init migration_init(void)
5970 void *cpu = (void *)(long)smp_processor_id();
5973 /* Initialize migration for the boot CPU */
5974 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5975 BUG_ON(err == NOTIFY_BAD);
5976 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5977 register_cpu_notifier(&migration_notifier);
5979 /* Register cpu active notifiers */
5980 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5981 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5985 early_initcall(migration_init);
5990 #ifdef CONFIG_SCHED_DEBUG
5992 static __read_mostly int sched_domain_debug_enabled;
5994 static int __init sched_domain_debug_setup(char *str)
5996 sched_domain_debug_enabled = 1;
6000 early_param("sched_debug", sched_domain_debug_setup);
6002 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6003 struct cpumask *groupmask)
6005 struct sched_group *group = sd->groups;
6008 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6009 cpumask_clear(groupmask);
6011 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6013 if (!(sd->flags & SD_LOAD_BALANCE)) {
6014 printk("does not load-balance\n");
6016 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6021 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6023 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6024 printk(KERN_ERR "ERROR: domain->span does not contain "
6027 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6028 printk(KERN_ERR "ERROR: domain->groups does not contain"
6032 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6036 printk(KERN_ERR "ERROR: group is NULL\n");
6040 if (!group->cpu_power) {
6041 printk(KERN_CONT "\n");
6042 printk(KERN_ERR "ERROR: domain->cpu_power not "
6047 if (!cpumask_weight(sched_group_cpus(group))) {
6048 printk(KERN_CONT "\n");
6049 printk(KERN_ERR "ERROR: empty group\n");
6053 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6054 printk(KERN_CONT "\n");
6055 printk(KERN_ERR "ERROR: repeated CPUs\n");
6059 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6061 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6063 printk(KERN_CONT " %s", str);
6064 if (group->cpu_power != SCHED_LOAD_SCALE) {
6065 printk(KERN_CONT " (cpu_power = %d)",
6069 group = group->next;
6070 } while (group != sd->groups);
6071 printk(KERN_CONT "\n");
6073 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6074 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6077 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6078 printk(KERN_ERR "ERROR: parent span is not a superset "
6079 "of domain->span\n");
6083 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6085 cpumask_var_t groupmask;
6088 if (!sched_domain_debug_enabled)
6092 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6096 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6098 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6099 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6104 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6111 free_cpumask_var(groupmask);
6113 #else /* !CONFIG_SCHED_DEBUG */
6114 # define sched_domain_debug(sd, cpu) do { } while (0)
6115 #endif /* CONFIG_SCHED_DEBUG */
6117 static int sd_degenerate(struct sched_domain *sd)
6119 if (cpumask_weight(sched_domain_span(sd)) == 1)
6122 /* Following flags need at least 2 groups */
6123 if (sd->flags & (SD_LOAD_BALANCE |
6124 SD_BALANCE_NEWIDLE |
6128 SD_SHARE_PKG_RESOURCES)) {
6129 if (sd->groups != sd->groups->next)
6133 /* Following flags don't use groups */
6134 if (sd->flags & (SD_WAKE_AFFINE))
6141 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6143 unsigned long cflags = sd->flags, pflags = parent->flags;
6145 if (sd_degenerate(parent))
6148 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6151 /* Flags needing groups don't count if only 1 group in parent */
6152 if (parent->groups == parent->groups->next) {
6153 pflags &= ~(SD_LOAD_BALANCE |
6154 SD_BALANCE_NEWIDLE |
6158 SD_SHARE_PKG_RESOURCES);
6159 if (nr_node_ids == 1)
6160 pflags &= ~SD_SERIALIZE;
6162 if (~cflags & pflags)
6168 static void free_rootdomain(struct root_domain *rd)
6170 synchronize_sched();
6172 cpupri_cleanup(&rd->cpupri);
6174 free_cpumask_var(rd->rto_mask);
6175 free_cpumask_var(rd->online);
6176 free_cpumask_var(rd->span);
6180 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6182 struct root_domain *old_rd = NULL;
6183 unsigned long flags;
6185 raw_spin_lock_irqsave(&rq->lock, flags);
6190 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6193 cpumask_clear_cpu(rq->cpu, old_rd->span);
6196 * If we dont want to free the old_rt yet then
6197 * set old_rd to NULL to skip the freeing later
6200 if (!atomic_dec_and_test(&old_rd->refcount))
6204 atomic_inc(&rd->refcount);
6207 cpumask_set_cpu(rq->cpu, rd->span);
6208 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6211 raw_spin_unlock_irqrestore(&rq->lock, flags);
6214 free_rootdomain(old_rd);
6217 static int init_rootdomain(struct root_domain *rd)
6219 memset(rd, 0, sizeof(*rd));
6221 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
6223 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
6225 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6228 if (cpupri_init(&rd->cpupri) != 0)
6233 free_cpumask_var(rd->rto_mask);
6235 free_cpumask_var(rd->online);
6237 free_cpumask_var(rd->span);
6242 static void init_defrootdomain(void)
6244 init_rootdomain(&def_root_domain);
6246 atomic_set(&def_root_domain.refcount, 1);
6249 static struct root_domain *alloc_rootdomain(void)
6251 struct root_domain *rd;
6253 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6257 if (init_rootdomain(rd) != 0) {
6266 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6267 * hold the hotplug lock.
6270 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6272 struct rq *rq = cpu_rq(cpu);
6273 struct sched_domain *tmp;
6275 for (tmp = sd; tmp; tmp = tmp->parent)
6276 tmp->span_weight = cpumask_weight(sched_domain_span(tmp));
6278 /* Remove the sched domains which do not contribute to scheduling. */
6279 for (tmp = sd; tmp; ) {
6280 struct sched_domain *parent = tmp->parent;
6284 if (sd_parent_degenerate(tmp, parent)) {
6285 tmp->parent = parent->parent;
6287 parent->parent->child = tmp;
6292 if (sd && sd_degenerate(sd)) {
6298 sched_domain_debug(sd, cpu);
6300 rq_attach_root(rq, rd);
6301 rcu_assign_pointer(rq->sd, sd);
6304 /* cpus with isolated domains */
6305 static cpumask_var_t cpu_isolated_map;
6307 /* Setup the mask of cpus configured for isolated domains */
6308 static int __init isolated_cpu_setup(char *str)
6310 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6311 cpulist_parse(str, cpu_isolated_map);
6315 __setup("isolcpus=", isolated_cpu_setup);
6318 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6319 * to a function which identifies what group(along with sched group) a CPU
6320 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6321 * (due to the fact that we keep track of groups covered with a struct cpumask).
6323 * init_sched_build_groups will build a circular linked list of the groups
6324 * covered by the given span, and will set each group's ->cpumask correctly,
6325 * and ->cpu_power to 0.
6328 init_sched_build_groups(const struct cpumask *span,
6329 const struct cpumask *cpu_map,
6330 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
6331 struct sched_group **sg,
6332 struct cpumask *tmpmask),
6333 struct cpumask *covered, struct cpumask *tmpmask)
6335 struct sched_group *first = NULL, *last = NULL;
6338 cpumask_clear(covered);
6340 for_each_cpu(i, span) {
6341 struct sched_group *sg;
6342 int group = group_fn(i, cpu_map, &sg, tmpmask);
6345 if (cpumask_test_cpu(i, covered))
6348 cpumask_clear(sched_group_cpus(sg));
6351 for_each_cpu(j, span) {
6352 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6355 cpumask_set_cpu(j, covered);
6356 cpumask_set_cpu(j, sched_group_cpus(sg));
6367 #define SD_NODES_PER_DOMAIN 16
6372 * find_next_best_node - find the next node to include in a sched_domain
6373 * @node: node whose sched_domain we're building
6374 * @used_nodes: nodes already in the sched_domain
6376 * Find the next node to include in a given scheduling domain. Simply
6377 * finds the closest node not already in the @used_nodes map.
6379 * Should use nodemask_t.
6381 static int find_next_best_node(int node, nodemask_t *used_nodes)
6383 int i, n, val, min_val, best_node = 0;
6387 for (i = 0; i < nr_node_ids; i++) {
6388 /* Start at @node */
6389 n = (node + i) % nr_node_ids;
6391 if (!nr_cpus_node(n))
6394 /* Skip already used nodes */
6395 if (node_isset(n, *used_nodes))
6398 /* Simple min distance search */
6399 val = node_distance(node, n);
6401 if (val < min_val) {
6407 node_set(best_node, *used_nodes);
6412 * sched_domain_node_span - get a cpumask for a node's sched_domain
6413 * @node: node whose cpumask we're constructing
6414 * @span: resulting cpumask
6416 * Given a node, construct a good cpumask for its sched_domain to span. It
6417 * should be one that prevents unnecessary balancing, but also spreads tasks
6420 static void sched_domain_node_span(int node, struct cpumask *span)
6422 nodemask_t used_nodes;
6425 cpumask_clear(span);
6426 nodes_clear(used_nodes);
6428 cpumask_or(span, span, cpumask_of_node(node));
6429 node_set(node, used_nodes);
6431 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6432 int next_node = find_next_best_node(node, &used_nodes);
6434 cpumask_or(span, span, cpumask_of_node(next_node));
6437 #endif /* CONFIG_NUMA */
6439 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6442 * The cpus mask in sched_group and sched_domain hangs off the end.
6444 * ( See the the comments in include/linux/sched.h:struct sched_group
6445 * and struct sched_domain. )
6447 struct static_sched_group {
6448 struct sched_group sg;
6449 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
6452 struct static_sched_domain {
6453 struct sched_domain sd;
6454 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
6460 cpumask_var_t domainspan;
6461 cpumask_var_t covered;
6462 cpumask_var_t notcovered;
6464 cpumask_var_t nodemask;
6465 cpumask_var_t this_sibling_map;
6466 cpumask_var_t this_core_map;
6467 cpumask_var_t this_book_map;
6468 cpumask_var_t send_covered;
6469 cpumask_var_t tmpmask;
6470 struct sched_group **sched_group_nodes;
6471 struct root_domain *rd;
6475 sa_sched_groups = 0,
6481 sa_this_sibling_map,
6483 sa_sched_group_nodes,
6493 * SMT sched-domains:
6495 #ifdef CONFIG_SCHED_SMT
6496 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
6497 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
6500 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
6501 struct sched_group **sg, struct cpumask *unused)
6504 *sg = &per_cpu(sched_groups, cpu).sg;
6507 #endif /* CONFIG_SCHED_SMT */
6510 * multi-core sched-domains:
6512 #ifdef CONFIG_SCHED_MC
6513 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
6514 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
6517 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6518 struct sched_group **sg, struct cpumask *mask)
6521 #ifdef CONFIG_SCHED_SMT
6522 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6523 group = cpumask_first(mask);
6528 *sg = &per_cpu(sched_group_core, group).sg;
6531 #endif /* CONFIG_SCHED_MC */
6534 * book sched-domains:
6536 #ifdef CONFIG_SCHED_BOOK
6537 static DEFINE_PER_CPU(struct static_sched_domain, book_domains);
6538 static DEFINE_PER_CPU(struct static_sched_group, sched_group_book);
6541 cpu_to_book_group(int cpu, const struct cpumask *cpu_map,
6542 struct sched_group **sg, struct cpumask *mask)
6545 #ifdef CONFIG_SCHED_MC
6546 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6547 group = cpumask_first(mask);
6548 #elif defined(CONFIG_SCHED_SMT)
6549 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6550 group = cpumask_first(mask);
6553 *sg = &per_cpu(sched_group_book, group).sg;
6556 #endif /* CONFIG_SCHED_BOOK */
6558 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
6559 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
6562 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
6563 struct sched_group **sg, struct cpumask *mask)
6566 #ifdef CONFIG_SCHED_BOOK
6567 cpumask_and(mask, cpu_book_mask(cpu), cpu_map);
6568 group = cpumask_first(mask);
6569 #elif defined(CONFIG_SCHED_MC)
6570 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6571 group = cpumask_first(mask);
6572 #elif defined(CONFIG_SCHED_SMT)
6573 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6574 group = cpumask_first(mask);
6579 *sg = &per_cpu(sched_group_phys, group).sg;
6585 * The init_sched_build_groups can't handle what we want to do with node
6586 * groups, so roll our own. Now each node has its own list of groups which
6587 * gets dynamically allocated.
6589 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
6590 static struct sched_group ***sched_group_nodes_bycpu;
6592 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
6593 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
6595 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
6596 struct sched_group **sg,
6597 struct cpumask *nodemask)
6601 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
6602 group = cpumask_first(nodemask);
6605 *sg = &per_cpu(sched_group_allnodes, group).sg;
6609 static void init_numa_sched_groups_power(struct sched_group *group_head)
6611 struct sched_group *sg = group_head;
6617 for_each_cpu(j, sched_group_cpus(sg)) {
6618 struct sched_domain *sd;
6620 sd = &per_cpu(phys_domains, j).sd;
6621 if (j != group_first_cpu(sd->groups)) {
6623 * Only add "power" once for each
6629 sg->cpu_power += sd->groups->cpu_power;
6632 } while (sg != group_head);
6635 static int build_numa_sched_groups(struct s_data *d,
6636 const struct cpumask *cpu_map, int num)
6638 struct sched_domain *sd;
6639 struct sched_group *sg, *prev;
6642 cpumask_clear(d->covered);
6643 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
6644 if (cpumask_empty(d->nodemask)) {
6645 d->sched_group_nodes[num] = NULL;
6649 sched_domain_node_span(num, d->domainspan);
6650 cpumask_and(d->domainspan, d->domainspan, cpu_map);
6652 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6655 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
6659 d->sched_group_nodes[num] = sg;
6661 for_each_cpu(j, d->nodemask) {
6662 sd = &per_cpu(node_domains, j).sd;
6667 cpumask_copy(sched_group_cpus(sg), d->nodemask);
6669 cpumask_or(d->covered, d->covered, d->nodemask);
6672 for (j = 0; j < nr_node_ids; j++) {
6673 n = (num + j) % nr_node_ids;
6674 cpumask_complement(d->notcovered, d->covered);
6675 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
6676 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
6677 if (cpumask_empty(d->tmpmask))
6679 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
6680 if (cpumask_empty(d->tmpmask))
6682 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6686 "Can not alloc domain group for node %d\n", j);
6690 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
6691 sg->next = prev->next;
6692 cpumask_or(d->covered, d->covered, d->tmpmask);
6699 #endif /* CONFIG_NUMA */
6702 /* Free memory allocated for various sched_group structures */
6703 static void free_sched_groups(const struct cpumask *cpu_map,
6704 struct cpumask *nodemask)
6708 for_each_cpu(cpu, cpu_map) {
6709 struct sched_group **sched_group_nodes
6710 = sched_group_nodes_bycpu[cpu];
6712 if (!sched_group_nodes)
6715 for (i = 0; i < nr_node_ids; i++) {
6716 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6718 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
6719 if (cpumask_empty(nodemask))
6729 if (oldsg != sched_group_nodes[i])
6732 kfree(sched_group_nodes);
6733 sched_group_nodes_bycpu[cpu] = NULL;
6736 #else /* !CONFIG_NUMA */
6737 static void free_sched_groups(const struct cpumask *cpu_map,
6738 struct cpumask *nodemask)
6741 #endif /* CONFIG_NUMA */
6744 * Initialize sched groups cpu_power.
6746 * cpu_power indicates the capacity of sched group, which is used while
6747 * distributing the load between different sched groups in a sched domain.
6748 * Typically cpu_power for all the groups in a sched domain will be same unless
6749 * there are asymmetries in the topology. If there are asymmetries, group
6750 * having more cpu_power will pickup more load compared to the group having
6753 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6755 struct sched_domain *child;
6756 struct sched_group *group;
6760 WARN_ON(!sd || !sd->groups);
6762 if (cpu != group_first_cpu(sd->groups))
6765 sd->groups->group_weight = cpumask_weight(sched_group_cpus(sd->groups));
6769 sd->groups->cpu_power = 0;
6772 power = SCHED_LOAD_SCALE;
6773 weight = cpumask_weight(sched_domain_span(sd));
6775 * SMT siblings share the power of a single core.
6776 * Usually multiple threads get a better yield out of
6777 * that one core than a single thread would have,
6778 * reflect that in sd->smt_gain.
6780 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
6781 power *= sd->smt_gain;
6783 power >>= SCHED_LOAD_SHIFT;
6785 sd->groups->cpu_power += power;
6790 * Add cpu_power of each child group to this groups cpu_power.
6792 group = child->groups;
6794 sd->groups->cpu_power += group->cpu_power;
6795 group = group->next;
6796 } while (group != child->groups);
6800 * Initializers for schedule domains
6801 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6804 #ifdef CONFIG_SCHED_DEBUG
6805 # define SD_INIT_NAME(sd, type) sd->name = #type
6807 # define SD_INIT_NAME(sd, type) do { } while (0)
6810 #define SD_INIT(sd, type) sd_init_##type(sd)
6812 #define SD_INIT_FUNC(type) \
6813 static noinline void sd_init_##type(struct sched_domain *sd) \
6815 memset(sd, 0, sizeof(*sd)); \
6816 *sd = SD_##type##_INIT; \
6817 sd->level = SD_LV_##type; \
6818 SD_INIT_NAME(sd, type); \
6823 SD_INIT_FUNC(ALLNODES)
6826 #ifdef CONFIG_SCHED_SMT
6827 SD_INIT_FUNC(SIBLING)
6829 #ifdef CONFIG_SCHED_MC
6832 #ifdef CONFIG_SCHED_BOOK
6836 static int default_relax_domain_level = -1;
6838 static int __init setup_relax_domain_level(char *str)
6842 val = simple_strtoul(str, NULL, 0);
6843 if (val < SD_LV_MAX)
6844 default_relax_domain_level = val;
6848 __setup("relax_domain_level=", setup_relax_domain_level);
6850 static void set_domain_attribute(struct sched_domain *sd,
6851 struct sched_domain_attr *attr)
6855 if (!attr || attr->relax_domain_level < 0) {
6856 if (default_relax_domain_level < 0)
6859 request = default_relax_domain_level;
6861 request = attr->relax_domain_level;
6862 if (request < sd->level) {
6863 /* turn off idle balance on this domain */
6864 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6866 /* turn on idle balance on this domain */
6867 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6871 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6872 const struct cpumask *cpu_map)
6875 case sa_sched_groups:
6876 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
6877 d->sched_group_nodes = NULL;
6879 free_rootdomain(d->rd); /* fall through */
6881 free_cpumask_var(d->tmpmask); /* fall through */
6882 case sa_send_covered:
6883 free_cpumask_var(d->send_covered); /* fall through */
6884 case sa_this_book_map:
6885 free_cpumask_var(d->this_book_map); /* fall through */
6886 case sa_this_core_map:
6887 free_cpumask_var(d->this_core_map); /* fall through */
6888 case sa_this_sibling_map:
6889 free_cpumask_var(d->this_sibling_map); /* fall through */
6891 free_cpumask_var(d->nodemask); /* fall through */
6892 case sa_sched_group_nodes:
6894 kfree(d->sched_group_nodes); /* fall through */
6896 free_cpumask_var(d->notcovered); /* fall through */
6898 free_cpumask_var(d->covered); /* fall through */
6900 free_cpumask_var(d->domainspan); /* fall through */
6907 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6908 const struct cpumask *cpu_map)
6911 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
6913 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
6914 return sa_domainspan;
6915 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
6917 /* Allocate the per-node list of sched groups */
6918 d->sched_group_nodes = kcalloc(nr_node_ids,
6919 sizeof(struct sched_group *), GFP_KERNEL);
6920 if (!d->sched_group_nodes) {
6921 printk(KERN_WARNING "Can not alloc sched group node list\n");
6922 return sa_notcovered;
6924 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
6926 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
6927 return sa_sched_group_nodes;
6928 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
6930 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
6931 return sa_this_sibling_map;
6932 if (!alloc_cpumask_var(&d->this_book_map, GFP_KERNEL))
6933 return sa_this_core_map;
6934 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
6935 return sa_this_book_map;
6936 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
6937 return sa_send_covered;
6938 d->rd = alloc_rootdomain();
6940 printk(KERN_WARNING "Cannot alloc root domain\n");
6943 return sa_rootdomain;
6946 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
6947 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
6949 struct sched_domain *sd = NULL;
6951 struct sched_domain *parent;
6954 if (cpumask_weight(cpu_map) >
6955 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
6956 sd = &per_cpu(allnodes_domains, i).sd;
6957 SD_INIT(sd, ALLNODES);
6958 set_domain_attribute(sd, attr);
6959 cpumask_copy(sched_domain_span(sd), cpu_map);
6960 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
6965 sd = &per_cpu(node_domains, i).sd;
6967 set_domain_attribute(sd, attr);
6968 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
6969 sd->parent = parent;
6972 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
6977 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
6978 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6979 struct sched_domain *parent, int i)
6981 struct sched_domain *sd;
6982 sd = &per_cpu(phys_domains, i).sd;
6984 set_domain_attribute(sd, attr);
6985 cpumask_copy(sched_domain_span(sd), d->nodemask);
6986 sd->parent = parent;
6989 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
6993 static struct sched_domain *__build_book_sched_domain(struct s_data *d,
6994 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6995 struct sched_domain *parent, int i)
6997 struct sched_domain *sd = parent;
6998 #ifdef CONFIG_SCHED_BOOK
6999 sd = &per_cpu(book_domains, i).sd;
7001 set_domain_attribute(sd, attr);
7002 cpumask_and(sched_domain_span(sd), cpu_map, cpu_book_mask(i));
7003 sd->parent = parent;
7005 cpu_to_book_group(i, cpu_map, &sd->groups, d->tmpmask);
7010 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
7011 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7012 struct sched_domain *parent, int i)
7014 struct sched_domain *sd = parent;
7015 #ifdef CONFIG_SCHED_MC
7016 sd = &per_cpu(core_domains, i).sd;
7018 set_domain_attribute(sd, attr);
7019 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
7020 sd->parent = parent;
7022 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
7027 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
7028 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7029 struct sched_domain *parent, int i)
7031 struct sched_domain *sd = parent;
7032 #ifdef CONFIG_SCHED_SMT
7033 sd = &per_cpu(cpu_domains, i).sd;
7034 SD_INIT(sd, SIBLING);
7035 set_domain_attribute(sd, attr);
7036 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
7037 sd->parent = parent;
7039 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
7044 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
7045 const struct cpumask *cpu_map, int cpu)
7048 #ifdef CONFIG_SCHED_SMT
7049 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
7050 cpumask_and(d->this_sibling_map, cpu_map,
7051 topology_thread_cpumask(cpu));
7052 if (cpu == cpumask_first(d->this_sibling_map))
7053 init_sched_build_groups(d->this_sibling_map, cpu_map,
7055 d->send_covered, d->tmpmask);
7058 #ifdef CONFIG_SCHED_MC
7059 case SD_LV_MC: /* set up multi-core groups */
7060 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
7061 if (cpu == cpumask_first(d->this_core_map))
7062 init_sched_build_groups(d->this_core_map, cpu_map,
7064 d->send_covered, d->tmpmask);
7067 #ifdef CONFIG_SCHED_BOOK
7068 case SD_LV_BOOK: /* set up book groups */
7069 cpumask_and(d->this_book_map, cpu_map, cpu_book_mask(cpu));
7070 if (cpu == cpumask_first(d->this_book_map))
7071 init_sched_build_groups(d->this_book_map, cpu_map,
7073 d->send_covered, d->tmpmask);
7076 case SD_LV_CPU: /* set up physical groups */
7077 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
7078 if (!cpumask_empty(d->nodemask))
7079 init_sched_build_groups(d->nodemask, cpu_map,
7081 d->send_covered, d->tmpmask);
7084 case SD_LV_ALLNODES:
7085 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
7086 d->send_covered, d->tmpmask);
7095 * Build sched domains for a given set of cpus and attach the sched domains
7096 * to the individual cpus
7098 static int __build_sched_domains(const struct cpumask *cpu_map,
7099 struct sched_domain_attr *attr)
7101 enum s_alloc alloc_state = sa_none;
7103 struct sched_domain *sd;
7109 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7110 if (alloc_state != sa_rootdomain)
7112 alloc_state = sa_sched_groups;
7115 * Set up domains for cpus specified by the cpu_map.
7117 for_each_cpu(i, cpu_map) {
7118 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
7121 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
7122 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
7123 sd = __build_book_sched_domain(&d, cpu_map, attr, sd, i);
7124 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
7125 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
7128 for_each_cpu(i, cpu_map) {
7129 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
7130 build_sched_groups(&d, SD_LV_BOOK, cpu_map, i);
7131 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
7134 /* Set up physical groups */
7135 for (i = 0; i < nr_node_ids; i++)
7136 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
7139 /* Set up node groups */
7141 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
7143 for (i = 0; i < nr_node_ids; i++)
7144 if (build_numa_sched_groups(&d, cpu_map, i))
7148 /* Calculate CPU power for physical packages and nodes */
7149 #ifdef CONFIG_SCHED_SMT
7150 for_each_cpu(i, cpu_map) {
7151 sd = &per_cpu(cpu_domains, i).sd;
7152 init_sched_groups_power(i, sd);
7155 #ifdef CONFIG_SCHED_MC
7156 for_each_cpu(i, cpu_map) {
7157 sd = &per_cpu(core_domains, i).sd;
7158 init_sched_groups_power(i, sd);
7161 #ifdef CONFIG_SCHED_BOOK
7162 for_each_cpu(i, cpu_map) {
7163 sd = &per_cpu(book_domains, i).sd;
7164 init_sched_groups_power(i, sd);
7168 for_each_cpu(i, cpu_map) {
7169 sd = &per_cpu(phys_domains, i).sd;
7170 init_sched_groups_power(i, sd);
7174 for (i = 0; i < nr_node_ids; i++)
7175 init_numa_sched_groups_power(d.sched_group_nodes[i]);
7177 if (d.sd_allnodes) {
7178 struct sched_group *sg;
7180 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7182 init_numa_sched_groups_power(sg);
7186 /* Attach the domains */
7187 for_each_cpu(i, cpu_map) {
7188 #ifdef CONFIG_SCHED_SMT
7189 sd = &per_cpu(cpu_domains, i).sd;
7190 #elif defined(CONFIG_SCHED_MC)
7191 sd = &per_cpu(core_domains, i).sd;
7192 #elif defined(CONFIG_SCHED_BOOK)
7193 sd = &per_cpu(book_domains, i).sd;
7195 sd = &per_cpu(phys_domains, i).sd;
7197 cpu_attach_domain(sd, d.rd, i);
7200 d.sched_group_nodes = NULL; /* don't free this we still need it */
7201 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
7205 __free_domain_allocs(&d, alloc_state, cpu_map);
7209 static int build_sched_domains(const struct cpumask *cpu_map)
7211 return __build_sched_domains(cpu_map, NULL);
7214 static cpumask_var_t *doms_cur; /* current sched domains */
7215 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7216 static struct sched_domain_attr *dattr_cur;
7217 /* attribues of custom domains in 'doms_cur' */
7220 * Special case: If a kmalloc of a doms_cur partition (array of
7221 * cpumask) fails, then fallback to a single sched domain,
7222 * as determined by the single cpumask fallback_doms.
7224 static cpumask_var_t fallback_doms;
7227 * arch_update_cpu_topology lets virtualized architectures update the
7228 * cpu core maps. It is supposed to return 1 if the topology changed
7229 * or 0 if it stayed the same.
7231 int __attribute__((weak)) arch_update_cpu_topology(void)
7236 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7239 cpumask_var_t *doms;
7241 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7244 for (i = 0; i < ndoms; i++) {
7245 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7246 free_sched_domains(doms, i);
7253 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7256 for (i = 0; i < ndoms; i++)
7257 free_cpumask_var(doms[i]);
7262 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7263 * For now this just excludes isolated cpus, but could be used to
7264 * exclude other special cases in the future.
7266 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7270 arch_update_cpu_topology();
7272 doms_cur = alloc_sched_domains(ndoms_cur);
7274 doms_cur = &fallback_doms;
7275 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7277 err = build_sched_domains(doms_cur[0]);
7278 register_sched_domain_sysctl();
7283 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7284 struct cpumask *tmpmask)
7286 free_sched_groups(cpu_map, tmpmask);
7290 * Detach sched domains from a group of cpus specified in cpu_map
7291 * These cpus will now be attached to the NULL domain
7293 static void detach_destroy_domains(const struct cpumask *cpu_map)
7295 /* Save because hotplug lock held. */
7296 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7299 for_each_cpu(i, cpu_map)
7300 cpu_attach_domain(NULL, &def_root_domain, i);
7301 synchronize_sched();
7302 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7305 /* handle null as "default" */
7306 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7307 struct sched_domain_attr *new, int idx_new)
7309 struct sched_domain_attr tmp;
7316 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7317 new ? (new + idx_new) : &tmp,
7318 sizeof(struct sched_domain_attr));
7322 * Partition sched domains as specified by the 'ndoms_new'
7323 * cpumasks in the array doms_new[] of cpumasks. This compares
7324 * doms_new[] to the current sched domain partitioning, doms_cur[].
7325 * It destroys each deleted domain and builds each new domain.
7327 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7328 * The masks don't intersect (don't overlap.) We should setup one
7329 * sched domain for each mask. CPUs not in any of the cpumasks will
7330 * not be load balanced. If the same cpumask appears both in the
7331 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7334 * The passed in 'doms_new' should be allocated using
7335 * alloc_sched_domains. This routine takes ownership of it and will
7336 * free_sched_domains it when done with it. If the caller failed the
7337 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7338 * and partition_sched_domains() will fallback to the single partition
7339 * 'fallback_doms', it also forces the domains to be rebuilt.
7341 * If doms_new == NULL it will be replaced with cpu_online_mask.
7342 * ndoms_new == 0 is a special case for destroying existing domains,
7343 * and it will not create the default domain.
7345 * Call with hotplug lock held
7347 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7348 struct sched_domain_attr *dattr_new)
7353 mutex_lock(&sched_domains_mutex);
7355 /* always unregister in case we don't destroy any domains */
7356 unregister_sched_domain_sysctl();
7358 /* Let architecture update cpu core mappings. */
7359 new_topology = arch_update_cpu_topology();
7361 n = doms_new ? ndoms_new : 0;
7363 /* Destroy deleted domains */
7364 for (i = 0; i < ndoms_cur; i++) {
7365 for (j = 0; j < n && !new_topology; j++) {
7366 if (cpumask_equal(doms_cur[i], doms_new[j])
7367 && dattrs_equal(dattr_cur, i, dattr_new, j))
7370 /* no match - a current sched domain not in new doms_new[] */
7371 detach_destroy_domains(doms_cur[i]);
7376 if (doms_new == NULL) {
7378 doms_new = &fallback_doms;
7379 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7380 WARN_ON_ONCE(dattr_new);
7383 /* Build new domains */
7384 for (i = 0; i < ndoms_new; i++) {
7385 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7386 if (cpumask_equal(doms_new[i], doms_cur[j])
7387 && dattrs_equal(dattr_new, i, dattr_cur, j))
7390 /* no match - add a new doms_new */
7391 __build_sched_domains(doms_new[i],
7392 dattr_new ? dattr_new + i : NULL);
7397 /* Remember the new sched domains */
7398 if (doms_cur != &fallback_doms)
7399 free_sched_domains(doms_cur, ndoms_cur);
7400 kfree(dattr_cur); /* kfree(NULL) is safe */
7401 doms_cur = doms_new;
7402 dattr_cur = dattr_new;
7403 ndoms_cur = ndoms_new;
7405 register_sched_domain_sysctl();
7407 mutex_unlock(&sched_domains_mutex);
7410 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7411 static void arch_reinit_sched_domains(void)
7415 /* Destroy domains first to force the rebuild */
7416 partition_sched_domains(0, NULL, NULL);
7418 rebuild_sched_domains();
7422 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7424 unsigned int level = 0;
7426 if (sscanf(buf, "%u", &level) != 1)
7430 * level is always be positive so don't check for
7431 * level < POWERSAVINGS_BALANCE_NONE which is 0
7432 * What happens on 0 or 1 byte write,
7433 * need to check for count as well?
7436 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7440 sched_smt_power_savings = level;
7442 sched_mc_power_savings = level;
7444 arch_reinit_sched_domains();
7449 #ifdef CONFIG_SCHED_MC
7450 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7451 struct sysdev_class_attribute *attr,
7454 return sprintf(page, "%u\n", sched_mc_power_savings);
7456 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7457 struct sysdev_class_attribute *attr,
7458 const char *buf, size_t count)
7460 return sched_power_savings_store(buf, count, 0);
7462 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7463 sched_mc_power_savings_show,
7464 sched_mc_power_savings_store);
7467 #ifdef CONFIG_SCHED_SMT
7468 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7469 struct sysdev_class_attribute *attr,
7472 return sprintf(page, "%u\n", sched_smt_power_savings);
7474 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7475 struct sysdev_class_attribute *attr,
7476 const char *buf, size_t count)
7478 return sched_power_savings_store(buf, count, 1);
7480 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7481 sched_smt_power_savings_show,
7482 sched_smt_power_savings_store);
7485 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7489 #ifdef CONFIG_SCHED_SMT
7491 err = sysfs_create_file(&cls->kset.kobj,
7492 &attr_sched_smt_power_savings.attr);
7494 #ifdef CONFIG_SCHED_MC
7495 if (!err && mc_capable())
7496 err = sysfs_create_file(&cls->kset.kobj,
7497 &attr_sched_mc_power_savings.attr);
7501 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7504 * Update cpusets according to cpu_active mask. If cpusets are
7505 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7506 * around partition_sched_domains().
7508 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7511 switch (action & ~CPU_TASKS_FROZEN) {
7513 case CPU_DOWN_FAILED:
7514 cpuset_update_active_cpus();
7521 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7524 switch (action & ~CPU_TASKS_FROZEN) {
7525 case CPU_DOWN_PREPARE:
7526 cpuset_update_active_cpus();
7533 static int update_runtime(struct notifier_block *nfb,
7534 unsigned long action, void *hcpu)
7536 int cpu = (int)(long)hcpu;
7539 case CPU_DOWN_PREPARE:
7540 case CPU_DOWN_PREPARE_FROZEN:
7541 disable_runtime(cpu_rq(cpu));
7544 case CPU_DOWN_FAILED:
7545 case CPU_DOWN_FAILED_FROZEN:
7547 case CPU_ONLINE_FROZEN:
7548 enable_runtime(cpu_rq(cpu));
7556 void __init sched_init_smp(void)
7558 cpumask_var_t non_isolated_cpus;
7560 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7561 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7563 #if defined(CONFIG_NUMA)
7564 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7566 BUG_ON(sched_group_nodes_bycpu == NULL);
7569 mutex_lock(&sched_domains_mutex);
7570 arch_init_sched_domains(cpu_active_mask);
7571 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7572 if (cpumask_empty(non_isolated_cpus))
7573 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7574 mutex_unlock(&sched_domains_mutex);
7577 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7578 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7580 /* RT runtime code needs to handle some hotplug events */
7581 hotcpu_notifier(update_runtime, 0);
7585 /* Move init over to a non-isolated CPU */
7586 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7588 sched_init_granularity();
7589 free_cpumask_var(non_isolated_cpus);
7591 init_sched_rt_class();
7594 void __init sched_init_smp(void)
7596 sched_init_granularity();
7598 #endif /* CONFIG_SMP */
7600 const_debug unsigned int sysctl_timer_migration = 1;
7602 int in_sched_functions(unsigned long addr)
7604 return in_lock_functions(addr) ||
7605 (addr >= (unsigned long)__sched_text_start
7606 && addr < (unsigned long)__sched_text_end);
7609 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7611 cfs_rq->tasks_timeline = RB_ROOT;
7612 INIT_LIST_HEAD(&cfs_rq->tasks);
7613 #ifdef CONFIG_FAIR_GROUP_SCHED
7616 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7619 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7621 struct rt_prio_array *array;
7624 array = &rt_rq->active;
7625 for (i = 0; i < MAX_RT_PRIO; i++) {
7626 INIT_LIST_HEAD(array->queue + i);
7627 __clear_bit(i, array->bitmap);
7629 /* delimiter for bitsearch: */
7630 __set_bit(MAX_RT_PRIO, array->bitmap);
7632 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7633 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7635 rt_rq->highest_prio.next = MAX_RT_PRIO;
7639 rt_rq->rt_nr_migratory = 0;
7640 rt_rq->overloaded = 0;
7641 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
7645 rt_rq->rt_throttled = 0;
7646 rt_rq->rt_runtime = 0;
7647 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7649 #ifdef CONFIG_RT_GROUP_SCHED
7650 rt_rq->rt_nr_boosted = 0;
7655 #ifdef CONFIG_FAIR_GROUP_SCHED
7656 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7657 struct sched_entity *se, int cpu,
7658 struct sched_entity *parent)
7660 struct rq *rq = cpu_rq(cpu);
7661 tg->cfs_rq[cpu] = cfs_rq;
7662 init_cfs_rq(cfs_rq, rq);
7666 /* se could be NULL for init_task_group */
7671 se->cfs_rq = &rq->cfs;
7673 se->cfs_rq = parent->my_q;
7676 update_load_set(&se->load, 0);
7677 se->parent = parent;
7681 #ifdef CONFIG_RT_GROUP_SCHED
7682 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7683 struct sched_rt_entity *rt_se, int cpu,
7684 struct sched_rt_entity *parent)
7686 struct rq *rq = cpu_rq(cpu);
7688 tg->rt_rq[cpu] = rt_rq;
7689 init_rt_rq(rt_rq, rq);
7691 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7693 tg->rt_se[cpu] = rt_se;
7698 rt_se->rt_rq = &rq->rt;
7700 rt_se->rt_rq = parent->my_q;
7702 rt_se->my_q = rt_rq;
7703 rt_se->parent = parent;
7704 INIT_LIST_HEAD(&rt_se->run_list);
7708 void __init sched_init(void)
7711 unsigned long alloc_size = 0, ptr;
7713 #ifdef CONFIG_FAIR_GROUP_SCHED
7714 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7716 #ifdef CONFIG_RT_GROUP_SCHED
7717 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7719 #ifdef CONFIG_CPUMASK_OFFSTACK
7720 alloc_size += num_possible_cpus() * cpumask_size();
7723 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7725 #ifdef CONFIG_FAIR_GROUP_SCHED
7726 init_task_group.se = (struct sched_entity **)ptr;
7727 ptr += nr_cpu_ids * sizeof(void **);
7729 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7730 ptr += nr_cpu_ids * sizeof(void **);
7732 #endif /* CONFIG_FAIR_GROUP_SCHED */
7733 #ifdef CONFIG_RT_GROUP_SCHED
7734 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7735 ptr += nr_cpu_ids * sizeof(void **);
7737 init_task_group.rt_rq = (struct rt_rq **)ptr;
7738 ptr += nr_cpu_ids * sizeof(void **);
7740 #endif /* CONFIG_RT_GROUP_SCHED */
7741 #ifdef CONFIG_CPUMASK_OFFSTACK
7742 for_each_possible_cpu(i) {
7743 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
7744 ptr += cpumask_size();
7746 #endif /* CONFIG_CPUMASK_OFFSTACK */
7750 init_defrootdomain();
7753 init_rt_bandwidth(&def_rt_bandwidth,
7754 global_rt_period(), global_rt_runtime());
7756 #ifdef CONFIG_RT_GROUP_SCHED
7757 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7758 global_rt_period(), global_rt_runtime());
7759 #endif /* CONFIG_RT_GROUP_SCHED */
7761 #ifdef CONFIG_CGROUP_SCHED
7762 list_add(&init_task_group.list, &task_groups);
7763 INIT_LIST_HEAD(&init_task_group.children);
7764 autogroup_init(&init_task);
7765 #endif /* CONFIG_CGROUP_SCHED */
7767 for_each_possible_cpu(i) {
7771 raw_spin_lock_init(&rq->lock);
7773 rq->calc_load_active = 0;
7774 rq->calc_load_update = jiffies + LOAD_FREQ;
7775 init_cfs_rq(&rq->cfs, rq);
7776 init_rt_rq(&rq->rt, rq);
7777 #ifdef CONFIG_FAIR_GROUP_SCHED
7778 init_task_group.shares = init_task_group_load;
7779 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7780 #ifdef CONFIG_CGROUP_SCHED
7782 * How much cpu bandwidth does init_task_group get?
7784 * In case of task-groups formed thr' the cgroup filesystem, it
7785 * gets 100% of the cpu resources in the system. This overall
7786 * system cpu resource is divided among the tasks of
7787 * init_task_group and its child task-groups in a fair manner,
7788 * based on each entity's (task or task-group's) weight
7789 * (se->load.weight).
7791 * In other words, if init_task_group has 10 tasks of weight
7792 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7793 * then A0's share of the cpu resource is:
7795 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7797 * We achieve this by letting init_task_group's tasks sit
7798 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7800 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, NULL);
7802 #endif /* CONFIG_FAIR_GROUP_SCHED */
7804 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7805 #ifdef CONFIG_RT_GROUP_SCHED
7806 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7807 #ifdef CONFIG_CGROUP_SCHED
7808 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, NULL);
7812 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7813 rq->cpu_load[j] = 0;
7815 rq->last_load_update_tick = jiffies;
7820 rq->cpu_power = SCHED_LOAD_SCALE;
7821 rq->post_schedule = 0;
7822 rq->active_balance = 0;
7823 rq->next_balance = jiffies;
7828 rq->avg_idle = 2*sysctl_sched_migration_cost;
7829 rq_attach_root(rq, &def_root_domain);
7831 rq->nohz_balance_kick = 0;
7832 init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb, i));
7836 atomic_set(&rq->nr_iowait, 0);
7839 set_load_weight(&init_task);
7841 #ifdef CONFIG_PREEMPT_NOTIFIERS
7842 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7846 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7849 #ifdef CONFIG_RT_MUTEXES
7850 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
7854 * The boot idle thread does lazy MMU switching as well:
7856 atomic_inc(&init_mm.mm_count);
7857 enter_lazy_tlb(&init_mm, current);
7860 * Make us the idle thread. Technically, schedule() should not be
7861 * called from this thread, however somewhere below it might be,
7862 * but because we are the idle thread, we just pick up running again
7863 * when this runqueue becomes "idle".
7865 init_idle(current, smp_processor_id());
7867 calc_load_update = jiffies + LOAD_FREQ;
7870 * During early bootup we pretend to be a normal task:
7872 current->sched_class = &fair_sched_class;
7874 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
7875 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
7878 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
7879 alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT);
7880 atomic_set(&nohz.load_balancer, nr_cpu_ids);
7881 atomic_set(&nohz.first_pick_cpu, nr_cpu_ids);
7882 atomic_set(&nohz.second_pick_cpu, nr_cpu_ids);
7884 /* May be allocated at isolcpus cmdline parse time */
7885 if (cpu_isolated_map == NULL)
7886 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7891 scheduler_running = 1;
7894 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7895 static inline int preempt_count_equals(int preempt_offset)
7897 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7899 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
7902 void __might_sleep(const char *file, int line, int preempt_offset)
7905 static unsigned long prev_jiffy; /* ratelimiting */
7907 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7908 system_state != SYSTEM_RUNNING || oops_in_progress)
7910 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7912 prev_jiffy = jiffies;
7915 "BUG: sleeping function called from invalid context at %s:%d\n",
7918 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7919 in_atomic(), irqs_disabled(),
7920 current->pid, current->comm);
7922 debug_show_held_locks(current);
7923 if (irqs_disabled())
7924 print_irqtrace_events(current);
7928 EXPORT_SYMBOL(__might_sleep);
7931 #ifdef CONFIG_MAGIC_SYSRQ
7932 static void normalize_task(struct rq *rq, struct task_struct *p)
7936 on_rq = p->se.on_rq;
7938 deactivate_task(rq, p, 0);
7939 __setscheduler(rq, p, SCHED_NORMAL, 0);
7941 activate_task(rq, p, 0);
7942 resched_task(rq->curr);
7946 void normalize_rt_tasks(void)
7948 struct task_struct *g, *p;
7949 unsigned long flags;
7952 read_lock_irqsave(&tasklist_lock, flags);
7953 do_each_thread(g, p) {
7955 * Only normalize user tasks:
7960 p->se.exec_start = 0;
7961 #ifdef CONFIG_SCHEDSTATS
7962 p->se.statistics.wait_start = 0;
7963 p->se.statistics.sleep_start = 0;
7964 p->se.statistics.block_start = 0;
7969 * Renice negative nice level userspace
7972 if (TASK_NICE(p) < 0 && p->mm)
7973 set_user_nice(p, 0);
7977 raw_spin_lock(&p->pi_lock);
7978 rq = __task_rq_lock(p);
7980 normalize_task(rq, p);
7982 __task_rq_unlock(rq);
7983 raw_spin_unlock(&p->pi_lock);
7984 } while_each_thread(g, p);
7986 read_unlock_irqrestore(&tasklist_lock, flags);
7989 #endif /* CONFIG_MAGIC_SYSRQ */
7991 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7993 * These functions are only useful for the IA64 MCA handling, or kdb.
7995 * They can only be called when the whole system has been
7996 * stopped - every CPU needs to be quiescent, and no scheduling
7997 * activity can take place. Using them for anything else would
7998 * be a serious bug, and as a result, they aren't even visible
7999 * under any other configuration.
8003 * curr_task - return the current task for a given cpu.
8004 * @cpu: the processor in question.
8006 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8008 struct task_struct *curr_task(int cpu)
8010 return cpu_curr(cpu);
8013 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8017 * set_curr_task - set the current task for a given cpu.
8018 * @cpu: the processor in question.
8019 * @p: the task pointer to set.
8021 * Description: This function must only be used when non-maskable interrupts
8022 * are serviced on a separate stack. It allows the architecture to switch the
8023 * notion of the current task on a cpu in a non-blocking manner. This function
8024 * must be called with all CPU's synchronized, and interrupts disabled, the
8025 * and caller must save the original value of the current task (see
8026 * curr_task() above) and restore that value before reenabling interrupts and
8027 * re-starting the system.
8029 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8031 void set_curr_task(int cpu, struct task_struct *p)
8038 #ifdef CONFIG_FAIR_GROUP_SCHED
8039 static void free_fair_sched_group(struct task_group *tg)
8043 for_each_possible_cpu(i) {
8045 kfree(tg->cfs_rq[i]);
8055 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8057 struct cfs_rq *cfs_rq;
8058 struct sched_entity *se;
8062 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8065 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8069 tg->shares = NICE_0_LOAD;
8071 for_each_possible_cpu(i) {
8074 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8075 GFP_KERNEL, cpu_to_node(i));
8079 se = kzalloc_node(sizeof(struct sched_entity),
8080 GFP_KERNEL, cpu_to_node(i));
8084 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8095 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8097 struct rq *rq = cpu_rq(cpu);
8098 unsigned long flags;
8101 * Only empty task groups can be destroyed; so we can speculatively
8102 * check on_list without danger of it being re-added.
8104 if (!tg->cfs_rq[cpu]->on_list)
8107 raw_spin_lock_irqsave(&rq->lock, flags);
8108 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8109 raw_spin_unlock_irqrestore(&rq->lock, flags);
8111 #else /* !CONFG_FAIR_GROUP_SCHED */
8112 static inline void free_fair_sched_group(struct task_group *tg)
8117 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8122 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8125 #endif /* CONFIG_FAIR_GROUP_SCHED */
8127 #ifdef CONFIG_RT_GROUP_SCHED
8128 static void free_rt_sched_group(struct task_group *tg)
8132 destroy_rt_bandwidth(&tg->rt_bandwidth);
8134 for_each_possible_cpu(i) {
8136 kfree(tg->rt_rq[i]);
8138 kfree(tg->rt_se[i]);
8146 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8148 struct rt_rq *rt_rq;
8149 struct sched_rt_entity *rt_se;
8153 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8156 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8160 init_rt_bandwidth(&tg->rt_bandwidth,
8161 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8163 for_each_possible_cpu(i) {
8166 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8167 GFP_KERNEL, cpu_to_node(i));
8171 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8172 GFP_KERNEL, cpu_to_node(i));
8176 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
8186 #else /* !CONFIG_RT_GROUP_SCHED */
8187 static inline void free_rt_sched_group(struct task_group *tg)
8192 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8196 #endif /* CONFIG_RT_GROUP_SCHED */
8198 #ifdef CONFIG_CGROUP_SCHED
8199 static void free_sched_group(struct task_group *tg)
8201 free_fair_sched_group(tg);
8202 free_rt_sched_group(tg);
8206 /* allocate runqueue etc for a new task group */
8207 struct task_group *sched_create_group(struct task_group *parent)
8209 struct task_group *tg;
8210 unsigned long flags;
8212 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8214 return ERR_PTR(-ENOMEM);
8216 if (!alloc_fair_sched_group(tg, parent))
8219 if (!alloc_rt_sched_group(tg, parent))
8222 spin_lock_irqsave(&task_group_lock, flags);
8223 list_add_rcu(&tg->list, &task_groups);
8225 WARN_ON(!parent); /* root should already exist */
8227 tg->parent = parent;
8228 INIT_LIST_HEAD(&tg->children);
8229 list_add_rcu(&tg->siblings, &parent->children);
8230 spin_unlock_irqrestore(&task_group_lock, flags);
8235 free_sched_group(tg);
8236 return ERR_PTR(-ENOMEM);
8239 /* rcu callback to free various structures associated with a task group */
8240 static void free_sched_group_rcu(struct rcu_head *rhp)
8242 /* now it should be safe to free those cfs_rqs */
8243 free_sched_group(container_of(rhp, struct task_group, rcu));
8246 /* Destroy runqueue etc associated with a task group */
8247 void sched_destroy_group(struct task_group *tg)
8249 unsigned long flags;
8252 /* end participation in shares distribution */
8253 for_each_possible_cpu(i)
8254 unregister_fair_sched_group(tg, i);
8256 spin_lock_irqsave(&task_group_lock, flags);
8257 list_del_rcu(&tg->list);
8258 list_del_rcu(&tg->siblings);
8259 spin_unlock_irqrestore(&task_group_lock, flags);
8261 /* wait for possible concurrent references to cfs_rqs complete */
8262 call_rcu(&tg->rcu, free_sched_group_rcu);
8265 /* change task's runqueue when it moves between groups.
8266 * The caller of this function should have put the task in its new group
8267 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8268 * reflect its new group.
8270 void sched_move_task(struct task_struct *tsk)
8273 unsigned long flags;
8276 rq = task_rq_lock(tsk, &flags);
8278 running = task_current(rq, tsk);
8279 on_rq = tsk->se.on_rq;
8282 dequeue_task(rq, tsk, 0);
8283 if (unlikely(running))
8284 tsk->sched_class->put_prev_task(rq, tsk);
8286 #ifdef CONFIG_FAIR_GROUP_SCHED
8287 if (tsk->sched_class->task_move_group)
8288 tsk->sched_class->task_move_group(tsk, on_rq);
8291 set_task_rq(tsk, task_cpu(tsk));
8293 if (unlikely(running))
8294 tsk->sched_class->set_curr_task(rq);
8296 enqueue_task(rq, tsk, 0);
8298 task_rq_unlock(rq, &flags);
8300 #endif /* CONFIG_CGROUP_SCHED */
8302 #ifdef CONFIG_FAIR_GROUP_SCHED
8303 static DEFINE_MUTEX(shares_mutex);
8305 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8308 unsigned long flags;
8311 * We can't change the weight of the root cgroup.
8316 if (shares < MIN_SHARES)
8317 shares = MIN_SHARES;
8318 else if (shares > MAX_SHARES)
8319 shares = MAX_SHARES;
8321 mutex_lock(&shares_mutex);
8322 if (tg->shares == shares)
8325 tg->shares = shares;
8326 for_each_possible_cpu(i) {
8327 struct rq *rq = cpu_rq(i);
8328 struct sched_entity *se;
8331 /* Propagate contribution to hierarchy */
8332 raw_spin_lock_irqsave(&rq->lock, flags);
8333 for_each_sched_entity(se)
8334 update_cfs_shares(group_cfs_rq(se), 0);
8335 raw_spin_unlock_irqrestore(&rq->lock, flags);
8339 mutex_unlock(&shares_mutex);
8343 unsigned long sched_group_shares(struct task_group *tg)
8349 #ifdef CONFIG_RT_GROUP_SCHED
8351 * Ensure that the real time constraints are schedulable.
8353 static DEFINE_MUTEX(rt_constraints_mutex);
8355 static unsigned long to_ratio(u64 period, u64 runtime)
8357 if (runtime == RUNTIME_INF)
8360 return div64_u64(runtime << 20, period);
8363 /* Must be called with tasklist_lock held */
8364 static inline int tg_has_rt_tasks(struct task_group *tg)
8366 struct task_struct *g, *p;
8368 do_each_thread(g, p) {
8369 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8371 } while_each_thread(g, p);
8376 struct rt_schedulable_data {
8377 struct task_group *tg;
8382 static int tg_schedulable(struct task_group *tg, void *data)
8384 struct rt_schedulable_data *d = data;
8385 struct task_group *child;
8386 unsigned long total, sum = 0;
8387 u64 period, runtime;
8389 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8390 runtime = tg->rt_bandwidth.rt_runtime;
8393 period = d->rt_period;
8394 runtime = d->rt_runtime;
8398 * Cannot have more runtime than the period.
8400 if (runtime > period && runtime != RUNTIME_INF)
8404 * Ensure we don't starve existing RT tasks.
8406 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8409 total = to_ratio(period, runtime);
8412 * Nobody can have more than the global setting allows.
8414 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8418 * The sum of our children's runtime should not exceed our own.
8420 list_for_each_entry_rcu(child, &tg->children, siblings) {
8421 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8422 runtime = child->rt_bandwidth.rt_runtime;
8424 if (child == d->tg) {
8425 period = d->rt_period;
8426 runtime = d->rt_runtime;
8429 sum += to_ratio(period, runtime);
8438 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8440 struct rt_schedulable_data data = {
8442 .rt_period = period,
8443 .rt_runtime = runtime,
8446 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8449 static int tg_set_bandwidth(struct task_group *tg,
8450 u64 rt_period, u64 rt_runtime)
8454 mutex_lock(&rt_constraints_mutex);
8455 read_lock(&tasklist_lock);
8456 err = __rt_schedulable(tg, rt_period, rt_runtime);
8460 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8461 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8462 tg->rt_bandwidth.rt_runtime = rt_runtime;
8464 for_each_possible_cpu(i) {
8465 struct rt_rq *rt_rq = tg->rt_rq[i];
8467 raw_spin_lock(&rt_rq->rt_runtime_lock);
8468 rt_rq->rt_runtime = rt_runtime;
8469 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8471 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8473 read_unlock(&tasklist_lock);
8474 mutex_unlock(&rt_constraints_mutex);
8479 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8481 u64 rt_runtime, rt_period;
8483 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8484 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8485 if (rt_runtime_us < 0)
8486 rt_runtime = RUNTIME_INF;
8488 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8491 long sched_group_rt_runtime(struct task_group *tg)
8495 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8498 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8499 do_div(rt_runtime_us, NSEC_PER_USEC);
8500 return rt_runtime_us;
8503 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8505 u64 rt_runtime, rt_period;
8507 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8508 rt_runtime = tg->rt_bandwidth.rt_runtime;
8513 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8516 long sched_group_rt_period(struct task_group *tg)
8520 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8521 do_div(rt_period_us, NSEC_PER_USEC);
8522 return rt_period_us;
8525 static int sched_rt_global_constraints(void)
8527 u64 runtime, period;
8530 if (sysctl_sched_rt_period <= 0)
8533 runtime = global_rt_runtime();
8534 period = global_rt_period();
8537 * Sanity check on the sysctl variables.
8539 if (runtime > period && runtime != RUNTIME_INF)
8542 mutex_lock(&rt_constraints_mutex);
8543 read_lock(&tasklist_lock);
8544 ret = __rt_schedulable(NULL, 0, 0);
8545 read_unlock(&tasklist_lock);
8546 mutex_unlock(&rt_constraints_mutex);
8551 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8553 /* Don't accept realtime tasks when there is no way for them to run */
8554 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8560 #else /* !CONFIG_RT_GROUP_SCHED */
8561 static int sched_rt_global_constraints(void)
8563 unsigned long flags;
8566 if (sysctl_sched_rt_period <= 0)
8570 * There's always some RT tasks in the root group
8571 * -- migration, kstopmachine etc..
8573 if (sysctl_sched_rt_runtime == 0)
8576 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8577 for_each_possible_cpu(i) {
8578 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8580 raw_spin_lock(&rt_rq->rt_runtime_lock);
8581 rt_rq->rt_runtime = global_rt_runtime();
8582 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8584 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8588 #endif /* CONFIG_RT_GROUP_SCHED */
8590 int sched_rt_handler(struct ctl_table *table, int write,
8591 void __user *buffer, size_t *lenp,
8595 int old_period, old_runtime;
8596 static DEFINE_MUTEX(mutex);
8599 old_period = sysctl_sched_rt_period;
8600 old_runtime = sysctl_sched_rt_runtime;
8602 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8604 if (!ret && write) {
8605 ret = sched_rt_global_constraints();
8607 sysctl_sched_rt_period = old_period;
8608 sysctl_sched_rt_runtime = old_runtime;
8610 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8611 def_rt_bandwidth.rt_period =
8612 ns_to_ktime(global_rt_period());
8615 mutex_unlock(&mutex);
8620 #ifdef CONFIG_CGROUP_SCHED
8622 /* return corresponding task_group object of a cgroup */
8623 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8625 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8626 struct task_group, css);
8629 static struct cgroup_subsys_state *
8630 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8632 struct task_group *tg, *parent;
8634 if (!cgrp->parent) {
8635 /* This is early initialization for the top cgroup */
8636 return &init_task_group.css;
8639 parent = cgroup_tg(cgrp->parent);
8640 tg = sched_create_group(parent);
8642 return ERR_PTR(-ENOMEM);
8648 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8650 struct task_group *tg = cgroup_tg(cgrp);
8652 sched_destroy_group(tg);
8656 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8658 #ifdef CONFIG_RT_GROUP_SCHED
8659 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
8662 /* We don't support RT-tasks being in separate groups */
8663 if (tsk->sched_class != &fair_sched_class)
8670 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8671 struct task_struct *tsk, bool threadgroup)
8673 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
8677 struct task_struct *c;
8679 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8680 retval = cpu_cgroup_can_attach_task(cgrp, c);
8692 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8693 struct cgroup *old_cont, struct task_struct *tsk,
8696 sched_move_task(tsk);
8698 struct task_struct *c;
8700 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8707 #ifdef CONFIG_FAIR_GROUP_SCHED
8708 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8711 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8714 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8716 struct task_group *tg = cgroup_tg(cgrp);
8718 return (u64) tg->shares;
8720 #endif /* CONFIG_FAIR_GROUP_SCHED */
8722 #ifdef CONFIG_RT_GROUP_SCHED
8723 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8726 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8729 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8731 return sched_group_rt_runtime(cgroup_tg(cgrp));
8734 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8737 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8740 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8742 return sched_group_rt_period(cgroup_tg(cgrp));
8744 #endif /* CONFIG_RT_GROUP_SCHED */
8746 static struct cftype cpu_files[] = {
8747 #ifdef CONFIG_FAIR_GROUP_SCHED
8750 .read_u64 = cpu_shares_read_u64,
8751 .write_u64 = cpu_shares_write_u64,
8754 #ifdef CONFIG_RT_GROUP_SCHED
8756 .name = "rt_runtime_us",
8757 .read_s64 = cpu_rt_runtime_read,
8758 .write_s64 = cpu_rt_runtime_write,
8761 .name = "rt_period_us",
8762 .read_u64 = cpu_rt_period_read_uint,
8763 .write_u64 = cpu_rt_period_write_uint,
8768 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8770 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8773 struct cgroup_subsys cpu_cgroup_subsys = {
8775 .create = cpu_cgroup_create,
8776 .destroy = cpu_cgroup_destroy,
8777 .can_attach = cpu_cgroup_can_attach,
8778 .attach = cpu_cgroup_attach,
8779 .populate = cpu_cgroup_populate,
8780 .subsys_id = cpu_cgroup_subsys_id,
8784 #endif /* CONFIG_CGROUP_SCHED */
8786 #ifdef CONFIG_CGROUP_CPUACCT
8789 * CPU accounting code for task groups.
8791 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8792 * (balbir@in.ibm.com).
8795 /* track cpu usage of a group of tasks and its child groups */
8797 struct cgroup_subsys_state css;
8798 /* cpuusage holds pointer to a u64-type object on every cpu */
8799 u64 __percpu *cpuusage;
8800 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
8801 struct cpuacct *parent;
8804 struct cgroup_subsys cpuacct_subsys;
8806 /* return cpu accounting group corresponding to this container */
8807 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8809 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8810 struct cpuacct, css);
8813 /* return cpu accounting group to which this task belongs */
8814 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8816 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8817 struct cpuacct, css);
8820 /* create a new cpu accounting group */
8821 static struct cgroup_subsys_state *cpuacct_create(
8822 struct cgroup_subsys *ss, struct cgroup *cgrp)
8824 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8830 ca->cpuusage = alloc_percpu(u64);
8834 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8835 if (percpu_counter_init(&ca->cpustat[i], 0))
8836 goto out_free_counters;
8839 ca->parent = cgroup_ca(cgrp->parent);
8845 percpu_counter_destroy(&ca->cpustat[i]);
8846 free_percpu(ca->cpuusage);
8850 return ERR_PTR(-ENOMEM);
8853 /* destroy an existing cpu accounting group */
8855 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8857 struct cpuacct *ca = cgroup_ca(cgrp);
8860 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8861 percpu_counter_destroy(&ca->cpustat[i]);
8862 free_percpu(ca->cpuusage);
8866 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
8868 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8871 #ifndef CONFIG_64BIT
8873 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8875 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8877 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8885 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
8887 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8889 #ifndef CONFIG_64BIT
8891 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8893 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8895 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8901 /* return total cpu usage (in nanoseconds) of a group */
8902 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8904 struct cpuacct *ca = cgroup_ca(cgrp);
8905 u64 totalcpuusage = 0;
8908 for_each_present_cpu(i)
8909 totalcpuusage += cpuacct_cpuusage_read(ca, i);
8911 return totalcpuusage;
8914 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8917 struct cpuacct *ca = cgroup_ca(cgrp);
8926 for_each_present_cpu(i)
8927 cpuacct_cpuusage_write(ca, i, 0);
8933 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
8936 struct cpuacct *ca = cgroup_ca(cgroup);
8940 for_each_present_cpu(i) {
8941 percpu = cpuacct_cpuusage_read(ca, i);
8942 seq_printf(m, "%llu ", (unsigned long long) percpu);
8944 seq_printf(m, "\n");
8948 static const char *cpuacct_stat_desc[] = {
8949 [CPUACCT_STAT_USER] = "user",
8950 [CPUACCT_STAT_SYSTEM] = "system",
8953 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
8954 struct cgroup_map_cb *cb)
8956 struct cpuacct *ca = cgroup_ca(cgrp);
8959 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
8960 s64 val = percpu_counter_read(&ca->cpustat[i]);
8961 val = cputime64_to_clock_t(val);
8962 cb->fill(cb, cpuacct_stat_desc[i], val);
8967 static struct cftype files[] = {
8970 .read_u64 = cpuusage_read,
8971 .write_u64 = cpuusage_write,
8974 .name = "usage_percpu",
8975 .read_seq_string = cpuacct_percpu_seq_read,
8979 .read_map = cpuacct_stats_show,
8983 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
8985 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
8989 * charge this task's execution time to its accounting group.
8991 * called with rq->lock held.
8993 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8998 if (unlikely(!cpuacct_subsys.active))
9001 cpu = task_cpu(tsk);
9007 for (; ca; ca = ca->parent) {
9008 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9009 *cpuusage += cputime;
9016 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9017 * in cputime_t units. As a result, cpuacct_update_stats calls
9018 * percpu_counter_add with values large enough to always overflow the
9019 * per cpu batch limit causing bad SMP scalability.
9021 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9022 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9023 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9026 #define CPUACCT_BATCH \
9027 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9029 #define CPUACCT_BATCH 0
9033 * Charge the system/user time to the task's accounting group.
9035 static void cpuacct_update_stats(struct task_struct *tsk,
9036 enum cpuacct_stat_index idx, cputime_t val)
9039 int batch = CPUACCT_BATCH;
9041 if (unlikely(!cpuacct_subsys.active))
9048 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9054 struct cgroup_subsys cpuacct_subsys = {
9056 .create = cpuacct_create,
9057 .destroy = cpuacct_destroy,
9058 .populate = cpuacct_populate,
9059 .subsys_id = cpuacct_subsys_id,
9061 #endif /* CONFIG_CGROUP_CPUACCT */
9065 void synchronize_sched_expedited(void)
9069 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9071 #else /* #ifndef CONFIG_SMP */
9073 static atomic_t synchronize_sched_expedited_count = ATOMIC_INIT(0);
9075 static int synchronize_sched_expedited_cpu_stop(void *data)
9078 * There must be a full memory barrier on each affected CPU
9079 * between the time that try_stop_cpus() is called and the
9080 * time that it returns.
9082 * In the current initial implementation of cpu_stop, the
9083 * above condition is already met when the control reaches
9084 * this point and the following smp_mb() is not strictly
9085 * necessary. Do smp_mb() anyway for documentation and
9086 * robustness against future implementation changes.
9088 smp_mb(); /* See above comment block. */
9093 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
9094 * approach to force grace period to end quickly. This consumes
9095 * significant time on all CPUs, and is thus not recommended for
9096 * any sort of common-case code.
9098 * Note that it is illegal to call this function while holding any
9099 * lock that is acquired by a CPU-hotplug notifier. Failing to
9100 * observe this restriction will result in deadlock.
9102 void synchronize_sched_expedited(void)
9104 int snap, trycount = 0;
9106 smp_mb(); /* ensure prior mod happens before capturing snap. */
9107 snap = atomic_read(&synchronize_sched_expedited_count) + 1;
9109 while (try_stop_cpus(cpu_online_mask,
9110 synchronize_sched_expedited_cpu_stop,
9113 if (trycount++ < 10)
9114 udelay(trycount * num_online_cpus());
9116 synchronize_sched();
9119 if (atomic_read(&synchronize_sched_expedited_count) - snap > 0) {
9120 smp_mb(); /* ensure test happens before caller kfree */
9125 atomic_inc(&synchronize_sched_expedited_count);
9126 smp_mb__after_atomic_inc(); /* ensure post-GP actions seen after GP. */
9129 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9131 #endif /* #else #ifndef CONFIG_SMP */