4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/perf_event.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/kthread.h>
59 #include <linux/proc_fs.h>
60 #include <linux/seq_file.h>
61 #include <linux/sysctl.h>
62 #include <linux/syscalls.h>
63 #include <linux/times.h>
64 #include <linux/tsacct_kern.h>
65 #include <linux/kprobes.h>
66 #include <linux/delayacct.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
73 #include <linux/ftrace.h>
74 #include <linux/slab.h>
77 #include <asm/irq_regs.h>
79 #include "sched_cpupri.h"
81 #define CREATE_TRACE_POINTS
82 #include <trace/events/sched.h>
85 * Convert user-nice values [ -20 ... 0 ... 19 ]
86 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
89 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
90 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
91 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
94 * 'User priority' is the nice value converted to something we
95 * can work with better when scaling various scheduler parameters,
96 * it's a [ 0 ... 39 ] range.
98 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
99 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
100 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
103 * Helpers for converting nanosecond timing to jiffy resolution
105 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
107 #define NICE_0_LOAD SCHED_LOAD_SCALE
108 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
111 * These are the 'tuning knobs' of the scheduler:
113 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
114 * Timeslices get refilled after they expire.
116 #define DEF_TIMESLICE (100 * HZ / 1000)
119 * single value that denotes runtime == period, ie unlimited time.
121 #define RUNTIME_INF ((u64)~0ULL)
123 static inline int rt_policy(int policy)
125 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
130 static inline int task_has_rt_policy(struct task_struct *p)
132 return rt_policy(p->policy);
136 * This is the priority-queue data structure of the RT scheduling class:
138 struct rt_prio_array {
139 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
140 struct list_head queue[MAX_RT_PRIO];
143 struct rt_bandwidth {
144 /* nests inside the rq lock: */
145 raw_spinlock_t rt_runtime_lock;
148 struct hrtimer rt_period_timer;
151 static struct rt_bandwidth def_rt_bandwidth;
153 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
155 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
157 struct rt_bandwidth *rt_b =
158 container_of(timer, struct rt_bandwidth, rt_period_timer);
164 now = hrtimer_cb_get_time(timer);
165 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
170 idle = do_sched_rt_period_timer(rt_b, overrun);
173 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
177 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
179 rt_b->rt_period = ns_to_ktime(period);
180 rt_b->rt_runtime = runtime;
182 raw_spin_lock_init(&rt_b->rt_runtime_lock);
184 hrtimer_init(&rt_b->rt_period_timer,
185 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
186 rt_b->rt_period_timer.function = sched_rt_period_timer;
189 static inline int rt_bandwidth_enabled(void)
191 return sysctl_sched_rt_runtime >= 0;
194 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
198 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
201 if (hrtimer_active(&rt_b->rt_period_timer))
204 raw_spin_lock(&rt_b->rt_runtime_lock);
209 if (hrtimer_active(&rt_b->rt_period_timer))
212 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
213 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
215 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
216 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
217 delta = ktime_to_ns(ktime_sub(hard, soft));
218 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
219 HRTIMER_MODE_ABS_PINNED, 0);
221 raw_spin_unlock(&rt_b->rt_runtime_lock);
224 #ifdef CONFIG_RT_GROUP_SCHED
225 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
227 hrtimer_cancel(&rt_b->rt_period_timer);
232 * sched_domains_mutex serializes calls to arch_init_sched_domains,
233 * detach_destroy_domains and partition_sched_domains.
235 static DEFINE_MUTEX(sched_domains_mutex);
237 #ifdef CONFIG_CGROUP_SCHED
239 #include <linux/cgroup.h>
243 static LIST_HEAD(task_groups);
245 /* task group related information */
247 struct cgroup_subsys_state css;
249 #ifdef CONFIG_FAIR_GROUP_SCHED
250 /* schedulable entities of this group on each cpu */
251 struct sched_entity **se;
252 /* runqueue "owned" by this group on each cpu */
253 struct cfs_rq **cfs_rq;
254 unsigned long shares;
257 #ifdef CONFIG_RT_GROUP_SCHED
258 struct sched_rt_entity **rt_se;
259 struct rt_rq **rt_rq;
261 struct rt_bandwidth rt_bandwidth;
265 struct list_head list;
267 struct task_group *parent;
268 struct list_head siblings;
269 struct list_head children;
272 #define root_task_group init_task_group
274 /* task_group_lock serializes add/remove of task groups and also changes to
275 * a task group's cpu shares.
277 static DEFINE_SPINLOCK(task_group_lock);
279 #ifdef CONFIG_FAIR_GROUP_SCHED
282 static int root_task_group_empty(void)
284 return list_empty(&root_task_group.children);
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 /* return group to which a task belongs */
310 static inline struct task_group *task_group(struct task_struct *p)
312 struct task_group *tg;
314 #ifdef CONFIG_CGROUP_SCHED
315 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
316 struct task_group, css);
318 tg = &init_task_group;
323 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
324 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
327 * Strictly speaking this rcu_read_lock() is not needed since the
328 * task_group is tied to the cgroup, which in turn can never go away
329 * as long as there are tasks attached to it.
331 * However since task_group() uses task_subsys_state() which is an
332 * rcu_dereference() user, this quiets CONFIG_PROVE_RCU.
335 #ifdef CONFIG_FAIR_GROUP_SCHED
336 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
337 p->se.parent = task_group(p)->se[cpu];
340 #ifdef CONFIG_RT_GROUP_SCHED
341 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
342 p->rt.parent = task_group(p)->rt_se[cpu];
349 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
350 static inline struct task_group *task_group(struct task_struct *p)
355 #endif /* CONFIG_CGROUP_SCHED */
357 /* CFS-related fields in a runqueue */
359 struct load_weight load;
360 unsigned long nr_running;
365 struct rb_root tasks_timeline;
366 struct rb_node *rb_leftmost;
368 struct list_head tasks;
369 struct list_head *balance_iterator;
372 * 'curr' points to currently running entity on this cfs_rq.
373 * It is set to NULL otherwise (i.e when none are currently running).
375 struct sched_entity *curr, *next, *last;
377 unsigned int nr_spread_over;
379 #ifdef CONFIG_FAIR_GROUP_SCHED
380 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
383 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
384 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
385 * (like users, containers etc.)
387 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
388 * list is used during load balance.
390 struct list_head leaf_cfs_rq_list;
391 struct task_group *tg; /* group that "owns" this runqueue */
395 * the part of load.weight contributed by tasks
397 unsigned long task_weight;
400 * h_load = weight * f(tg)
402 * Where f(tg) is the recursive weight fraction assigned to
405 unsigned long h_load;
408 * this cpu's part of tg->shares
410 unsigned long shares;
413 * load.weight at the time we set shares
415 unsigned long rq_weight;
420 /* Real-Time classes' related field in a runqueue: */
422 struct rt_prio_array active;
423 unsigned long rt_nr_running;
424 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
426 int curr; /* highest queued rt task prio */
428 int next; /* next highest */
433 unsigned long rt_nr_migratory;
434 unsigned long rt_nr_total;
436 struct plist_head pushable_tasks;
441 /* Nests inside the rq lock: */
442 raw_spinlock_t rt_runtime_lock;
444 #ifdef CONFIG_RT_GROUP_SCHED
445 unsigned long rt_nr_boosted;
448 struct list_head leaf_rt_rq_list;
449 struct task_group *tg;
456 * We add the notion of a root-domain which will be used to define per-domain
457 * variables. Each exclusive cpuset essentially defines an island domain by
458 * fully partitioning the member cpus from any other cpuset. Whenever a new
459 * exclusive cpuset is created, we also create and attach a new root-domain
466 cpumask_var_t online;
469 * The "RT overload" flag: it gets set if a CPU has more than
470 * one runnable RT task.
472 cpumask_var_t rto_mask;
475 struct cpupri cpupri;
480 * By default the system creates a single root-domain with all cpus as
481 * members (mimicking the global state we have today).
483 static struct root_domain def_root_domain;
488 * This is the main, per-CPU runqueue data structure.
490 * Locking rule: those places that want to lock multiple runqueues
491 * (such as the load balancing or the thread migration code), lock
492 * acquire operations must be ordered by ascending &runqueue.
499 * nr_running and cpu_load should be in the same cacheline because
500 * remote CPUs use both these fields when doing load calculation.
502 unsigned long nr_running;
503 #define CPU_LOAD_IDX_MAX 5
504 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
506 unsigned char in_nohz_recently;
508 /* capture load from *all* tasks on this cpu: */
509 struct load_weight load;
510 unsigned long nr_load_updates;
516 #ifdef CONFIG_FAIR_GROUP_SCHED
517 /* list of leaf cfs_rq on this cpu: */
518 struct list_head leaf_cfs_rq_list;
520 #ifdef CONFIG_RT_GROUP_SCHED
521 struct list_head leaf_rt_rq_list;
525 * This is part of a global counter where only the total sum
526 * over all CPUs matters. A task can increase this counter on
527 * one CPU and if it got migrated afterwards it may decrease
528 * it on another CPU. Always updated under the runqueue lock:
530 unsigned long nr_uninterruptible;
532 struct task_struct *curr, *idle;
533 unsigned long next_balance;
534 struct mm_struct *prev_mm;
541 struct root_domain *rd;
542 struct sched_domain *sd;
544 unsigned char idle_at_tick;
545 /* For active balancing */
549 /* cpu of this runqueue: */
553 unsigned long avg_load_per_task;
555 struct task_struct *migration_thread;
556 struct list_head migration_queue;
564 /* calc_load related fields */
565 unsigned long calc_load_update;
566 long calc_load_active;
568 #ifdef CONFIG_SCHED_HRTICK
570 int hrtick_csd_pending;
571 struct call_single_data hrtick_csd;
573 struct hrtimer hrtick_timer;
576 #ifdef CONFIG_SCHEDSTATS
578 struct sched_info rq_sched_info;
579 unsigned long long rq_cpu_time;
580 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
582 /* sys_sched_yield() stats */
583 unsigned int yld_count;
585 /* schedule() stats */
586 unsigned int sched_switch;
587 unsigned int sched_count;
588 unsigned int sched_goidle;
590 /* try_to_wake_up() stats */
591 unsigned int ttwu_count;
592 unsigned int ttwu_local;
595 unsigned int bkl_count;
599 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
602 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
604 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
607 static inline int cpu_of(struct rq *rq)
616 #define rcu_dereference_check_sched_domain(p) \
617 rcu_dereference_check((p), \
618 rcu_read_lock_sched_held() || \
619 lockdep_is_held(&sched_domains_mutex))
622 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
623 * See detach_destroy_domains: synchronize_sched for details.
625 * The domain tree of any CPU may only be accessed from within
626 * preempt-disabled sections.
628 #define for_each_domain(cpu, __sd) \
629 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
631 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
632 #define this_rq() (&__get_cpu_var(runqueues))
633 #define task_rq(p) cpu_rq(task_cpu(p))
634 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
635 #define raw_rq() (&__raw_get_cpu_var(runqueues))
637 inline void update_rq_clock(struct rq *rq)
639 rq->clock = sched_clock_cpu(cpu_of(rq));
643 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
645 #ifdef CONFIG_SCHED_DEBUG
646 # define const_debug __read_mostly
648 # define const_debug static const
653 * @cpu: the processor in question.
655 * Returns true if the current cpu runqueue is locked.
656 * This interface allows printk to be called with the runqueue lock
657 * held and know whether or not it is OK to wake up the klogd.
659 int runqueue_is_locked(int cpu)
661 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
665 * Debugging: various feature bits
668 #define SCHED_FEAT(name, enabled) \
669 __SCHED_FEAT_##name ,
672 #include "sched_features.h"
677 #define SCHED_FEAT(name, enabled) \
678 (1UL << __SCHED_FEAT_##name) * enabled |
680 const_debug unsigned int sysctl_sched_features =
681 #include "sched_features.h"
686 #ifdef CONFIG_SCHED_DEBUG
687 #define SCHED_FEAT(name, enabled) \
690 static __read_mostly char *sched_feat_names[] = {
691 #include "sched_features.h"
697 static int sched_feat_show(struct seq_file *m, void *v)
701 for (i = 0; sched_feat_names[i]; i++) {
702 if (!(sysctl_sched_features & (1UL << i)))
704 seq_printf(m, "%s ", sched_feat_names[i]);
712 sched_feat_write(struct file *filp, const char __user *ubuf,
713 size_t cnt, loff_t *ppos)
723 if (copy_from_user(&buf, ubuf, cnt))
728 if (strncmp(buf, "NO_", 3) == 0) {
733 for (i = 0; sched_feat_names[i]; i++) {
734 int len = strlen(sched_feat_names[i]);
736 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
738 sysctl_sched_features &= ~(1UL << i);
740 sysctl_sched_features |= (1UL << i);
745 if (!sched_feat_names[i])
753 static int sched_feat_open(struct inode *inode, struct file *filp)
755 return single_open(filp, sched_feat_show, NULL);
758 static const struct file_operations sched_feat_fops = {
759 .open = sched_feat_open,
760 .write = sched_feat_write,
763 .release = single_release,
766 static __init int sched_init_debug(void)
768 debugfs_create_file("sched_features", 0644, NULL, NULL,
773 late_initcall(sched_init_debug);
777 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
780 * Number of tasks to iterate in a single balance run.
781 * Limited because this is done with IRQs disabled.
783 const_debug unsigned int sysctl_sched_nr_migrate = 32;
786 * ratelimit for updating the group shares.
789 unsigned int sysctl_sched_shares_ratelimit = 250000;
790 unsigned int normalized_sysctl_sched_shares_ratelimit = 250000;
793 * Inject some fuzzyness into changing the per-cpu group shares
794 * this avoids remote rq-locks at the expense of fairness.
797 unsigned int sysctl_sched_shares_thresh = 4;
800 * period over which we average the RT time consumption, measured
805 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
808 * period over which we measure -rt task cpu usage in us.
811 unsigned int sysctl_sched_rt_period = 1000000;
813 static __read_mostly int scheduler_running;
816 * part of the period that we allow rt tasks to run in us.
819 int sysctl_sched_rt_runtime = 950000;
821 static inline u64 global_rt_period(void)
823 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
826 static inline u64 global_rt_runtime(void)
828 if (sysctl_sched_rt_runtime < 0)
831 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
834 #ifndef prepare_arch_switch
835 # define prepare_arch_switch(next) do { } while (0)
837 #ifndef finish_arch_switch
838 # define finish_arch_switch(prev) do { } while (0)
841 static inline int task_current(struct rq *rq, struct task_struct *p)
843 return rq->curr == p;
846 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
847 static inline int task_running(struct rq *rq, struct task_struct *p)
849 return task_current(rq, p);
852 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
856 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
858 #ifdef CONFIG_DEBUG_SPINLOCK
859 /* this is a valid case when another task releases the spinlock */
860 rq->lock.owner = current;
863 * If we are tracking spinlock dependencies then we have to
864 * fix up the runqueue lock - which gets 'carried over' from
867 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
869 raw_spin_unlock_irq(&rq->lock);
872 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
873 static inline int task_running(struct rq *rq, struct task_struct *p)
878 return task_current(rq, p);
882 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
886 * We can optimise this out completely for !SMP, because the
887 * SMP rebalancing from interrupt is the only thing that cares
892 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
893 raw_spin_unlock_irq(&rq->lock);
895 raw_spin_unlock(&rq->lock);
899 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
903 * After ->oncpu is cleared, the task can be moved to a different CPU.
904 * We must ensure this doesn't happen until the switch is completely
910 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
914 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
917 * Check whether the task is waking, we use this to synchronize against
918 * ttwu() so that task_cpu() reports a stable number.
920 * We need to make an exception for PF_STARTING tasks because the fork
921 * path might require task_rq_lock() to work, eg. it can call
922 * set_cpus_allowed_ptr() from the cpuset clone_ns code.
924 static inline int task_is_waking(struct task_struct *p)
926 return unlikely((p->state == TASK_WAKING) && !(p->flags & PF_STARTING));
930 * __task_rq_lock - lock the runqueue a given task resides on.
931 * Must be called interrupts disabled.
933 static inline struct rq *__task_rq_lock(struct task_struct *p)
939 while (task_is_waking(p))
942 raw_spin_lock(&rq->lock);
943 if (likely(rq == task_rq(p) && !task_is_waking(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 while (task_is_waking(p))
962 local_irq_save(*flags);
964 raw_spin_lock(&rq->lock);
965 if (likely(rq == task_rq(p) && !task_is_waking(p)))
967 raw_spin_unlock_irqrestore(&rq->lock, *flags);
971 void task_rq_unlock_wait(struct task_struct *p)
973 struct rq *rq = task_rq(p);
975 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
976 raw_spin_unlock_wait(&rq->lock);
979 static void __task_rq_unlock(struct rq *rq)
982 raw_spin_unlock(&rq->lock);
985 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
988 raw_spin_unlock_irqrestore(&rq->lock, *flags);
992 * this_rq_lock - lock this runqueue and disable interrupts.
994 static struct rq *this_rq_lock(void)
1001 raw_spin_lock(&rq->lock);
1006 #ifdef CONFIG_SCHED_HRTICK
1008 * Use HR-timers to deliver accurate preemption points.
1010 * Its all a bit involved since we cannot program an hrt while holding the
1011 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1014 * When we get rescheduled we reprogram the hrtick_timer outside of the
1020 * - enabled by features
1021 * - hrtimer is actually high res
1023 static inline int hrtick_enabled(struct rq *rq)
1025 if (!sched_feat(HRTICK))
1027 if (!cpu_active(cpu_of(rq)))
1029 return hrtimer_is_hres_active(&rq->hrtick_timer);
1032 static void hrtick_clear(struct rq *rq)
1034 if (hrtimer_active(&rq->hrtick_timer))
1035 hrtimer_cancel(&rq->hrtick_timer);
1039 * High-resolution timer tick.
1040 * Runs from hardirq context with interrupts disabled.
1042 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1044 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1046 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1048 raw_spin_lock(&rq->lock);
1049 update_rq_clock(rq);
1050 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1051 raw_spin_unlock(&rq->lock);
1053 return HRTIMER_NORESTART;
1058 * called from hardirq (IPI) context
1060 static void __hrtick_start(void *arg)
1062 struct rq *rq = arg;
1064 raw_spin_lock(&rq->lock);
1065 hrtimer_restart(&rq->hrtick_timer);
1066 rq->hrtick_csd_pending = 0;
1067 raw_spin_unlock(&rq->lock);
1071 * Called to set the hrtick timer state.
1073 * called with rq->lock held and irqs disabled
1075 static void hrtick_start(struct rq *rq, u64 delay)
1077 struct hrtimer *timer = &rq->hrtick_timer;
1078 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1080 hrtimer_set_expires(timer, time);
1082 if (rq == this_rq()) {
1083 hrtimer_restart(timer);
1084 } else if (!rq->hrtick_csd_pending) {
1085 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1086 rq->hrtick_csd_pending = 1;
1091 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1093 int cpu = (int)(long)hcpu;
1096 case CPU_UP_CANCELED:
1097 case CPU_UP_CANCELED_FROZEN:
1098 case CPU_DOWN_PREPARE:
1099 case CPU_DOWN_PREPARE_FROZEN:
1101 case CPU_DEAD_FROZEN:
1102 hrtick_clear(cpu_rq(cpu));
1109 static __init void init_hrtick(void)
1111 hotcpu_notifier(hotplug_hrtick, 0);
1115 * Called to set the hrtick timer state.
1117 * called with rq->lock held and irqs disabled
1119 static void hrtick_start(struct rq *rq, u64 delay)
1121 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1122 HRTIMER_MODE_REL_PINNED, 0);
1125 static inline void init_hrtick(void)
1128 #endif /* CONFIG_SMP */
1130 static void init_rq_hrtick(struct rq *rq)
1133 rq->hrtick_csd_pending = 0;
1135 rq->hrtick_csd.flags = 0;
1136 rq->hrtick_csd.func = __hrtick_start;
1137 rq->hrtick_csd.info = rq;
1140 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1141 rq->hrtick_timer.function = hrtick;
1143 #else /* CONFIG_SCHED_HRTICK */
1144 static inline void hrtick_clear(struct rq *rq)
1148 static inline void init_rq_hrtick(struct rq *rq)
1152 static inline void init_hrtick(void)
1155 #endif /* CONFIG_SCHED_HRTICK */
1158 * resched_task - mark a task 'to be rescheduled now'.
1160 * On UP this means the setting of the need_resched flag, on SMP it
1161 * might also involve a cross-CPU call to trigger the scheduler on
1166 #ifndef tsk_is_polling
1167 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1170 static void resched_task(struct task_struct *p)
1174 assert_raw_spin_locked(&task_rq(p)->lock);
1176 if (test_tsk_need_resched(p))
1179 set_tsk_need_resched(p);
1182 if (cpu == smp_processor_id())
1185 /* NEED_RESCHED must be visible before we test polling */
1187 if (!tsk_is_polling(p))
1188 smp_send_reschedule(cpu);
1191 static void resched_cpu(int cpu)
1193 struct rq *rq = cpu_rq(cpu);
1194 unsigned long flags;
1196 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1198 resched_task(cpu_curr(cpu));
1199 raw_spin_unlock_irqrestore(&rq->lock, flags);
1204 * When add_timer_on() enqueues a timer into the timer wheel of an
1205 * idle CPU then this timer might expire before the next timer event
1206 * which is scheduled to wake up that CPU. In case of a completely
1207 * idle system the next event might even be infinite time into the
1208 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1209 * leaves the inner idle loop so the newly added timer is taken into
1210 * account when the CPU goes back to idle and evaluates the timer
1211 * wheel for the next timer event.
1213 void wake_up_idle_cpu(int cpu)
1215 struct rq *rq = cpu_rq(cpu);
1217 if (cpu == smp_processor_id())
1221 * This is safe, as this function is called with the timer
1222 * wheel base lock of (cpu) held. When the CPU is on the way
1223 * to idle and has not yet set rq->curr to idle then it will
1224 * be serialized on the timer wheel base lock and take the new
1225 * timer into account automatically.
1227 if (rq->curr != rq->idle)
1231 * We can set TIF_RESCHED on the idle task of the other CPU
1232 * lockless. The worst case is that the other CPU runs the
1233 * idle task through an additional NOOP schedule()
1235 set_tsk_need_resched(rq->idle);
1237 /* NEED_RESCHED must be visible before we test polling */
1239 if (!tsk_is_polling(rq->idle))
1240 smp_send_reschedule(cpu);
1242 #endif /* CONFIG_NO_HZ */
1244 static u64 sched_avg_period(void)
1246 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1249 static void sched_avg_update(struct rq *rq)
1251 s64 period = sched_avg_period();
1253 while ((s64)(rq->clock - rq->age_stamp) > period) {
1254 rq->age_stamp += period;
1259 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1261 rq->rt_avg += rt_delta;
1262 sched_avg_update(rq);
1265 #else /* !CONFIG_SMP */
1266 static void resched_task(struct task_struct *p)
1268 assert_raw_spin_locked(&task_rq(p)->lock);
1269 set_tsk_need_resched(p);
1272 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1275 #endif /* CONFIG_SMP */
1277 #if BITS_PER_LONG == 32
1278 # define WMULT_CONST (~0UL)
1280 # define WMULT_CONST (1UL << 32)
1283 #define WMULT_SHIFT 32
1286 * Shift right and round:
1288 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1291 * delta *= weight / lw
1293 static unsigned long
1294 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1295 struct load_weight *lw)
1299 if (!lw->inv_weight) {
1300 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1303 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1307 tmp = (u64)delta_exec * weight;
1309 * Check whether we'd overflow the 64-bit multiplication:
1311 if (unlikely(tmp > WMULT_CONST))
1312 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1315 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1317 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1320 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1326 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1333 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1334 * of tasks with abnormal "nice" values across CPUs the contribution that
1335 * each task makes to its run queue's load is weighted according to its
1336 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1337 * scaled version of the new time slice allocation that they receive on time
1341 #define WEIGHT_IDLEPRIO 3
1342 #define WMULT_IDLEPRIO 1431655765
1345 * Nice levels are multiplicative, with a gentle 10% change for every
1346 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1347 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1348 * that remained on nice 0.
1350 * The "10% effect" is relative and cumulative: from _any_ nice level,
1351 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1352 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1353 * If a task goes up by ~10% and another task goes down by ~10% then
1354 * the relative distance between them is ~25%.)
1356 static const int prio_to_weight[40] = {
1357 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1358 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1359 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1360 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1361 /* 0 */ 1024, 820, 655, 526, 423,
1362 /* 5 */ 335, 272, 215, 172, 137,
1363 /* 10 */ 110, 87, 70, 56, 45,
1364 /* 15 */ 36, 29, 23, 18, 15,
1368 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1370 * In cases where the weight does not change often, we can use the
1371 * precalculated inverse to speed up arithmetics by turning divisions
1372 * into multiplications:
1374 static const u32 prio_to_wmult[40] = {
1375 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1376 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1377 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1378 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1379 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1380 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1381 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1382 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1385 /* Time spent by the tasks of the cpu accounting group executing in ... */
1386 enum cpuacct_stat_index {
1387 CPUACCT_STAT_USER, /* ... user mode */
1388 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1390 CPUACCT_STAT_NSTATS,
1393 #ifdef CONFIG_CGROUP_CPUACCT
1394 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1395 static void cpuacct_update_stats(struct task_struct *tsk,
1396 enum cpuacct_stat_index idx, cputime_t val);
1398 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1399 static inline void cpuacct_update_stats(struct task_struct *tsk,
1400 enum cpuacct_stat_index idx, cputime_t val) {}
1403 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1405 update_load_add(&rq->load, load);
1408 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1410 update_load_sub(&rq->load, load);
1413 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1414 typedef int (*tg_visitor)(struct task_group *, void *);
1417 * Iterate the full tree, calling @down when first entering a node and @up when
1418 * leaving it for the final time.
1420 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1422 struct task_group *parent, *child;
1426 parent = &root_task_group;
1428 ret = (*down)(parent, data);
1431 list_for_each_entry_rcu(child, &parent->children, siblings) {
1438 ret = (*up)(parent, data);
1443 parent = parent->parent;
1452 static int tg_nop(struct task_group *tg, void *data)
1459 /* Used instead of source_load when we know the type == 0 */
1460 static unsigned long weighted_cpuload(const int cpu)
1462 return cpu_rq(cpu)->load.weight;
1466 * Return a low guess at the load of a migration-source cpu weighted
1467 * according to the scheduling class and "nice" value.
1469 * We want to under-estimate the load of migration sources, to
1470 * balance conservatively.
1472 static unsigned long source_load(int cpu, int type)
1474 struct rq *rq = cpu_rq(cpu);
1475 unsigned long total = weighted_cpuload(cpu);
1477 if (type == 0 || !sched_feat(LB_BIAS))
1480 return min(rq->cpu_load[type-1], total);
1484 * Return a high guess at the load of a migration-target cpu weighted
1485 * according to the scheduling class and "nice" value.
1487 static unsigned long target_load(int cpu, int type)
1489 struct rq *rq = cpu_rq(cpu);
1490 unsigned long total = weighted_cpuload(cpu);
1492 if (type == 0 || !sched_feat(LB_BIAS))
1495 return max(rq->cpu_load[type-1], total);
1498 static struct sched_group *group_of(int cpu)
1500 struct sched_domain *sd = rcu_dereference_sched(cpu_rq(cpu)->sd);
1508 static unsigned long power_of(int cpu)
1510 struct sched_group *group = group_of(cpu);
1513 return SCHED_LOAD_SCALE;
1515 return group->cpu_power;
1518 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1520 static unsigned long cpu_avg_load_per_task(int cpu)
1522 struct rq *rq = cpu_rq(cpu);
1523 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1526 rq->avg_load_per_task = rq->load.weight / nr_running;
1528 rq->avg_load_per_task = 0;
1530 return rq->avg_load_per_task;
1533 #ifdef CONFIG_FAIR_GROUP_SCHED
1535 static __read_mostly unsigned long __percpu *update_shares_data;
1537 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1540 * Calculate and set the cpu's group shares.
1542 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1543 unsigned long sd_shares,
1544 unsigned long sd_rq_weight,
1545 unsigned long *usd_rq_weight)
1547 unsigned long shares, rq_weight;
1550 rq_weight = usd_rq_weight[cpu];
1553 rq_weight = NICE_0_LOAD;
1557 * \Sum_j shares_j * rq_weight_i
1558 * shares_i = -----------------------------
1559 * \Sum_j rq_weight_j
1561 shares = (sd_shares * rq_weight) / sd_rq_weight;
1562 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1564 if (abs(shares - tg->se[cpu]->load.weight) >
1565 sysctl_sched_shares_thresh) {
1566 struct rq *rq = cpu_rq(cpu);
1567 unsigned long flags;
1569 raw_spin_lock_irqsave(&rq->lock, flags);
1570 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1571 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1572 __set_se_shares(tg->se[cpu], shares);
1573 raw_spin_unlock_irqrestore(&rq->lock, flags);
1578 * Re-compute the task group their per cpu shares over the given domain.
1579 * This needs to be done in a bottom-up fashion because the rq weight of a
1580 * parent group depends on the shares of its child groups.
1582 static int tg_shares_up(struct task_group *tg, void *data)
1584 unsigned long weight, rq_weight = 0, sum_weight = 0, shares = 0;
1585 unsigned long *usd_rq_weight;
1586 struct sched_domain *sd = data;
1587 unsigned long flags;
1593 local_irq_save(flags);
1594 usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id());
1596 for_each_cpu(i, sched_domain_span(sd)) {
1597 weight = tg->cfs_rq[i]->load.weight;
1598 usd_rq_weight[i] = weight;
1600 rq_weight += weight;
1602 * If there are currently no tasks on the cpu pretend there
1603 * is one of average load so that when a new task gets to
1604 * run here it will not get delayed by group starvation.
1607 weight = NICE_0_LOAD;
1609 sum_weight += weight;
1610 shares += tg->cfs_rq[i]->shares;
1614 rq_weight = sum_weight;
1616 if ((!shares && rq_weight) || shares > tg->shares)
1617 shares = tg->shares;
1619 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1620 shares = tg->shares;
1622 for_each_cpu(i, sched_domain_span(sd))
1623 update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight);
1625 local_irq_restore(flags);
1631 * Compute the cpu's hierarchical load factor for each task group.
1632 * This needs to be done in a top-down fashion because the load of a child
1633 * group is a fraction of its parents load.
1635 static int tg_load_down(struct task_group *tg, void *data)
1638 long cpu = (long)data;
1641 load = cpu_rq(cpu)->load.weight;
1643 load = tg->parent->cfs_rq[cpu]->h_load;
1644 load *= tg->cfs_rq[cpu]->shares;
1645 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1648 tg->cfs_rq[cpu]->h_load = load;
1653 static void update_shares(struct sched_domain *sd)
1658 if (root_task_group_empty())
1661 now = cpu_clock(raw_smp_processor_id());
1662 elapsed = now - sd->last_update;
1664 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1665 sd->last_update = now;
1666 walk_tg_tree(tg_nop, tg_shares_up, sd);
1670 static void update_h_load(long cpu)
1672 if (root_task_group_empty())
1675 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1680 static inline void update_shares(struct sched_domain *sd)
1686 #ifdef CONFIG_PREEMPT
1688 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1691 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1692 * way at the expense of forcing extra atomic operations in all
1693 * invocations. This assures that the double_lock is acquired using the
1694 * same underlying policy as the spinlock_t on this architecture, which
1695 * reduces latency compared to the unfair variant below. However, it
1696 * also adds more overhead and therefore may reduce throughput.
1698 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1699 __releases(this_rq->lock)
1700 __acquires(busiest->lock)
1701 __acquires(this_rq->lock)
1703 raw_spin_unlock(&this_rq->lock);
1704 double_rq_lock(this_rq, busiest);
1711 * Unfair double_lock_balance: Optimizes throughput at the expense of
1712 * latency by eliminating extra atomic operations when the locks are
1713 * already in proper order on entry. This favors lower cpu-ids and will
1714 * grant the double lock to lower cpus over higher ids under contention,
1715 * regardless of entry order into the function.
1717 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1718 __releases(this_rq->lock)
1719 __acquires(busiest->lock)
1720 __acquires(this_rq->lock)
1724 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1725 if (busiest < this_rq) {
1726 raw_spin_unlock(&this_rq->lock);
1727 raw_spin_lock(&busiest->lock);
1728 raw_spin_lock_nested(&this_rq->lock,
1729 SINGLE_DEPTH_NESTING);
1732 raw_spin_lock_nested(&busiest->lock,
1733 SINGLE_DEPTH_NESTING);
1738 #endif /* CONFIG_PREEMPT */
1741 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1743 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1745 if (unlikely(!irqs_disabled())) {
1746 /* printk() doesn't work good under rq->lock */
1747 raw_spin_unlock(&this_rq->lock);
1751 return _double_lock_balance(this_rq, busiest);
1754 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1755 __releases(busiest->lock)
1757 raw_spin_unlock(&busiest->lock);
1758 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1762 * double_rq_lock - safely lock two runqueues
1764 * Note this does not disable interrupts like task_rq_lock,
1765 * you need to do so manually before calling.
1767 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1768 __acquires(rq1->lock)
1769 __acquires(rq2->lock)
1771 BUG_ON(!irqs_disabled());
1773 raw_spin_lock(&rq1->lock);
1774 __acquire(rq2->lock); /* Fake it out ;) */
1777 raw_spin_lock(&rq1->lock);
1778 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1780 raw_spin_lock(&rq2->lock);
1781 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1784 update_rq_clock(rq1);
1785 update_rq_clock(rq2);
1789 * double_rq_unlock - safely unlock two runqueues
1791 * Note this does not restore interrupts like task_rq_unlock,
1792 * you need to do so manually after calling.
1794 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1795 __releases(rq1->lock)
1796 __releases(rq2->lock)
1798 raw_spin_unlock(&rq1->lock);
1800 raw_spin_unlock(&rq2->lock);
1802 __release(rq2->lock);
1807 #ifdef CONFIG_FAIR_GROUP_SCHED
1808 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1811 cfs_rq->shares = shares;
1816 static void calc_load_account_active(struct rq *this_rq);
1817 static void update_sysctl(void);
1818 static int get_update_sysctl_factor(void);
1820 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1822 set_task_rq(p, cpu);
1825 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1826 * successfuly executed on another CPU. We must ensure that updates of
1827 * per-task data have been completed by this moment.
1830 task_thread_info(p)->cpu = cpu;
1834 static const struct sched_class rt_sched_class;
1836 #define sched_class_highest (&rt_sched_class)
1837 #define for_each_class(class) \
1838 for (class = sched_class_highest; class; class = class->next)
1840 #include "sched_stats.h"
1842 static void inc_nr_running(struct rq *rq)
1847 static void dec_nr_running(struct rq *rq)
1852 static void set_load_weight(struct task_struct *p)
1854 if (task_has_rt_policy(p)) {
1855 p->se.load.weight = prio_to_weight[0] * 2;
1856 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1861 * SCHED_IDLE tasks get minimal weight:
1863 if (p->policy == SCHED_IDLE) {
1864 p->se.load.weight = WEIGHT_IDLEPRIO;
1865 p->se.load.inv_weight = WMULT_IDLEPRIO;
1869 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1870 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1873 static void update_avg(u64 *avg, u64 sample)
1875 s64 diff = sample - *avg;
1880 enqueue_task(struct rq *rq, struct task_struct *p, int wakeup, bool head)
1883 p->se.start_runtime = p->se.sum_exec_runtime;
1885 sched_info_queued(p);
1886 p->sched_class->enqueue_task(rq, p, wakeup, head);
1890 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1893 if (p->se.last_wakeup) {
1894 update_avg(&p->se.avg_overlap,
1895 p->se.sum_exec_runtime - p->se.last_wakeup);
1896 p->se.last_wakeup = 0;
1898 update_avg(&p->se.avg_wakeup,
1899 sysctl_sched_wakeup_granularity);
1903 sched_info_dequeued(p);
1904 p->sched_class->dequeue_task(rq, p, sleep);
1909 * activate_task - move a task to the runqueue.
1911 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1913 if (task_contributes_to_load(p))
1914 rq->nr_uninterruptible--;
1916 enqueue_task(rq, p, wakeup, false);
1921 * deactivate_task - remove a task from the runqueue.
1923 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1925 if (task_contributes_to_load(p))
1926 rq->nr_uninterruptible++;
1928 dequeue_task(rq, p, sleep);
1932 #include "sched_idletask.c"
1933 #include "sched_fair.c"
1934 #include "sched_rt.c"
1935 #ifdef CONFIG_SCHED_DEBUG
1936 # include "sched_debug.c"
1940 * __normal_prio - return the priority that is based on the static prio
1942 static inline int __normal_prio(struct task_struct *p)
1944 return p->static_prio;
1948 * Calculate the expected normal priority: i.e. priority
1949 * without taking RT-inheritance into account. Might be
1950 * boosted by interactivity modifiers. Changes upon fork,
1951 * setprio syscalls, and whenever the interactivity
1952 * estimator recalculates.
1954 static inline int normal_prio(struct task_struct *p)
1958 if (task_has_rt_policy(p))
1959 prio = MAX_RT_PRIO-1 - p->rt_priority;
1961 prio = __normal_prio(p);
1966 * Calculate the current priority, i.e. the priority
1967 * taken into account by the scheduler. This value might
1968 * be boosted by RT tasks, or might be boosted by
1969 * interactivity modifiers. Will be RT if the task got
1970 * RT-boosted. If not then it returns p->normal_prio.
1972 static int effective_prio(struct task_struct *p)
1974 p->normal_prio = normal_prio(p);
1976 * If we are RT tasks or we were boosted to RT priority,
1977 * keep the priority unchanged. Otherwise, update priority
1978 * to the normal priority:
1980 if (!rt_prio(p->prio))
1981 return p->normal_prio;
1986 * task_curr - is this task currently executing on a CPU?
1987 * @p: the task in question.
1989 inline int task_curr(const struct task_struct *p)
1991 return cpu_curr(task_cpu(p)) == p;
1994 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1995 const struct sched_class *prev_class,
1996 int oldprio, int running)
1998 if (prev_class != p->sched_class) {
1999 if (prev_class->switched_from)
2000 prev_class->switched_from(rq, p, running);
2001 p->sched_class->switched_to(rq, p, running);
2003 p->sched_class->prio_changed(rq, p, oldprio, running);
2008 * Is this task likely cache-hot:
2011 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2015 if (p->sched_class != &fair_sched_class)
2019 * Buddy candidates are cache hot:
2021 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2022 (&p->se == cfs_rq_of(&p->se)->next ||
2023 &p->se == cfs_rq_of(&p->se)->last))
2026 if (sysctl_sched_migration_cost == -1)
2028 if (sysctl_sched_migration_cost == 0)
2031 delta = now - p->se.exec_start;
2033 return delta < (s64)sysctl_sched_migration_cost;
2036 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2038 #ifdef CONFIG_SCHED_DEBUG
2040 * We should never call set_task_cpu() on a blocked task,
2041 * ttwu() will sort out the placement.
2043 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2044 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2047 trace_sched_migrate_task(p, new_cpu);
2049 if (task_cpu(p) != new_cpu) {
2050 p->se.nr_migrations++;
2051 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2054 __set_task_cpu(p, new_cpu);
2057 struct migration_req {
2058 struct list_head list;
2060 struct task_struct *task;
2063 struct completion done;
2067 * The task's runqueue lock must be held.
2068 * Returns true if you have to wait for migration thread.
2071 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2073 struct rq *rq = task_rq(p);
2076 * If the task is not on a runqueue (and not running), then
2077 * the next wake-up will properly place the task.
2079 if (!p->se.on_rq && !task_running(rq, p))
2082 init_completion(&req->done);
2084 req->dest_cpu = dest_cpu;
2085 list_add(&req->list, &rq->migration_queue);
2091 * wait_task_context_switch - wait for a thread to complete at least one
2094 * @p must not be current.
2096 void wait_task_context_switch(struct task_struct *p)
2098 unsigned long nvcsw, nivcsw, flags;
2106 * The runqueue is assigned before the actual context
2107 * switch. We need to take the runqueue lock.
2109 * We could check initially without the lock but it is
2110 * very likely that we need to take the lock in every
2113 rq = task_rq_lock(p, &flags);
2114 running = task_running(rq, p);
2115 task_rq_unlock(rq, &flags);
2117 if (likely(!running))
2120 * The switch count is incremented before the actual
2121 * context switch. We thus wait for two switches to be
2122 * sure at least one completed.
2124 if ((p->nvcsw - nvcsw) > 1)
2126 if ((p->nivcsw - nivcsw) > 1)
2134 * wait_task_inactive - wait for a thread to unschedule.
2136 * If @match_state is nonzero, it's the @p->state value just checked and
2137 * not expected to change. If it changes, i.e. @p might have woken up,
2138 * then return zero. When we succeed in waiting for @p to be off its CPU,
2139 * we return a positive number (its total switch count). If a second call
2140 * a short while later returns the same number, the caller can be sure that
2141 * @p has remained unscheduled the whole time.
2143 * The caller must ensure that the task *will* unschedule sometime soon,
2144 * else this function might spin for a *long* time. This function can't
2145 * be called with interrupts off, or it may introduce deadlock with
2146 * smp_call_function() if an IPI is sent by the same process we are
2147 * waiting to become inactive.
2149 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2151 unsigned long flags;
2158 * We do the initial early heuristics without holding
2159 * any task-queue locks at all. We'll only try to get
2160 * the runqueue lock when things look like they will
2166 * If the task is actively running on another CPU
2167 * still, just relax and busy-wait without holding
2170 * NOTE! Since we don't hold any locks, it's not
2171 * even sure that "rq" stays as the right runqueue!
2172 * But we don't care, since "task_running()" will
2173 * return false if the runqueue has changed and p
2174 * is actually now running somewhere else!
2176 while (task_running(rq, p)) {
2177 if (match_state && unlikely(p->state != match_state))
2183 * Ok, time to look more closely! We need the rq
2184 * lock now, to be *sure*. If we're wrong, we'll
2185 * just go back and repeat.
2187 rq = task_rq_lock(p, &flags);
2188 trace_sched_wait_task(rq, p);
2189 running = task_running(rq, p);
2190 on_rq = p->se.on_rq;
2192 if (!match_state || p->state == match_state)
2193 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2194 task_rq_unlock(rq, &flags);
2197 * If it changed from the expected state, bail out now.
2199 if (unlikely(!ncsw))
2203 * Was it really running after all now that we
2204 * checked with the proper locks actually held?
2206 * Oops. Go back and try again..
2208 if (unlikely(running)) {
2214 * It's not enough that it's not actively running,
2215 * it must be off the runqueue _entirely_, and not
2218 * So if it was still runnable (but just not actively
2219 * running right now), it's preempted, and we should
2220 * yield - it could be a while.
2222 if (unlikely(on_rq)) {
2223 schedule_timeout_uninterruptible(1);
2228 * Ahh, all good. It wasn't running, and it wasn't
2229 * runnable, which means that it will never become
2230 * running in the future either. We're all done!
2239 * kick_process - kick a running thread to enter/exit the kernel
2240 * @p: the to-be-kicked thread
2242 * Cause a process which is running on another CPU to enter
2243 * kernel-mode, without any delay. (to get signals handled.)
2245 * NOTE: this function doesnt have to take the runqueue lock,
2246 * because all it wants to ensure is that the remote task enters
2247 * the kernel. If the IPI races and the task has been migrated
2248 * to another CPU then no harm is done and the purpose has been
2251 void kick_process(struct task_struct *p)
2257 if ((cpu != smp_processor_id()) && task_curr(p))
2258 smp_send_reschedule(cpu);
2261 EXPORT_SYMBOL_GPL(kick_process);
2262 #endif /* CONFIG_SMP */
2265 * task_oncpu_function_call - call a function on the cpu on which a task runs
2266 * @p: the task to evaluate
2267 * @func: the function to be called
2268 * @info: the function call argument
2270 * Calls the function @func when the task is currently running. This might
2271 * be on the current CPU, which just calls the function directly
2273 void task_oncpu_function_call(struct task_struct *p,
2274 void (*func) (void *info), void *info)
2281 smp_call_function_single(cpu, func, info, 1);
2286 static int select_fallback_rq(int cpu, struct task_struct *p)
2289 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2291 /* Look for allowed, online CPU in same node. */
2292 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2293 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2296 /* Any allowed, online CPU? */
2297 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2298 if (dest_cpu < nr_cpu_ids)
2301 /* No more Mr. Nice Guy. */
2302 if (dest_cpu >= nr_cpu_ids) {
2304 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
2306 dest_cpu = cpumask_any_and(cpu_active_mask, &p->cpus_allowed);
2309 * Don't tell them about moving exiting tasks or
2310 * kernel threads (both mm NULL), since they never
2313 if (p->mm && printk_ratelimit()) {
2314 printk(KERN_INFO "process %d (%s) no "
2315 "longer affine to cpu%d\n",
2316 task_pid_nr(p), p->comm, cpu);
2324 * Gets called from 3 sites (exec, fork, wakeup), since it is called without
2325 * holding rq->lock we need to ensure ->cpus_allowed is stable, this is done
2328 * exec: is unstable, retry loop
2329 * fork & wake-up: serialize ->cpus_allowed against TASK_WAKING
2332 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2334 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2337 * In order not to call set_task_cpu() on a blocking task we need
2338 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2341 * Since this is common to all placement strategies, this lives here.
2343 * [ this allows ->select_task() to simply return task_cpu(p) and
2344 * not worry about this generic constraint ]
2346 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2348 cpu = select_fallback_rq(task_cpu(p), p);
2355 * try_to_wake_up - wake up a thread
2356 * @p: the to-be-woken-up thread
2357 * @state: the mask of task states that can be woken
2358 * @sync: do a synchronous wakeup?
2360 * Put it on the run-queue if it's not already there. The "current"
2361 * thread is always on the run-queue (except when the actual
2362 * re-schedule is in progress), and as such you're allowed to do
2363 * the simpler "current->state = TASK_RUNNING" to mark yourself
2364 * runnable without the overhead of this.
2366 * returns failure only if the task is already active.
2368 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2371 int cpu, orig_cpu, this_cpu, success = 0;
2372 unsigned long flags;
2375 if (!sched_feat(SYNC_WAKEUPS))
2376 wake_flags &= ~WF_SYNC;
2378 this_cpu = get_cpu();
2381 rq = task_rq_lock(p, &flags);
2382 update_rq_clock(rq);
2383 if (!(p->state & state))
2393 if (unlikely(task_running(rq, p)))
2397 * In order to handle concurrent wakeups and release the rq->lock
2398 * we put the task in TASK_WAKING state.
2400 * First fix up the nr_uninterruptible count:
2402 if (task_contributes_to_load(p))
2403 rq->nr_uninterruptible--;
2404 p->state = TASK_WAKING;
2406 if (p->sched_class->task_waking)
2407 p->sched_class->task_waking(rq, p);
2409 __task_rq_unlock(rq);
2411 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2412 if (cpu != orig_cpu) {
2414 * Since we migrate the task without holding any rq->lock,
2415 * we need to be careful with task_rq_lock(), since that
2416 * might end up locking an invalid rq.
2418 set_task_cpu(p, cpu);
2422 raw_spin_lock(&rq->lock);
2423 update_rq_clock(rq);
2426 * We migrated the task without holding either rq->lock, however
2427 * since the task is not on the task list itself, nobody else
2428 * will try and migrate the task, hence the rq should match the
2429 * cpu we just moved it to.
2431 WARN_ON(task_cpu(p) != cpu);
2432 WARN_ON(p->state != TASK_WAKING);
2434 #ifdef CONFIG_SCHEDSTATS
2435 schedstat_inc(rq, ttwu_count);
2436 if (cpu == this_cpu)
2437 schedstat_inc(rq, ttwu_local);
2439 struct sched_domain *sd;
2440 for_each_domain(this_cpu, sd) {
2441 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2442 schedstat_inc(sd, ttwu_wake_remote);
2447 #endif /* CONFIG_SCHEDSTATS */
2450 #endif /* CONFIG_SMP */
2451 schedstat_inc(p, se.nr_wakeups);
2452 if (wake_flags & WF_SYNC)
2453 schedstat_inc(p, se.nr_wakeups_sync);
2454 if (orig_cpu != cpu)
2455 schedstat_inc(p, se.nr_wakeups_migrate);
2456 if (cpu == this_cpu)
2457 schedstat_inc(p, se.nr_wakeups_local);
2459 schedstat_inc(p, se.nr_wakeups_remote);
2460 activate_task(rq, p, 1);
2464 * Only attribute actual wakeups done by this task.
2466 if (!in_interrupt()) {
2467 struct sched_entity *se = ¤t->se;
2468 u64 sample = se->sum_exec_runtime;
2470 if (se->last_wakeup)
2471 sample -= se->last_wakeup;
2473 sample -= se->start_runtime;
2474 update_avg(&se->avg_wakeup, sample);
2476 se->last_wakeup = se->sum_exec_runtime;
2480 trace_sched_wakeup(rq, p, success);
2481 check_preempt_curr(rq, p, wake_flags);
2483 p->state = TASK_RUNNING;
2485 if (p->sched_class->task_woken)
2486 p->sched_class->task_woken(rq, p);
2488 if (unlikely(rq->idle_stamp)) {
2489 u64 delta = rq->clock - rq->idle_stamp;
2490 u64 max = 2*sysctl_sched_migration_cost;
2495 update_avg(&rq->avg_idle, delta);
2500 task_rq_unlock(rq, &flags);
2507 * wake_up_process - Wake up a specific process
2508 * @p: The process to be woken up.
2510 * Attempt to wake up the nominated process and move it to the set of runnable
2511 * processes. Returns 1 if the process was woken up, 0 if it was already
2514 * It may be assumed that this function implies a write memory barrier before
2515 * changing the task state if and only if any tasks are woken up.
2517 int wake_up_process(struct task_struct *p)
2519 return try_to_wake_up(p, TASK_ALL, 0);
2521 EXPORT_SYMBOL(wake_up_process);
2523 int wake_up_state(struct task_struct *p, unsigned int state)
2525 return try_to_wake_up(p, state, 0);
2529 * Perform scheduler related setup for a newly forked process p.
2530 * p is forked by current.
2532 * __sched_fork() is basic setup used by init_idle() too:
2534 static void __sched_fork(struct task_struct *p)
2536 p->se.exec_start = 0;
2537 p->se.sum_exec_runtime = 0;
2538 p->se.prev_sum_exec_runtime = 0;
2539 p->se.nr_migrations = 0;
2540 p->se.last_wakeup = 0;
2541 p->se.avg_overlap = 0;
2542 p->se.start_runtime = 0;
2543 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2545 #ifdef CONFIG_SCHEDSTATS
2546 p->se.wait_start = 0;
2548 p->se.wait_count = 0;
2551 p->se.sleep_start = 0;
2552 p->se.sleep_max = 0;
2553 p->se.sum_sleep_runtime = 0;
2555 p->se.block_start = 0;
2556 p->se.block_max = 0;
2558 p->se.slice_max = 0;
2560 p->se.nr_migrations_cold = 0;
2561 p->se.nr_failed_migrations_affine = 0;
2562 p->se.nr_failed_migrations_running = 0;
2563 p->se.nr_failed_migrations_hot = 0;
2564 p->se.nr_forced_migrations = 0;
2566 p->se.nr_wakeups = 0;
2567 p->se.nr_wakeups_sync = 0;
2568 p->se.nr_wakeups_migrate = 0;
2569 p->se.nr_wakeups_local = 0;
2570 p->se.nr_wakeups_remote = 0;
2571 p->se.nr_wakeups_affine = 0;
2572 p->se.nr_wakeups_affine_attempts = 0;
2573 p->se.nr_wakeups_passive = 0;
2574 p->se.nr_wakeups_idle = 0;
2578 INIT_LIST_HEAD(&p->rt.run_list);
2580 INIT_LIST_HEAD(&p->se.group_node);
2582 #ifdef CONFIG_PREEMPT_NOTIFIERS
2583 INIT_HLIST_HEAD(&p->preempt_notifiers);
2588 * fork()/clone()-time setup:
2590 void sched_fork(struct task_struct *p, int clone_flags)
2592 int cpu = get_cpu();
2596 * We mark the process as waking here. This guarantees that
2597 * nobody will actually run it, and a signal or other external
2598 * event cannot wake it up and insert it on the runqueue either.
2600 p->state = TASK_WAKING;
2603 * Revert to default priority/policy on fork if requested.
2605 if (unlikely(p->sched_reset_on_fork)) {
2606 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2607 p->policy = SCHED_NORMAL;
2608 p->normal_prio = p->static_prio;
2611 if (PRIO_TO_NICE(p->static_prio) < 0) {
2612 p->static_prio = NICE_TO_PRIO(0);
2613 p->normal_prio = p->static_prio;
2618 * We don't need the reset flag anymore after the fork. It has
2619 * fulfilled its duty:
2621 p->sched_reset_on_fork = 0;
2625 * Make sure we do not leak PI boosting priority to the child.
2627 p->prio = current->normal_prio;
2629 if (!rt_prio(p->prio))
2630 p->sched_class = &fair_sched_class;
2632 if (p->sched_class->task_fork)
2633 p->sched_class->task_fork(p);
2635 set_task_cpu(p, cpu);
2637 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2638 if (likely(sched_info_on()))
2639 memset(&p->sched_info, 0, sizeof(p->sched_info));
2641 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2644 #ifdef CONFIG_PREEMPT
2645 /* Want to start with kernel preemption disabled. */
2646 task_thread_info(p)->preempt_count = 1;
2648 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2654 * wake_up_new_task - wake up a newly created task for the first time.
2656 * This function will do some initial scheduler statistics housekeeping
2657 * that must be done for every newly created context, then puts the task
2658 * on the runqueue and wakes it.
2660 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2662 unsigned long flags;
2664 int cpu __maybe_unused = get_cpu();
2668 * Fork balancing, do it here and not earlier because:
2669 * - cpus_allowed can change in the fork path
2670 * - any previously selected cpu might disappear through hotplug
2672 * We still have TASK_WAKING but PF_STARTING is gone now, meaning
2673 * ->cpus_allowed is stable, we have preemption disabled, meaning
2674 * cpu_online_mask is stable.
2676 cpu = select_task_rq(p, SD_BALANCE_FORK, 0);
2677 set_task_cpu(p, cpu);
2681 * Since the task is not on the rq and we still have TASK_WAKING set
2682 * nobody else will migrate this task.
2685 raw_spin_lock_irqsave(&rq->lock, flags);
2687 BUG_ON(p->state != TASK_WAKING);
2688 p->state = TASK_RUNNING;
2689 update_rq_clock(rq);
2690 activate_task(rq, p, 0);
2691 trace_sched_wakeup_new(rq, p, 1);
2692 check_preempt_curr(rq, p, WF_FORK);
2694 if (p->sched_class->task_woken)
2695 p->sched_class->task_woken(rq, p);
2697 task_rq_unlock(rq, &flags);
2701 #ifdef CONFIG_PREEMPT_NOTIFIERS
2704 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2705 * @notifier: notifier struct to register
2707 void preempt_notifier_register(struct preempt_notifier *notifier)
2709 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2711 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2714 * preempt_notifier_unregister - no longer interested in preemption notifications
2715 * @notifier: notifier struct to unregister
2717 * This is safe to call from within a preemption notifier.
2719 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2721 hlist_del(¬ifier->link);
2723 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2725 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2727 struct preempt_notifier *notifier;
2728 struct hlist_node *node;
2730 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2731 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2735 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2736 struct task_struct *next)
2738 struct preempt_notifier *notifier;
2739 struct hlist_node *node;
2741 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2742 notifier->ops->sched_out(notifier, next);
2745 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2747 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2752 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2753 struct task_struct *next)
2757 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2760 * prepare_task_switch - prepare to switch tasks
2761 * @rq: the runqueue preparing to switch
2762 * @prev: the current task that is being switched out
2763 * @next: the task we are going to switch to.
2765 * This is called with the rq lock held and interrupts off. It must
2766 * be paired with a subsequent finish_task_switch after the context
2769 * prepare_task_switch sets up locking and calls architecture specific
2773 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2774 struct task_struct *next)
2776 fire_sched_out_preempt_notifiers(prev, next);
2777 prepare_lock_switch(rq, next);
2778 prepare_arch_switch(next);
2782 * finish_task_switch - clean up after a task-switch
2783 * @rq: runqueue associated with task-switch
2784 * @prev: the thread we just switched away from.
2786 * finish_task_switch must be called after the context switch, paired
2787 * with a prepare_task_switch call before the context switch.
2788 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2789 * and do any other architecture-specific cleanup actions.
2791 * Note that we may have delayed dropping an mm in context_switch(). If
2792 * so, we finish that here outside of the runqueue lock. (Doing it
2793 * with the lock held can cause deadlocks; see schedule() for
2796 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2797 __releases(rq->lock)
2799 struct mm_struct *mm = rq->prev_mm;
2805 * A task struct has one reference for the use as "current".
2806 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2807 * schedule one last time. The schedule call will never return, and
2808 * the scheduled task must drop that reference.
2809 * The test for TASK_DEAD must occur while the runqueue locks are
2810 * still held, otherwise prev could be scheduled on another cpu, die
2811 * there before we look at prev->state, and then the reference would
2813 * Manfred Spraul <manfred@colorfullife.com>
2815 prev_state = prev->state;
2816 finish_arch_switch(prev);
2817 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2818 local_irq_disable();
2819 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2820 perf_event_task_sched_in(current);
2821 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2823 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2824 finish_lock_switch(rq, prev);
2826 fire_sched_in_preempt_notifiers(current);
2829 if (unlikely(prev_state == TASK_DEAD)) {
2831 * Remove function-return probe instances associated with this
2832 * task and put them back on the free list.
2834 kprobe_flush_task(prev);
2835 put_task_struct(prev);
2841 /* assumes rq->lock is held */
2842 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2844 if (prev->sched_class->pre_schedule)
2845 prev->sched_class->pre_schedule(rq, prev);
2848 /* rq->lock is NOT held, but preemption is disabled */
2849 static inline void post_schedule(struct rq *rq)
2851 if (rq->post_schedule) {
2852 unsigned long flags;
2854 raw_spin_lock_irqsave(&rq->lock, flags);
2855 if (rq->curr->sched_class->post_schedule)
2856 rq->curr->sched_class->post_schedule(rq);
2857 raw_spin_unlock_irqrestore(&rq->lock, flags);
2859 rq->post_schedule = 0;
2865 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2869 static inline void post_schedule(struct rq *rq)
2876 * schedule_tail - first thing a freshly forked thread must call.
2877 * @prev: the thread we just switched away from.
2879 asmlinkage void schedule_tail(struct task_struct *prev)
2880 __releases(rq->lock)
2882 struct rq *rq = this_rq();
2884 finish_task_switch(rq, prev);
2887 * FIXME: do we need to worry about rq being invalidated by the
2892 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2893 /* In this case, finish_task_switch does not reenable preemption */
2896 if (current->set_child_tid)
2897 put_user(task_pid_vnr(current), current->set_child_tid);
2901 * context_switch - switch to the new MM and the new
2902 * thread's register state.
2905 context_switch(struct rq *rq, struct task_struct *prev,
2906 struct task_struct *next)
2908 struct mm_struct *mm, *oldmm;
2910 prepare_task_switch(rq, prev, next);
2911 trace_sched_switch(rq, prev, next);
2913 oldmm = prev->active_mm;
2915 * For paravirt, this is coupled with an exit in switch_to to
2916 * combine the page table reload and the switch backend into
2919 arch_start_context_switch(prev);
2922 next->active_mm = oldmm;
2923 atomic_inc(&oldmm->mm_count);
2924 enter_lazy_tlb(oldmm, next);
2926 switch_mm(oldmm, mm, next);
2928 if (likely(!prev->mm)) {
2929 prev->active_mm = NULL;
2930 rq->prev_mm = oldmm;
2933 * Since the runqueue lock will be released by the next
2934 * task (which is an invalid locking op but in the case
2935 * of the scheduler it's an obvious special-case), so we
2936 * do an early lockdep release here:
2938 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2939 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2942 /* Here we just switch the register state and the stack. */
2943 switch_to(prev, next, prev);
2947 * this_rq must be evaluated again because prev may have moved
2948 * CPUs since it called schedule(), thus the 'rq' on its stack
2949 * frame will be invalid.
2951 finish_task_switch(this_rq(), prev);
2955 * nr_running, nr_uninterruptible and nr_context_switches:
2957 * externally visible scheduler statistics: current number of runnable
2958 * threads, current number of uninterruptible-sleeping threads, total
2959 * number of context switches performed since bootup.
2961 unsigned long nr_running(void)
2963 unsigned long i, sum = 0;
2965 for_each_online_cpu(i)
2966 sum += cpu_rq(i)->nr_running;
2971 unsigned long nr_uninterruptible(void)
2973 unsigned long i, sum = 0;
2975 for_each_possible_cpu(i)
2976 sum += cpu_rq(i)->nr_uninterruptible;
2979 * Since we read the counters lockless, it might be slightly
2980 * inaccurate. Do not allow it to go below zero though:
2982 if (unlikely((long)sum < 0))
2988 unsigned long long nr_context_switches(void)
2991 unsigned long long sum = 0;
2993 for_each_possible_cpu(i)
2994 sum += cpu_rq(i)->nr_switches;
2999 unsigned long nr_iowait(void)
3001 unsigned long i, sum = 0;
3003 for_each_possible_cpu(i)
3004 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3009 unsigned long nr_iowait_cpu(void)
3011 struct rq *this = this_rq();
3012 return atomic_read(&this->nr_iowait);
3015 unsigned long this_cpu_load(void)
3017 struct rq *this = this_rq();
3018 return this->cpu_load[0];
3022 /* Variables and functions for calc_load */
3023 static atomic_long_t calc_load_tasks;
3024 static unsigned long calc_load_update;
3025 unsigned long avenrun[3];
3026 EXPORT_SYMBOL(avenrun);
3029 * get_avenrun - get the load average array
3030 * @loads: pointer to dest load array
3031 * @offset: offset to add
3032 * @shift: shift count to shift the result left
3034 * These values are estimates at best, so no need for locking.
3036 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3038 loads[0] = (avenrun[0] + offset) << shift;
3039 loads[1] = (avenrun[1] + offset) << shift;
3040 loads[2] = (avenrun[2] + offset) << shift;
3043 static unsigned long
3044 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3047 load += active * (FIXED_1 - exp);
3048 return load >> FSHIFT;
3052 * calc_load - update the avenrun load estimates 10 ticks after the
3053 * CPUs have updated calc_load_tasks.
3055 void calc_global_load(void)
3057 unsigned long upd = calc_load_update + 10;
3060 if (time_before(jiffies, upd))
3063 active = atomic_long_read(&calc_load_tasks);
3064 active = active > 0 ? active * FIXED_1 : 0;
3066 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3067 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3068 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3070 calc_load_update += LOAD_FREQ;
3074 * Either called from update_cpu_load() or from a cpu going idle
3076 static void calc_load_account_active(struct rq *this_rq)
3078 long nr_active, delta;
3080 nr_active = this_rq->nr_running;
3081 nr_active += (long) this_rq->nr_uninterruptible;
3083 if (nr_active != this_rq->calc_load_active) {
3084 delta = nr_active - this_rq->calc_load_active;
3085 this_rq->calc_load_active = nr_active;
3086 atomic_long_add(delta, &calc_load_tasks);
3091 * Update rq->cpu_load[] statistics. This function is usually called every
3092 * scheduler tick (TICK_NSEC).
3094 static void update_cpu_load(struct rq *this_rq)
3096 unsigned long this_load = this_rq->load.weight;
3099 this_rq->nr_load_updates++;
3101 /* Update our load: */
3102 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3103 unsigned long old_load, new_load;
3105 /* scale is effectively 1 << i now, and >> i divides by scale */
3107 old_load = this_rq->cpu_load[i];
3108 new_load = this_load;
3110 * Round up the averaging division if load is increasing. This
3111 * prevents us from getting stuck on 9 if the load is 10, for
3114 if (new_load > old_load)
3115 new_load += scale-1;
3116 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3119 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3120 this_rq->calc_load_update += LOAD_FREQ;
3121 calc_load_account_active(this_rq);
3128 * sched_exec - execve() is a valuable balancing opportunity, because at
3129 * this point the task has the smallest effective memory and cache footprint.
3131 void sched_exec(void)
3133 struct task_struct *p = current;
3134 struct migration_req req;
3135 int dest_cpu, this_cpu;
3136 unsigned long flags;
3140 this_cpu = get_cpu();
3141 dest_cpu = select_task_rq(p, SD_BALANCE_EXEC, 0);
3142 if (dest_cpu == this_cpu) {
3147 rq = task_rq_lock(p, &flags);
3151 * select_task_rq() can race against ->cpus_allowed
3153 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
3154 || unlikely(!cpu_active(dest_cpu))) {
3155 task_rq_unlock(rq, &flags);
3159 /* force the process onto the specified CPU */
3160 if (migrate_task(p, dest_cpu, &req)) {
3161 /* Need to wait for migration thread (might exit: take ref). */
3162 struct task_struct *mt = rq->migration_thread;
3164 get_task_struct(mt);
3165 task_rq_unlock(rq, &flags);
3166 wake_up_process(mt);
3167 put_task_struct(mt);
3168 wait_for_completion(&req.done);
3172 task_rq_unlock(rq, &flags);
3177 DEFINE_PER_CPU(struct kernel_stat, kstat);
3179 EXPORT_PER_CPU_SYMBOL(kstat);
3182 * Return any ns on the sched_clock that have not yet been accounted in
3183 * @p in case that task is currently running.
3185 * Called with task_rq_lock() held on @rq.
3187 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3191 if (task_current(rq, p)) {
3192 update_rq_clock(rq);
3193 ns = rq->clock - p->se.exec_start;
3201 unsigned long long task_delta_exec(struct task_struct *p)
3203 unsigned long flags;
3207 rq = task_rq_lock(p, &flags);
3208 ns = do_task_delta_exec(p, rq);
3209 task_rq_unlock(rq, &flags);
3215 * Return accounted runtime for the task.
3216 * In case the task is currently running, return the runtime plus current's
3217 * pending runtime that have not been accounted yet.
3219 unsigned long long task_sched_runtime(struct task_struct *p)
3221 unsigned long flags;
3225 rq = task_rq_lock(p, &flags);
3226 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3227 task_rq_unlock(rq, &flags);
3233 * Return sum_exec_runtime for the thread group.
3234 * In case the task is currently running, return the sum plus current's
3235 * pending runtime that have not been accounted yet.
3237 * Note that the thread group might have other running tasks as well,
3238 * so the return value not includes other pending runtime that other
3239 * running tasks might have.
3241 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3243 struct task_cputime totals;
3244 unsigned long flags;
3248 rq = task_rq_lock(p, &flags);
3249 thread_group_cputime(p, &totals);
3250 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3251 task_rq_unlock(rq, &flags);
3257 * Account user cpu time to a process.
3258 * @p: the process that the cpu time gets accounted to
3259 * @cputime: the cpu time spent in user space since the last update
3260 * @cputime_scaled: cputime scaled by cpu frequency
3262 void account_user_time(struct task_struct *p, cputime_t cputime,
3263 cputime_t cputime_scaled)
3265 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3268 /* Add user time to process. */
3269 p->utime = cputime_add(p->utime, cputime);
3270 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3271 account_group_user_time(p, cputime);
3273 /* Add user time to cpustat. */
3274 tmp = cputime_to_cputime64(cputime);
3275 if (TASK_NICE(p) > 0)
3276 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3278 cpustat->user = cputime64_add(cpustat->user, tmp);
3280 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3281 /* Account for user time used */
3282 acct_update_integrals(p);
3286 * Account guest cpu time to a process.
3287 * @p: the process that the cpu time gets accounted to
3288 * @cputime: the cpu time spent in virtual machine since the last update
3289 * @cputime_scaled: cputime scaled by cpu frequency
3291 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3292 cputime_t cputime_scaled)
3295 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3297 tmp = cputime_to_cputime64(cputime);
3299 /* Add guest time to process. */
3300 p->utime = cputime_add(p->utime, cputime);
3301 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3302 account_group_user_time(p, cputime);
3303 p->gtime = cputime_add(p->gtime, cputime);
3305 /* Add guest time to cpustat. */
3306 if (TASK_NICE(p) > 0) {
3307 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3308 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3310 cpustat->user = cputime64_add(cpustat->user, tmp);
3311 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3316 * Account system cpu time to a process.
3317 * @p: the process that the cpu time gets accounted to
3318 * @hardirq_offset: the offset to subtract from hardirq_count()
3319 * @cputime: the cpu time spent in kernel space since the last update
3320 * @cputime_scaled: cputime scaled by cpu frequency
3322 void account_system_time(struct task_struct *p, int hardirq_offset,
3323 cputime_t cputime, cputime_t cputime_scaled)
3325 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3328 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3329 account_guest_time(p, cputime, cputime_scaled);
3333 /* Add system time to process. */
3334 p->stime = cputime_add(p->stime, cputime);
3335 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3336 account_group_system_time(p, cputime);
3338 /* Add system time to cpustat. */
3339 tmp = cputime_to_cputime64(cputime);
3340 if (hardirq_count() - hardirq_offset)
3341 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3342 else if (softirq_count())
3343 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3345 cpustat->system = cputime64_add(cpustat->system, tmp);
3347 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3349 /* Account for system time used */
3350 acct_update_integrals(p);
3354 * Account for involuntary wait time.
3355 * @steal: the cpu time spent in involuntary wait
3357 void account_steal_time(cputime_t cputime)
3359 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3360 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3362 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3366 * Account for idle time.
3367 * @cputime: the cpu time spent in idle wait
3369 void account_idle_time(cputime_t cputime)
3371 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3372 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3373 struct rq *rq = this_rq();
3375 if (atomic_read(&rq->nr_iowait) > 0)
3376 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3378 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3381 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3384 * Account a single tick of cpu time.
3385 * @p: the process that the cpu time gets accounted to
3386 * @user_tick: indicates if the tick is a user or a system tick
3388 void account_process_tick(struct task_struct *p, int user_tick)
3390 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3391 struct rq *rq = this_rq();
3394 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3395 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3396 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3399 account_idle_time(cputime_one_jiffy);
3403 * Account multiple ticks of steal time.
3404 * @p: the process from which the cpu time has been stolen
3405 * @ticks: number of stolen ticks
3407 void account_steal_ticks(unsigned long ticks)
3409 account_steal_time(jiffies_to_cputime(ticks));
3413 * Account multiple ticks of idle time.
3414 * @ticks: number of stolen ticks
3416 void account_idle_ticks(unsigned long ticks)
3418 account_idle_time(jiffies_to_cputime(ticks));
3424 * Use precise platform statistics if available:
3426 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3427 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3433 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3435 struct task_cputime cputime;
3437 thread_group_cputime(p, &cputime);
3439 *ut = cputime.utime;
3440 *st = cputime.stime;
3444 #ifndef nsecs_to_cputime
3445 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3448 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3450 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
3453 * Use CFS's precise accounting:
3455 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3460 temp = (u64)(rtime * utime);
3461 do_div(temp, total);
3462 utime = (cputime_t)temp;
3467 * Compare with previous values, to keep monotonicity:
3469 p->prev_utime = max(p->prev_utime, utime);
3470 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
3472 *ut = p->prev_utime;
3473 *st = p->prev_stime;
3477 * Must be called with siglock held.
3479 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3481 struct signal_struct *sig = p->signal;
3482 struct task_cputime cputime;
3483 cputime_t rtime, utime, total;
3485 thread_group_cputime(p, &cputime);
3487 total = cputime_add(cputime.utime, cputime.stime);
3488 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3493 temp = (u64)(rtime * cputime.utime);
3494 do_div(temp, total);
3495 utime = (cputime_t)temp;
3499 sig->prev_utime = max(sig->prev_utime, utime);
3500 sig->prev_stime = max(sig->prev_stime,
3501 cputime_sub(rtime, sig->prev_utime));
3503 *ut = sig->prev_utime;
3504 *st = sig->prev_stime;
3509 * This function gets called by the timer code, with HZ frequency.
3510 * We call it with interrupts disabled.
3512 * It also gets called by the fork code, when changing the parent's
3515 void scheduler_tick(void)
3517 int cpu = smp_processor_id();
3518 struct rq *rq = cpu_rq(cpu);
3519 struct task_struct *curr = rq->curr;
3523 raw_spin_lock(&rq->lock);
3524 update_rq_clock(rq);
3525 update_cpu_load(rq);
3526 curr->sched_class->task_tick(rq, curr, 0);
3527 raw_spin_unlock(&rq->lock);
3529 perf_event_task_tick(curr);
3532 rq->idle_at_tick = idle_cpu(cpu);
3533 trigger_load_balance(rq, cpu);
3537 notrace unsigned long get_parent_ip(unsigned long addr)
3539 if (in_lock_functions(addr)) {
3540 addr = CALLER_ADDR2;
3541 if (in_lock_functions(addr))
3542 addr = CALLER_ADDR3;
3547 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3548 defined(CONFIG_PREEMPT_TRACER))
3550 void __kprobes add_preempt_count(int val)
3552 #ifdef CONFIG_DEBUG_PREEMPT
3556 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3559 preempt_count() += val;
3560 #ifdef CONFIG_DEBUG_PREEMPT
3562 * Spinlock count overflowing soon?
3564 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3567 if (preempt_count() == val)
3568 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3570 EXPORT_SYMBOL(add_preempt_count);
3572 void __kprobes sub_preempt_count(int val)
3574 #ifdef CONFIG_DEBUG_PREEMPT
3578 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3581 * Is the spinlock portion underflowing?
3583 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3584 !(preempt_count() & PREEMPT_MASK)))
3588 if (preempt_count() == val)
3589 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3590 preempt_count() -= val;
3592 EXPORT_SYMBOL(sub_preempt_count);
3597 * Print scheduling while atomic bug:
3599 static noinline void __schedule_bug(struct task_struct *prev)
3601 struct pt_regs *regs = get_irq_regs();
3603 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3604 prev->comm, prev->pid, preempt_count());
3606 debug_show_held_locks(prev);
3608 if (irqs_disabled())
3609 print_irqtrace_events(prev);
3618 * Various schedule()-time debugging checks and statistics:
3620 static inline void schedule_debug(struct task_struct *prev)
3623 * Test if we are atomic. Since do_exit() needs to call into
3624 * schedule() atomically, we ignore that path for now.
3625 * Otherwise, whine if we are scheduling when we should not be.
3627 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3628 __schedule_bug(prev);
3630 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3632 schedstat_inc(this_rq(), sched_count);
3633 #ifdef CONFIG_SCHEDSTATS
3634 if (unlikely(prev->lock_depth >= 0)) {
3635 schedstat_inc(this_rq(), bkl_count);
3636 schedstat_inc(prev, sched_info.bkl_count);
3641 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3643 if (prev->state == TASK_RUNNING) {
3644 u64 runtime = prev->se.sum_exec_runtime;
3646 runtime -= prev->se.prev_sum_exec_runtime;
3647 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
3650 * In order to avoid avg_overlap growing stale when we are
3651 * indeed overlapping and hence not getting put to sleep, grow
3652 * the avg_overlap on preemption.
3654 * We use the average preemption runtime because that
3655 * correlates to the amount of cache footprint a task can
3658 update_avg(&prev->se.avg_overlap, runtime);
3660 prev->sched_class->put_prev_task(rq, prev);
3664 * Pick up the highest-prio task:
3666 static inline struct task_struct *
3667 pick_next_task(struct rq *rq)
3669 const struct sched_class *class;
3670 struct task_struct *p;
3673 * Optimization: we know that if all tasks are in
3674 * the fair class we can call that function directly:
3676 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3677 p = fair_sched_class.pick_next_task(rq);
3682 class = sched_class_highest;
3684 p = class->pick_next_task(rq);
3688 * Will never be NULL as the idle class always
3689 * returns a non-NULL p:
3691 class = class->next;
3696 * schedule() is the main scheduler function.
3698 asmlinkage void __sched schedule(void)
3700 struct task_struct *prev, *next;
3701 unsigned long *switch_count;
3707 cpu = smp_processor_id();
3711 switch_count = &prev->nivcsw;
3713 release_kernel_lock(prev);
3714 need_resched_nonpreemptible:
3716 schedule_debug(prev);
3718 if (sched_feat(HRTICK))
3721 raw_spin_lock_irq(&rq->lock);
3722 update_rq_clock(rq);
3723 clear_tsk_need_resched(prev);
3725 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3726 if (unlikely(signal_pending_state(prev->state, prev)))
3727 prev->state = TASK_RUNNING;
3729 deactivate_task(rq, prev, 1);
3730 switch_count = &prev->nvcsw;
3733 pre_schedule(rq, prev);
3735 if (unlikely(!rq->nr_running))
3736 idle_balance(cpu, rq);
3738 put_prev_task(rq, prev);
3739 next = pick_next_task(rq);
3741 if (likely(prev != next)) {
3742 sched_info_switch(prev, next);
3743 perf_event_task_sched_out(prev, next);
3749 context_switch(rq, prev, next); /* unlocks the rq */
3751 * the context switch might have flipped the stack from under
3752 * us, hence refresh the local variables.
3754 cpu = smp_processor_id();
3757 raw_spin_unlock_irq(&rq->lock);
3761 if (unlikely(reacquire_kernel_lock(current) < 0)) {
3763 switch_count = &prev->nivcsw;
3764 goto need_resched_nonpreemptible;
3767 preempt_enable_no_resched();
3771 EXPORT_SYMBOL(schedule);
3773 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3775 * Look out! "owner" is an entirely speculative pointer
3776 * access and not reliable.
3778 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
3783 if (!sched_feat(OWNER_SPIN))
3786 #ifdef CONFIG_DEBUG_PAGEALLOC
3788 * Need to access the cpu field knowing that
3789 * DEBUG_PAGEALLOC could have unmapped it if
3790 * the mutex owner just released it and exited.
3792 if (probe_kernel_address(&owner->cpu, cpu))
3799 * Even if the access succeeded (likely case),
3800 * the cpu field may no longer be valid.
3802 if (cpu >= nr_cpumask_bits)
3806 * We need to validate that we can do a
3807 * get_cpu() and that we have the percpu area.
3809 if (!cpu_online(cpu))
3816 * Owner changed, break to re-assess state.
3818 if (lock->owner != owner)
3822 * Is that owner really running on that cpu?
3824 if (task_thread_info(rq->curr) != owner || need_resched())
3834 #ifdef CONFIG_PREEMPT
3836 * this is the entry point to schedule() from in-kernel preemption
3837 * off of preempt_enable. Kernel preemptions off return from interrupt
3838 * occur there and call schedule directly.
3840 asmlinkage void __sched preempt_schedule(void)
3842 struct thread_info *ti = current_thread_info();
3845 * If there is a non-zero preempt_count or interrupts are disabled,
3846 * we do not want to preempt the current task. Just return..
3848 if (likely(ti->preempt_count || irqs_disabled()))
3852 add_preempt_count(PREEMPT_ACTIVE);
3854 sub_preempt_count(PREEMPT_ACTIVE);
3857 * Check again in case we missed a preemption opportunity
3858 * between schedule and now.
3861 } while (need_resched());
3863 EXPORT_SYMBOL(preempt_schedule);
3866 * this is the entry point to schedule() from kernel preemption
3867 * off of irq context.
3868 * Note, that this is called and return with irqs disabled. This will
3869 * protect us against recursive calling from irq.
3871 asmlinkage void __sched preempt_schedule_irq(void)
3873 struct thread_info *ti = current_thread_info();
3875 /* Catch callers which need to be fixed */
3876 BUG_ON(ti->preempt_count || !irqs_disabled());
3879 add_preempt_count(PREEMPT_ACTIVE);
3882 local_irq_disable();
3883 sub_preempt_count(PREEMPT_ACTIVE);
3886 * Check again in case we missed a preemption opportunity
3887 * between schedule and now.
3890 } while (need_resched());
3893 #endif /* CONFIG_PREEMPT */
3895 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3898 return try_to_wake_up(curr->private, mode, wake_flags);
3900 EXPORT_SYMBOL(default_wake_function);
3903 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3904 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3905 * number) then we wake all the non-exclusive tasks and one exclusive task.
3907 * There are circumstances in which we can try to wake a task which has already
3908 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3909 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3911 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3912 int nr_exclusive, int wake_flags, void *key)
3914 wait_queue_t *curr, *next;
3916 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3917 unsigned flags = curr->flags;
3919 if (curr->func(curr, mode, wake_flags, key) &&
3920 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3926 * __wake_up - wake up threads blocked on a waitqueue.
3928 * @mode: which threads
3929 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3930 * @key: is directly passed to the wakeup function
3932 * It may be assumed that this function implies a write memory barrier before
3933 * changing the task state if and only if any tasks are woken up.
3935 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3936 int nr_exclusive, void *key)
3938 unsigned long flags;
3940 spin_lock_irqsave(&q->lock, flags);
3941 __wake_up_common(q, mode, nr_exclusive, 0, key);
3942 spin_unlock_irqrestore(&q->lock, flags);
3944 EXPORT_SYMBOL(__wake_up);
3947 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3949 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3951 __wake_up_common(q, mode, 1, 0, NULL);
3954 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3956 __wake_up_common(q, mode, 1, 0, key);
3960 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3962 * @mode: which threads
3963 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3964 * @key: opaque value to be passed to wakeup targets
3966 * The sync wakeup differs that the waker knows that it will schedule
3967 * away soon, so while the target thread will be woken up, it will not
3968 * be migrated to another CPU - ie. the two threads are 'synchronized'
3969 * with each other. This can prevent needless bouncing between CPUs.
3971 * On UP it can prevent extra preemption.
3973 * It may be assumed that this function implies a write memory barrier before
3974 * changing the task state if and only if any tasks are woken up.
3976 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3977 int nr_exclusive, void *key)
3979 unsigned long flags;
3980 int wake_flags = WF_SYNC;
3985 if (unlikely(!nr_exclusive))
3988 spin_lock_irqsave(&q->lock, flags);
3989 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3990 spin_unlock_irqrestore(&q->lock, flags);
3992 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3995 * __wake_up_sync - see __wake_up_sync_key()
3997 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3999 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4001 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4004 * complete: - signals a single thread waiting on this completion
4005 * @x: holds the state of this particular completion
4007 * This will wake up a single thread waiting on this completion. Threads will be
4008 * awakened in the same order in which they were queued.
4010 * See also complete_all(), wait_for_completion() and related routines.
4012 * It may be assumed that this function implies a write memory barrier before
4013 * changing the task state if and only if any tasks are woken up.
4015 void complete(struct completion *x)
4017 unsigned long flags;
4019 spin_lock_irqsave(&x->wait.lock, flags);
4021 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4022 spin_unlock_irqrestore(&x->wait.lock, flags);
4024 EXPORT_SYMBOL(complete);
4027 * complete_all: - signals all threads waiting on this completion
4028 * @x: holds the state of this particular completion
4030 * This will wake up all threads waiting on this particular completion event.
4032 * It may be assumed that this function implies a write memory barrier before
4033 * changing the task state if and only if any tasks are woken up.
4035 void complete_all(struct completion *x)
4037 unsigned long flags;
4039 spin_lock_irqsave(&x->wait.lock, flags);
4040 x->done += UINT_MAX/2;
4041 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4042 spin_unlock_irqrestore(&x->wait.lock, flags);
4044 EXPORT_SYMBOL(complete_all);
4046 static inline long __sched
4047 do_wait_for_common(struct completion *x, long timeout, int state)
4050 DECLARE_WAITQUEUE(wait, current);
4052 wait.flags |= WQ_FLAG_EXCLUSIVE;
4053 __add_wait_queue_tail(&x->wait, &wait);
4055 if (signal_pending_state(state, current)) {
4056 timeout = -ERESTARTSYS;
4059 __set_current_state(state);
4060 spin_unlock_irq(&x->wait.lock);
4061 timeout = schedule_timeout(timeout);
4062 spin_lock_irq(&x->wait.lock);
4063 } while (!x->done && timeout);
4064 __remove_wait_queue(&x->wait, &wait);
4069 return timeout ?: 1;
4073 wait_for_common(struct completion *x, long timeout, int state)
4077 spin_lock_irq(&x->wait.lock);
4078 timeout = do_wait_for_common(x, timeout, state);
4079 spin_unlock_irq(&x->wait.lock);
4084 * wait_for_completion: - waits for completion of a task
4085 * @x: holds the state of this particular completion
4087 * This waits to be signaled for completion of a specific task. It is NOT
4088 * interruptible and there is no timeout.
4090 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4091 * and interrupt capability. Also see complete().
4093 void __sched wait_for_completion(struct completion *x)
4095 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4097 EXPORT_SYMBOL(wait_for_completion);
4100 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4101 * @x: holds the state of this particular completion
4102 * @timeout: timeout value in jiffies
4104 * This waits for either a completion of a specific task to be signaled or for a
4105 * specified timeout to expire. The timeout is in jiffies. It is not
4108 unsigned long __sched
4109 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4111 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4113 EXPORT_SYMBOL(wait_for_completion_timeout);
4116 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4117 * @x: holds the state of this particular completion
4119 * This waits for completion of a specific task to be signaled. It is
4122 int __sched wait_for_completion_interruptible(struct completion *x)
4124 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4125 if (t == -ERESTARTSYS)
4129 EXPORT_SYMBOL(wait_for_completion_interruptible);
4132 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4133 * @x: holds the state of this particular completion
4134 * @timeout: timeout value in jiffies
4136 * This waits for either a completion of a specific task to be signaled or for a
4137 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4139 unsigned long __sched
4140 wait_for_completion_interruptible_timeout(struct completion *x,
4141 unsigned long timeout)
4143 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4145 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4148 * wait_for_completion_killable: - waits for completion of a task (killable)
4149 * @x: holds the state of this particular completion
4151 * This waits to be signaled for completion of a specific task. It can be
4152 * interrupted by a kill signal.
4154 int __sched wait_for_completion_killable(struct completion *x)
4156 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4157 if (t == -ERESTARTSYS)
4161 EXPORT_SYMBOL(wait_for_completion_killable);
4164 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4165 * @x: holds the state of this particular completion
4166 * @timeout: timeout value in jiffies
4168 * This waits for either a completion of a specific task to be
4169 * signaled or for a specified timeout to expire. It can be
4170 * interrupted by a kill signal. The timeout is in jiffies.
4172 unsigned long __sched
4173 wait_for_completion_killable_timeout(struct completion *x,
4174 unsigned long timeout)
4176 return wait_for_common(x, timeout, TASK_KILLABLE);
4178 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
4181 * try_wait_for_completion - try to decrement a completion without blocking
4182 * @x: completion structure
4184 * Returns: 0 if a decrement cannot be done without blocking
4185 * 1 if a decrement succeeded.
4187 * If a completion is being used as a counting completion,
4188 * attempt to decrement the counter without blocking. This
4189 * enables us to avoid waiting if the resource the completion
4190 * is protecting is not available.
4192 bool try_wait_for_completion(struct completion *x)
4194 unsigned long flags;
4197 spin_lock_irqsave(&x->wait.lock, flags);
4202 spin_unlock_irqrestore(&x->wait.lock, flags);
4205 EXPORT_SYMBOL(try_wait_for_completion);
4208 * completion_done - Test to see if a completion has any waiters
4209 * @x: completion structure
4211 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4212 * 1 if there are no waiters.
4215 bool completion_done(struct completion *x)
4217 unsigned long flags;
4220 spin_lock_irqsave(&x->wait.lock, flags);
4223 spin_unlock_irqrestore(&x->wait.lock, flags);
4226 EXPORT_SYMBOL(completion_done);
4229 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4231 unsigned long flags;
4234 init_waitqueue_entry(&wait, current);
4236 __set_current_state(state);
4238 spin_lock_irqsave(&q->lock, flags);
4239 __add_wait_queue(q, &wait);
4240 spin_unlock(&q->lock);
4241 timeout = schedule_timeout(timeout);
4242 spin_lock_irq(&q->lock);
4243 __remove_wait_queue(q, &wait);
4244 spin_unlock_irqrestore(&q->lock, flags);
4249 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4251 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4253 EXPORT_SYMBOL(interruptible_sleep_on);
4256 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4258 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4260 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4262 void __sched sleep_on(wait_queue_head_t *q)
4264 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4266 EXPORT_SYMBOL(sleep_on);
4268 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4270 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4272 EXPORT_SYMBOL(sleep_on_timeout);
4274 #ifdef CONFIG_RT_MUTEXES
4277 * rt_mutex_setprio - set the current priority of a task
4279 * @prio: prio value (kernel-internal form)
4281 * This function changes the 'effective' priority of a task. It does
4282 * not touch ->normal_prio like __setscheduler().
4284 * Used by the rt_mutex code to implement priority inheritance logic.
4286 void rt_mutex_setprio(struct task_struct *p, int prio)
4288 unsigned long flags;
4289 int oldprio, on_rq, running;
4291 const struct sched_class *prev_class;
4293 BUG_ON(prio < 0 || prio > MAX_PRIO);
4295 rq = task_rq_lock(p, &flags);
4296 update_rq_clock(rq);
4299 prev_class = p->sched_class;
4300 on_rq = p->se.on_rq;
4301 running = task_current(rq, p);
4303 dequeue_task(rq, p, 0);
4305 p->sched_class->put_prev_task(rq, p);
4308 p->sched_class = &rt_sched_class;
4310 p->sched_class = &fair_sched_class;
4315 p->sched_class->set_curr_task(rq);
4317 enqueue_task(rq, p, 0, oldprio < prio);
4319 check_class_changed(rq, p, prev_class, oldprio, running);
4321 task_rq_unlock(rq, &flags);
4326 void set_user_nice(struct task_struct *p, long nice)
4328 int old_prio, delta, on_rq;
4329 unsigned long flags;
4332 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4335 * We have to be careful, if called from sys_setpriority(),
4336 * the task might be in the middle of scheduling on another CPU.
4338 rq = task_rq_lock(p, &flags);
4339 update_rq_clock(rq);
4341 * The RT priorities are set via sched_setscheduler(), but we still
4342 * allow the 'normal' nice value to be set - but as expected
4343 * it wont have any effect on scheduling until the task is
4344 * SCHED_FIFO/SCHED_RR:
4346 if (task_has_rt_policy(p)) {
4347 p->static_prio = NICE_TO_PRIO(nice);
4350 on_rq = p->se.on_rq;
4352 dequeue_task(rq, p, 0);
4354 p->static_prio = NICE_TO_PRIO(nice);
4357 p->prio = effective_prio(p);
4358 delta = p->prio - old_prio;
4361 enqueue_task(rq, p, 0, false);
4363 * If the task increased its priority or is running and
4364 * lowered its priority, then reschedule its CPU:
4366 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4367 resched_task(rq->curr);
4370 task_rq_unlock(rq, &flags);
4372 EXPORT_SYMBOL(set_user_nice);
4375 * can_nice - check if a task can reduce its nice value
4379 int can_nice(const struct task_struct *p, const int nice)
4381 /* convert nice value [19,-20] to rlimit style value [1,40] */
4382 int nice_rlim = 20 - nice;
4384 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4385 capable(CAP_SYS_NICE));
4388 #ifdef __ARCH_WANT_SYS_NICE
4391 * sys_nice - change the priority of the current process.
4392 * @increment: priority increment
4394 * sys_setpriority is a more generic, but much slower function that
4395 * does similar things.
4397 SYSCALL_DEFINE1(nice, int, increment)
4402 * Setpriority might change our priority at the same moment.
4403 * We don't have to worry. Conceptually one call occurs first
4404 * and we have a single winner.
4406 if (increment < -40)
4411 nice = TASK_NICE(current) + increment;
4417 if (increment < 0 && !can_nice(current, nice))
4420 retval = security_task_setnice(current, nice);
4424 set_user_nice(current, nice);
4431 * task_prio - return the priority value of a given task.
4432 * @p: the task in question.
4434 * This is the priority value as seen by users in /proc.
4435 * RT tasks are offset by -200. Normal tasks are centered
4436 * around 0, value goes from -16 to +15.
4438 int task_prio(const struct task_struct *p)
4440 return p->prio - MAX_RT_PRIO;
4444 * task_nice - return the nice value of a given task.
4445 * @p: the task in question.
4447 int task_nice(const struct task_struct *p)
4449 return TASK_NICE(p);
4451 EXPORT_SYMBOL(task_nice);
4454 * idle_cpu - is a given cpu idle currently?
4455 * @cpu: the processor in question.
4457 int idle_cpu(int cpu)
4459 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4463 * idle_task - return the idle task for a given cpu.
4464 * @cpu: the processor in question.
4466 struct task_struct *idle_task(int cpu)
4468 return cpu_rq(cpu)->idle;
4472 * find_process_by_pid - find a process with a matching PID value.
4473 * @pid: the pid in question.
4475 static struct task_struct *find_process_by_pid(pid_t pid)
4477 return pid ? find_task_by_vpid(pid) : current;
4480 /* Actually do priority change: must hold rq lock. */
4482 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4484 BUG_ON(p->se.on_rq);
4487 p->rt_priority = prio;
4488 p->normal_prio = normal_prio(p);
4489 /* we are holding p->pi_lock already */
4490 p->prio = rt_mutex_getprio(p);
4491 if (rt_prio(p->prio))
4492 p->sched_class = &rt_sched_class;
4494 p->sched_class = &fair_sched_class;
4499 * check the target process has a UID that matches the current process's
4501 static bool check_same_owner(struct task_struct *p)
4503 const struct cred *cred = current_cred(), *pcred;
4507 pcred = __task_cred(p);
4508 match = (cred->euid == pcred->euid ||
4509 cred->euid == pcred->uid);
4514 static int __sched_setscheduler(struct task_struct *p, int policy,
4515 struct sched_param *param, bool user)
4517 int retval, oldprio, oldpolicy = -1, on_rq, running;
4518 unsigned long flags;
4519 const struct sched_class *prev_class;
4523 /* may grab non-irq protected spin_locks */
4524 BUG_ON(in_interrupt());
4526 /* double check policy once rq lock held */
4528 reset_on_fork = p->sched_reset_on_fork;
4529 policy = oldpolicy = p->policy;
4531 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4532 policy &= ~SCHED_RESET_ON_FORK;
4534 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4535 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4536 policy != SCHED_IDLE)
4541 * Valid priorities for SCHED_FIFO and SCHED_RR are
4542 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4543 * SCHED_BATCH and SCHED_IDLE is 0.
4545 if (param->sched_priority < 0 ||
4546 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4547 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4549 if (rt_policy(policy) != (param->sched_priority != 0))
4553 * Allow unprivileged RT tasks to decrease priority:
4555 if (user && !capable(CAP_SYS_NICE)) {
4556 if (rt_policy(policy)) {
4557 unsigned long rlim_rtprio;
4559 if (!lock_task_sighand(p, &flags))
4561 rlim_rtprio = task_rlimit(p, RLIMIT_RTPRIO);
4562 unlock_task_sighand(p, &flags);
4564 /* can't set/change the rt policy */
4565 if (policy != p->policy && !rlim_rtprio)
4568 /* can't increase priority */
4569 if (param->sched_priority > p->rt_priority &&
4570 param->sched_priority > rlim_rtprio)
4574 * Like positive nice levels, dont allow tasks to
4575 * move out of SCHED_IDLE either:
4577 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4580 /* can't change other user's priorities */
4581 if (!check_same_owner(p))
4584 /* Normal users shall not reset the sched_reset_on_fork flag */
4585 if (p->sched_reset_on_fork && !reset_on_fork)
4590 #ifdef CONFIG_RT_GROUP_SCHED
4592 * Do not allow realtime tasks into groups that have no runtime
4595 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4596 task_group(p)->rt_bandwidth.rt_runtime == 0)
4600 retval = security_task_setscheduler(p, policy, param);
4606 * make sure no PI-waiters arrive (or leave) while we are
4607 * changing the priority of the task:
4609 raw_spin_lock_irqsave(&p->pi_lock, flags);
4611 * To be able to change p->policy safely, the apropriate
4612 * runqueue lock must be held.
4614 rq = __task_rq_lock(p);
4615 /* recheck policy now with rq lock held */
4616 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4617 policy = oldpolicy = -1;
4618 __task_rq_unlock(rq);
4619 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4622 update_rq_clock(rq);
4623 on_rq = p->se.on_rq;
4624 running = task_current(rq, p);
4626 deactivate_task(rq, p, 0);
4628 p->sched_class->put_prev_task(rq, p);
4630 p->sched_reset_on_fork = reset_on_fork;
4633 prev_class = p->sched_class;
4634 __setscheduler(rq, p, policy, param->sched_priority);
4637 p->sched_class->set_curr_task(rq);
4639 activate_task(rq, p, 0);
4641 check_class_changed(rq, p, prev_class, oldprio, running);
4643 __task_rq_unlock(rq);
4644 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4646 rt_mutex_adjust_pi(p);
4652 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4653 * @p: the task in question.
4654 * @policy: new policy.
4655 * @param: structure containing the new RT priority.
4657 * NOTE that the task may be already dead.
4659 int sched_setscheduler(struct task_struct *p, int policy,
4660 struct sched_param *param)
4662 return __sched_setscheduler(p, policy, param, true);
4664 EXPORT_SYMBOL_GPL(sched_setscheduler);
4667 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4668 * @p: the task in question.
4669 * @policy: new policy.
4670 * @param: structure containing the new RT priority.
4672 * Just like sched_setscheduler, only don't bother checking if the
4673 * current context has permission. For example, this is needed in
4674 * stop_machine(): we create temporary high priority worker threads,
4675 * but our caller might not have that capability.
4677 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4678 struct sched_param *param)
4680 return __sched_setscheduler(p, policy, param, false);
4684 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4686 struct sched_param lparam;
4687 struct task_struct *p;
4690 if (!param || pid < 0)
4692 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4697 p = find_process_by_pid(pid);
4699 retval = sched_setscheduler(p, policy, &lparam);
4706 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4707 * @pid: the pid in question.
4708 * @policy: new policy.
4709 * @param: structure containing the new RT priority.
4711 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4712 struct sched_param __user *, param)
4714 /* negative values for policy are not valid */
4718 return do_sched_setscheduler(pid, policy, param);
4722 * sys_sched_setparam - set/change the RT priority of a thread
4723 * @pid: the pid in question.
4724 * @param: structure containing the new RT priority.
4726 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4728 return do_sched_setscheduler(pid, -1, param);
4732 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4733 * @pid: the pid in question.
4735 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4737 struct task_struct *p;
4745 p = find_process_by_pid(pid);
4747 retval = security_task_getscheduler(p);
4750 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4757 * sys_sched_getparam - get the RT priority of a thread
4758 * @pid: the pid in question.
4759 * @param: structure containing the RT priority.
4761 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4763 struct sched_param lp;
4764 struct task_struct *p;
4767 if (!param || pid < 0)
4771 p = find_process_by_pid(pid);
4776 retval = security_task_getscheduler(p);
4780 lp.sched_priority = p->rt_priority;
4784 * This one might sleep, we cannot do it with a spinlock held ...
4786 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4795 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4797 cpumask_var_t cpus_allowed, new_mask;
4798 struct task_struct *p;
4804 p = find_process_by_pid(pid);
4811 /* Prevent p going away */
4815 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4819 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4821 goto out_free_cpus_allowed;
4824 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
4827 retval = security_task_setscheduler(p, 0, NULL);
4831 cpuset_cpus_allowed(p, cpus_allowed);
4832 cpumask_and(new_mask, in_mask, cpus_allowed);
4834 retval = set_cpus_allowed_ptr(p, new_mask);
4837 cpuset_cpus_allowed(p, cpus_allowed);
4838 if (!cpumask_subset(new_mask, cpus_allowed)) {
4840 * We must have raced with a concurrent cpuset
4841 * update. Just reset the cpus_allowed to the
4842 * cpuset's cpus_allowed
4844 cpumask_copy(new_mask, cpus_allowed);
4849 free_cpumask_var(new_mask);
4850 out_free_cpus_allowed:
4851 free_cpumask_var(cpus_allowed);
4858 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4859 struct cpumask *new_mask)
4861 if (len < cpumask_size())
4862 cpumask_clear(new_mask);
4863 else if (len > cpumask_size())
4864 len = cpumask_size();
4866 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4870 * sys_sched_setaffinity - set the cpu affinity of a process
4871 * @pid: pid of the process
4872 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4873 * @user_mask_ptr: user-space pointer to the new cpu mask
4875 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4876 unsigned long __user *, user_mask_ptr)
4878 cpumask_var_t new_mask;
4881 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4884 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4886 retval = sched_setaffinity(pid, new_mask);
4887 free_cpumask_var(new_mask);
4891 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4893 struct task_struct *p;
4894 unsigned long flags;
4902 p = find_process_by_pid(pid);
4906 retval = security_task_getscheduler(p);
4910 rq = task_rq_lock(p, &flags);
4911 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4912 task_rq_unlock(rq, &flags);
4922 * sys_sched_getaffinity - get the cpu affinity of a process
4923 * @pid: pid of the process
4924 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4925 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4927 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4928 unsigned long __user *, user_mask_ptr)
4933 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4935 if (len & (sizeof(unsigned long)-1))
4938 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4941 ret = sched_getaffinity(pid, mask);
4943 size_t retlen = min_t(size_t, len, cpumask_size());
4945 if (copy_to_user(user_mask_ptr, mask, retlen))
4950 free_cpumask_var(mask);
4956 * sys_sched_yield - yield the current processor to other threads.
4958 * This function yields the current CPU to other tasks. If there are no
4959 * other threads running on this CPU then this function will return.
4961 SYSCALL_DEFINE0(sched_yield)
4963 struct rq *rq = this_rq_lock();
4965 schedstat_inc(rq, yld_count);
4966 current->sched_class->yield_task(rq);
4969 * Since we are going to call schedule() anyway, there's
4970 * no need to preempt or enable interrupts:
4972 __release(rq->lock);
4973 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4974 do_raw_spin_unlock(&rq->lock);
4975 preempt_enable_no_resched();
4982 static inline int should_resched(void)
4984 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4987 static void __cond_resched(void)
4989 add_preempt_count(PREEMPT_ACTIVE);
4991 sub_preempt_count(PREEMPT_ACTIVE);
4994 int __sched _cond_resched(void)
4996 if (should_resched()) {
5002 EXPORT_SYMBOL(_cond_resched);
5005 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5006 * call schedule, and on return reacquire the lock.
5008 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5009 * operations here to prevent schedule() from being called twice (once via
5010 * spin_unlock(), once by hand).
5012 int __cond_resched_lock(spinlock_t *lock)
5014 int resched = should_resched();
5017 lockdep_assert_held(lock);
5019 if (spin_needbreak(lock) || resched) {
5030 EXPORT_SYMBOL(__cond_resched_lock);
5032 int __sched __cond_resched_softirq(void)
5034 BUG_ON(!in_softirq());
5036 if (should_resched()) {
5044 EXPORT_SYMBOL(__cond_resched_softirq);
5047 * yield - yield the current processor to other threads.
5049 * This is a shortcut for kernel-space yielding - it marks the
5050 * thread runnable and calls sys_sched_yield().
5052 void __sched yield(void)
5054 set_current_state(TASK_RUNNING);
5057 EXPORT_SYMBOL(yield);
5060 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5061 * that process accounting knows that this is a task in IO wait state.
5063 void __sched io_schedule(void)
5065 struct rq *rq = raw_rq();
5067 delayacct_blkio_start();
5068 atomic_inc(&rq->nr_iowait);
5069 current->in_iowait = 1;
5071 current->in_iowait = 0;
5072 atomic_dec(&rq->nr_iowait);
5073 delayacct_blkio_end();
5075 EXPORT_SYMBOL(io_schedule);
5077 long __sched io_schedule_timeout(long timeout)
5079 struct rq *rq = raw_rq();
5082 delayacct_blkio_start();
5083 atomic_inc(&rq->nr_iowait);
5084 current->in_iowait = 1;
5085 ret = schedule_timeout(timeout);
5086 current->in_iowait = 0;
5087 atomic_dec(&rq->nr_iowait);
5088 delayacct_blkio_end();
5093 * sys_sched_get_priority_max - return maximum RT priority.
5094 * @policy: scheduling class.
5096 * this syscall returns the maximum rt_priority that can be used
5097 * by a given scheduling class.
5099 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5106 ret = MAX_USER_RT_PRIO-1;
5118 * sys_sched_get_priority_min - return minimum RT priority.
5119 * @policy: scheduling class.
5121 * this syscall returns the minimum rt_priority that can be used
5122 * by a given scheduling class.
5124 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5142 * sys_sched_rr_get_interval - return the default timeslice of a process.
5143 * @pid: pid of the process.
5144 * @interval: userspace pointer to the timeslice value.
5146 * this syscall writes the default timeslice value of a given process
5147 * into the user-space timespec buffer. A value of '0' means infinity.
5149 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5150 struct timespec __user *, interval)
5152 struct task_struct *p;
5153 unsigned int time_slice;
5154 unsigned long flags;
5164 p = find_process_by_pid(pid);
5168 retval = security_task_getscheduler(p);
5172 rq = task_rq_lock(p, &flags);
5173 time_slice = p->sched_class->get_rr_interval(rq, p);
5174 task_rq_unlock(rq, &flags);
5177 jiffies_to_timespec(time_slice, &t);
5178 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5186 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5188 void sched_show_task(struct task_struct *p)
5190 unsigned long free = 0;
5193 state = p->state ? __ffs(p->state) + 1 : 0;
5194 printk(KERN_INFO "%-13.13s %c", p->comm,
5195 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5196 #if BITS_PER_LONG == 32
5197 if (state == TASK_RUNNING)
5198 printk(KERN_CONT " running ");
5200 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5202 if (state == TASK_RUNNING)
5203 printk(KERN_CONT " running task ");
5205 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5207 #ifdef CONFIG_DEBUG_STACK_USAGE
5208 free = stack_not_used(p);
5210 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5211 task_pid_nr(p), task_pid_nr(p->real_parent),
5212 (unsigned long)task_thread_info(p)->flags);
5214 show_stack(p, NULL);
5217 void show_state_filter(unsigned long state_filter)
5219 struct task_struct *g, *p;
5221 #if BITS_PER_LONG == 32
5223 " task PC stack pid father\n");
5226 " task PC stack pid father\n");
5228 read_lock(&tasklist_lock);
5229 do_each_thread(g, p) {
5231 * reset the NMI-timeout, listing all files on a slow
5232 * console might take alot of time:
5234 touch_nmi_watchdog();
5235 if (!state_filter || (p->state & state_filter))
5237 } while_each_thread(g, p);
5239 touch_all_softlockup_watchdogs();
5241 #ifdef CONFIG_SCHED_DEBUG
5242 sysrq_sched_debug_show();
5244 read_unlock(&tasklist_lock);
5246 * Only show locks if all tasks are dumped:
5249 debug_show_all_locks();
5252 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5254 idle->sched_class = &idle_sched_class;
5258 * init_idle - set up an idle thread for a given CPU
5259 * @idle: task in question
5260 * @cpu: cpu the idle task belongs to
5262 * NOTE: this function does not set the idle thread's NEED_RESCHED
5263 * flag, to make booting more robust.
5265 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5267 struct rq *rq = cpu_rq(cpu);
5268 unsigned long flags;
5270 raw_spin_lock_irqsave(&rq->lock, flags);
5273 idle->state = TASK_RUNNING;
5274 idle->se.exec_start = sched_clock();
5276 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
5277 __set_task_cpu(idle, cpu);
5279 rq->curr = rq->idle = idle;
5280 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5283 raw_spin_unlock_irqrestore(&rq->lock, flags);
5285 /* Set the preempt count _outside_ the spinlocks! */
5286 #if defined(CONFIG_PREEMPT)
5287 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5289 task_thread_info(idle)->preempt_count = 0;
5292 * The idle tasks have their own, simple scheduling class:
5294 idle->sched_class = &idle_sched_class;
5295 ftrace_graph_init_task(idle);
5299 * In a system that switches off the HZ timer nohz_cpu_mask
5300 * indicates which cpus entered this state. This is used
5301 * in the rcu update to wait only for active cpus. For system
5302 * which do not switch off the HZ timer nohz_cpu_mask should
5303 * always be CPU_BITS_NONE.
5305 cpumask_var_t nohz_cpu_mask;
5308 * Increase the granularity value when there are more CPUs,
5309 * because with more CPUs the 'effective latency' as visible
5310 * to users decreases. But the relationship is not linear,
5311 * so pick a second-best guess by going with the log2 of the
5314 * This idea comes from the SD scheduler of Con Kolivas:
5316 static int get_update_sysctl_factor(void)
5318 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5319 unsigned int factor;
5321 switch (sysctl_sched_tunable_scaling) {
5322 case SCHED_TUNABLESCALING_NONE:
5325 case SCHED_TUNABLESCALING_LINEAR:
5328 case SCHED_TUNABLESCALING_LOG:
5330 factor = 1 + ilog2(cpus);
5337 static void update_sysctl(void)
5339 unsigned int factor = get_update_sysctl_factor();
5341 #define SET_SYSCTL(name) \
5342 (sysctl_##name = (factor) * normalized_sysctl_##name)
5343 SET_SYSCTL(sched_min_granularity);
5344 SET_SYSCTL(sched_latency);
5345 SET_SYSCTL(sched_wakeup_granularity);
5346 SET_SYSCTL(sched_shares_ratelimit);
5350 static inline void sched_init_granularity(void)
5357 * This is how migration works:
5359 * 1) we queue a struct migration_req structure in the source CPU's
5360 * runqueue and wake up that CPU's migration thread.
5361 * 2) we down() the locked semaphore => thread blocks.
5362 * 3) migration thread wakes up (implicitly it forces the migrated
5363 * thread off the CPU)
5364 * 4) it gets the migration request and checks whether the migrated
5365 * task is still in the wrong runqueue.
5366 * 5) if it's in the wrong runqueue then the migration thread removes
5367 * it and puts it into the right queue.
5368 * 6) migration thread up()s the semaphore.
5369 * 7) we wake up and the migration is done.
5373 * Change a given task's CPU affinity. Migrate the thread to a
5374 * proper CPU and schedule it away if the CPU it's executing on
5375 * is removed from the allowed bitmask.
5377 * NOTE: the caller must have a valid reference to the task, the
5378 * task must not exit() & deallocate itself prematurely. The
5379 * call is not atomic; no spinlocks may be held.
5381 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5383 struct migration_req req;
5384 unsigned long flags;
5388 rq = task_rq_lock(p, &flags);
5390 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5395 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5396 !cpumask_equal(&p->cpus_allowed, new_mask))) {
5401 if (p->sched_class->set_cpus_allowed)
5402 p->sched_class->set_cpus_allowed(p, new_mask);
5404 cpumask_copy(&p->cpus_allowed, new_mask);
5405 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5408 /* Can the task run on the task's current CPU? If so, we're done */
5409 if (cpumask_test_cpu(task_cpu(p), new_mask))
5412 if (migrate_task(p, cpumask_any_and(cpu_active_mask, new_mask), &req)) {
5413 /* Need help from migration thread: drop lock and wait. */
5414 struct task_struct *mt = rq->migration_thread;
5416 get_task_struct(mt);
5417 task_rq_unlock(rq, &flags);
5418 wake_up_process(mt);
5419 put_task_struct(mt);
5420 wait_for_completion(&req.done);
5421 tlb_migrate_finish(p->mm);
5425 task_rq_unlock(rq, &flags);
5429 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5432 * Move (not current) task off this cpu, onto dest cpu. We're doing
5433 * this because either it can't run here any more (set_cpus_allowed()
5434 * away from this CPU, or CPU going down), or because we're
5435 * attempting to rebalance this task on exec (sched_exec).
5437 * So we race with normal scheduler movements, but that's OK, as long
5438 * as the task is no longer on this CPU.
5440 * Returns non-zero if task was successfully migrated.
5442 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5444 struct rq *rq_dest, *rq_src;
5447 if (unlikely(!cpu_active(dest_cpu)))
5450 rq_src = cpu_rq(src_cpu);
5451 rq_dest = cpu_rq(dest_cpu);
5453 double_rq_lock(rq_src, rq_dest);
5454 /* Already moved. */
5455 if (task_cpu(p) != src_cpu)
5457 /* Affinity changed (again). */
5458 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
5462 * If we're not on a rq, the next wake-up will ensure we're
5466 deactivate_task(rq_src, p, 0);
5467 set_task_cpu(p, dest_cpu);
5468 activate_task(rq_dest, p, 0);
5469 check_preempt_curr(rq_dest, p, 0);
5474 double_rq_unlock(rq_src, rq_dest);
5478 #define RCU_MIGRATION_IDLE 0
5479 #define RCU_MIGRATION_NEED_QS 1
5480 #define RCU_MIGRATION_GOT_QS 2
5481 #define RCU_MIGRATION_MUST_SYNC 3
5484 * migration_thread - this is a highprio system thread that performs
5485 * thread migration by bumping thread off CPU then 'pushing' onto
5488 static int migration_thread(void *data)
5491 int cpu = (long)data;
5495 BUG_ON(rq->migration_thread != current);
5497 set_current_state(TASK_INTERRUPTIBLE);
5498 while (!kthread_should_stop()) {
5499 struct migration_req *req;
5500 struct list_head *head;
5502 raw_spin_lock_irq(&rq->lock);
5504 if (cpu_is_offline(cpu)) {
5505 raw_spin_unlock_irq(&rq->lock);
5509 if (rq->active_balance) {
5510 active_load_balance(rq, cpu);
5511 rq->active_balance = 0;
5514 head = &rq->migration_queue;
5516 if (list_empty(head)) {
5517 raw_spin_unlock_irq(&rq->lock);
5519 set_current_state(TASK_INTERRUPTIBLE);
5522 req = list_entry(head->next, struct migration_req, list);
5523 list_del_init(head->next);
5525 if (req->task != NULL) {
5526 raw_spin_unlock(&rq->lock);
5527 __migrate_task(req->task, cpu, req->dest_cpu);
5528 } else if (likely(cpu == (badcpu = smp_processor_id()))) {
5529 req->dest_cpu = RCU_MIGRATION_GOT_QS;
5530 raw_spin_unlock(&rq->lock);
5532 req->dest_cpu = RCU_MIGRATION_MUST_SYNC;
5533 raw_spin_unlock(&rq->lock);
5534 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu, cpu);
5538 complete(&req->done);
5540 __set_current_state(TASK_RUNNING);
5545 #ifdef CONFIG_HOTPLUG_CPU
5547 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5551 local_irq_disable();
5552 ret = __migrate_task(p, src_cpu, dest_cpu);
5558 * Figure out where task on dead CPU should go, use force if necessary.
5560 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5565 dest_cpu = select_fallback_rq(dead_cpu, p);
5567 /* It can have affinity changed while we were choosing. */
5568 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
5573 * While a dead CPU has no uninterruptible tasks queued at this point,
5574 * it might still have a nonzero ->nr_uninterruptible counter, because
5575 * for performance reasons the counter is not stricly tracking tasks to
5576 * their home CPUs. So we just add the counter to another CPU's counter,
5577 * to keep the global sum constant after CPU-down:
5579 static void migrate_nr_uninterruptible(struct rq *rq_src)
5581 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5582 unsigned long flags;
5584 local_irq_save(flags);
5585 double_rq_lock(rq_src, rq_dest);
5586 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5587 rq_src->nr_uninterruptible = 0;
5588 double_rq_unlock(rq_src, rq_dest);
5589 local_irq_restore(flags);
5592 /* Run through task list and migrate tasks from the dead cpu. */
5593 static void migrate_live_tasks(int src_cpu)
5595 struct task_struct *p, *t;
5597 read_lock(&tasklist_lock);
5599 do_each_thread(t, p) {
5603 if (task_cpu(p) == src_cpu)
5604 move_task_off_dead_cpu(src_cpu, p);
5605 } while_each_thread(t, p);
5607 read_unlock(&tasklist_lock);
5611 * Schedules idle task to be the next runnable task on current CPU.
5612 * It does so by boosting its priority to highest possible.
5613 * Used by CPU offline code.
5615 void sched_idle_next(void)
5617 int this_cpu = smp_processor_id();
5618 struct rq *rq = cpu_rq(this_cpu);
5619 struct task_struct *p = rq->idle;
5620 unsigned long flags;
5622 /* cpu has to be offline */
5623 BUG_ON(cpu_online(this_cpu));
5626 * Strictly not necessary since rest of the CPUs are stopped by now
5627 * and interrupts disabled on the current cpu.
5629 raw_spin_lock_irqsave(&rq->lock, flags);
5631 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5633 update_rq_clock(rq);
5634 activate_task(rq, p, 0);
5636 raw_spin_unlock_irqrestore(&rq->lock, flags);
5640 * Ensures that the idle task is using init_mm right before its cpu goes
5643 void idle_task_exit(void)
5645 struct mm_struct *mm = current->active_mm;
5647 BUG_ON(cpu_online(smp_processor_id()));
5650 switch_mm(mm, &init_mm, current);
5654 /* called under rq->lock with disabled interrupts */
5655 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5657 struct rq *rq = cpu_rq(dead_cpu);
5659 /* Must be exiting, otherwise would be on tasklist. */
5660 BUG_ON(!p->exit_state);
5662 /* Cannot have done final schedule yet: would have vanished. */
5663 BUG_ON(p->state == TASK_DEAD);
5668 * Drop lock around migration; if someone else moves it,
5669 * that's OK. No task can be added to this CPU, so iteration is
5672 raw_spin_unlock_irq(&rq->lock);
5673 move_task_off_dead_cpu(dead_cpu, p);
5674 raw_spin_lock_irq(&rq->lock);
5679 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5680 static void migrate_dead_tasks(unsigned int dead_cpu)
5682 struct rq *rq = cpu_rq(dead_cpu);
5683 struct task_struct *next;
5686 if (!rq->nr_running)
5688 update_rq_clock(rq);
5689 next = pick_next_task(rq);
5692 next->sched_class->put_prev_task(rq, next);
5693 migrate_dead(dead_cpu, next);
5699 * remove the tasks which were accounted by rq from calc_load_tasks.
5701 static void calc_global_load_remove(struct rq *rq)
5703 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5704 rq->calc_load_active = 0;
5706 #endif /* CONFIG_HOTPLUG_CPU */
5708 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5710 static struct ctl_table sd_ctl_dir[] = {
5712 .procname = "sched_domain",
5718 static struct ctl_table sd_ctl_root[] = {
5720 .procname = "kernel",
5722 .child = sd_ctl_dir,
5727 static struct ctl_table *sd_alloc_ctl_entry(int n)
5729 struct ctl_table *entry =
5730 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5735 static void sd_free_ctl_entry(struct ctl_table **tablep)
5737 struct ctl_table *entry;
5740 * In the intermediate directories, both the child directory and
5741 * procname are dynamically allocated and could fail but the mode
5742 * will always be set. In the lowest directory the names are
5743 * static strings and all have proc handlers.
5745 for (entry = *tablep; entry->mode; entry++) {
5747 sd_free_ctl_entry(&entry->child);
5748 if (entry->proc_handler == NULL)
5749 kfree(entry->procname);
5757 set_table_entry(struct ctl_table *entry,
5758 const char *procname, void *data, int maxlen,
5759 mode_t mode, proc_handler *proc_handler)
5761 entry->procname = procname;
5763 entry->maxlen = maxlen;
5765 entry->proc_handler = proc_handler;
5768 static struct ctl_table *
5769 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5771 struct ctl_table *table = sd_alloc_ctl_entry(13);
5776 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5777 sizeof(long), 0644, proc_doulongvec_minmax);
5778 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5779 sizeof(long), 0644, proc_doulongvec_minmax);
5780 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5781 sizeof(int), 0644, proc_dointvec_minmax);
5782 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5783 sizeof(int), 0644, proc_dointvec_minmax);
5784 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5785 sizeof(int), 0644, proc_dointvec_minmax);
5786 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5787 sizeof(int), 0644, proc_dointvec_minmax);
5788 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5789 sizeof(int), 0644, proc_dointvec_minmax);
5790 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5791 sizeof(int), 0644, proc_dointvec_minmax);
5792 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5793 sizeof(int), 0644, proc_dointvec_minmax);
5794 set_table_entry(&table[9], "cache_nice_tries",
5795 &sd->cache_nice_tries,
5796 sizeof(int), 0644, proc_dointvec_minmax);
5797 set_table_entry(&table[10], "flags", &sd->flags,
5798 sizeof(int), 0644, proc_dointvec_minmax);
5799 set_table_entry(&table[11], "name", sd->name,
5800 CORENAME_MAX_SIZE, 0444, proc_dostring);
5801 /* &table[12] is terminator */
5806 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5808 struct ctl_table *entry, *table;
5809 struct sched_domain *sd;
5810 int domain_num = 0, i;
5813 for_each_domain(cpu, sd)
5815 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5820 for_each_domain(cpu, sd) {
5821 snprintf(buf, 32, "domain%d", i);
5822 entry->procname = kstrdup(buf, GFP_KERNEL);
5824 entry->child = sd_alloc_ctl_domain_table(sd);
5831 static struct ctl_table_header *sd_sysctl_header;
5832 static void register_sched_domain_sysctl(void)
5834 int i, cpu_num = num_possible_cpus();
5835 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5838 WARN_ON(sd_ctl_dir[0].child);
5839 sd_ctl_dir[0].child = entry;
5844 for_each_possible_cpu(i) {
5845 snprintf(buf, 32, "cpu%d", i);
5846 entry->procname = kstrdup(buf, GFP_KERNEL);
5848 entry->child = sd_alloc_ctl_cpu_table(i);
5852 WARN_ON(sd_sysctl_header);
5853 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5856 /* may be called multiple times per register */
5857 static void unregister_sched_domain_sysctl(void)
5859 if (sd_sysctl_header)
5860 unregister_sysctl_table(sd_sysctl_header);
5861 sd_sysctl_header = NULL;
5862 if (sd_ctl_dir[0].child)
5863 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5866 static void register_sched_domain_sysctl(void)
5869 static void unregister_sched_domain_sysctl(void)
5874 static void set_rq_online(struct rq *rq)
5877 const struct sched_class *class;
5879 cpumask_set_cpu(rq->cpu, rq->rd->online);
5882 for_each_class(class) {
5883 if (class->rq_online)
5884 class->rq_online(rq);
5889 static void set_rq_offline(struct rq *rq)
5892 const struct sched_class *class;
5894 for_each_class(class) {
5895 if (class->rq_offline)
5896 class->rq_offline(rq);
5899 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5905 * migration_call - callback that gets triggered when a CPU is added.
5906 * Here we can start up the necessary migration thread for the new CPU.
5908 static int __cpuinit
5909 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5911 struct task_struct *p;
5912 int cpu = (long)hcpu;
5913 unsigned long flags;
5918 case CPU_UP_PREPARE:
5919 case CPU_UP_PREPARE_FROZEN:
5920 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
5923 kthread_bind(p, cpu);
5924 /* Must be high prio: stop_machine expects to yield to it. */
5925 rq = task_rq_lock(p, &flags);
5926 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5927 task_rq_unlock(rq, &flags);
5929 cpu_rq(cpu)->migration_thread = p;
5930 rq->calc_load_update = calc_load_update;
5934 case CPU_ONLINE_FROZEN:
5935 /* Strictly unnecessary, as first user will wake it. */
5936 wake_up_process(cpu_rq(cpu)->migration_thread);
5938 /* Update our root-domain */
5940 raw_spin_lock_irqsave(&rq->lock, flags);
5942 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5946 raw_spin_unlock_irqrestore(&rq->lock, flags);
5949 #ifdef CONFIG_HOTPLUG_CPU
5950 case CPU_UP_CANCELED:
5951 case CPU_UP_CANCELED_FROZEN:
5952 if (!cpu_rq(cpu)->migration_thread)
5954 /* Unbind it from offline cpu so it can run. Fall thru. */
5955 kthread_bind(cpu_rq(cpu)->migration_thread,
5956 cpumask_any(cpu_online_mask));
5957 kthread_stop(cpu_rq(cpu)->migration_thread);
5958 put_task_struct(cpu_rq(cpu)->migration_thread);
5959 cpu_rq(cpu)->migration_thread = NULL;
5963 case CPU_DEAD_FROZEN:
5964 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
5965 migrate_live_tasks(cpu);
5967 kthread_stop(rq->migration_thread);
5968 put_task_struct(rq->migration_thread);
5969 rq->migration_thread = NULL;
5970 /* Idle task back to normal (off runqueue, low prio) */
5971 raw_spin_lock_irq(&rq->lock);
5972 update_rq_clock(rq);
5973 deactivate_task(rq, rq->idle, 0);
5974 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5975 rq->idle->sched_class = &idle_sched_class;
5976 migrate_dead_tasks(cpu);
5977 raw_spin_unlock_irq(&rq->lock);
5979 migrate_nr_uninterruptible(rq);
5980 BUG_ON(rq->nr_running != 0);
5981 calc_global_load_remove(rq);
5983 * No need to migrate the tasks: it was best-effort if
5984 * they didn't take sched_hotcpu_mutex. Just wake up
5987 raw_spin_lock_irq(&rq->lock);
5988 while (!list_empty(&rq->migration_queue)) {
5989 struct migration_req *req;
5991 req = list_entry(rq->migration_queue.next,
5992 struct migration_req, list);
5993 list_del_init(&req->list);
5994 raw_spin_unlock_irq(&rq->lock);
5995 complete(&req->done);
5996 raw_spin_lock_irq(&rq->lock);
5998 raw_spin_unlock_irq(&rq->lock);
6002 case CPU_DYING_FROZEN:
6003 /* Update our root-domain */
6005 raw_spin_lock_irqsave(&rq->lock, flags);
6007 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6010 raw_spin_unlock_irqrestore(&rq->lock, flags);
6018 * Register at high priority so that task migration (migrate_all_tasks)
6019 * happens before everything else. This has to be lower priority than
6020 * the notifier in the perf_event subsystem, though.
6022 static struct notifier_block __cpuinitdata migration_notifier = {
6023 .notifier_call = migration_call,
6027 static int __init migration_init(void)
6029 void *cpu = (void *)(long)smp_processor_id();
6032 /* Start one for the boot CPU: */
6033 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6034 BUG_ON(err == NOTIFY_BAD);
6035 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6036 register_cpu_notifier(&migration_notifier);
6040 early_initcall(migration_init);
6045 #ifdef CONFIG_SCHED_DEBUG
6047 static __read_mostly int sched_domain_debug_enabled;
6049 static int __init sched_domain_debug_setup(char *str)
6051 sched_domain_debug_enabled = 1;
6055 early_param("sched_debug", sched_domain_debug_setup);
6057 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6058 struct cpumask *groupmask)
6060 struct sched_group *group = sd->groups;
6063 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6064 cpumask_clear(groupmask);
6066 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6068 if (!(sd->flags & SD_LOAD_BALANCE)) {
6069 printk("does not load-balance\n");
6071 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6076 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6078 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6079 printk(KERN_ERR "ERROR: domain->span does not contain "
6082 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6083 printk(KERN_ERR "ERROR: domain->groups does not contain"
6087 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6091 printk(KERN_ERR "ERROR: group is NULL\n");
6095 if (!group->cpu_power) {
6096 printk(KERN_CONT "\n");
6097 printk(KERN_ERR "ERROR: domain->cpu_power not "
6102 if (!cpumask_weight(sched_group_cpus(group))) {
6103 printk(KERN_CONT "\n");
6104 printk(KERN_ERR "ERROR: empty group\n");
6108 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6109 printk(KERN_CONT "\n");
6110 printk(KERN_ERR "ERROR: repeated CPUs\n");
6114 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6116 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6118 printk(KERN_CONT " %s", str);
6119 if (group->cpu_power != SCHED_LOAD_SCALE) {
6120 printk(KERN_CONT " (cpu_power = %d)",
6124 group = group->next;
6125 } while (group != sd->groups);
6126 printk(KERN_CONT "\n");
6128 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6129 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6132 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6133 printk(KERN_ERR "ERROR: parent span is not a superset "
6134 "of domain->span\n");
6138 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6140 cpumask_var_t groupmask;
6143 if (!sched_domain_debug_enabled)
6147 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6151 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6153 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6154 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6159 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6166 free_cpumask_var(groupmask);
6168 #else /* !CONFIG_SCHED_DEBUG */
6169 # define sched_domain_debug(sd, cpu) do { } while (0)
6170 #endif /* CONFIG_SCHED_DEBUG */
6172 static int sd_degenerate(struct sched_domain *sd)
6174 if (cpumask_weight(sched_domain_span(sd)) == 1)
6177 /* Following flags need at least 2 groups */
6178 if (sd->flags & (SD_LOAD_BALANCE |
6179 SD_BALANCE_NEWIDLE |
6183 SD_SHARE_PKG_RESOURCES)) {
6184 if (sd->groups != sd->groups->next)
6188 /* Following flags don't use groups */
6189 if (sd->flags & (SD_WAKE_AFFINE))
6196 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6198 unsigned long cflags = sd->flags, pflags = parent->flags;
6200 if (sd_degenerate(parent))
6203 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6206 /* Flags needing groups don't count if only 1 group in parent */
6207 if (parent->groups == parent->groups->next) {
6208 pflags &= ~(SD_LOAD_BALANCE |
6209 SD_BALANCE_NEWIDLE |
6213 SD_SHARE_PKG_RESOURCES);
6214 if (nr_node_ids == 1)
6215 pflags &= ~SD_SERIALIZE;
6217 if (~cflags & pflags)
6223 static void free_rootdomain(struct root_domain *rd)
6225 synchronize_sched();
6227 cpupri_cleanup(&rd->cpupri);
6229 free_cpumask_var(rd->rto_mask);
6230 free_cpumask_var(rd->online);
6231 free_cpumask_var(rd->span);
6235 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6237 struct root_domain *old_rd = NULL;
6238 unsigned long flags;
6240 raw_spin_lock_irqsave(&rq->lock, flags);
6245 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6248 cpumask_clear_cpu(rq->cpu, old_rd->span);
6251 * If we dont want to free the old_rt yet then
6252 * set old_rd to NULL to skip the freeing later
6255 if (!atomic_dec_and_test(&old_rd->refcount))
6259 atomic_inc(&rd->refcount);
6262 cpumask_set_cpu(rq->cpu, rd->span);
6263 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6266 raw_spin_unlock_irqrestore(&rq->lock, flags);
6269 free_rootdomain(old_rd);
6272 static int init_rootdomain(struct root_domain *rd, bool bootmem)
6274 gfp_t gfp = GFP_KERNEL;
6276 memset(rd, 0, sizeof(*rd));
6281 if (!alloc_cpumask_var(&rd->span, gfp))
6283 if (!alloc_cpumask_var(&rd->online, gfp))
6285 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
6288 if (cpupri_init(&rd->cpupri, bootmem) != 0)
6293 free_cpumask_var(rd->rto_mask);
6295 free_cpumask_var(rd->online);
6297 free_cpumask_var(rd->span);
6302 static void init_defrootdomain(void)
6304 init_rootdomain(&def_root_domain, true);
6306 atomic_set(&def_root_domain.refcount, 1);
6309 static struct root_domain *alloc_rootdomain(void)
6311 struct root_domain *rd;
6313 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6317 if (init_rootdomain(rd, false) != 0) {
6326 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6327 * hold the hotplug lock.
6330 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6332 struct rq *rq = cpu_rq(cpu);
6333 struct sched_domain *tmp;
6335 /* Remove the sched domains which do not contribute to scheduling. */
6336 for (tmp = sd; tmp; ) {
6337 struct sched_domain *parent = tmp->parent;
6341 if (sd_parent_degenerate(tmp, parent)) {
6342 tmp->parent = parent->parent;
6344 parent->parent->child = tmp;
6349 if (sd && sd_degenerate(sd)) {
6355 sched_domain_debug(sd, cpu);
6357 rq_attach_root(rq, rd);
6358 rcu_assign_pointer(rq->sd, sd);
6361 /* cpus with isolated domains */
6362 static cpumask_var_t cpu_isolated_map;
6364 /* Setup the mask of cpus configured for isolated domains */
6365 static int __init isolated_cpu_setup(char *str)
6367 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6368 cpulist_parse(str, cpu_isolated_map);
6372 __setup("isolcpus=", isolated_cpu_setup);
6375 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6376 * to a function which identifies what group(along with sched group) a CPU
6377 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6378 * (due to the fact that we keep track of groups covered with a struct cpumask).
6380 * init_sched_build_groups will build a circular linked list of the groups
6381 * covered by the given span, and will set each group's ->cpumask correctly,
6382 * and ->cpu_power to 0.
6385 init_sched_build_groups(const struct cpumask *span,
6386 const struct cpumask *cpu_map,
6387 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
6388 struct sched_group **sg,
6389 struct cpumask *tmpmask),
6390 struct cpumask *covered, struct cpumask *tmpmask)
6392 struct sched_group *first = NULL, *last = NULL;
6395 cpumask_clear(covered);
6397 for_each_cpu(i, span) {
6398 struct sched_group *sg;
6399 int group = group_fn(i, cpu_map, &sg, tmpmask);
6402 if (cpumask_test_cpu(i, covered))
6405 cpumask_clear(sched_group_cpus(sg));
6408 for_each_cpu(j, span) {
6409 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6412 cpumask_set_cpu(j, covered);
6413 cpumask_set_cpu(j, sched_group_cpus(sg));
6424 #define SD_NODES_PER_DOMAIN 16
6429 * find_next_best_node - find the next node to include in a sched_domain
6430 * @node: node whose sched_domain we're building
6431 * @used_nodes: nodes already in the sched_domain
6433 * Find the next node to include in a given scheduling domain. Simply
6434 * finds the closest node not already in the @used_nodes map.
6436 * Should use nodemask_t.
6438 static int find_next_best_node(int node, nodemask_t *used_nodes)
6440 int i, n, val, min_val, best_node = 0;
6444 for (i = 0; i < nr_node_ids; i++) {
6445 /* Start at @node */
6446 n = (node + i) % nr_node_ids;
6448 if (!nr_cpus_node(n))
6451 /* Skip already used nodes */
6452 if (node_isset(n, *used_nodes))
6455 /* Simple min distance search */
6456 val = node_distance(node, n);
6458 if (val < min_val) {
6464 node_set(best_node, *used_nodes);
6469 * sched_domain_node_span - get a cpumask for a node's sched_domain
6470 * @node: node whose cpumask we're constructing
6471 * @span: resulting cpumask
6473 * Given a node, construct a good cpumask for its sched_domain to span. It
6474 * should be one that prevents unnecessary balancing, but also spreads tasks
6477 static void sched_domain_node_span(int node, struct cpumask *span)
6479 nodemask_t used_nodes;
6482 cpumask_clear(span);
6483 nodes_clear(used_nodes);
6485 cpumask_or(span, span, cpumask_of_node(node));
6486 node_set(node, used_nodes);
6488 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6489 int next_node = find_next_best_node(node, &used_nodes);
6491 cpumask_or(span, span, cpumask_of_node(next_node));
6494 #endif /* CONFIG_NUMA */
6496 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6499 * The cpus mask in sched_group and sched_domain hangs off the end.
6501 * ( See the the comments in include/linux/sched.h:struct sched_group
6502 * and struct sched_domain. )
6504 struct static_sched_group {
6505 struct sched_group sg;
6506 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
6509 struct static_sched_domain {
6510 struct sched_domain sd;
6511 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
6517 cpumask_var_t domainspan;
6518 cpumask_var_t covered;
6519 cpumask_var_t notcovered;
6521 cpumask_var_t nodemask;
6522 cpumask_var_t this_sibling_map;
6523 cpumask_var_t this_core_map;
6524 cpumask_var_t send_covered;
6525 cpumask_var_t tmpmask;
6526 struct sched_group **sched_group_nodes;
6527 struct root_domain *rd;
6531 sa_sched_groups = 0,
6536 sa_this_sibling_map,
6538 sa_sched_group_nodes,
6548 * SMT sched-domains:
6550 #ifdef CONFIG_SCHED_SMT
6551 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
6552 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
6555 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
6556 struct sched_group **sg, struct cpumask *unused)
6559 *sg = &per_cpu(sched_groups, cpu).sg;
6562 #endif /* CONFIG_SCHED_SMT */
6565 * multi-core sched-domains:
6567 #ifdef CONFIG_SCHED_MC
6568 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
6569 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
6570 #endif /* CONFIG_SCHED_MC */
6572 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6574 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6575 struct sched_group **sg, struct cpumask *mask)
6579 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6580 group = cpumask_first(mask);
6582 *sg = &per_cpu(sched_group_core, group).sg;
6585 #elif defined(CONFIG_SCHED_MC)
6587 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6588 struct sched_group **sg, struct cpumask *unused)
6591 *sg = &per_cpu(sched_group_core, cpu).sg;
6596 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
6597 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
6600 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
6601 struct sched_group **sg, struct cpumask *mask)
6604 #ifdef CONFIG_SCHED_MC
6605 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6606 group = cpumask_first(mask);
6607 #elif defined(CONFIG_SCHED_SMT)
6608 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6609 group = cpumask_first(mask);
6614 *sg = &per_cpu(sched_group_phys, group).sg;
6620 * The init_sched_build_groups can't handle what we want to do with node
6621 * groups, so roll our own. Now each node has its own list of groups which
6622 * gets dynamically allocated.
6624 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
6625 static struct sched_group ***sched_group_nodes_bycpu;
6627 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
6628 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
6630 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
6631 struct sched_group **sg,
6632 struct cpumask *nodemask)
6636 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
6637 group = cpumask_first(nodemask);
6640 *sg = &per_cpu(sched_group_allnodes, group).sg;
6644 static void init_numa_sched_groups_power(struct sched_group *group_head)
6646 struct sched_group *sg = group_head;
6652 for_each_cpu(j, sched_group_cpus(sg)) {
6653 struct sched_domain *sd;
6655 sd = &per_cpu(phys_domains, j).sd;
6656 if (j != group_first_cpu(sd->groups)) {
6658 * Only add "power" once for each
6664 sg->cpu_power += sd->groups->cpu_power;
6667 } while (sg != group_head);
6670 static int build_numa_sched_groups(struct s_data *d,
6671 const struct cpumask *cpu_map, int num)
6673 struct sched_domain *sd;
6674 struct sched_group *sg, *prev;
6677 cpumask_clear(d->covered);
6678 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
6679 if (cpumask_empty(d->nodemask)) {
6680 d->sched_group_nodes[num] = NULL;
6684 sched_domain_node_span(num, d->domainspan);
6685 cpumask_and(d->domainspan, d->domainspan, cpu_map);
6687 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6690 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
6694 d->sched_group_nodes[num] = sg;
6696 for_each_cpu(j, d->nodemask) {
6697 sd = &per_cpu(node_domains, j).sd;
6702 cpumask_copy(sched_group_cpus(sg), d->nodemask);
6704 cpumask_or(d->covered, d->covered, d->nodemask);
6707 for (j = 0; j < nr_node_ids; j++) {
6708 n = (num + j) % nr_node_ids;
6709 cpumask_complement(d->notcovered, d->covered);
6710 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
6711 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
6712 if (cpumask_empty(d->tmpmask))
6714 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
6715 if (cpumask_empty(d->tmpmask))
6717 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6721 "Can not alloc domain group for node %d\n", j);
6725 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
6726 sg->next = prev->next;
6727 cpumask_or(d->covered, d->covered, d->tmpmask);
6734 #endif /* CONFIG_NUMA */
6737 /* Free memory allocated for various sched_group structures */
6738 static void free_sched_groups(const struct cpumask *cpu_map,
6739 struct cpumask *nodemask)
6743 for_each_cpu(cpu, cpu_map) {
6744 struct sched_group **sched_group_nodes
6745 = sched_group_nodes_bycpu[cpu];
6747 if (!sched_group_nodes)
6750 for (i = 0; i < nr_node_ids; i++) {
6751 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6753 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
6754 if (cpumask_empty(nodemask))
6764 if (oldsg != sched_group_nodes[i])
6767 kfree(sched_group_nodes);
6768 sched_group_nodes_bycpu[cpu] = NULL;
6771 #else /* !CONFIG_NUMA */
6772 static void free_sched_groups(const struct cpumask *cpu_map,
6773 struct cpumask *nodemask)
6776 #endif /* CONFIG_NUMA */
6779 * Initialize sched groups cpu_power.
6781 * cpu_power indicates the capacity of sched group, which is used while
6782 * distributing the load between different sched groups in a sched domain.
6783 * Typically cpu_power for all the groups in a sched domain will be same unless
6784 * there are asymmetries in the topology. If there are asymmetries, group
6785 * having more cpu_power will pickup more load compared to the group having
6788 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6790 struct sched_domain *child;
6791 struct sched_group *group;
6795 WARN_ON(!sd || !sd->groups);
6797 if (cpu != group_first_cpu(sd->groups))
6802 sd->groups->cpu_power = 0;
6805 power = SCHED_LOAD_SCALE;
6806 weight = cpumask_weight(sched_domain_span(sd));
6808 * SMT siblings share the power of a single core.
6809 * Usually multiple threads get a better yield out of
6810 * that one core than a single thread would have,
6811 * reflect that in sd->smt_gain.
6813 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
6814 power *= sd->smt_gain;
6816 power >>= SCHED_LOAD_SHIFT;
6818 sd->groups->cpu_power += power;
6823 * Add cpu_power of each child group to this groups cpu_power.
6825 group = child->groups;
6827 sd->groups->cpu_power += group->cpu_power;
6828 group = group->next;
6829 } while (group != child->groups);
6833 * Initializers for schedule domains
6834 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6837 #ifdef CONFIG_SCHED_DEBUG
6838 # define SD_INIT_NAME(sd, type) sd->name = #type
6840 # define SD_INIT_NAME(sd, type) do { } while (0)
6843 #define SD_INIT(sd, type) sd_init_##type(sd)
6845 #define SD_INIT_FUNC(type) \
6846 static noinline void sd_init_##type(struct sched_domain *sd) \
6848 memset(sd, 0, sizeof(*sd)); \
6849 *sd = SD_##type##_INIT; \
6850 sd->level = SD_LV_##type; \
6851 SD_INIT_NAME(sd, type); \
6856 SD_INIT_FUNC(ALLNODES)
6859 #ifdef CONFIG_SCHED_SMT
6860 SD_INIT_FUNC(SIBLING)
6862 #ifdef CONFIG_SCHED_MC
6866 static int default_relax_domain_level = -1;
6868 static int __init setup_relax_domain_level(char *str)
6872 val = simple_strtoul(str, NULL, 0);
6873 if (val < SD_LV_MAX)
6874 default_relax_domain_level = val;
6878 __setup("relax_domain_level=", setup_relax_domain_level);
6880 static void set_domain_attribute(struct sched_domain *sd,
6881 struct sched_domain_attr *attr)
6885 if (!attr || attr->relax_domain_level < 0) {
6886 if (default_relax_domain_level < 0)
6889 request = default_relax_domain_level;
6891 request = attr->relax_domain_level;
6892 if (request < sd->level) {
6893 /* turn off idle balance on this domain */
6894 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6896 /* turn on idle balance on this domain */
6897 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6901 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6902 const struct cpumask *cpu_map)
6905 case sa_sched_groups:
6906 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
6907 d->sched_group_nodes = NULL;
6909 free_rootdomain(d->rd); /* fall through */
6911 free_cpumask_var(d->tmpmask); /* fall through */
6912 case sa_send_covered:
6913 free_cpumask_var(d->send_covered); /* fall through */
6914 case sa_this_core_map:
6915 free_cpumask_var(d->this_core_map); /* fall through */
6916 case sa_this_sibling_map:
6917 free_cpumask_var(d->this_sibling_map); /* fall through */
6919 free_cpumask_var(d->nodemask); /* fall through */
6920 case sa_sched_group_nodes:
6922 kfree(d->sched_group_nodes); /* fall through */
6924 free_cpumask_var(d->notcovered); /* fall through */
6926 free_cpumask_var(d->covered); /* fall through */
6928 free_cpumask_var(d->domainspan); /* fall through */
6935 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6936 const struct cpumask *cpu_map)
6939 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
6941 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
6942 return sa_domainspan;
6943 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
6945 /* Allocate the per-node list of sched groups */
6946 d->sched_group_nodes = kcalloc(nr_node_ids,
6947 sizeof(struct sched_group *), GFP_KERNEL);
6948 if (!d->sched_group_nodes) {
6949 printk(KERN_WARNING "Can not alloc sched group node list\n");
6950 return sa_notcovered;
6952 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
6954 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
6955 return sa_sched_group_nodes;
6956 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
6958 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
6959 return sa_this_sibling_map;
6960 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
6961 return sa_this_core_map;
6962 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
6963 return sa_send_covered;
6964 d->rd = alloc_rootdomain();
6966 printk(KERN_WARNING "Cannot alloc root domain\n");
6969 return sa_rootdomain;
6972 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
6973 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
6975 struct sched_domain *sd = NULL;
6977 struct sched_domain *parent;
6980 if (cpumask_weight(cpu_map) >
6981 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
6982 sd = &per_cpu(allnodes_domains, i).sd;
6983 SD_INIT(sd, ALLNODES);
6984 set_domain_attribute(sd, attr);
6985 cpumask_copy(sched_domain_span(sd), cpu_map);
6986 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
6991 sd = &per_cpu(node_domains, i).sd;
6993 set_domain_attribute(sd, attr);
6994 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
6995 sd->parent = parent;
6998 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
7003 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
7004 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7005 struct sched_domain *parent, int i)
7007 struct sched_domain *sd;
7008 sd = &per_cpu(phys_domains, i).sd;
7010 set_domain_attribute(sd, attr);
7011 cpumask_copy(sched_domain_span(sd), d->nodemask);
7012 sd->parent = parent;
7015 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
7019 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
7020 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7021 struct sched_domain *parent, int i)
7023 struct sched_domain *sd = parent;
7024 #ifdef CONFIG_SCHED_MC
7025 sd = &per_cpu(core_domains, i).sd;
7027 set_domain_attribute(sd, attr);
7028 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
7029 sd->parent = parent;
7031 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
7036 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
7037 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7038 struct sched_domain *parent, int i)
7040 struct sched_domain *sd = parent;
7041 #ifdef CONFIG_SCHED_SMT
7042 sd = &per_cpu(cpu_domains, i).sd;
7043 SD_INIT(sd, SIBLING);
7044 set_domain_attribute(sd, attr);
7045 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
7046 sd->parent = parent;
7048 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
7053 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
7054 const struct cpumask *cpu_map, int cpu)
7057 #ifdef CONFIG_SCHED_SMT
7058 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
7059 cpumask_and(d->this_sibling_map, cpu_map,
7060 topology_thread_cpumask(cpu));
7061 if (cpu == cpumask_first(d->this_sibling_map))
7062 init_sched_build_groups(d->this_sibling_map, cpu_map,
7064 d->send_covered, d->tmpmask);
7067 #ifdef CONFIG_SCHED_MC
7068 case SD_LV_MC: /* set up multi-core groups */
7069 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
7070 if (cpu == cpumask_first(d->this_core_map))
7071 init_sched_build_groups(d->this_core_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_mc_sched_domain(&d, cpu_map, attr, sd, i);
7124 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
7127 for_each_cpu(i, cpu_map) {
7128 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
7129 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
7132 /* Set up physical groups */
7133 for (i = 0; i < nr_node_ids; i++)
7134 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
7137 /* Set up node groups */
7139 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
7141 for (i = 0; i < nr_node_ids; i++)
7142 if (build_numa_sched_groups(&d, cpu_map, i))
7146 /* Calculate CPU power for physical packages and nodes */
7147 #ifdef CONFIG_SCHED_SMT
7148 for_each_cpu(i, cpu_map) {
7149 sd = &per_cpu(cpu_domains, i).sd;
7150 init_sched_groups_power(i, sd);
7153 #ifdef CONFIG_SCHED_MC
7154 for_each_cpu(i, cpu_map) {
7155 sd = &per_cpu(core_domains, i).sd;
7156 init_sched_groups_power(i, sd);
7160 for_each_cpu(i, cpu_map) {
7161 sd = &per_cpu(phys_domains, i).sd;
7162 init_sched_groups_power(i, sd);
7166 for (i = 0; i < nr_node_ids; i++)
7167 init_numa_sched_groups_power(d.sched_group_nodes[i]);
7169 if (d.sd_allnodes) {
7170 struct sched_group *sg;
7172 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7174 init_numa_sched_groups_power(sg);
7178 /* Attach the domains */
7179 for_each_cpu(i, cpu_map) {
7180 #ifdef CONFIG_SCHED_SMT
7181 sd = &per_cpu(cpu_domains, i).sd;
7182 #elif defined(CONFIG_SCHED_MC)
7183 sd = &per_cpu(core_domains, i).sd;
7185 sd = &per_cpu(phys_domains, i).sd;
7187 cpu_attach_domain(sd, d.rd, i);
7190 d.sched_group_nodes = NULL; /* don't free this we still need it */
7191 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
7195 __free_domain_allocs(&d, alloc_state, cpu_map);
7199 static int build_sched_domains(const struct cpumask *cpu_map)
7201 return __build_sched_domains(cpu_map, NULL);
7204 static cpumask_var_t *doms_cur; /* current sched domains */
7205 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7206 static struct sched_domain_attr *dattr_cur;
7207 /* attribues of custom domains in 'doms_cur' */
7210 * Special case: If a kmalloc of a doms_cur partition (array of
7211 * cpumask) fails, then fallback to a single sched domain,
7212 * as determined by the single cpumask fallback_doms.
7214 static cpumask_var_t fallback_doms;
7217 * arch_update_cpu_topology lets virtualized architectures update the
7218 * cpu core maps. It is supposed to return 1 if the topology changed
7219 * or 0 if it stayed the same.
7221 int __attribute__((weak)) arch_update_cpu_topology(void)
7226 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7229 cpumask_var_t *doms;
7231 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7234 for (i = 0; i < ndoms; i++) {
7235 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7236 free_sched_domains(doms, i);
7243 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7246 for (i = 0; i < ndoms; i++)
7247 free_cpumask_var(doms[i]);
7252 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7253 * For now this just excludes isolated cpus, but could be used to
7254 * exclude other special cases in the future.
7256 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7260 arch_update_cpu_topology();
7262 doms_cur = alloc_sched_domains(ndoms_cur);
7264 doms_cur = &fallback_doms;
7265 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7267 err = build_sched_domains(doms_cur[0]);
7268 register_sched_domain_sysctl();
7273 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7274 struct cpumask *tmpmask)
7276 free_sched_groups(cpu_map, tmpmask);
7280 * Detach sched domains from a group of cpus specified in cpu_map
7281 * These cpus will now be attached to the NULL domain
7283 static void detach_destroy_domains(const struct cpumask *cpu_map)
7285 /* Save because hotplug lock held. */
7286 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7289 for_each_cpu(i, cpu_map)
7290 cpu_attach_domain(NULL, &def_root_domain, i);
7291 synchronize_sched();
7292 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7295 /* handle null as "default" */
7296 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7297 struct sched_domain_attr *new, int idx_new)
7299 struct sched_domain_attr tmp;
7306 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7307 new ? (new + idx_new) : &tmp,
7308 sizeof(struct sched_domain_attr));
7312 * Partition sched domains as specified by the 'ndoms_new'
7313 * cpumasks in the array doms_new[] of cpumasks. This compares
7314 * doms_new[] to the current sched domain partitioning, doms_cur[].
7315 * It destroys each deleted domain and builds each new domain.
7317 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7318 * The masks don't intersect (don't overlap.) We should setup one
7319 * sched domain for each mask. CPUs not in any of the cpumasks will
7320 * not be load balanced. If the same cpumask appears both in the
7321 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7324 * The passed in 'doms_new' should be allocated using
7325 * alloc_sched_domains. This routine takes ownership of it and will
7326 * free_sched_domains it when done with it. If the caller failed the
7327 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7328 * and partition_sched_domains() will fallback to the single partition
7329 * 'fallback_doms', it also forces the domains to be rebuilt.
7331 * If doms_new == NULL it will be replaced with cpu_online_mask.
7332 * ndoms_new == 0 is a special case for destroying existing domains,
7333 * and it will not create the default domain.
7335 * Call with hotplug lock held
7337 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7338 struct sched_domain_attr *dattr_new)
7343 mutex_lock(&sched_domains_mutex);
7345 /* always unregister in case we don't destroy any domains */
7346 unregister_sched_domain_sysctl();
7348 /* Let architecture update cpu core mappings. */
7349 new_topology = arch_update_cpu_topology();
7351 n = doms_new ? ndoms_new : 0;
7353 /* Destroy deleted domains */
7354 for (i = 0; i < ndoms_cur; i++) {
7355 for (j = 0; j < n && !new_topology; j++) {
7356 if (cpumask_equal(doms_cur[i], doms_new[j])
7357 && dattrs_equal(dattr_cur, i, dattr_new, j))
7360 /* no match - a current sched domain not in new doms_new[] */
7361 detach_destroy_domains(doms_cur[i]);
7366 if (doms_new == NULL) {
7368 doms_new = &fallback_doms;
7369 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7370 WARN_ON_ONCE(dattr_new);
7373 /* Build new domains */
7374 for (i = 0; i < ndoms_new; i++) {
7375 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7376 if (cpumask_equal(doms_new[i], doms_cur[j])
7377 && dattrs_equal(dattr_new, i, dattr_cur, j))
7380 /* no match - add a new doms_new */
7381 __build_sched_domains(doms_new[i],
7382 dattr_new ? dattr_new + i : NULL);
7387 /* Remember the new sched domains */
7388 if (doms_cur != &fallback_doms)
7389 free_sched_domains(doms_cur, ndoms_cur);
7390 kfree(dattr_cur); /* kfree(NULL) is safe */
7391 doms_cur = doms_new;
7392 dattr_cur = dattr_new;
7393 ndoms_cur = ndoms_new;
7395 register_sched_domain_sysctl();
7397 mutex_unlock(&sched_domains_mutex);
7400 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7401 static void arch_reinit_sched_domains(void)
7405 /* Destroy domains first to force the rebuild */
7406 partition_sched_domains(0, NULL, NULL);
7408 rebuild_sched_domains();
7412 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7414 unsigned int level = 0;
7416 if (sscanf(buf, "%u", &level) != 1)
7420 * level is always be positive so don't check for
7421 * level < POWERSAVINGS_BALANCE_NONE which is 0
7422 * What happens on 0 or 1 byte write,
7423 * need to check for count as well?
7426 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7430 sched_smt_power_savings = level;
7432 sched_mc_power_savings = level;
7434 arch_reinit_sched_domains();
7439 #ifdef CONFIG_SCHED_MC
7440 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7441 struct sysdev_class_attribute *attr,
7444 return sprintf(page, "%u\n", sched_mc_power_savings);
7446 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7447 struct sysdev_class_attribute *attr,
7448 const char *buf, size_t count)
7450 return sched_power_savings_store(buf, count, 0);
7452 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7453 sched_mc_power_savings_show,
7454 sched_mc_power_savings_store);
7457 #ifdef CONFIG_SCHED_SMT
7458 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7459 struct sysdev_class_attribute *attr,
7462 return sprintf(page, "%u\n", sched_smt_power_savings);
7464 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7465 struct sysdev_class_attribute *attr,
7466 const char *buf, size_t count)
7468 return sched_power_savings_store(buf, count, 1);
7470 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7471 sched_smt_power_savings_show,
7472 sched_smt_power_savings_store);
7475 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7479 #ifdef CONFIG_SCHED_SMT
7481 err = sysfs_create_file(&cls->kset.kobj,
7482 &attr_sched_smt_power_savings.attr);
7484 #ifdef CONFIG_SCHED_MC
7485 if (!err && mc_capable())
7486 err = sysfs_create_file(&cls->kset.kobj,
7487 &attr_sched_mc_power_savings.attr);
7491 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7493 #ifndef CONFIG_CPUSETS
7495 * Add online and remove offline CPUs from the scheduler domains.
7496 * When cpusets are enabled they take over this function.
7498 static int update_sched_domains(struct notifier_block *nfb,
7499 unsigned long action, void *hcpu)
7503 case CPU_ONLINE_FROZEN:
7504 case CPU_DOWN_PREPARE:
7505 case CPU_DOWN_PREPARE_FROZEN:
7506 case CPU_DOWN_FAILED:
7507 case CPU_DOWN_FAILED_FROZEN:
7508 partition_sched_domains(1, NULL, NULL);
7517 static int update_runtime(struct notifier_block *nfb,
7518 unsigned long action, void *hcpu)
7520 int cpu = (int)(long)hcpu;
7523 case CPU_DOWN_PREPARE:
7524 case CPU_DOWN_PREPARE_FROZEN:
7525 disable_runtime(cpu_rq(cpu));
7528 case CPU_DOWN_FAILED:
7529 case CPU_DOWN_FAILED_FROZEN:
7531 case CPU_ONLINE_FROZEN:
7532 enable_runtime(cpu_rq(cpu));
7540 void __init sched_init_smp(void)
7542 cpumask_var_t non_isolated_cpus;
7544 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7545 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7547 #if defined(CONFIG_NUMA)
7548 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7550 BUG_ON(sched_group_nodes_bycpu == NULL);
7553 mutex_lock(&sched_domains_mutex);
7554 arch_init_sched_domains(cpu_active_mask);
7555 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7556 if (cpumask_empty(non_isolated_cpus))
7557 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7558 mutex_unlock(&sched_domains_mutex);
7561 #ifndef CONFIG_CPUSETS
7562 /* XXX: Theoretical race here - CPU may be hotplugged now */
7563 hotcpu_notifier(update_sched_domains, 0);
7566 /* RT runtime code needs to handle some hotplug events */
7567 hotcpu_notifier(update_runtime, 0);
7571 /* Move init over to a non-isolated CPU */
7572 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7574 sched_init_granularity();
7575 free_cpumask_var(non_isolated_cpus);
7577 init_sched_rt_class();
7580 void __init sched_init_smp(void)
7582 sched_init_granularity();
7584 #endif /* CONFIG_SMP */
7586 const_debug unsigned int sysctl_timer_migration = 1;
7588 int in_sched_functions(unsigned long addr)
7590 return in_lock_functions(addr) ||
7591 (addr >= (unsigned long)__sched_text_start
7592 && addr < (unsigned long)__sched_text_end);
7595 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7597 cfs_rq->tasks_timeline = RB_ROOT;
7598 INIT_LIST_HEAD(&cfs_rq->tasks);
7599 #ifdef CONFIG_FAIR_GROUP_SCHED
7602 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7605 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7607 struct rt_prio_array *array;
7610 array = &rt_rq->active;
7611 for (i = 0; i < MAX_RT_PRIO; i++) {
7612 INIT_LIST_HEAD(array->queue + i);
7613 __clear_bit(i, array->bitmap);
7615 /* delimiter for bitsearch: */
7616 __set_bit(MAX_RT_PRIO, array->bitmap);
7618 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7619 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7621 rt_rq->highest_prio.next = MAX_RT_PRIO;
7625 rt_rq->rt_nr_migratory = 0;
7626 rt_rq->overloaded = 0;
7627 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
7631 rt_rq->rt_throttled = 0;
7632 rt_rq->rt_runtime = 0;
7633 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7635 #ifdef CONFIG_RT_GROUP_SCHED
7636 rt_rq->rt_nr_boosted = 0;
7641 #ifdef CONFIG_FAIR_GROUP_SCHED
7642 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7643 struct sched_entity *se, int cpu, int add,
7644 struct sched_entity *parent)
7646 struct rq *rq = cpu_rq(cpu);
7647 tg->cfs_rq[cpu] = cfs_rq;
7648 init_cfs_rq(cfs_rq, rq);
7651 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7654 /* se could be NULL for init_task_group */
7659 se->cfs_rq = &rq->cfs;
7661 se->cfs_rq = parent->my_q;
7664 se->load.weight = tg->shares;
7665 se->load.inv_weight = 0;
7666 se->parent = parent;
7670 #ifdef CONFIG_RT_GROUP_SCHED
7671 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7672 struct sched_rt_entity *rt_se, int cpu, int add,
7673 struct sched_rt_entity *parent)
7675 struct rq *rq = cpu_rq(cpu);
7677 tg->rt_rq[cpu] = rt_rq;
7678 init_rt_rq(rt_rq, rq);
7680 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7682 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7684 tg->rt_se[cpu] = rt_se;
7689 rt_se->rt_rq = &rq->rt;
7691 rt_se->rt_rq = parent->my_q;
7693 rt_se->my_q = rt_rq;
7694 rt_se->parent = parent;
7695 INIT_LIST_HEAD(&rt_se->run_list);
7699 void __init sched_init(void)
7702 unsigned long alloc_size = 0, ptr;
7704 #ifdef CONFIG_FAIR_GROUP_SCHED
7705 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7707 #ifdef CONFIG_RT_GROUP_SCHED
7708 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7710 #ifdef CONFIG_CPUMASK_OFFSTACK
7711 alloc_size += num_possible_cpus() * cpumask_size();
7714 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7716 #ifdef CONFIG_FAIR_GROUP_SCHED
7717 init_task_group.se = (struct sched_entity **)ptr;
7718 ptr += nr_cpu_ids * sizeof(void **);
7720 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7721 ptr += nr_cpu_ids * sizeof(void **);
7723 #endif /* CONFIG_FAIR_GROUP_SCHED */
7724 #ifdef CONFIG_RT_GROUP_SCHED
7725 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7726 ptr += nr_cpu_ids * sizeof(void **);
7728 init_task_group.rt_rq = (struct rt_rq **)ptr;
7729 ptr += nr_cpu_ids * sizeof(void **);
7731 #endif /* CONFIG_RT_GROUP_SCHED */
7732 #ifdef CONFIG_CPUMASK_OFFSTACK
7733 for_each_possible_cpu(i) {
7734 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
7735 ptr += cpumask_size();
7737 #endif /* CONFIG_CPUMASK_OFFSTACK */
7741 init_defrootdomain();
7744 init_rt_bandwidth(&def_rt_bandwidth,
7745 global_rt_period(), global_rt_runtime());
7747 #ifdef CONFIG_RT_GROUP_SCHED
7748 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7749 global_rt_period(), global_rt_runtime());
7750 #endif /* CONFIG_RT_GROUP_SCHED */
7752 #ifdef CONFIG_CGROUP_SCHED
7753 list_add(&init_task_group.list, &task_groups);
7754 INIT_LIST_HEAD(&init_task_group.children);
7756 #endif /* CONFIG_CGROUP_SCHED */
7758 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
7759 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
7760 __alignof__(unsigned long));
7762 for_each_possible_cpu(i) {
7766 raw_spin_lock_init(&rq->lock);
7768 rq->calc_load_active = 0;
7769 rq->calc_load_update = jiffies + LOAD_FREQ;
7770 init_cfs_rq(&rq->cfs, rq);
7771 init_rt_rq(&rq->rt, rq);
7772 #ifdef CONFIG_FAIR_GROUP_SCHED
7773 init_task_group.shares = init_task_group_load;
7774 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7775 #ifdef CONFIG_CGROUP_SCHED
7777 * How much cpu bandwidth does init_task_group get?
7779 * In case of task-groups formed thr' the cgroup filesystem, it
7780 * gets 100% of the cpu resources in the system. This overall
7781 * system cpu resource is divided among the tasks of
7782 * init_task_group and its child task-groups in a fair manner,
7783 * based on each entity's (task or task-group's) weight
7784 * (se->load.weight).
7786 * In other words, if init_task_group has 10 tasks of weight
7787 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7788 * then A0's share of the cpu resource is:
7790 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7792 * We achieve this by letting init_task_group's tasks sit
7793 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7795 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
7797 #endif /* CONFIG_FAIR_GROUP_SCHED */
7799 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7800 #ifdef CONFIG_RT_GROUP_SCHED
7801 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7802 #ifdef CONFIG_CGROUP_SCHED
7803 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
7807 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7808 rq->cpu_load[j] = 0;
7812 rq->post_schedule = 0;
7813 rq->active_balance = 0;
7814 rq->next_balance = jiffies;
7818 rq->migration_thread = NULL;
7820 rq->avg_idle = 2*sysctl_sched_migration_cost;
7821 INIT_LIST_HEAD(&rq->migration_queue);
7822 rq_attach_root(rq, &def_root_domain);
7825 atomic_set(&rq->nr_iowait, 0);
7828 set_load_weight(&init_task);
7830 #ifdef CONFIG_PREEMPT_NOTIFIERS
7831 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7835 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7838 #ifdef CONFIG_RT_MUTEXES
7839 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
7843 * The boot idle thread does lazy MMU switching as well:
7845 atomic_inc(&init_mm.mm_count);
7846 enter_lazy_tlb(&init_mm, current);
7849 * Make us the idle thread. Technically, schedule() should not be
7850 * called from this thread, however somewhere below it might be,
7851 * but because we are the idle thread, we just pick up running again
7852 * when this runqueue becomes "idle".
7854 init_idle(current, smp_processor_id());
7856 calc_load_update = jiffies + LOAD_FREQ;
7859 * During early bootup we pretend to be a normal task:
7861 current->sched_class = &fair_sched_class;
7863 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
7864 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
7867 zalloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
7868 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
7870 /* May be allocated at isolcpus cmdline parse time */
7871 if (cpu_isolated_map == NULL)
7872 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7877 scheduler_running = 1;
7880 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7881 static inline int preempt_count_equals(int preempt_offset)
7883 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7885 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
7888 void __might_sleep(const char *file, int line, int preempt_offset)
7891 static unsigned long prev_jiffy; /* ratelimiting */
7893 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7894 system_state != SYSTEM_RUNNING || oops_in_progress)
7896 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7898 prev_jiffy = jiffies;
7901 "BUG: sleeping function called from invalid context at %s:%d\n",
7904 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7905 in_atomic(), irqs_disabled(),
7906 current->pid, current->comm);
7908 debug_show_held_locks(current);
7909 if (irqs_disabled())
7910 print_irqtrace_events(current);
7914 EXPORT_SYMBOL(__might_sleep);
7917 #ifdef CONFIG_MAGIC_SYSRQ
7918 static void normalize_task(struct rq *rq, struct task_struct *p)
7922 update_rq_clock(rq);
7923 on_rq = p->se.on_rq;
7925 deactivate_task(rq, p, 0);
7926 __setscheduler(rq, p, SCHED_NORMAL, 0);
7928 activate_task(rq, p, 0);
7929 resched_task(rq->curr);
7933 void normalize_rt_tasks(void)
7935 struct task_struct *g, *p;
7936 unsigned long flags;
7939 read_lock_irqsave(&tasklist_lock, flags);
7940 do_each_thread(g, p) {
7942 * Only normalize user tasks:
7947 p->se.exec_start = 0;
7948 #ifdef CONFIG_SCHEDSTATS
7949 p->se.wait_start = 0;
7950 p->se.sleep_start = 0;
7951 p->se.block_start = 0;
7956 * Renice negative nice level userspace
7959 if (TASK_NICE(p) < 0 && p->mm)
7960 set_user_nice(p, 0);
7964 raw_spin_lock(&p->pi_lock);
7965 rq = __task_rq_lock(p);
7967 normalize_task(rq, p);
7969 __task_rq_unlock(rq);
7970 raw_spin_unlock(&p->pi_lock);
7971 } while_each_thread(g, p);
7973 read_unlock_irqrestore(&tasklist_lock, flags);
7976 #endif /* CONFIG_MAGIC_SYSRQ */
7980 * These functions are only useful for the IA64 MCA handling.
7982 * They can only be called when the whole system has been
7983 * stopped - every CPU needs to be quiescent, and no scheduling
7984 * activity can take place. Using them for anything else would
7985 * be a serious bug, and as a result, they aren't even visible
7986 * under any other configuration.
7990 * curr_task - return the current task for a given cpu.
7991 * @cpu: the processor in question.
7993 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7995 struct task_struct *curr_task(int cpu)
7997 return cpu_curr(cpu);
8001 * set_curr_task - set the current task for a given cpu.
8002 * @cpu: the processor in question.
8003 * @p: the task pointer to set.
8005 * Description: This function must only be used when non-maskable interrupts
8006 * are serviced on a separate stack. It allows the architecture to switch the
8007 * notion of the current task on a cpu in a non-blocking manner. This function
8008 * must be called with all CPU's synchronized, and interrupts disabled, the
8009 * and caller must save the original value of the current task (see
8010 * curr_task() above) and restore that value before reenabling interrupts and
8011 * re-starting the system.
8013 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8015 void set_curr_task(int cpu, struct task_struct *p)
8022 #ifdef CONFIG_FAIR_GROUP_SCHED
8023 static void free_fair_sched_group(struct task_group *tg)
8027 for_each_possible_cpu(i) {
8029 kfree(tg->cfs_rq[i]);
8039 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8041 struct cfs_rq *cfs_rq;
8042 struct sched_entity *se;
8046 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8049 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8053 tg->shares = NICE_0_LOAD;
8055 for_each_possible_cpu(i) {
8058 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8059 GFP_KERNEL, cpu_to_node(i));
8063 se = kzalloc_node(sizeof(struct sched_entity),
8064 GFP_KERNEL, cpu_to_node(i));
8068 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
8079 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8081 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8082 &cpu_rq(cpu)->leaf_cfs_rq_list);
8085 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8087 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8089 #else /* !CONFG_FAIR_GROUP_SCHED */
8090 static inline void free_fair_sched_group(struct task_group *tg)
8095 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8100 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8104 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8107 #endif /* CONFIG_FAIR_GROUP_SCHED */
8109 #ifdef CONFIG_RT_GROUP_SCHED
8110 static void free_rt_sched_group(struct task_group *tg)
8114 destroy_rt_bandwidth(&tg->rt_bandwidth);
8116 for_each_possible_cpu(i) {
8118 kfree(tg->rt_rq[i]);
8120 kfree(tg->rt_se[i]);
8128 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8130 struct rt_rq *rt_rq;
8131 struct sched_rt_entity *rt_se;
8135 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8138 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8142 init_rt_bandwidth(&tg->rt_bandwidth,
8143 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8145 for_each_possible_cpu(i) {
8148 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8149 GFP_KERNEL, cpu_to_node(i));
8153 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8154 GFP_KERNEL, cpu_to_node(i));
8158 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
8169 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8171 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8172 &cpu_rq(cpu)->leaf_rt_rq_list);
8175 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8177 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8179 #else /* !CONFIG_RT_GROUP_SCHED */
8180 static inline void free_rt_sched_group(struct task_group *tg)
8185 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8190 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8194 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8197 #endif /* CONFIG_RT_GROUP_SCHED */
8199 #ifdef CONFIG_CGROUP_SCHED
8200 static void free_sched_group(struct task_group *tg)
8202 free_fair_sched_group(tg);
8203 free_rt_sched_group(tg);
8207 /* allocate runqueue etc for a new task group */
8208 struct task_group *sched_create_group(struct task_group *parent)
8210 struct task_group *tg;
8211 unsigned long flags;
8214 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8216 return ERR_PTR(-ENOMEM);
8218 if (!alloc_fair_sched_group(tg, parent))
8221 if (!alloc_rt_sched_group(tg, parent))
8224 spin_lock_irqsave(&task_group_lock, flags);
8225 for_each_possible_cpu(i) {
8226 register_fair_sched_group(tg, i);
8227 register_rt_sched_group(tg, i);
8229 list_add_rcu(&tg->list, &task_groups);
8231 WARN_ON(!parent); /* root should already exist */
8233 tg->parent = parent;
8234 INIT_LIST_HEAD(&tg->children);
8235 list_add_rcu(&tg->siblings, &parent->children);
8236 spin_unlock_irqrestore(&task_group_lock, flags);
8241 free_sched_group(tg);
8242 return ERR_PTR(-ENOMEM);
8245 /* rcu callback to free various structures associated with a task group */
8246 static void free_sched_group_rcu(struct rcu_head *rhp)
8248 /* now it should be safe to free those cfs_rqs */
8249 free_sched_group(container_of(rhp, struct task_group, rcu));
8252 /* Destroy runqueue etc associated with a task group */
8253 void sched_destroy_group(struct task_group *tg)
8255 unsigned long flags;
8258 spin_lock_irqsave(&task_group_lock, flags);
8259 for_each_possible_cpu(i) {
8260 unregister_fair_sched_group(tg, i);
8261 unregister_rt_sched_group(tg, i);
8263 list_del_rcu(&tg->list);
8264 list_del_rcu(&tg->siblings);
8265 spin_unlock_irqrestore(&task_group_lock, flags);
8267 /* wait for possible concurrent references to cfs_rqs complete */
8268 call_rcu(&tg->rcu, free_sched_group_rcu);
8271 /* change task's runqueue when it moves between groups.
8272 * The caller of this function should have put the task in its new group
8273 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8274 * reflect its new group.
8276 void sched_move_task(struct task_struct *tsk)
8279 unsigned long flags;
8282 rq = task_rq_lock(tsk, &flags);
8284 update_rq_clock(rq);
8286 running = task_current(rq, tsk);
8287 on_rq = tsk->se.on_rq;
8290 dequeue_task(rq, tsk, 0);
8291 if (unlikely(running))
8292 tsk->sched_class->put_prev_task(rq, tsk);
8294 set_task_rq(tsk, task_cpu(tsk));
8296 #ifdef CONFIG_FAIR_GROUP_SCHED
8297 if (tsk->sched_class->moved_group)
8298 tsk->sched_class->moved_group(tsk, on_rq);
8301 if (unlikely(running))
8302 tsk->sched_class->set_curr_task(rq);
8304 enqueue_task(rq, tsk, 0, false);
8306 task_rq_unlock(rq, &flags);
8308 #endif /* CONFIG_CGROUP_SCHED */
8310 #ifdef CONFIG_FAIR_GROUP_SCHED
8311 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8313 struct cfs_rq *cfs_rq = se->cfs_rq;
8318 dequeue_entity(cfs_rq, se, 0);
8320 se->load.weight = shares;
8321 se->load.inv_weight = 0;
8324 enqueue_entity(cfs_rq, se, 0);
8327 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8329 struct cfs_rq *cfs_rq = se->cfs_rq;
8330 struct rq *rq = cfs_rq->rq;
8331 unsigned long flags;
8333 raw_spin_lock_irqsave(&rq->lock, flags);
8334 __set_se_shares(se, shares);
8335 raw_spin_unlock_irqrestore(&rq->lock, flags);
8338 static DEFINE_MUTEX(shares_mutex);
8340 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8343 unsigned long flags;
8346 * We can't change the weight of the root cgroup.
8351 if (shares < MIN_SHARES)
8352 shares = MIN_SHARES;
8353 else if (shares > MAX_SHARES)
8354 shares = MAX_SHARES;
8356 mutex_lock(&shares_mutex);
8357 if (tg->shares == shares)
8360 spin_lock_irqsave(&task_group_lock, flags);
8361 for_each_possible_cpu(i)
8362 unregister_fair_sched_group(tg, i);
8363 list_del_rcu(&tg->siblings);
8364 spin_unlock_irqrestore(&task_group_lock, flags);
8366 /* wait for any ongoing reference to this group to finish */
8367 synchronize_sched();
8370 * Now we are free to modify the group's share on each cpu
8371 * w/o tripping rebalance_share or load_balance_fair.
8373 tg->shares = shares;
8374 for_each_possible_cpu(i) {
8378 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8379 set_se_shares(tg->se[i], shares);
8383 * Enable load balance activity on this group, by inserting it back on
8384 * each cpu's rq->leaf_cfs_rq_list.
8386 spin_lock_irqsave(&task_group_lock, flags);
8387 for_each_possible_cpu(i)
8388 register_fair_sched_group(tg, i);
8389 list_add_rcu(&tg->siblings, &tg->parent->children);
8390 spin_unlock_irqrestore(&task_group_lock, flags);
8392 mutex_unlock(&shares_mutex);
8396 unsigned long sched_group_shares(struct task_group *tg)
8402 #ifdef CONFIG_RT_GROUP_SCHED
8404 * Ensure that the real time constraints are schedulable.
8406 static DEFINE_MUTEX(rt_constraints_mutex);
8408 static unsigned long to_ratio(u64 period, u64 runtime)
8410 if (runtime == RUNTIME_INF)
8413 return div64_u64(runtime << 20, period);
8416 /* Must be called with tasklist_lock held */
8417 static inline int tg_has_rt_tasks(struct task_group *tg)
8419 struct task_struct *g, *p;
8421 do_each_thread(g, p) {
8422 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8424 } while_each_thread(g, p);
8429 struct rt_schedulable_data {
8430 struct task_group *tg;
8435 static int tg_schedulable(struct task_group *tg, void *data)
8437 struct rt_schedulable_data *d = data;
8438 struct task_group *child;
8439 unsigned long total, sum = 0;
8440 u64 period, runtime;
8442 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8443 runtime = tg->rt_bandwidth.rt_runtime;
8446 period = d->rt_period;
8447 runtime = d->rt_runtime;
8451 * Cannot have more runtime than the period.
8453 if (runtime > period && runtime != RUNTIME_INF)
8457 * Ensure we don't starve existing RT tasks.
8459 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8462 total = to_ratio(period, runtime);
8465 * Nobody can have more than the global setting allows.
8467 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8471 * The sum of our children's runtime should not exceed our own.
8473 list_for_each_entry_rcu(child, &tg->children, siblings) {
8474 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8475 runtime = child->rt_bandwidth.rt_runtime;
8477 if (child == d->tg) {
8478 period = d->rt_period;
8479 runtime = d->rt_runtime;
8482 sum += to_ratio(period, runtime);
8491 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8493 struct rt_schedulable_data data = {
8495 .rt_period = period,
8496 .rt_runtime = runtime,
8499 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8502 static int tg_set_bandwidth(struct task_group *tg,
8503 u64 rt_period, u64 rt_runtime)
8507 mutex_lock(&rt_constraints_mutex);
8508 read_lock(&tasklist_lock);
8509 err = __rt_schedulable(tg, rt_period, rt_runtime);
8513 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8514 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8515 tg->rt_bandwidth.rt_runtime = rt_runtime;
8517 for_each_possible_cpu(i) {
8518 struct rt_rq *rt_rq = tg->rt_rq[i];
8520 raw_spin_lock(&rt_rq->rt_runtime_lock);
8521 rt_rq->rt_runtime = rt_runtime;
8522 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8524 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8526 read_unlock(&tasklist_lock);
8527 mutex_unlock(&rt_constraints_mutex);
8532 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8534 u64 rt_runtime, rt_period;
8536 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8537 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8538 if (rt_runtime_us < 0)
8539 rt_runtime = RUNTIME_INF;
8541 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8544 long sched_group_rt_runtime(struct task_group *tg)
8548 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8551 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8552 do_div(rt_runtime_us, NSEC_PER_USEC);
8553 return rt_runtime_us;
8556 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8558 u64 rt_runtime, rt_period;
8560 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8561 rt_runtime = tg->rt_bandwidth.rt_runtime;
8566 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8569 long sched_group_rt_period(struct task_group *tg)
8573 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8574 do_div(rt_period_us, NSEC_PER_USEC);
8575 return rt_period_us;
8578 static int sched_rt_global_constraints(void)
8580 u64 runtime, period;
8583 if (sysctl_sched_rt_period <= 0)
8586 runtime = global_rt_runtime();
8587 period = global_rt_period();
8590 * Sanity check on the sysctl variables.
8592 if (runtime > period && runtime != RUNTIME_INF)
8595 mutex_lock(&rt_constraints_mutex);
8596 read_lock(&tasklist_lock);
8597 ret = __rt_schedulable(NULL, 0, 0);
8598 read_unlock(&tasklist_lock);
8599 mutex_unlock(&rt_constraints_mutex);
8604 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8606 /* Don't accept realtime tasks when there is no way for them to run */
8607 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8613 #else /* !CONFIG_RT_GROUP_SCHED */
8614 static int sched_rt_global_constraints(void)
8616 unsigned long flags;
8619 if (sysctl_sched_rt_period <= 0)
8623 * There's always some RT tasks in the root group
8624 * -- migration, kstopmachine etc..
8626 if (sysctl_sched_rt_runtime == 0)
8629 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8630 for_each_possible_cpu(i) {
8631 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8633 raw_spin_lock(&rt_rq->rt_runtime_lock);
8634 rt_rq->rt_runtime = global_rt_runtime();
8635 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8637 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8641 #endif /* CONFIG_RT_GROUP_SCHED */
8643 int sched_rt_handler(struct ctl_table *table, int write,
8644 void __user *buffer, size_t *lenp,
8648 int old_period, old_runtime;
8649 static DEFINE_MUTEX(mutex);
8652 old_period = sysctl_sched_rt_period;
8653 old_runtime = sysctl_sched_rt_runtime;
8655 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8657 if (!ret && write) {
8658 ret = sched_rt_global_constraints();
8660 sysctl_sched_rt_period = old_period;
8661 sysctl_sched_rt_runtime = old_runtime;
8663 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8664 def_rt_bandwidth.rt_period =
8665 ns_to_ktime(global_rt_period());
8668 mutex_unlock(&mutex);
8673 #ifdef CONFIG_CGROUP_SCHED
8675 /* return corresponding task_group object of a cgroup */
8676 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8678 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8679 struct task_group, css);
8682 static struct cgroup_subsys_state *
8683 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8685 struct task_group *tg, *parent;
8687 if (!cgrp->parent) {
8688 /* This is early initialization for the top cgroup */
8689 return &init_task_group.css;
8692 parent = cgroup_tg(cgrp->parent);
8693 tg = sched_create_group(parent);
8695 return ERR_PTR(-ENOMEM);
8701 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8703 struct task_group *tg = cgroup_tg(cgrp);
8705 sched_destroy_group(tg);
8709 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8711 #ifdef CONFIG_RT_GROUP_SCHED
8712 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
8715 /* We don't support RT-tasks being in separate groups */
8716 if (tsk->sched_class != &fair_sched_class)
8723 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8724 struct task_struct *tsk, bool threadgroup)
8726 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
8730 struct task_struct *c;
8732 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8733 retval = cpu_cgroup_can_attach_task(cgrp, c);
8745 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8746 struct cgroup *old_cont, struct task_struct *tsk,
8749 sched_move_task(tsk);
8751 struct task_struct *c;
8753 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8760 #ifdef CONFIG_FAIR_GROUP_SCHED
8761 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8764 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8767 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8769 struct task_group *tg = cgroup_tg(cgrp);
8771 return (u64) tg->shares;
8773 #endif /* CONFIG_FAIR_GROUP_SCHED */
8775 #ifdef CONFIG_RT_GROUP_SCHED
8776 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8779 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8782 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8784 return sched_group_rt_runtime(cgroup_tg(cgrp));
8787 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8790 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8793 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8795 return sched_group_rt_period(cgroup_tg(cgrp));
8797 #endif /* CONFIG_RT_GROUP_SCHED */
8799 static struct cftype cpu_files[] = {
8800 #ifdef CONFIG_FAIR_GROUP_SCHED
8803 .read_u64 = cpu_shares_read_u64,
8804 .write_u64 = cpu_shares_write_u64,
8807 #ifdef CONFIG_RT_GROUP_SCHED
8809 .name = "rt_runtime_us",
8810 .read_s64 = cpu_rt_runtime_read,
8811 .write_s64 = cpu_rt_runtime_write,
8814 .name = "rt_period_us",
8815 .read_u64 = cpu_rt_period_read_uint,
8816 .write_u64 = cpu_rt_period_write_uint,
8821 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8823 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8826 struct cgroup_subsys cpu_cgroup_subsys = {
8828 .create = cpu_cgroup_create,
8829 .destroy = cpu_cgroup_destroy,
8830 .can_attach = cpu_cgroup_can_attach,
8831 .attach = cpu_cgroup_attach,
8832 .populate = cpu_cgroup_populate,
8833 .subsys_id = cpu_cgroup_subsys_id,
8837 #endif /* CONFIG_CGROUP_SCHED */
8839 #ifdef CONFIG_CGROUP_CPUACCT
8842 * CPU accounting code for task groups.
8844 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8845 * (balbir@in.ibm.com).
8848 /* track cpu usage of a group of tasks and its child groups */
8850 struct cgroup_subsys_state css;
8851 /* cpuusage holds pointer to a u64-type object on every cpu */
8852 u64 __percpu *cpuusage;
8853 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
8854 struct cpuacct *parent;
8857 struct cgroup_subsys cpuacct_subsys;
8859 /* return cpu accounting group corresponding to this container */
8860 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8862 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8863 struct cpuacct, css);
8866 /* return cpu accounting group to which this task belongs */
8867 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8869 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8870 struct cpuacct, css);
8873 /* create a new cpu accounting group */
8874 static struct cgroup_subsys_state *cpuacct_create(
8875 struct cgroup_subsys *ss, struct cgroup *cgrp)
8877 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8883 ca->cpuusage = alloc_percpu(u64);
8887 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8888 if (percpu_counter_init(&ca->cpustat[i], 0))
8889 goto out_free_counters;
8892 ca->parent = cgroup_ca(cgrp->parent);
8898 percpu_counter_destroy(&ca->cpustat[i]);
8899 free_percpu(ca->cpuusage);
8903 return ERR_PTR(-ENOMEM);
8906 /* destroy an existing cpu accounting group */
8908 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8910 struct cpuacct *ca = cgroup_ca(cgrp);
8913 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8914 percpu_counter_destroy(&ca->cpustat[i]);
8915 free_percpu(ca->cpuusage);
8919 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
8921 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8924 #ifndef CONFIG_64BIT
8926 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8928 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8930 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8938 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
8940 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8942 #ifndef CONFIG_64BIT
8944 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8946 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8948 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8954 /* return total cpu usage (in nanoseconds) of a group */
8955 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8957 struct cpuacct *ca = cgroup_ca(cgrp);
8958 u64 totalcpuusage = 0;
8961 for_each_present_cpu(i)
8962 totalcpuusage += cpuacct_cpuusage_read(ca, i);
8964 return totalcpuusage;
8967 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8970 struct cpuacct *ca = cgroup_ca(cgrp);
8979 for_each_present_cpu(i)
8980 cpuacct_cpuusage_write(ca, i, 0);
8986 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
8989 struct cpuacct *ca = cgroup_ca(cgroup);
8993 for_each_present_cpu(i) {
8994 percpu = cpuacct_cpuusage_read(ca, i);
8995 seq_printf(m, "%llu ", (unsigned long long) percpu);
8997 seq_printf(m, "\n");
9001 static const char *cpuacct_stat_desc[] = {
9002 [CPUACCT_STAT_USER] = "user",
9003 [CPUACCT_STAT_SYSTEM] = "system",
9006 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
9007 struct cgroup_map_cb *cb)
9009 struct cpuacct *ca = cgroup_ca(cgrp);
9012 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
9013 s64 val = percpu_counter_read(&ca->cpustat[i]);
9014 val = cputime64_to_clock_t(val);
9015 cb->fill(cb, cpuacct_stat_desc[i], val);
9020 static struct cftype files[] = {
9023 .read_u64 = cpuusage_read,
9024 .write_u64 = cpuusage_write,
9027 .name = "usage_percpu",
9028 .read_seq_string = cpuacct_percpu_seq_read,
9032 .read_map = cpuacct_stats_show,
9036 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9038 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9042 * charge this task's execution time to its accounting group.
9044 * called with rq->lock held.
9046 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9051 if (unlikely(!cpuacct_subsys.active))
9054 cpu = task_cpu(tsk);
9060 for (; ca; ca = ca->parent) {
9061 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9062 *cpuusage += cputime;
9069 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9070 * in cputime_t units. As a result, cpuacct_update_stats calls
9071 * percpu_counter_add with values large enough to always overflow the
9072 * per cpu batch limit causing bad SMP scalability.
9074 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9075 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9076 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9079 #define CPUACCT_BATCH \
9080 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9082 #define CPUACCT_BATCH 0
9086 * Charge the system/user time to the task's accounting group.
9088 static void cpuacct_update_stats(struct task_struct *tsk,
9089 enum cpuacct_stat_index idx, cputime_t val)
9092 int batch = CPUACCT_BATCH;
9094 if (unlikely(!cpuacct_subsys.active))
9101 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9107 struct cgroup_subsys cpuacct_subsys = {
9109 .create = cpuacct_create,
9110 .destroy = cpuacct_destroy,
9111 .populate = cpuacct_populate,
9112 .subsys_id = cpuacct_subsys_id,
9114 #endif /* CONFIG_CGROUP_CPUACCT */
9118 int rcu_expedited_torture_stats(char *page)
9122 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
9124 void synchronize_sched_expedited(void)
9127 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9129 #else /* #ifndef CONFIG_SMP */
9131 static DEFINE_PER_CPU(struct migration_req, rcu_migration_req);
9132 static DEFINE_MUTEX(rcu_sched_expedited_mutex);
9134 #define RCU_EXPEDITED_STATE_POST -2
9135 #define RCU_EXPEDITED_STATE_IDLE -1
9137 static int rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
9139 int rcu_expedited_torture_stats(char *page)
9144 cnt += sprintf(&page[cnt], "state: %d /", rcu_expedited_state);
9145 for_each_online_cpu(cpu) {
9146 cnt += sprintf(&page[cnt], " %d:%d",
9147 cpu, per_cpu(rcu_migration_req, cpu).dest_cpu);
9149 cnt += sprintf(&page[cnt], "\n");
9152 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
9154 static long synchronize_sched_expedited_count;
9157 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
9158 * approach to force grace period to end quickly. This consumes
9159 * significant time on all CPUs, and is thus not recommended for
9160 * any sort of common-case code.
9162 * Note that it is illegal to call this function while holding any
9163 * lock that is acquired by a CPU-hotplug notifier. Failing to
9164 * observe this restriction will result in deadlock.
9166 void synchronize_sched_expedited(void)
9169 unsigned long flags;
9170 bool need_full_sync = 0;
9172 struct migration_req *req;
9176 smp_mb(); /* ensure prior mod happens before capturing snap. */
9177 snap = ACCESS_ONCE(synchronize_sched_expedited_count) + 1;
9179 while (!mutex_trylock(&rcu_sched_expedited_mutex)) {
9181 if (trycount++ < 10)
9182 udelay(trycount * num_online_cpus());
9184 synchronize_sched();
9187 if (ACCESS_ONCE(synchronize_sched_expedited_count) - snap > 0) {
9188 smp_mb(); /* ensure test happens before caller kfree */
9193 rcu_expedited_state = RCU_EXPEDITED_STATE_POST;
9194 for_each_online_cpu(cpu) {
9196 req = &per_cpu(rcu_migration_req, cpu);
9197 init_completion(&req->done);
9199 req->dest_cpu = RCU_MIGRATION_NEED_QS;
9200 raw_spin_lock_irqsave(&rq->lock, flags);
9201 list_add(&req->list, &rq->migration_queue);
9202 raw_spin_unlock_irqrestore(&rq->lock, flags);
9203 wake_up_process(rq->migration_thread);
9205 for_each_online_cpu(cpu) {
9206 rcu_expedited_state = cpu;
9207 req = &per_cpu(rcu_migration_req, cpu);
9209 wait_for_completion(&req->done);
9210 raw_spin_lock_irqsave(&rq->lock, flags);
9211 if (unlikely(req->dest_cpu == RCU_MIGRATION_MUST_SYNC))
9213 req->dest_cpu = RCU_MIGRATION_IDLE;
9214 raw_spin_unlock_irqrestore(&rq->lock, flags);
9216 rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
9217 synchronize_sched_expedited_count++;
9218 mutex_unlock(&rcu_sched_expedited_mutex);
9221 synchronize_sched();
9223 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9225 #endif /* #else #ifndef CONFIG_SMP */