4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/perf_event.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/kthread.h>
59 #include <linux/proc_fs.h>
60 #include <linux/seq_file.h>
61 #include <linux/sysctl.h>
62 #include <linux/syscalls.h>
63 #include <linux/times.h>
64 #include <linux/tsacct_kern.h>
65 #include <linux/kprobes.h>
66 #include <linux/delayacct.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
73 #include <linux/ftrace.h>
76 #include <asm/irq_regs.h>
78 #include "sched_cpupri.h"
80 #define CREATE_TRACE_POINTS
81 #include <trace/events/sched.h>
84 * Convert user-nice values [ -20 ... 0 ... 19 ]
85 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
88 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
89 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
90 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
93 * 'User priority' is the nice value converted to something we
94 * can work with better when scaling various scheduler parameters,
95 * it's a [ 0 ... 39 ] range.
97 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
98 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
99 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
102 * Helpers for converting nanosecond timing to jiffy resolution
104 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
106 #define NICE_0_LOAD SCHED_LOAD_SCALE
107 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
110 * These are the 'tuning knobs' of the scheduler:
112 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
113 * Timeslices get refilled after they expire.
115 #define DEF_TIMESLICE (100 * HZ / 1000)
118 * single value that denotes runtime == period, ie unlimited time.
120 #define RUNTIME_INF ((u64)~0ULL)
122 static inline int rt_policy(int policy)
124 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
129 static inline int task_has_rt_policy(struct task_struct *p)
131 return rt_policy(p->policy);
135 * This is the priority-queue data structure of the RT scheduling class:
137 struct rt_prio_array {
138 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
139 struct list_head queue[MAX_RT_PRIO];
142 struct rt_bandwidth {
143 /* nests inside the rq lock: */
144 raw_spinlock_t rt_runtime_lock;
147 struct hrtimer rt_period_timer;
150 static struct rt_bandwidth def_rt_bandwidth;
152 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
154 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
156 struct rt_bandwidth *rt_b =
157 container_of(timer, struct rt_bandwidth, rt_period_timer);
163 now = hrtimer_cb_get_time(timer);
164 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
169 idle = do_sched_rt_period_timer(rt_b, overrun);
172 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
176 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
178 rt_b->rt_period = ns_to_ktime(period);
179 rt_b->rt_runtime = runtime;
181 raw_spin_lock_init(&rt_b->rt_runtime_lock);
183 hrtimer_init(&rt_b->rt_period_timer,
184 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
185 rt_b->rt_period_timer.function = sched_rt_period_timer;
188 static inline int rt_bandwidth_enabled(void)
190 return sysctl_sched_rt_runtime >= 0;
193 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
197 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
200 if (hrtimer_active(&rt_b->rt_period_timer))
203 raw_spin_lock(&rt_b->rt_runtime_lock);
208 if (hrtimer_active(&rt_b->rt_period_timer))
211 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
212 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
214 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
215 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
216 delta = ktime_to_ns(ktime_sub(hard, soft));
217 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
218 HRTIMER_MODE_ABS_PINNED, 0);
220 raw_spin_unlock(&rt_b->rt_runtime_lock);
223 #ifdef CONFIG_RT_GROUP_SCHED
224 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
226 hrtimer_cancel(&rt_b->rt_period_timer);
231 * sched_domains_mutex serializes calls to arch_init_sched_domains,
232 * detach_destroy_domains and partition_sched_domains.
234 static DEFINE_MUTEX(sched_domains_mutex);
236 #ifdef CONFIG_CGROUP_SCHED
238 #include <linux/cgroup.h>
242 static LIST_HEAD(task_groups);
244 /* task group related information */
246 struct cgroup_subsys_state css;
248 #ifdef CONFIG_FAIR_GROUP_SCHED
249 /* schedulable entities of this group on each cpu */
250 struct sched_entity **se;
251 /* runqueue "owned" by this group on each cpu */
252 struct cfs_rq **cfs_rq;
253 unsigned long shares;
256 #ifdef CONFIG_RT_GROUP_SCHED
257 struct sched_rt_entity **rt_se;
258 struct rt_rq **rt_rq;
260 struct rt_bandwidth rt_bandwidth;
264 struct list_head list;
266 struct task_group *parent;
267 struct list_head siblings;
268 struct list_head children;
271 #define root_task_group init_task_group
273 /* task_group_lock serializes add/remove of task groups and also changes to
274 * a task group's cpu shares.
276 static DEFINE_SPINLOCK(task_group_lock);
278 #ifdef CONFIG_FAIR_GROUP_SCHED
281 static int root_task_group_empty(void)
283 return list_empty(&root_task_group.children);
287 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
290 * A weight of 0 or 1 can cause arithmetics problems.
291 * A weight of a cfs_rq is the sum of weights of which entities
292 * are queued on this cfs_rq, so a weight of a entity should not be
293 * too large, so as the shares value of a task group.
294 * (The default weight is 1024 - so there's no practical
295 * limitation from this.)
298 #define MAX_SHARES (1UL << 18)
300 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
303 /* Default task group.
304 * Every task in system belong to this group at bootup.
306 struct task_group init_task_group;
308 /* return group to which a task belongs */
309 static inline struct task_group *task_group(struct task_struct *p)
311 struct task_group *tg;
313 #ifdef CONFIG_CGROUP_SCHED
314 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
315 struct task_group, css);
317 tg = &init_task_group;
322 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
323 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
325 #ifdef CONFIG_FAIR_GROUP_SCHED
326 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
327 p->se.parent = task_group(p)->se[cpu];
330 #ifdef CONFIG_RT_GROUP_SCHED
331 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
332 p->rt.parent = task_group(p)->rt_se[cpu];
338 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
339 static inline struct task_group *task_group(struct task_struct *p)
344 #endif /* CONFIG_CGROUP_SCHED */
346 /* CFS-related fields in a runqueue */
348 struct load_weight load;
349 unsigned long nr_running;
354 struct rb_root tasks_timeline;
355 struct rb_node *rb_leftmost;
357 struct list_head tasks;
358 struct list_head *balance_iterator;
361 * 'curr' points to currently running entity on this cfs_rq.
362 * It is set to NULL otherwise (i.e when none are currently running).
364 struct sched_entity *curr, *next, *last;
366 unsigned int nr_spread_over;
368 #ifdef CONFIG_FAIR_GROUP_SCHED
369 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
372 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
373 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
374 * (like users, containers etc.)
376 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
377 * list is used during load balance.
379 struct list_head leaf_cfs_rq_list;
380 struct task_group *tg; /* group that "owns" this runqueue */
384 * the part of load.weight contributed by tasks
386 unsigned long task_weight;
389 * h_load = weight * f(tg)
391 * Where f(tg) is the recursive weight fraction assigned to
394 unsigned long h_load;
397 * this cpu's part of tg->shares
399 unsigned long shares;
402 * load.weight at the time we set shares
404 unsigned long rq_weight;
409 /* Real-Time classes' related field in a runqueue: */
411 struct rt_prio_array active;
412 unsigned long rt_nr_running;
413 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
415 int curr; /* highest queued rt task prio */
417 int next; /* next highest */
422 unsigned long rt_nr_migratory;
423 unsigned long rt_nr_total;
425 struct plist_head pushable_tasks;
430 /* Nests inside the rq lock: */
431 raw_spinlock_t rt_runtime_lock;
433 #ifdef CONFIG_RT_GROUP_SCHED
434 unsigned long rt_nr_boosted;
437 struct list_head leaf_rt_rq_list;
438 struct task_group *tg;
445 * We add the notion of a root-domain which will be used to define per-domain
446 * variables. Each exclusive cpuset essentially defines an island domain by
447 * fully partitioning the member cpus from any other cpuset. Whenever a new
448 * exclusive cpuset is created, we also create and attach a new root-domain
455 cpumask_var_t online;
458 * The "RT overload" flag: it gets set if a CPU has more than
459 * one runnable RT task.
461 cpumask_var_t rto_mask;
464 struct cpupri cpupri;
469 * By default the system creates a single root-domain with all cpus as
470 * members (mimicking the global state we have today).
472 static struct root_domain def_root_domain;
477 * This is the main, per-CPU runqueue data structure.
479 * Locking rule: those places that want to lock multiple runqueues
480 * (such as the load balancing or the thread migration code), lock
481 * acquire operations must be ordered by ascending &runqueue.
488 * nr_running and cpu_load should be in the same cacheline because
489 * remote CPUs use both these fields when doing load calculation.
491 unsigned long nr_running;
492 #define CPU_LOAD_IDX_MAX 5
493 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
495 unsigned char in_nohz_recently;
497 /* capture load from *all* tasks on this cpu: */
498 struct load_weight load;
499 unsigned long nr_load_updates;
505 #ifdef CONFIG_FAIR_GROUP_SCHED
506 /* list of leaf cfs_rq on this cpu: */
507 struct list_head leaf_cfs_rq_list;
509 #ifdef CONFIG_RT_GROUP_SCHED
510 struct list_head leaf_rt_rq_list;
514 * This is part of a global counter where only the total sum
515 * over all CPUs matters. A task can increase this counter on
516 * one CPU and if it got migrated afterwards it may decrease
517 * it on another CPU. Always updated under the runqueue lock:
519 unsigned long nr_uninterruptible;
521 struct task_struct *curr, *idle;
522 unsigned long next_balance;
523 struct mm_struct *prev_mm;
530 struct root_domain *rd;
531 struct sched_domain *sd;
533 unsigned char idle_at_tick;
534 /* For active balancing */
538 /* cpu of this runqueue: */
542 unsigned long avg_load_per_task;
544 struct task_struct *migration_thread;
545 struct list_head migration_queue;
553 /* calc_load related fields */
554 unsigned long calc_load_update;
555 long calc_load_active;
557 #ifdef CONFIG_SCHED_HRTICK
559 int hrtick_csd_pending;
560 struct call_single_data hrtick_csd;
562 struct hrtimer hrtick_timer;
565 #ifdef CONFIG_SCHEDSTATS
567 struct sched_info rq_sched_info;
568 unsigned long long rq_cpu_time;
569 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
571 /* sys_sched_yield() stats */
572 unsigned int yld_count;
574 /* schedule() stats */
575 unsigned int sched_switch;
576 unsigned int sched_count;
577 unsigned int sched_goidle;
579 /* try_to_wake_up() stats */
580 unsigned int ttwu_count;
581 unsigned int ttwu_local;
584 unsigned int bkl_count;
588 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
591 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
593 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
596 static inline int cpu_of(struct rq *rq)
605 #define rcu_dereference_check_sched_domain(p) \
606 rcu_dereference_check((p), \
607 rcu_read_lock_sched_held() || \
608 lockdep_is_held(&sched_domains_mutex))
611 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
612 * See detach_destroy_domains: synchronize_sched for details.
614 * The domain tree of any CPU may only be accessed from within
615 * preempt-disabled sections.
617 #define for_each_domain(cpu, __sd) \
618 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
620 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
621 #define this_rq() (&__get_cpu_var(runqueues))
622 #define task_rq(p) cpu_rq(task_cpu(p))
623 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
624 #define raw_rq() (&__raw_get_cpu_var(runqueues))
626 inline void update_rq_clock(struct rq *rq)
628 rq->clock = sched_clock_cpu(cpu_of(rq));
632 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
634 #ifdef CONFIG_SCHED_DEBUG
635 # define const_debug __read_mostly
637 # define const_debug static const
642 * @cpu: the processor in question.
644 * Returns true if the current cpu runqueue is locked.
645 * This interface allows printk to be called with the runqueue lock
646 * held and know whether or not it is OK to wake up the klogd.
648 int runqueue_is_locked(int cpu)
650 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
654 * Debugging: various feature bits
657 #define SCHED_FEAT(name, enabled) \
658 __SCHED_FEAT_##name ,
661 #include "sched_features.h"
666 #define SCHED_FEAT(name, enabled) \
667 (1UL << __SCHED_FEAT_##name) * enabled |
669 const_debug unsigned int sysctl_sched_features =
670 #include "sched_features.h"
675 #ifdef CONFIG_SCHED_DEBUG
676 #define SCHED_FEAT(name, enabled) \
679 static __read_mostly char *sched_feat_names[] = {
680 #include "sched_features.h"
686 static int sched_feat_show(struct seq_file *m, void *v)
690 for (i = 0; sched_feat_names[i]; i++) {
691 if (!(sysctl_sched_features & (1UL << i)))
693 seq_printf(m, "%s ", sched_feat_names[i]);
701 sched_feat_write(struct file *filp, const char __user *ubuf,
702 size_t cnt, loff_t *ppos)
712 if (copy_from_user(&buf, ubuf, cnt))
717 if (strncmp(buf, "NO_", 3) == 0) {
722 for (i = 0; sched_feat_names[i]; i++) {
723 int len = strlen(sched_feat_names[i]);
725 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
727 sysctl_sched_features &= ~(1UL << i);
729 sysctl_sched_features |= (1UL << i);
734 if (!sched_feat_names[i])
742 static int sched_feat_open(struct inode *inode, struct file *filp)
744 return single_open(filp, sched_feat_show, NULL);
747 static const struct file_operations sched_feat_fops = {
748 .open = sched_feat_open,
749 .write = sched_feat_write,
752 .release = single_release,
755 static __init int sched_init_debug(void)
757 debugfs_create_file("sched_features", 0644, NULL, NULL,
762 late_initcall(sched_init_debug);
766 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
769 * Number of tasks to iterate in a single balance run.
770 * Limited because this is done with IRQs disabled.
772 const_debug unsigned int sysctl_sched_nr_migrate = 32;
775 * ratelimit for updating the group shares.
778 unsigned int sysctl_sched_shares_ratelimit = 250000;
779 unsigned int normalized_sysctl_sched_shares_ratelimit = 250000;
782 * Inject some fuzzyness into changing the per-cpu group shares
783 * this avoids remote rq-locks at the expense of fairness.
786 unsigned int sysctl_sched_shares_thresh = 4;
789 * period over which we average the RT time consumption, measured
794 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
797 * period over which we measure -rt task cpu usage in us.
800 unsigned int sysctl_sched_rt_period = 1000000;
802 static __read_mostly int scheduler_running;
805 * part of the period that we allow rt tasks to run in us.
808 int sysctl_sched_rt_runtime = 950000;
810 static inline u64 global_rt_period(void)
812 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
815 static inline u64 global_rt_runtime(void)
817 if (sysctl_sched_rt_runtime < 0)
820 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
823 #ifndef prepare_arch_switch
824 # define prepare_arch_switch(next) do { } while (0)
826 #ifndef finish_arch_switch
827 # define finish_arch_switch(prev) do { } while (0)
830 static inline int task_current(struct rq *rq, struct task_struct *p)
832 return rq->curr == p;
835 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
836 static inline int task_running(struct rq *rq, struct task_struct *p)
838 return task_current(rq, p);
841 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
845 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
847 #ifdef CONFIG_DEBUG_SPINLOCK
848 /* this is a valid case when another task releases the spinlock */
849 rq->lock.owner = current;
852 * If we are tracking spinlock dependencies then we have to
853 * fix up the runqueue lock - which gets 'carried over' from
856 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
858 raw_spin_unlock_irq(&rq->lock);
861 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
862 static inline int task_running(struct rq *rq, struct task_struct *p)
867 return task_current(rq, p);
871 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
875 * We can optimise this out completely for !SMP, because the
876 * SMP rebalancing from interrupt is the only thing that cares
881 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
882 raw_spin_unlock_irq(&rq->lock);
884 raw_spin_unlock(&rq->lock);
888 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
892 * After ->oncpu is cleared, the task can be moved to a different CPU.
893 * We must ensure this doesn't happen until the switch is completely
899 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
903 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
906 * Check whether the task is waking, we use this to synchronize against
907 * ttwu() so that task_cpu() reports a stable number.
909 * We need to make an exception for PF_STARTING tasks because the fork
910 * path might require task_rq_lock() to work, eg. it can call
911 * set_cpus_allowed_ptr() from the cpuset clone_ns code.
913 static inline int task_is_waking(struct task_struct *p)
915 return unlikely((p->state == TASK_WAKING) && !(p->flags & PF_STARTING));
919 * __task_rq_lock - lock the runqueue a given task resides on.
920 * Must be called interrupts disabled.
922 static inline struct rq *__task_rq_lock(struct task_struct *p)
928 while (task_is_waking(p))
931 raw_spin_lock(&rq->lock);
932 if (likely(rq == task_rq(p) && !task_is_waking(p)))
934 raw_spin_unlock(&rq->lock);
939 * task_rq_lock - lock the runqueue a given task resides on and disable
940 * interrupts. Note the ordering: we can safely lookup the task_rq without
941 * explicitly disabling preemption.
943 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
949 while (task_is_waking(p))
951 local_irq_save(*flags);
953 raw_spin_lock(&rq->lock);
954 if (likely(rq == task_rq(p) && !task_is_waking(p)))
956 raw_spin_unlock_irqrestore(&rq->lock, *flags);
960 void task_rq_unlock_wait(struct task_struct *p)
962 struct rq *rq = task_rq(p);
964 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
965 raw_spin_unlock_wait(&rq->lock);
968 static void __task_rq_unlock(struct rq *rq)
971 raw_spin_unlock(&rq->lock);
974 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
977 raw_spin_unlock_irqrestore(&rq->lock, *flags);
981 * this_rq_lock - lock this runqueue and disable interrupts.
983 static struct rq *this_rq_lock(void)
990 raw_spin_lock(&rq->lock);
995 #ifdef CONFIG_SCHED_HRTICK
997 * Use HR-timers to deliver accurate preemption points.
999 * Its all a bit involved since we cannot program an hrt while holding the
1000 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1003 * When we get rescheduled we reprogram the hrtick_timer outside of the
1009 * - enabled by features
1010 * - hrtimer is actually high res
1012 static inline int hrtick_enabled(struct rq *rq)
1014 if (!sched_feat(HRTICK))
1016 if (!cpu_active(cpu_of(rq)))
1018 return hrtimer_is_hres_active(&rq->hrtick_timer);
1021 static void hrtick_clear(struct rq *rq)
1023 if (hrtimer_active(&rq->hrtick_timer))
1024 hrtimer_cancel(&rq->hrtick_timer);
1028 * High-resolution timer tick.
1029 * Runs from hardirq context with interrupts disabled.
1031 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1033 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1035 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1037 raw_spin_lock(&rq->lock);
1038 update_rq_clock(rq);
1039 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1040 raw_spin_unlock(&rq->lock);
1042 return HRTIMER_NORESTART;
1047 * called from hardirq (IPI) context
1049 static void __hrtick_start(void *arg)
1051 struct rq *rq = arg;
1053 raw_spin_lock(&rq->lock);
1054 hrtimer_restart(&rq->hrtick_timer);
1055 rq->hrtick_csd_pending = 0;
1056 raw_spin_unlock(&rq->lock);
1060 * Called to set the hrtick timer state.
1062 * called with rq->lock held and irqs disabled
1064 static void hrtick_start(struct rq *rq, u64 delay)
1066 struct hrtimer *timer = &rq->hrtick_timer;
1067 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1069 hrtimer_set_expires(timer, time);
1071 if (rq == this_rq()) {
1072 hrtimer_restart(timer);
1073 } else if (!rq->hrtick_csd_pending) {
1074 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1075 rq->hrtick_csd_pending = 1;
1080 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1082 int cpu = (int)(long)hcpu;
1085 case CPU_UP_CANCELED:
1086 case CPU_UP_CANCELED_FROZEN:
1087 case CPU_DOWN_PREPARE:
1088 case CPU_DOWN_PREPARE_FROZEN:
1090 case CPU_DEAD_FROZEN:
1091 hrtick_clear(cpu_rq(cpu));
1098 static __init void init_hrtick(void)
1100 hotcpu_notifier(hotplug_hrtick, 0);
1104 * Called to set the hrtick timer state.
1106 * called with rq->lock held and irqs disabled
1108 static void hrtick_start(struct rq *rq, u64 delay)
1110 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1111 HRTIMER_MODE_REL_PINNED, 0);
1114 static inline void init_hrtick(void)
1117 #endif /* CONFIG_SMP */
1119 static void init_rq_hrtick(struct rq *rq)
1122 rq->hrtick_csd_pending = 0;
1124 rq->hrtick_csd.flags = 0;
1125 rq->hrtick_csd.func = __hrtick_start;
1126 rq->hrtick_csd.info = rq;
1129 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1130 rq->hrtick_timer.function = hrtick;
1132 #else /* CONFIG_SCHED_HRTICK */
1133 static inline void hrtick_clear(struct rq *rq)
1137 static inline void init_rq_hrtick(struct rq *rq)
1141 static inline void init_hrtick(void)
1144 #endif /* CONFIG_SCHED_HRTICK */
1147 * resched_task - mark a task 'to be rescheduled now'.
1149 * On UP this means the setting of the need_resched flag, on SMP it
1150 * might also involve a cross-CPU call to trigger the scheduler on
1155 #ifndef tsk_is_polling
1156 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1159 static void resched_task(struct task_struct *p)
1163 assert_raw_spin_locked(&task_rq(p)->lock);
1165 if (test_tsk_need_resched(p))
1168 set_tsk_need_resched(p);
1171 if (cpu == smp_processor_id())
1174 /* NEED_RESCHED must be visible before we test polling */
1176 if (!tsk_is_polling(p))
1177 smp_send_reschedule(cpu);
1180 static void resched_cpu(int cpu)
1182 struct rq *rq = cpu_rq(cpu);
1183 unsigned long flags;
1185 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1187 resched_task(cpu_curr(cpu));
1188 raw_spin_unlock_irqrestore(&rq->lock, flags);
1193 * When add_timer_on() enqueues a timer into the timer wheel of an
1194 * idle CPU then this timer might expire before the next timer event
1195 * which is scheduled to wake up that CPU. In case of a completely
1196 * idle system the next event might even be infinite time into the
1197 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1198 * leaves the inner idle loop so the newly added timer is taken into
1199 * account when the CPU goes back to idle and evaluates the timer
1200 * wheel for the next timer event.
1202 void wake_up_idle_cpu(int cpu)
1204 struct rq *rq = cpu_rq(cpu);
1206 if (cpu == smp_processor_id())
1210 * This is safe, as this function is called with the timer
1211 * wheel base lock of (cpu) held. When the CPU is on the way
1212 * to idle and has not yet set rq->curr to idle then it will
1213 * be serialized on the timer wheel base lock and take the new
1214 * timer into account automatically.
1216 if (rq->curr != rq->idle)
1220 * We can set TIF_RESCHED on the idle task of the other CPU
1221 * lockless. The worst case is that the other CPU runs the
1222 * idle task through an additional NOOP schedule()
1224 set_tsk_need_resched(rq->idle);
1226 /* NEED_RESCHED must be visible before we test polling */
1228 if (!tsk_is_polling(rq->idle))
1229 smp_send_reschedule(cpu);
1231 #endif /* CONFIG_NO_HZ */
1233 static u64 sched_avg_period(void)
1235 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1238 static void sched_avg_update(struct rq *rq)
1240 s64 period = sched_avg_period();
1242 while ((s64)(rq->clock - rq->age_stamp) > period) {
1243 rq->age_stamp += period;
1248 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1250 rq->rt_avg += rt_delta;
1251 sched_avg_update(rq);
1254 #else /* !CONFIG_SMP */
1255 static void resched_task(struct task_struct *p)
1257 assert_raw_spin_locked(&task_rq(p)->lock);
1258 set_tsk_need_resched(p);
1261 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1264 #endif /* CONFIG_SMP */
1266 #if BITS_PER_LONG == 32
1267 # define WMULT_CONST (~0UL)
1269 # define WMULT_CONST (1UL << 32)
1272 #define WMULT_SHIFT 32
1275 * Shift right and round:
1277 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1280 * delta *= weight / lw
1282 static unsigned long
1283 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1284 struct load_weight *lw)
1288 if (!lw->inv_weight) {
1289 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1292 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1296 tmp = (u64)delta_exec * weight;
1298 * Check whether we'd overflow the 64-bit multiplication:
1300 if (unlikely(tmp > WMULT_CONST))
1301 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1304 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1306 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1309 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1315 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1322 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1323 * of tasks with abnormal "nice" values across CPUs the contribution that
1324 * each task makes to its run queue's load is weighted according to its
1325 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1326 * scaled version of the new time slice allocation that they receive on time
1330 #define WEIGHT_IDLEPRIO 3
1331 #define WMULT_IDLEPRIO 1431655765
1334 * Nice levels are multiplicative, with a gentle 10% change for every
1335 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1336 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1337 * that remained on nice 0.
1339 * The "10% effect" is relative and cumulative: from _any_ nice level,
1340 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1341 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1342 * If a task goes up by ~10% and another task goes down by ~10% then
1343 * the relative distance between them is ~25%.)
1345 static const int prio_to_weight[40] = {
1346 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1347 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1348 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1349 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1350 /* 0 */ 1024, 820, 655, 526, 423,
1351 /* 5 */ 335, 272, 215, 172, 137,
1352 /* 10 */ 110, 87, 70, 56, 45,
1353 /* 15 */ 36, 29, 23, 18, 15,
1357 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1359 * In cases where the weight does not change often, we can use the
1360 * precalculated inverse to speed up arithmetics by turning divisions
1361 * into multiplications:
1363 static const u32 prio_to_wmult[40] = {
1364 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1365 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1366 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1367 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1368 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1369 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1370 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1371 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1374 /* Time spent by the tasks of the cpu accounting group executing in ... */
1375 enum cpuacct_stat_index {
1376 CPUACCT_STAT_USER, /* ... user mode */
1377 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1379 CPUACCT_STAT_NSTATS,
1382 #ifdef CONFIG_CGROUP_CPUACCT
1383 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1384 static void cpuacct_update_stats(struct task_struct *tsk,
1385 enum cpuacct_stat_index idx, cputime_t val);
1387 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1388 static inline void cpuacct_update_stats(struct task_struct *tsk,
1389 enum cpuacct_stat_index idx, cputime_t val) {}
1392 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1394 update_load_add(&rq->load, load);
1397 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1399 update_load_sub(&rq->load, load);
1402 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1403 typedef int (*tg_visitor)(struct task_group *, void *);
1406 * Iterate the full tree, calling @down when first entering a node and @up when
1407 * leaving it for the final time.
1409 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1411 struct task_group *parent, *child;
1415 parent = &root_task_group;
1417 ret = (*down)(parent, data);
1420 list_for_each_entry_rcu(child, &parent->children, siblings) {
1427 ret = (*up)(parent, data);
1432 parent = parent->parent;
1441 static int tg_nop(struct task_group *tg, void *data)
1448 /* Used instead of source_load when we know the type == 0 */
1449 static unsigned long weighted_cpuload(const int cpu)
1451 return cpu_rq(cpu)->load.weight;
1455 * Return a low guess at the load of a migration-source cpu weighted
1456 * according to the scheduling class and "nice" value.
1458 * We want to under-estimate the load of migration sources, to
1459 * balance conservatively.
1461 static unsigned long source_load(int cpu, int type)
1463 struct rq *rq = cpu_rq(cpu);
1464 unsigned long total = weighted_cpuload(cpu);
1466 if (type == 0 || !sched_feat(LB_BIAS))
1469 return min(rq->cpu_load[type-1], total);
1473 * Return a high guess at the load of a migration-target cpu weighted
1474 * according to the scheduling class and "nice" value.
1476 static unsigned long target_load(int cpu, int type)
1478 struct rq *rq = cpu_rq(cpu);
1479 unsigned long total = weighted_cpuload(cpu);
1481 if (type == 0 || !sched_feat(LB_BIAS))
1484 return max(rq->cpu_load[type-1], total);
1487 static struct sched_group *group_of(int cpu)
1489 struct sched_domain *sd = rcu_dereference_sched(cpu_rq(cpu)->sd);
1497 static unsigned long power_of(int cpu)
1499 struct sched_group *group = group_of(cpu);
1502 return SCHED_LOAD_SCALE;
1504 return group->cpu_power;
1507 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1509 static unsigned long cpu_avg_load_per_task(int cpu)
1511 struct rq *rq = cpu_rq(cpu);
1512 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1515 rq->avg_load_per_task = rq->load.weight / nr_running;
1517 rq->avg_load_per_task = 0;
1519 return rq->avg_load_per_task;
1522 #ifdef CONFIG_FAIR_GROUP_SCHED
1524 static __read_mostly unsigned long __percpu *update_shares_data;
1526 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1529 * Calculate and set the cpu's group shares.
1531 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1532 unsigned long sd_shares,
1533 unsigned long sd_rq_weight,
1534 unsigned long *usd_rq_weight)
1536 unsigned long shares, rq_weight;
1539 rq_weight = usd_rq_weight[cpu];
1542 rq_weight = NICE_0_LOAD;
1546 * \Sum_j shares_j * rq_weight_i
1547 * shares_i = -----------------------------
1548 * \Sum_j rq_weight_j
1550 shares = (sd_shares * rq_weight) / sd_rq_weight;
1551 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1553 if (abs(shares - tg->se[cpu]->load.weight) >
1554 sysctl_sched_shares_thresh) {
1555 struct rq *rq = cpu_rq(cpu);
1556 unsigned long flags;
1558 raw_spin_lock_irqsave(&rq->lock, flags);
1559 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1560 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1561 __set_se_shares(tg->se[cpu], shares);
1562 raw_spin_unlock_irqrestore(&rq->lock, flags);
1567 * Re-compute the task group their per cpu shares over the given domain.
1568 * This needs to be done in a bottom-up fashion because the rq weight of a
1569 * parent group depends on the shares of its child groups.
1571 static int tg_shares_up(struct task_group *tg, void *data)
1573 unsigned long weight, rq_weight = 0, sum_weight = 0, shares = 0;
1574 unsigned long *usd_rq_weight;
1575 struct sched_domain *sd = data;
1576 unsigned long flags;
1582 local_irq_save(flags);
1583 usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id());
1585 for_each_cpu(i, sched_domain_span(sd)) {
1586 weight = tg->cfs_rq[i]->load.weight;
1587 usd_rq_weight[i] = weight;
1589 rq_weight += weight;
1591 * If there are currently no tasks on the cpu pretend there
1592 * is one of average load so that when a new task gets to
1593 * run here it will not get delayed by group starvation.
1596 weight = NICE_0_LOAD;
1598 sum_weight += weight;
1599 shares += tg->cfs_rq[i]->shares;
1603 rq_weight = sum_weight;
1605 if ((!shares && rq_weight) || shares > tg->shares)
1606 shares = tg->shares;
1608 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1609 shares = tg->shares;
1611 for_each_cpu(i, sched_domain_span(sd))
1612 update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight);
1614 local_irq_restore(flags);
1620 * Compute the cpu's hierarchical load factor for each task group.
1621 * This needs to be done in a top-down fashion because the load of a child
1622 * group is a fraction of its parents load.
1624 static int tg_load_down(struct task_group *tg, void *data)
1627 long cpu = (long)data;
1630 load = cpu_rq(cpu)->load.weight;
1632 load = tg->parent->cfs_rq[cpu]->h_load;
1633 load *= tg->cfs_rq[cpu]->shares;
1634 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1637 tg->cfs_rq[cpu]->h_load = load;
1642 static void update_shares(struct sched_domain *sd)
1647 if (root_task_group_empty())
1650 now = cpu_clock(raw_smp_processor_id());
1651 elapsed = now - sd->last_update;
1653 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1654 sd->last_update = now;
1655 walk_tg_tree(tg_nop, tg_shares_up, sd);
1659 static void update_h_load(long cpu)
1661 if (root_task_group_empty())
1664 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1669 static inline void update_shares(struct sched_domain *sd)
1675 #ifdef CONFIG_PREEMPT
1677 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1680 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1681 * way at the expense of forcing extra atomic operations in all
1682 * invocations. This assures that the double_lock is acquired using the
1683 * same underlying policy as the spinlock_t on this architecture, which
1684 * reduces latency compared to the unfair variant below. However, it
1685 * also adds more overhead and therefore may reduce throughput.
1687 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1688 __releases(this_rq->lock)
1689 __acquires(busiest->lock)
1690 __acquires(this_rq->lock)
1692 raw_spin_unlock(&this_rq->lock);
1693 double_rq_lock(this_rq, busiest);
1700 * Unfair double_lock_balance: Optimizes throughput at the expense of
1701 * latency by eliminating extra atomic operations when the locks are
1702 * already in proper order on entry. This favors lower cpu-ids and will
1703 * grant the double lock to lower cpus over higher ids under contention,
1704 * regardless of entry order into the function.
1706 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1707 __releases(this_rq->lock)
1708 __acquires(busiest->lock)
1709 __acquires(this_rq->lock)
1713 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1714 if (busiest < this_rq) {
1715 raw_spin_unlock(&this_rq->lock);
1716 raw_spin_lock(&busiest->lock);
1717 raw_spin_lock_nested(&this_rq->lock,
1718 SINGLE_DEPTH_NESTING);
1721 raw_spin_lock_nested(&busiest->lock,
1722 SINGLE_DEPTH_NESTING);
1727 #endif /* CONFIG_PREEMPT */
1730 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1732 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1734 if (unlikely(!irqs_disabled())) {
1735 /* printk() doesn't work good under rq->lock */
1736 raw_spin_unlock(&this_rq->lock);
1740 return _double_lock_balance(this_rq, busiest);
1743 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1744 __releases(busiest->lock)
1746 raw_spin_unlock(&busiest->lock);
1747 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1751 * double_rq_lock - safely lock two runqueues
1753 * Note this does not disable interrupts like task_rq_lock,
1754 * you need to do so manually before calling.
1756 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1757 __acquires(rq1->lock)
1758 __acquires(rq2->lock)
1760 BUG_ON(!irqs_disabled());
1762 raw_spin_lock(&rq1->lock);
1763 __acquire(rq2->lock); /* Fake it out ;) */
1766 raw_spin_lock(&rq1->lock);
1767 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1769 raw_spin_lock(&rq2->lock);
1770 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1773 update_rq_clock(rq1);
1774 update_rq_clock(rq2);
1778 * double_rq_unlock - safely unlock two runqueues
1780 * Note this does not restore interrupts like task_rq_unlock,
1781 * you need to do so manually after calling.
1783 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1784 __releases(rq1->lock)
1785 __releases(rq2->lock)
1787 raw_spin_unlock(&rq1->lock);
1789 raw_spin_unlock(&rq2->lock);
1791 __release(rq2->lock);
1796 #ifdef CONFIG_FAIR_GROUP_SCHED
1797 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1800 cfs_rq->shares = shares;
1805 static void calc_load_account_active(struct rq *this_rq);
1806 static void update_sysctl(void);
1807 static int get_update_sysctl_factor(void);
1809 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1811 set_task_rq(p, cpu);
1814 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1815 * successfuly executed on another CPU. We must ensure that updates of
1816 * per-task data have been completed by this moment.
1819 task_thread_info(p)->cpu = cpu;
1823 static const struct sched_class rt_sched_class;
1825 #define sched_class_highest (&rt_sched_class)
1826 #define for_each_class(class) \
1827 for (class = sched_class_highest; class; class = class->next)
1829 #include "sched_stats.h"
1831 static void inc_nr_running(struct rq *rq)
1836 static void dec_nr_running(struct rq *rq)
1841 static void set_load_weight(struct task_struct *p)
1843 if (task_has_rt_policy(p)) {
1844 p->se.load.weight = prio_to_weight[0] * 2;
1845 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1850 * SCHED_IDLE tasks get minimal weight:
1852 if (p->policy == SCHED_IDLE) {
1853 p->se.load.weight = WEIGHT_IDLEPRIO;
1854 p->se.load.inv_weight = WMULT_IDLEPRIO;
1858 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1859 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1862 static void update_avg(u64 *avg, u64 sample)
1864 s64 diff = sample - *avg;
1869 enqueue_task(struct rq *rq, struct task_struct *p, int wakeup, bool head)
1872 p->se.start_runtime = p->se.sum_exec_runtime;
1874 sched_info_queued(p);
1875 p->sched_class->enqueue_task(rq, p, wakeup, head);
1879 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1882 if (p->se.last_wakeup) {
1883 update_avg(&p->se.avg_overlap,
1884 p->se.sum_exec_runtime - p->se.last_wakeup);
1885 p->se.last_wakeup = 0;
1887 update_avg(&p->se.avg_wakeup,
1888 sysctl_sched_wakeup_granularity);
1892 sched_info_dequeued(p);
1893 p->sched_class->dequeue_task(rq, p, sleep);
1898 * activate_task - move a task to the runqueue.
1900 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1902 if (task_contributes_to_load(p))
1903 rq->nr_uninterruptible--;
1905 enqueue_task(rq, p, wakeup, false);
1910 * deactivate_task - remove a task from the runqueue.
1912 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1914 if (task_contributes_to_load(p))
1915 rq->nr_uninterruptible++;
1917 dequeue_task(rq, p, sleep);
1921 #include "sched_idletask.c"
1922 #include "sched_fair.c"
1923 #include "sched_rt.c"
1924 #ifdef CONFIG_SCHED_DEBUG
1925 # include "sched_debug.c"
1929 * __normal_prio - return the priority that is based on the static prio
1931 static inline int __normal_prio(struct task_struct *p)
1933 return p->static_prio;
1937 * Calculate the expected normal priority: i.e. priority
1938 * without taking RT-inheritance into account. Might be
1939 * boosted by interactivity modifiers. Changes upon fork,
1940 * setprio syscalls, and whenever the interactivity
1941 * estimator recalculates.
1943 static inline int normal_prio(struct task_struct *p)
1947 if (task_has_rt_policy(p))
1948 prio = MAX_RT_PRIO-1 - p->rt_priority;
1950 prio = __normal_prio(p);
1955 * Calculate the current priority, i.e. the priority
1956 * taken into account by the scheduler. This value might
1957 * be boosted by RT tasks, or might be boosted by
1958 * interactivity modifiers. Will be RT if the task got
1959 * RT-boosted. If not then it returns p->normal_prio.
1961 static int effective_prio(struct task_struct *p)
1963 p->normal_prio = normal_prio(p);
1965 * If we are RT tasks or we were boosted to RT priority,
1966 * keep the priority unchanged. Otherwise, update priority
1967 * to the normal priority:
1969 if (!rt_prio(p->prio))
1970 return p->normal_prio;
1975 * task_curr - is this task currently executing on a CPU?
1976 * @p: the task in question.
1978 inline int task_curr(const struct task_struct *p)
1980 return cpu_curr(task_cpu(p)) == p;
1983 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1984 const struct sched_class *prev_class,
1985 int oldprio, int running)
1987 if (prev_class != p->sched_class) {
1988 if (prev_class->switched_from)
1989 prev_class->switched_from(rq, p, running);
1990 p->sched_class->switched_to(rq, p, running);
1992 p->sched_class->prio_changed(rq, p, oldprio, running);
1997 * Is this task likely cache-hot:
2000 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2004 if (p->sched_class != &fair_sched_class)
2008 * Buddy candidates are cache hot:
2010 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2011 (&p->se == cfs_rq_of(&p->se)->next ||
2012 &p->se == cfs_rq_of(&p->se)->last))
2015 if (sysctl_sched_migration_cost == -1)
2017 if (sysctl_sched_migration_cost == 0)
2020 delta = now - p->se.exec_start;
2022 return delta < (s64)sysctl_sched_migration_cost;
2025 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2027 #ifdef CONFIG_SCHED_DEBUG
2029 * We should never call set_task_cpu() on a blocked task,
2030 * ttwu() will sort out the placement.
2032 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2033 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2036 trace_sched_migrate_task(p, new_cpu);
2038 if (task_cpu(p) != new_cpu) {
2039 p->se.nr_migrations++;
2040 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2043 __set_task_cpu(p, new_cpu);
2046 struct migration_req {
2047 struct list_head list;
2049 struct task_struct *task;
2052 struct completion done;
2056 * The task's runqueue lock must be held.
2057 * Returns true if you have to wait for migration thread.
2060 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2062 struct rq *rq = task_rq(p);
2065 * If the task is not on a runqueue (and not running), then
2066 * the next wake-up will properly place the task.
2068 if (!p->se.on_rq && !task_running(rq, p))
2071 init_completion(&req->done);
2073 req->dest_cpu = dest_cpu;
2074 list_add(&req->list, &rq->migration_queue);
2080 * wait_task_inactive - wait for a thread to unschedule.
2082 * If @match_state is nonzero, it's the @p->state value just checked and
2083 * not expected to change. If it changes, i.e. @p might have woken up,
2084 * then return zero. When we succeed in waiting for @p to be off its CPU,
2085 * we return a positive number (its total switch count). If a second call
2086 * a short while later returns the same number, the caller can be sure that
2087 * @p has remained unscheduled the whole time.
2089 * The caller must ensure that the task *will* unschedule sometime soon,
2090 * else this function might spin for a *long* time. This function can't
2091 * be called with interrupts off, or it may introduce deadlock with
2092 * smp_call_function() if an IPI is sent by the same process we are
2093 * waiting to become inactive.
2095 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2097 unsigned long flags;
2104 * We do the initial early heuristics without holding
2105 * any task-queue locks at all. We'll only try to get
2106 * the runqueue lock when things look like they will
2112 * If the task is actively running on another CPU
2113 * still, just relax and busy-wait without holding
2116 * NOTE! Since we don't hold any locks, it's not
2117 * even sure that "rq" stays as the right runqueue!
2118 * But we don't care, since "task_running()" will
2119 * return false if the runqueue has changed and p
2120 * is actually now running somewhere else!
2122 while (task_running(rq, p)) {
2123 if (match_state && unlikely(p->state != match_state))
2129 * Ok, time to look more closely! We need the rq
2130 * lock now, to be *sure*. If we're wrong, we'll
2131 * just go back and repeat.
2133 rq = task_rq_lock(p, &flags);
2134 trace_sched_wait_task(rq, p);
2135 running = task_running(rq, p);
2136 on_rq = p->se.on_rq;
2138 if (!match_state || p->state == match_state)
2139 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2140 task_rq_unlock(rq, &flags);
2143 * If it changed from the expected state, bail out now.
2145 if (unlikely(!ncsw))
2149 * Was it really running after all now that we
2150 * checked with the proper locks actually held?
2152 * Oops. Go back and try again..
2154 if (unlikely(running)) {
2160 * It's not enough that it's not actively running,
2161 * it must be off the runqueue _entirely_, and not
2164 * So if it was still runnable (but just not actively
2165 * running right now), it's preempted, and we should
2166 * yield - it could be a while.
2168 if (unlikely(on_rq)) {
2169 schedule_timeout_uninterruptible(1);
2174 * Ahh, all good. It wasn't running, and it wasn't
2175 * runnable, which means that it will never become
2176 * running in the future either. We're all done!
2185 * kick_process - kick a running thread to enter/exit the kernel
2186 * @p: the to-be-kicked thread
2188 * Cause a process which is running on another CPU to enter
2189 * kernel-mode, without any delay. (to get signals handled.)
2191 * NOTE: this function doesnt have to take the runqueue lock,
2192 * because all it wants to ensure is that the remote task enters
2193 * the kernel. If the IPI races and the task has been migrated
2194 * to another CPU then no harm is done and the purpose has been
2197 void kick_process(struct task_struct *p)
2203 if ((cpu != smp_processor_id()) && task_curr(p))
2204 smp_send_reschedule(cpu);
2207 EXPORT_SYMBOL_GPL(kick_process);
2208 #endif /* CONFIG_SMP */
2211 * task_oncpu_function_call - call a function on the cpu on which a task runs
2212 * @p: the task to evaluate
2213 * @func: the function to be called
2214 * @info: the function call argument
2216 * Calls the function @func when the task is currently running. This might
2217 * be on the current CPU, which just calls the function directly
2219 void task_oncpu_function_call(struct task_struct *p,
2220 void (*func) (void *info), void *info)
2227 smp_call_function_single(cpu, func, info, 1);
2232 static int select_fallback_rq(int cpu, struct task_struct *p)
2235 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2237 /* Look for allowed, online CPU in same node. */
2238 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2239 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2242 /* Any allowed, online CPU? */
2243 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2244 if (dest_cpu < nr_cpu_ids)
2247 /* No more Mr. Nice Guy. */
2248 if (dest_cpu >= nr_cpu_ids) {
2250 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
2252 dest_cpu = cpumask_any_and(cpu_active_mask, &p->cpus_allowed);
2255 * Don't tell them about moving exiting tasks or
2256 * kernel threads (both mm NULL), since they never
2259 if (p->mm && printk_ratelimit()) {
2260 printk(KERN_INFO "process %d (%s) no "
2261 "longer affine to cpu%d\n",
2262 task_pid_nr(p), p->comm, cpu);
2270 * Gets called from 3 sites (exec, fork, wakeup), since it is called without
2271 * holding rq->lock we need to ensure ->cpus_allowed is stable, this is done
2274 * exec: is unstable, retry loop
2275 * fork & wake-up: serialize ->cpus_allowed against TASK_WAKING
2278 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2280 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2283 * In order not to call set_task_cpu() on a blocking task we need
2284 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2287 * Since this is common to all placement strategies, this lives here.
2289 * [ this allows ->select_task() to simply return task_cpu(p) and
2290 * not worry about this generic constraint ]
2292 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2294 cpu = select_fallback_rq(task_cpu(p), p);
2301 * try_to_wake_up - wake up a thread
2302 * @p: the to-be-woken-up thread
2303 * @state: the mask of task states that can be woken
2304 * @sync: do a synchronous wakeup?
2306 * Put it on the run-queue if it's not already there. The "current"
2307 * thread is always on the run-queue (except when the actual
2308 * re-schedule is in progress), and as such you're allowed to do
2309 * the simpler "current->state = TASK_RUNNING" to mark yourself
2310 * runnable without the overhead of this.
2312 * returns failure only if the task is already active.
2314 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2317 int cpu, orig_cpu, this_cpu, success = 0;
2318 unsigned long flags;
2321 if (!sched_feat(SYNC_WAKEUPS))
2322 wake_flags &= ~WF_SYNC;
2324 this_cpu = get_cpu();
2327 rq = task_rq_lock(p, &flags);
2328 update_rq_clock(rq);
2329 if (!(p->state & state))
2339 if (unlikely(task_running(rq, p)))
2343 * In order to handle concurrent wakeups and release the rq->lock
2344 * we put the task in TASK_WAKING state.
2346 * First fix up the nr_uninterruptible count:
2348 if (task_contributes_to_load(p))
2349 rq->nr_uninterruptible--;
2350 p->state = TASK_WAKING;
2352 if (p->sched_class->task_waking)
2353 p->sched_class->task_waking(rq, p);
2355 __task_rq_unlock(rq);
2357 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2358 if (cpu != orig_cpu) {
2360 * Since we migrate the task without holding any rq->lock,
2361 * we need to be careful with task_rq_lock(), since that
2362 * might end up locking an invalid rq.
2364 set_task_cpu(p, cpu);
2368 raw_spin_lock(&rq->lock);
2369 update_rq_clock(rq);
2372 * We migrated the task without holding either rq->lock, however
2373 * since the task is not on the task list itself, nobody else
2374 * will try and migrate the task, hence the rq should match the
2375 * cpu we just moved it to.
2377 WARN_ON(task_cpu(p) != cpu);
2378 WARN_ON(p->state != TASK_WAKING);
2380 #ifdef CONFIG_SCHEDSTATS
2381 schedstat_inc(rq, ttwu_count);
2382 if (cpu == this_cpu)
2383 schedstat_inc(rq, ttwu_local);
2385 struct sched_domain *sd;
2386 for_each_domain(this_cpu, sd) {
2387 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2388 schedstat_inc(sd, ttwu_wake_remote);
2393 #endif /* CONFIG_SCHEDSTATS */
2396 #endif /* CONFIG_SMP */
2397 schedstat_inc(p, se.nr_wakeups);
2398 if (wake_flags & WF_SYNC)
2399 schedstat_inc(p, se.nr_wakeups_sync);
2400 if (orig_cpu != cpu)
2401 schedstat_inc(p, se.nr_wakeups_migrate);
2402 if (cpu == this_cpu)
2403 schedstat_inc(p, se.nr_wakeups_local);
2405 schedstat_inc(p, se.nr_wakeups_remote);
2406 activate_task(rq, p, 1);
2410 * Only attribute actual wakeups done by this task.
2412 if (!in_interrupt()) {
2413 struct sched_entity *se = ¤t->se;
2414 u64 sample = se->sum_exec_runtime;
2416 if (se->last_wakeup)
2417 sample -= se->last_wakeup;
2419 sample -= se->start_runtime;
2420 update_avg(&se->avg_wakeup, sample);
2422 se->last_wakeup = se->sum_exec_runtime;
2426 trace_sched_wakeup(rq, p, success);
2427 check_preempt_curr(rq, p, wake_flags);
2429 p->state = TASK_RUNNING;
2431 if (p->sched_class->task_woken)
2432 p->sched_class->task_woken(rq, p);
2434 if (unlikely(rq->idle_stamp)) {
2435 u64 delta = rq->clock - rq->idle_stamp;
2436 u64 max = 2*sysctl_sched_migration_cost;
2441 update_avg(&rq->avg_idle, delta);
2446 task_rq_unlock(rq, &flags);
2453 * wake_up_process - Wake up a specific process
2454 * @p: The process to be woken up.
2456 * Attempt to wake up the nominated process and move it to the set of runnable
2457 * processes. Returns 1 if the process was woken up, 0 if it was already
2460 * It may be assumed that this function implies a write memory barrier before
2461 * changing the task state if and only if any tasks are woken up.
2463 int wake_up_process(struct task_struct *p)
2465 return try_to_wake_up(p, TASK_ALL, 0);
2467 EXPORT_SYMBOL(wake_up_process);
2469 int wake_up_state(struct task_struct *p, unsigned int state)
2471 return try_to_wake_up(p, state, 0);
2475 * Perform scheduler related setup for a newly forked process p.
2476 * p is forked by current.
2478 * __sched_fork() is basic setup used by init_idle() too:
2480 static void __sched_fork(struct task_struct *p)
2482 p->se.exec_start = 0;
2483 p->se.sum_exec_runtime = 0;
2484 p->se.prev_sum_exec_runtime = 0;
2485 p->se.nr_migrations = 0;
2486 p->se.last_wakeup = 0;
2487 p->se.avg_overlap = 0;
2488 p->se.start_runtime = 0;
2489 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2491 #ifdef CONFIG_SCHEDSTATS
2492 p->se.wait_start = 0;
2494 p->se.wait_count = 0;
2497 p->se.sleep_start = 0;
2498 p->se.sleep_max = 0;
2499 p->se.sum_sleep_runtime = 0;
2501 p->se.block_start = 0;
2502 p->se.block_max = 0;
2504 p->se.slice_max = 0;
2506 p->se.nr_migrations_cold = 0;
2507 p->se.nr_failed_migrations_affine = 0;
2508 p->se.nr_failed_migrations_running = 0;
2509 p->se.nr_failed_migrations_hot = 0;
2510 p->se.nr_forced_migrations = 0;
2512 p->se.nr_wakeups = 0;
2513 p->se.nr_wakeups_sync = 0;
2514 p->se.nr_wakeups_migrate = 0;
2515 p->se.nr_wakeups_local = 0;
2516 p->se.nr_wakeups_remote = 0;
2517 p->se.nr_wakeups_affine = 0;
2518 p->se.nr_wakeups_affine_attempts = 0;
2519 p->se.nr_wakeups_passive = 0;
2520 p->se.nr_wakeups_idle = 0;
2524 INIT_LIST_HEAD(&p->rt.run_list);
2526 INIT_LIST_HEAD(&p->se.group_node);
2528 #ifdef CONFIG_PREEMPT_NOTIFIERS
2529 INIT_HLIST_HEAD(&p->preempt_notifiers);
2534 * fork()/clone()-time setup:
2536 void sched_fork(struct task_struct *p, int clone_flags)
2538 int cpu = get_cpu();
2542 * We mark the process as waking here. This guarantees that
2543 * nobody will actually run it, and a signal or other external
2544 * event cannot wake it up and insert it on the runqueue either.
2546 p->state = TASK_WAKING;
2549 * Revert to default priority/policy on fork if requested.
2551 if (unlikely(p->sched_reset_on_fork)) {
2552 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2553 p->policy = SCHED_NORMAL;
2554 p->normal_prio = p->static_prio;
2557 if (PRIO_TO_NICE(p->static_prio) < 0) {
2558 p->static_prio = NICE_TO_PRIO(0);
2559 p->normal_prio = p->static_prio;
2564 * We don't need the reset flag anymore after the fork. It has
2565 * fulfilled its duty:
2567 p->sched_reset_on_fork = 0;
2571 * Make sure we do not leak PI boosting priority to the child.
2573 p->prio = current->normal_prio;
2575 if (!rt_prio(p->prio))
2576 p->sched_class = &fair_sched_class;
2578 if (p->sched_class->task_fork)
2579 p->sched_class->task_fork(p);
2581 set_task_cpu(p, cpu);
2583 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2584 if (likely(sched_info_on()))
2585 memset(&p->sched_info, 0, sizeof(p->sched_info));
2587 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2590 #ifdef CONFIG_PREEMPT
2591 /* Want to start with kernel preemption disabled. */
2592 task_thread_info(p)->preempt_count = 1;
2594 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2600 * wake_up_new_task - wake up a newly created task for the first time.
2602 * This function will do some initial scheduler statistics housekeeping
2603 * that must be done for every newly created context, then puts the task
2604 * on the runqueue and wakes it.
2606 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2608 unsigned long flags;
2610 int cpu = get_cpu();
2614 * Fork balancing, do it here and not earlier because:
2615 * - cpus_allowed can change in the fork path
2616 * - any previously selected cpu might disappear through hotplug
2618 * We still have TASK_WAKING but PF_STARTING is gone now, meaning
2619 * ->cpus_allowed is stable, we have preemption disabled, meaning
2620 * cpu_online_mask is stable.
2622 cpu = select_task_rq(p, SD_BALANCE_FORK, 0);
2623 set_task_cpu(p, cpu);
2627 * Since the task is not on the rq and we still have TASK_WAKING set
2628 * nobody else will migrate this task.
2631 raw_spin_lock_irqsave(&rq->lock, flags);
2633 BUG_ON(p->state != TASK_WAKING);
2634 p->state = TASK_RUNNING;
2635 update_rq_clock(rq);
2636 activate_task(rq, p, 0);
2637 trace_sched_wakeup_new(rq, p, 1);
2638 check_preempt_curr(rq, p, WF_FORK);
2640 if (p->sched_class->task_woken)
2641 p->sched_class->task_woken(rq, p);
2643 task_rq_unlock(rq, &flags);
2647 #ifdef CONFIG_PREEMPT_NOTIFIERS
2650 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2651 * @notifier: notifier struct to register
2653 void preempt_notifier_register(struct preempt_notifier *notifier)
2655 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2657 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2660 * preempt_notifier_unregister - no longer interested in preemption notifications
2661 * @notifier: notifier struct to unregister
2663 * This is safe to call from within a preemption notifier.
2665 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2667 hlist_del(¬ifier->link);
2669 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2671 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2673 struct preempt_notifier *notifier;
2674 struct hlist_node *node;
2676 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2677 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2681 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2682 struct task_struct *next)
2684 struct preempt_notifier *notifier;
2685 struct hlist_node *node;
2687 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2688 notifier->ops->sched_out(notifier, next);
2691 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2693 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2698 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2699 struct task_struct *next)
2703 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2706 * prepare_task_switch - prepare to switch tasks
2707 * @rq: the runqueue preparing to switch
2708 * @prev: the current task that is being switched out
2709 * @next: the task we are going to switch to.
2711 * This is called with the rq lock held and interrupts off. It must
2712 * be paired with a subsequent finish_task_switch after the context
2715 * prepare_task_switch sets up locking and calls architecture specific
2719 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2720 struct task_struct *next)
2722 fire_sched_out_preempt_notifiers(prev, next);
2723 prepare_lock_switch(rq, next);
2724 prepare_arch_switch(next);
2728 * finish_task_switch - clean up after a task-switch
2729 * @rq: runqueue associated with task-switch
2730 * @prev: the thread we just switched away from.
2732 * finish_task_switch must be called after the context switch, paired
2733 * with a prepare_task_switch call before the context switch.
2734 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2735 * and do any other architecture-specific cleanup actions.
2737 * Note that we may have delayed dropping an mm in context_switch(). If
2738 * so, we finish that here outside of the runqueue lock. (Doing it
2739 * with the lock held can cause deadlocks; see schedule() for
2742 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2743 __releases(rq->lock)
2745 struct mm_struct *mm = rq->prev_mm;
2751 * A task struct has one reference for the use as "current".
2752 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2753 * schedule one last time. The schedule call will never return, and
2754 * the scheduled task must drop that reference.
2755 * The test for TASK_DEAD must occur while the runqueue locks are
2756 * still held, otherwise prev could be scheduled on another cpu, die
2757 * there before we look at prev->state, and then the reference would
2759 * Manfred Spraul <manfred@colorfullife.com>
2761 prev_state = prev->state;
2762 finish_arch_switch(prev);
2763 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2764 local_irq_disable();
2765 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2766 perf_event_task_sched_in(current);
2767 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2769 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2770 finish_lock_switch(rq, prev);
2772 fire_sched_in_preempt_notifiers(current);
2775 if (unlikely(prev_state == TASK_DEAD)) {
2777 * Remove function-return probe instances associated with this
2778 * task and put them back on the free list.
2780 kprobe_flush_task(prev);
2781 put_task_struct(prev);
2787 /* assumes rq->lock is held */
2788 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2790 if (prev->sched_class->pre_schedule)
2791 prev->sched_class->pre_schedule(rq, prev);
2794 /* rq->lock is NOT held, but preemption is disabled */
2795 static inline void post_schedule(struct rq *rq)
2797 if (rq->post_schedule) {
2798 unsigned long flags;
2800 raw_spin_lock_irqsave(&rq->lock, flags);
2801 if (rq->curr->sched_class->post_schedule)
2802 rq->curr->sched_class->post_schedule(rq);
2803 raw_spin_unlock_irqrestore(&rq->lock, flags);
2805 rq->post_schedule = 0;
2811 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2815 static inline void post_schedule(struct rq *rq)
2822 * schedule_tail - first thing a freshly forked thread must call.
2823 * @prev: the thread we just switched away from.
2825 asmlinkage void schedule_tail(struct task_struct *prev)
2826 __releases(rq->lock)
2828 struct rq *rq = this_rq();
2830 finish_task_switch(rq, prev);
2833 * FIXME: do we need to worry about rq being invalidated by the
2838 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2839 /* In this case, finish_task_switch does not reenable preemption */
2842 if (current->set_child_tid)
2843 put_user(task_pid_vnr(current), current->set_child_tid);
2847 * context_switch - switch to the new MM and the new
2848 * thread's register state.
2851 context_switch(struct rq *rq, struct task_struct *prev,
2852 struct task_struct *next)
2854 struct mm_struct *mm, *oldmm;
2856 prepare_task_switch(rq, prev, next);
2857 trace_sched_switch(rq, prev, next);
2859 oldmm = prev->active_mm;
2861 * For paravirt, this is coupled with an exit in switch_to to
2862 * combine the page table reload and the switch backend into
2865 arch_start_context_switch(prev);
2868 next->active_mm = oldmm;
2869 atomic_inc(&oldmm->mm_count);
2870 enter_lazy_tlb(oldmm, next);
2872 switch_mm(oldmm, mm, next);
2874 if (likely(!prev->mm)) {
2875 prev->active_mm = NULL;
2876 rq->prev_mm = oldmm;
2879 * Since the runqueue lock will be released by the next
2880 * task (which is an invalid locking op but in the case
2881 * of the scheduler it's an obvious special-case), so we
2882 * do an early lockdep release here:
2884 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2885 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2888 /* Here we just switch the register state and the stack. */
2889 switch_to(prev, next, prev);
2893 * this_rq must be evaluated again because prev may have moved
2894 * CPUs since it called schedule(), thus the 'rq' on its stack
2895 * frame will be invalid.
2897 finish_task_switch(this_rq(), prev);
2901 * nr_running, nr_uninterruptible and nr_context_switches:
2903 * externally visible scheduler statistics: current number of runnable
2904 * threads, current number of uninterruptible-sleeping threads, total
2905 * number of context switches performed since bootup.
2907 unsigned long nr_running(void)
2909 unsigned long i, sum = 0;
2911 for_each_online_cpu(i)
2912 sum += cpu_rq(i)->nr_running;
2917 unsigned long nr_uninterruptible(void)
2919 unsigned long i, sum = 0;
2921 for_each_possible_cpu(i)
2922 sum += cpu_rq(i)->nr_uninterruptible;
2925 * Since we read the counters lockless, it might be slightly
2926 * inaccurate. Do not allow it to go below zero though:
2928 if (unlikely((long)sum < 0))
2934 unsigned long long nr_context_switches(void)
2937 unsigned long long sum = 0;
2939 for_each_possible_cpu(i)
2940 sum += cpu_rq(i)->nr_switches;
2945 unsigned long nr_iowait(void)
2947 unsigned long i, sum = 0;
2949 for_each_possible_cpu(i)
2950 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2955 unsigned long nr_iowait_cpu(void)
2957 struct rq *this = this_rq();
2958 return atomic_read(&this->nr_iowait);
2961 unsigned long this_cpu_load(void)
2963 struct rq *this = this_rq();
2964 return this->cpu_load[0];
2968 /* Variables and functions for calc_load */
2969 static atomic_long_t calc_load_tasks;
2970 static unsigned long calc_load_update;
2971 unsigned long avenrun[3];
2972 EXPORT_SYMBOL(avenrun);
2975 * get_avenrun - get the load average array
2976 * @loads: pointer to dest load array
2977 * @offset: offset to add
2978 * @shift: shift count to shift the result left
2980 * These values are estimates at best, so no need for locking.
2982 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2984 loads[0] = (avenrun[0] + offset) << shift;
2985 loads[1] = (avenrun[1] + offset) << shift;
2986 loads[2] = (avenrun[2] + offset) << shift;
2989 static unsigned long
2990 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2993 load += active * (FIXED_1 - exp);
2994 return load >> FSHIFT;
2998 * calc_load - update the avenrun load estimates 10 ticks after the
2999 * CPUs have updated calc_load_tasks.
3001 void calc_global_load(void)
3003 unsigned long upd = calc_load_update + 10;
3006 if (time_before(jiffies, upd))
3009 active = atomic_long_read(&calc_load_tasks);
3010 active = active > 0 ? active * FIXED_1 : 0;
3012 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3013 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3014 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3016 calc_load_update += LOAD_FREQ;
3020 * Either called from update_cpu_load() or from a cpu going idle
3022 static void calc_load_account_active(struct rq *this_rq)
3024 long nr_active, delta;
3026 nr_active = this_rq->nr_running;
3027 nr_active += (long) this_rq->nr_uninterruptible;
3029 if (nr_active != this_rq->calc_load_active) {
3030 delta = nr_active - this_rq->calc_load_active;
3031 this_rq->calc_load_active = nr_active;
3032 atomic_long_add(delta, &calc_load_tasks);
3037 * Update rq->cpu_load[] statistics. This function is usually called every
3038 * scheduler tick (TICK_NSEC).
3040 static void update_cpu_load(struct rq *this_rq)
3042 unsigned long this_load = this_rq->load.weight;
3045 this_rq->nr_load_updates++;
3047 /* Update our load: */
3048 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3049 unsigned long old_load, new_load;
3051 /* scale is effectively 1 << i now, and >> i divides by scale */
3053 old_load = this_rq->cpu_load[i];
3054 new_load = this_load;
3056 * Round up the averaging division if load is increasing. This
3057 * prevents us from getting stuck on 9 if the load is 10, for
3060 if (new_load > old_load)
3061 new_load += scale-1;
3062 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3065 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3066 this_rq->calc_load_update += LOAD_FREQ;
3067 calc_load_account_active(this_rq);
3074 * sched_exec - execve() is a valuable balancing opportunity, because at
3075 * this point the task has the smallest effective memory and cache footprint.
3077 void sched_exec(void)
3079 struct task_struct *p = current;
3080 struct migration_req req;
3081 int dest_cpu, this_cpu;
3082 unsigned long flags;
3086 this_cpu = get_cpu();
3087 dest_cpu = select_task_rq(p, SD_BALANCE_EXEC, 0);
3088 if (dest_cpu == this_cpu) {
3093 rq = task_rq_lock(p, &flags);
3097 * select_task_rq() can race against ->cpus_allowed
3099 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
3100 || unlikely(!cpu_active(dest_cpu))) {
3101 task_rq_unlock(rq, &flags);
3105 /* force the process onto the specified CPU */
3106 if (migrate_task(p, dest_cpu, &req)) {
3107 /* Need to wait for migration thread (might exit: take ref). */
3108 struct task_struct *mt = rq->migration_thread;
3110 get_task_struct(mt);
3111 task_rq_unlock(rq, &flags);
3112 wake_up_process(mt);
3113 put_task_struct(mt);
3114 wait_for_completion(&req.done);
3118 task_rq_unlock(rq, &flags);
3123 DEFINE_PER_CPU(struct kernel_stat, kstat);
3125 EXPORT_PER_CPU_SYMBOL(kstat);
3128 * Return any ns on the sched_clock that have not yet been accounted in
3129 * @p in case that task is currently running.
3131 * Called with task_rq_lock() held on @rq.
3133 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3137 if (task_current(rq, p)) {
3138 update_rq_clock(rq);
3139 ns = rq->clock - p->se.exec_start;
3147 unsigned long long task_delta_exec(struct task_struct *p)
3149 unsigned long flags;
3153 rq = task_rq_lock(p, &flags);
3154 ns = do_task_delta_exec(p, rq);
3155 task_rq_unlock(rq, &flags);
3161 * Return accounted runtime for the task.
3162 * In case the task is currently running, return the runtime plus current's
3163 * pending runtime that have not been accounted yet.
3165 unsigned long long task_sched_runtime(struct task_struct *p)
3167 unsigned long flags;
3171 rq = task_rq_lock(p, &flags);
3172 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3173 task_rq_unlock(rq, &flags);
3179 * Return sum_exec_runtime for the thread group.
3180 * In case the task is currently running, return the sum plus current's
3181 * pending runtime that have not been accounted yet.
3183 * Note that the thread group might have other running tasks as well,
3184 * so the return value not includes other pending runtime that other
3185 * running tasks might have.
3187 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3189 struct task_cputime totals;
3190 unsigned long flags;
3194 rq = task_rq_lock(p, &flags);
3195 thread_group_cputime(p, &totals);
3196 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3197 task_rq_unlock(rq, &flags);
3203 * Account user cpu time to a process.
3204 * @p: the process that the cpu time gets accounted to
3205 * @cputime: the cpu time spent in user space since the last update
3206 * @cputime_scaled: cputime scaled by cpu frequency
3208 void account_user_time(struct task_struct *p, cputime_t cputime,
3209 cputime_t cputime_scaled)
3211 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3214 /* Add user time to process. */
3215 p->utime = cputime_add(p->utime, cputime);
3216 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3217 account_group_user_time(p, cputime);
3219 /* Add user time to cpustat. */
3220 tmp = cputime_to_cputime64(cputime);
3221 if (TASK_NICE(p) > 0)
3222 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3224 cpustat->user = cputime64_add(cpustat->user, tmp);
3226 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3227 /* Account for user time used */
3228 acct_update_integrals(p);
3232 * Account guest cpu time to a process.
3233 * @p: the process that the cpu time gets accounted to
3234 * @cputime: the cpu time spent in virtual machine since the last update
3235 * @cputime_scaled: cputime scaled by cpu frequency
3237 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3238 cputime_t cputime_scaled)
3241 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3243 tmp = cputime_to_cputime64(cputime);
3245 /* Add guest time to process. */
3246 p->utime = cputime_add(p->utime, cputime);
3247 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3248 account_group_user_time(p, cputime);
3249 p->gtime = cputime_add(p->gtime, cputime);
3251 /* Add guest time to cpustat. */
3252 if (TASK_NICE(p) > 0) {
3253 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3254 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3256 cpustat->user = cputime64_add(cpustat->user, tmp);
3257 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3262 * Account system cpu time to a process.
3263 * @p: the process that the cpu time gets accounted to
3264 * @hardirq_offset: the offset to subtract from hardirq_count()
3265 * @cputime: the cpu time spent in kernel space since the last update
3266 * @cputime_scaled: cputime scaled by cpu frequency
3268 void account_system_time(struct task_struct *p, int hardirq_offset,
3269 cputime_t cputime, cputime_t cputime_scaled)
3271 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3274 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3275 account_guest_time(p, cputime, cputime_scaled);
3279 /* Add system time to process. */
3280 p->stime = cputime_add(p->stime, cputime);
3281 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3282 account_group_system_time(p, cputime);
3284 /* Add system time to cpustat. */
3285 tmp = cputime_to_cputime64(cputime);
3286 if (hardirq_count() - hardirq_offset)
3287 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3288 else if (softirq_count())
3289 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3291 cpustat->system = cputime64_add(cpustat->system, tmp);
3293 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3295 /* Account for system time used */
3296 acct_update_integrals(p);
3300 * Account for involuntary wait time.
3301 * @steal: the cpu time spent in involuntary wait
3303 void account_steal_time(cputime_t cputime)
3305 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3306 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3308 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3312 * Account for idle time.
3313 * @cputime: the cpu time spent in idle wait
3315 void account_idle_time(cputime_t cputime)
3317 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3318 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3319 struct rq *rq = this_rq();
3321 if (atomic_read(&rq->nr_iowait) > 0)
3322 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3324 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3327 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3330 * Account a single tick of cpu time.
3331 * @p: the process that the cpu time gets accounted to
3332 * @user_tick: indicates if the tick is a user or a system tick
3334 void account_process_tick(struct task_struct *p, int user_tick)
3336 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3337 struct rq *rq = this_rq();
3340 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3341 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3342 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3345 account_idle_time(cputime_one_jiffy);
3349 * Account multiple ticks of steal time.
3350 * @p: the process from which the cpu time has been stolen
3351 * @ticks: number of stolen ticks
3353 void account_steal_ticks(unsigned long ticks)
3355 account_steal_time(jiffies_to_cputime(ticks));
3359 * Account multiple ticks of idle time.
3360 * @ticks: number of stolen ticks
3362 void account_idle_ticks(unsigned long ticks)
3364 account_idle_time(jiffies_to_cputime(ticks));
3370 * Use precise platform statistics if available:
3372 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3373 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3379 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3381 struct task_cputime cputime;
3383 thread_group_cputime(p, &cputime);
3385 *ut = cputime.utime;
3386 *st = cputime.stime;
3390 #ifndef nsecs_to_cputime
3391 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3394 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3396 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
3399 * Use CFS's precise accounting:
3401 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3406 temp = (u64)(rtime * utime);
3407 do_div(temp, total);
3408 utime = (cputime_t)temp;
3413 * Compare with previous values, to keep monotonicity:
3415 p->prev_utime = max(p->prev_utime, utime);
3416 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
3418 *ut = p->prev_utime;
3419 *st = p->prev_stime;
3423 * Must be called with siglock held.
3425 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3427 struct signal_struct *sig = p->signal;
3428 struct task_cputime cputime;
3429 cputime_t rtime, utime, total;
3431 thread_group_cputime(p, &cputime);
3433 total = cputime_add(cputime.utime, cputime.stime);
3434 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3439 temp = (u64)(rtime * cputime.utime);
3440 do_div(temp, total);
3441 utime = (cputime_t)temp;
3445 sig->prev_utime = max(sig->prev_utime, utime);
3446 sig->prev_stime = max(sig->prev_stime,
3447 cputime_sub(rtime, sig->prev_utime));
3449 *ut = sig->prev_utime;
3450 *st = sig->prev_stime;
3455 * This function gets called by the timer code, with HZ frequency.
3456 * We call it with interrupts disabled.
3458 * It also gets called by the fork code, when changing the parent's
3461 void scheduler_tick(void)
3463 int cpu = smp_processor_id();
3464 struct rq *rq = cpu_rq(cpu);
3465 struct task_struct *curr = rq->curr;
3469 raw_spin_lock(&rq->lock);
3470 update_rq_clock(rq);
3471 update_cpu_load(rq);
3472 curr->sched_class->task_tick(rq, curr, 0);
3473 raw_spin_unlock(&rq->lock);
3475 perf_event_task_tick(curr);
3478 rq->idle_at_tick = idle_cpu(cpu);
3479 trigger_load_balance(rq, cpu);
3483 notrace unsigned long get_parent_ip(unsigned long addr)
3485 if (in_lock_functions(addr)) {
3486 addr = CALLER_ADDR2;
3487 if (in_lock_functions(addr))
3488 addr = CALLER_ADDR3;
3493 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3494 defined(CONFIG_PREEMPT_TRACER))
3496 void __kprobes add_preempt_count(int val)
3498 #ifdef CONFIG_DEBUG_PREEMPT
3502 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3505 preempt_count() += val;
3506 #ifdef CONFIG_DEBUG_PREEMPT
3508 * Spinlock count overflowing soon?
3510 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3513 if (preempt_count() == val)
3514 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3516 EXPORT_SYMBOL(add_preempt_count);
3518 void __kprobes sub_preempt_count(int val)
3520 #ifdef CONFIG_DEBUG_PREEMPT
3524 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3527 * Is the spinlock portion underflowing?
3529 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3530 !(preempt_count() & PREEMPT_MASK)))
3534 if (preempt_count() == val)
3535 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3536 preempt_count() -= val;
3538 EXPORT_SYMBOL(sub_preempt_count);
3543 * Print scheduling while atomic bug:
3545 static noinline void __schedule_bug(struct task_struct *prev)
3547 struct pt_regs *regs = get_irq_regs();
3549 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3550 prev->comm, prev->pid, preempt_count());
3552 debug_show_held_locks(prev);
3554 if (irqs_disabled())
3555 print_irqtrace_events(prev);
3564 * Various schedule()-time debugging checks and statistics:
3566 static inline void schedule_debug(struct task_struct *prev)
3569 * Test if we are atomic. Since do_exit() needs to call into
3570 * schedule() atomically, we ignore that path for now.
3571 * Otherwise, whine if we are scheduling when we should not be.
3573 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3574 __schedule_bug(prev);
3576 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3578 schedstat_inc(this_rq(), sched_count);
3579 #ifdef CONFIG_SCHEDSTATS
3580 if (unlikely(prev->lock_depth >= 0)) {
3581 schedstat_inc(this_rq(), bkl_count);
3582 schedstat_inc(prev, sched_info.bkl_count);
3587 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3589 if (prev->state == TASK_RUNNING) {
3590 u64 runtime = prev->se.sum_exec_runtime;
3592 runtime -= prev->se.prev_sum_exec_runtime;
3593 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
3596 * In order to avoid avg_overlap growing stale when we are
3597 * indeed overlapping and hence not getting put to sleep, grow
3598 * the avg_overlap on preemption.
3600 * We use the average preemption runtime because that
3601 * correlates to the amount of cache footprint a task can
3604 update_avg(&prev->se.avg_overlap, runtime);
3606 prev->sched_class->put_prev_task(rq, prev);
3610 * Pick up the highest-prio task:
3612 static inline struct task_struct *
3613 pick_next_task(struct rq *rq)
3615 const struct sched_class *class;
3616 struct task_struct *p;
3619 * Optimization: we know that if all tasks are in
3620 * the fair class we can call that function directly:
3622 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3623 p = fair_sched_class.pick_next_task(rq);
3628 class = sched_class_highest;
3630 p = class->pick_next_task(rq);
3634 * Will never be NULL as the idle class always
3635 * returns a non-NULL p:
3637 class = class->next;
3642 * schedule() is the main scheduler function.
3644 asmlinkage void __sched schedule(void)
3646 struct task_struct *prev, *next;
3647 unsigned long *switch_count;
3653 cpu = smp_processor_id();
3657 switch_count = &prev->nivcsw;
3659 release_kernel_lock(prev);
3660 need_resched_nonpreemptible:
3662 schedule_debug(prev);
3664 if (sched_feat(HRTICK))
3667 raw_spin_lock_irq(&rq->lock);
3668 update_rq_clock(rq);
3669 clear_tsk_need_resched(prev);
3671 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3672 if (unlikely(signal_pending_state(prev->state, prev)))
3673 prev->state = TASK_RUNNING;
3675 deactivate_task(rq, prev, 1);
3676 switch_count = &prev->nvcsw;
3679 pre_schedule(rq, prev);
3681 if (unlikely(!rq->nr_running))
3682 idle_balance(cpu, rq);
3684 put_prev_task(rq, prev);
3685 next = pick_next_task(rq);
3687 if (likely(prev != next)) {
3688 sched_info_switch(prev, next);
3689 perf_event_task_sched_out(prev, next);
3695 context_switch(rq, prev, next); /* unlocks the rq */
3697 * the context switch might have flipped the stack from under
3698 * us, hence refresh the local variables.
3700 cpu = smp_processor_id();
3703 raw_spin_unlock_irq(&rq->lock);
3707 if (unlikely(reacquire_kernel_lock(current) < 0)) {
3709 switch_count = &prev->nivcsw;
3710 goto need_resched_nonpreemptible;
3713 preempt_enable_no_resched();
3717 EXPORT_SYMBOL(schedule);
3719 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3721 * Look out! "owner" is an entirely speculative pointer
3722 * access and not reliable.
3724 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
3729 if (!sched_feat(OWNER_SPIN))
3732 #ifdef CONFIG_DEBUG_PAGEALLOC
3734 * Need to access the cpu field knowing that
3735 * DEBUG_PAGEALLOC could have unmapped it if
3736 * the mutex owner just released it and exited.
3738 if (probe_kernel_address(&owner->cpu, cpu))
3745 * Even if the access succeeded (likely case),
3746 * the cpu field may no longer be valid.
3748 if (cpu >= nr_cpumask_bits)
3752 * We need to validate that we can do a
3753 * get_cpu() and that we have the percpu area.
3755 if (!cpu_online(cpu))
3762 * Owner changed, break to re-assess state.
3764 if (lock->owner != owner)
3768 * Is that owner really running on that cpu?
3770 if (task_thread_info(rq->curr) != owner || need_resched())
3780 #ifdef CONFIG_PREEMPT
3782 * this is the entry point to schedule() from in-kernel preemption
3783 * off of preempt_enable. Kernel preemptions off return from interrupt
3784 * occur there and call schedule directly.
3786 asmlinkage void __sched preempt_schedule(void)
3788 struct thread_info *ti = current_thread_info();
3791 * If there is a non-zero preempt_count or interrupts are disabled,
3792 * we do not want to preempt the current task. Just return..
3794 if (likely(ti->preempt_count || irqs_disabled()))
3798 add_preempt_count(PREEMPT_ACTIVE);
3800 sub_preempt_count(PREEMPT_ACTIVE);
3803 * Check again in case we missed a preemption opportunity
3804 * between schedule and now.
3807 } while (need_resched());
3809 EXPORT_SYMBOL(preempt_schedule);
3812 * this is the entry point to schedule() from kernel preemption
3813 * off of irq context.
3814 * Note, that this is called and return with irqs disabled. This will
3815 * protect us against recursive calling from irq.
3817 asmlinkage void __sched preempt_schedule_irq(void)
3819 struct thread_info *ti = current_thread_info();
3821 /* Catch callers which need to be fixed */
3822 BUG_ON(ti->preempt_count || !irqs_disabled());
3825 add_preempt_count(PREEMPT_ACTIVE);
3828 local_irq_disable();
3829 sub_preempt_count(PREEMPT_ACTIVE);
3832 * Check again in case we missed a preemption opportunity
3833 * between schedule and now.
3836 } while (need_resched());
3839 #endif /* CONFIG_PREEMPT */
3841 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3844 return try_to_wake_up(curr->private, mode, wake_flags);
3846 EXPORT_SYMBOL(default_wake_function);
3849 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3850 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3851 * number) then we wake all the non-exclusive tasks and one exclusive task.
3853 * There are circumstances in which we can try to wake a task which has already
3854 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3855 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3857 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3858 int nr_exclusive, int wake_flags, void *key)
3860 wait_queue_t *curr, *next;
3862 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3863 unsigned flags = curr->flags;
3865 if (curr->func(curr, mode, wake_flags, key) &&
3866 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3872 * __wake_up - wake up threads blocked on a waitqueue.
3874 * @mode: which threads
3875 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3876 * @key: is directly passed to the wakeup function
3878 * It may be assumed that this function implies a write memory barrier before
3879 * changing the task state if and only if any tasks are woken up.
3881 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3882 int nr_exclusive, void *key)
3884 unsigned long flags;
3886 spin_lock_irqsave(&q->lock, flags);
3887 __wake_up_common(q, mode, nr_exclusive, 0, key);
3888 spin_unlock_irqrestore(&q->lock, flags);
3890 EXPORT_SYMBOL(__wake_up);
3893 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3895 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3897 __wake_up_common(q, mode, 1, 0, NULL);
3900 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3902 __wake_up_common(q, mode, 1, 0, key);
3906 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3908 * @mode: which threads
3909 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3910 * @key: opaque value to be passed to wakeup targets
3912 * The sync wakeup differs that the waker knows that it will schedule
3913 * away soon, so while the target thread will be woken up, it will not
3914 * be migrated to another CPU - ie. the two threads are 'synchronized'
3915 * with each other. This can prevent needless bouncing between CPUs.
3917 * On UP it can prevent extra preemption.
3919 * It may be assumed that this function implies a write memory barrier before
3920 * changing the task state if and only if any tasks are woken up.
3922 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3923 int nr_exclusive, void *key)
3925 unsigned long flags;
3926 int wake_flags = WF_SYNC;
3931 if (unlikely(!nr_exclusive))
3934 spin_lock_irqsave(&q->lock, flags);
3935 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3936 spin_unlock_irqrestore(&q->lock, flags);
3938 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3941 * __wake_up_sync - see __wake_up_sync_key()
3943 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3945 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
3947 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3950 * complete: - signals a single thread waiting on this completion
3951 * @x: holds the state of this particular completion
3953 * This will wake up a single thread waiting on this completion. Threads will be
3954 * awakened in the same order in which they were queued.
3956 * See also complete_all(), wait_for_completion() and related routines.
3958 * It may be assumed that this function implies a write memory barrier before
3959 * changing the task state if and only if any tasks are woken up.
3961 void complete(struct completion *x)
3963 unsigned long flags;
3965 spin_lock_irqsave(&x->wait.lock, flags);
3967 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3968 spin_unlock_irqrestore(&x->wait.lock, flags);
3970 EXPORT_SYMBOL(complete);
3973 * complete_all: - signals all threads waiting on this completion
3974 * @x: holds the state of this particular completion
3976 * This will wake up all threads waiting on this particular completion event.
3978 * It may be assumed that this function implies a write memory barrier before
3979 * changing the task state if and only if any tasks are woken up.
3981 void complete_all(struct completion *x)
3983 unsigned long flags;
3985 spin_lock_irqsave(&x->wait.lock, flags);
3986 x->done += UINT_MAX/2;
3987 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3988 spin_unlock_irqrestore(&x->wait.lock, flags);
3990 EXPORT_SYMBOL(complete_all);
3992 static inline long __sched
3993 do_wait_for_common(struct completion *x, long timeout, int state)
3996 DECLARE_WAITQUEUE(wait, current);
3998 wait.flags |= WQ_FLAG_EXCLUSIVE;
3999 __add_wait_queue_tail(&x->wait, &wait);
4001 if (signal_pending_state(state, current)) {
4002 timeout = -ERESTARTSYS;
4005 __set_current_state(state);
4006 spin_unlock_irq(&x->wait.lock);
4007 timeout = schedule_timeout(timeout);
4008 spin_lock_irq(&x->wait.lock);
4009 } while (!x->done && timeout);
4010 __remove_wait_queue(&x->wait, &wait);
4015 return timeout ?: 1;
4019 wait_for_common(struct completion *x, long timeout, int state)
4023 spin_lock_irq(&x->wait.lock);
4024 timeout = do_wait_for_common(x, timeout, state);
4025 spin_unlock_irq(&x->wait.lock);
4030 * wait_for_completion: - waits for completion of a task
4031 * @x: holds the state of this particular completion
4033 * This waits to be signaled for completion of a specific task. It is NOT
4034 * interruptible and there is no timeout.
4036 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4037 * and interrupt capability. Also see complete().
4039 void __sched wait_for_completion(struct completion *x)
4041 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4043 EXPORT_SYMBOL(wait_for_completion);
4046 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4047 * @x: holds the state of this particular completion
4048 * @timeout: timeout value in jiffies
4050 * This waits for either a completion of a specific task to be signaled or for a
4051 * specified timeout to expire. The timeout is in jiffies. It is not
4054 unsigned long __sched
4055 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4057 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4059 EXPORT_SYMBOL(wait_for_completion_timeout);
4062 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4063 * @x: holds the state of this particular completion
4065 * This waits for completion of a specific task to be signaled. It is
4068 int __sched wait_for_completion_interruptible(struct completion *x)
4070 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4071 if (t == -ERESTARTSYS)
4075 EXPORT_SYMBOL(wait_for_completion_interruptible);
4078 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4079 * @x: holds the state of this particular completion
4080 * @timeout: timeout value in jiffies
4082 * This waits for either a completion of a specific task to be signaled or for a
4083 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4085 unsigned long __sched
4086 wait_for_completion_interruptible_timeout(struct completion *x,
4087 unsigned long timeout)
4089 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4091 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4094 * wait_for_completion_killable: - waits for completion of a task (killable)
4095 * @x: holds the state of this particular completion
4097 * This waits to be signaled for completion of a specific task. It can be
4098 * interrupted by a kill signal.
4100 int __sched wait_for_completion_killable(struct completion *x)
4102 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4103 if (t == -ERESTARTSYS)
4107 EXPORT_SYMBOL(wait_for_completion_killable);
4110 * try_wait_for_completion - try to decrement a completion without blocking
4111 * @x: completion structure
4113 * Returns: 0 if a decrement cannot be done without blocking
4114 * 1 if a decrement succeeded.
4116 * If a completion is being used as a counting completion,
4117 * attempt to decrement the counter without blocking. This
4118 * enables us to avoid waiting if the resource the completion
4119 * is protecting is not available.
4121 bool try_wait_for_completion(struct completion *x)
4123 unsigned long flags;
4126 spin_lock_irqsave(&x->wait.lock, flags);
4131 spin_unlock_irqrestore(&x->wait.lock, flags);
4134 EXPORT_SYMBOL(try_wait_for_completion);
4137 * completion_done - Test to see if a completion has any waiters
4138 * @x: completion structure
4140 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4141 * 1 if there are no waiters.
4144 bool completion_done(struct completion *x)
4146 unsigned long flags;
4149 spin_lock_irqsave(&x->wait.lock, flags);
4152 spin_unlock_irqrestore(&x->wait.lock, flags);
4155 EXPORT_SYMBOL(completion_done);
4158 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4160 unsigned long flags;
4163 init_waitqueue_entry(&wait, current);
4165 __set_current_state(state);
4167 spin_lock_irqsave(&q->lock, flags);
4168 __add_wait_queue(q, &wait);
4169 spin_unlock(&q->lock);
4170 timeout = schedule_timeout(timeout);
4171 spin_lock_irq(&q->lock);
4172 __remove_wait_queue(q, &wait);
4173 spin_unlock_irqrestore(&q->lock, flags);
4178 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4180 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4182 EXPORT_SYMBOL(interruptible_sleep_on);
4185 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4187 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4189 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4191 void __sched sleep_on(wait_queue_head_t *q)
4193 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4195 EXPORT_SYMBOL(sleep_on);
4197 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4199 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4201 EXPORT_SYMBOL(sleep_on_timeout);
4203 #ifdef CONFIG_RT_MUTEXES
4206 * rt_mutex_setprio - set the current priority of a task
4208 * @prio: prio value (kernel-internal form)
4210 * This function changes the 'effective' priority of a task. It does
4211 * not touch ->normal_prio like __setscheduler().
4213 * Used by the rt_mutex code to implement priority inheritance logic.
4215 void rt_mutex_setprio(struct task_struct *p, int prio)
4217 unsigned long flags;
4218 int oldprio, on_rq, running;
4220 const struct sched_class *prev_class;
4222 BUG_ON(prio < 0 || prio > MAX_PRIO);
4224 rq = task_rq_lock(p, &flags);
4225 update_rq_clock(rq);
4228 prev_class = p->sched_class;
4229 on_rq = p->se.on_rq;
4230 running = task_current(rq, p);
4232 dequeue_task(rq, p, 0);
4234 p->sched_class->put_prev_task(rq, p);
4237 p->sched_class = &rt_sched_class;
4239 p->sched_class = &fair_sched_class;
4244 p->sched_class->set_curr_task(rq);
4246 enqueue_task(rq, p, 0, oldprio < prio);
4248 check_class_changed(rq, p, prev_class, oldprio, running);
4250 task_rq_unlock(rq, &flags);
4255 void set_user_nice(struct task_struct *p, long nice)
4257 int old_prio, delta, on_rq;
4258 unsigned long flags;
4261 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4264 * We have to be careful, if called from sys_setpriority(),
4265 * the task might be in the middle of scheduling on another CPU.
4267 rq = task_rq_lock(p, &flags);
4268 update_rq_clock(rq);
4270 * The RT priorities are set via sched_setscheduler(), but we still
4271 * allow the 'normal' nice value to be set - but as expected
4272 * it wont have any effect on scheduling until the task is
4273 * SCHED_FIFO/SCHED_RR:
4275 if (task_has_rt_policy(p)) {
4276 p->static_prio = NICE_TO_PRIO(nice);
4279 on_rq = p->se.on_rq;
4281 dequeue_task(rq, p, 0);
4283 p->static_prio = NICE_TO_PRIO(nice);
4286 p->prio = effective_prio(p);
4287 delta = p->prio - old_prio;
4290 enqueue_task(rq, p, 0, false);
4292 * If the task increased its priority or is running and
4293 * lowered its priority, then reschedule its CPU:
4295 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4296 resched_task(rq->curr);
4299 task_rq_unlock(rq, &flags);
4301 EXPORT_SYMBOL(set_user_nice);
4304 * can_nice - check if a task can reduce its nice value
4308 int can_nice(const struct task_struct *p, const int nice)
4310 /* convert nice value [19,-20] to rlimit style value [1,40] */
4311 int nice_rlim = 20 - nice;
4313 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4314 capable(CAP_SYS_NICE));
4317 #ifdef __ARCH_WANT_SYS_NICE
4320 * sys_nice - change the priority of the current process.
4321 * @increment: priority increment
4323 * sys_setpriority is a more generic, but much slower function that
4324 * does similar things.
4326 SYSCALL_DEFINE1(nice, int, increment)
4331 * Setpriority might change our priority at the same moment.
4332 * We don't have to worry. Conceptually one call occurs first
4333 * and we have a single winner.
4335 if (increment < -40)
4340 nice = TASK_NICE(current) + increment;
4346 if (increment < 0 && !can_nice(current, nice))
4349 retval = security_task_setnice(current, nice);
4353 set_user_nice(current, nice);
4360 * task_prio - return the priority value of a given task.
4361 * @p: the task in question.
4363 * This is the priority value as seen by users in /proc.
4364 * RT tasks are offset by -200. Normal tasks are centered
4365 * around 0, value goes from -16 to +15.
4367 int task_prio(const struct task_struct *p)
4369 return p->prio - MAX_RT_PRIO;
4373 * task_nice - return the nice value of a given task.
4374 * @p: the task in question.
4376 int task_nice(const struct task_struct *p)
4378 return TASK_NICE(p);
4380 EXPORT_SYMBOL(task_nice);
4383 * idle_cpu - is a given cpu idle currently?
4384 * @cpu: the processor in question.
4386 int idle_cpu(int cpu)
4388 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4392 * idle_task - return the idle task for a given cpu.
4393 * @cpu: the processor in question.
4395 struct task_struct *idle_task(int cpu)
4397 return cpu_rq(cpu)->idle;
4401 * find_process_by_pid - find a process with a matching PID value.
4402 * @pid: the pid in question.
4404 static struct task_struct *find_process_by_pid(pid_t pid)
4406 return pid ? find_task_by_vpid(pid) : current;
4409 /* Actually do priority change: must hold rq lock. */
4411 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4413 BUG_ON(p->se.on_rq);
4416 p->rt_priority = prio;
4417 p->normal_prio = normal_prio(p);
4418 /* we are holding p->pi_lock already */
4419 p->prio = rt_mutex_getprio(p);
4420 if (rt_prio(p->prio))
4421 p->sched_class = &rt_sched_class;
4423 p->sched_class = &fair_sched_class;
4428 * check the target process has a UID that matches the current process's
4430 static bool check_same_owner(struct task_struct *p)
4432 const struct cred *cred = current_cred(), *pcred;
4436 pcred = __task_cred(p);
4437 match = (cred->euid == pcred->euid ||
4438 cred->euid == pcred->uid);
4443 static int __sched_setscheduler(struct task_struct *p, int policy,
4444 struct sched_param *param, bool user)
4446 int retval, oldprio, oldpolicy = -1, on_rq, running;
4447 unsigned long flags;
4448 const struct sched_class *prev_class;
4452 /* may grab non-irq protected spin_locks */
4453 BUG_ON(in_interrupt());
4455 /* double check policy once rq lock held */
4457 reset_on_fork = p->sched_reset_on_fork;
4458 policy = oldpolicy = p->policy;
4460 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4461 policy &= ~SCHED_RESET_ON_FORK;
4463 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4464 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4465 policy != SCHED_IDLE)
4470 * Valid priorities for SCHED_FIFO and SCHED_RR are
4471 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4472 * SCHED_BATCH and SCHED_IDLE is 0.
4474 if (param->sched_priority < 0 ||
4475 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4476 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4478 if (rt_policy(policy) != (param->sched_priority != 0))
4482 * Allow unprivileged RT tasks to decrease priority:
4484 if (user && !capable(CAP_SYS_NICE)) {
4485 if (rt_policy(policy)) {
4486 unsigned long rlim_rtprio;
4488 if (!lock_task_sighand(p, &flags))
4490 rlim_rtprio = task_rlimit(p, RLIMIT_RTPRIO);
4491 unlock_task_sighand(p, &flags);
4493 /* can't set/change the rt policy */
4494 if (policy != p->policy && !rlim_rtprio)
4497 /* can't increase priority */
4498 if (param->sched_priority > p->rt_priority &&
4499 param->sched_priority > rlim_rtprio)
4503 * Like positive nice levels, dont allow tasks to
4504 * move out of SCHED_IDLE either:
4506 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4509 /* can't change other user's priorities */
4510 if (!check_same_owner(p))
4513 /* Normal users shall not reset the sched_reset_on_fork flag */
4514 if (p->sched_reset_on_fork && !reset_on_fork)
4519 #ifdef CONFIG_RT_GROUP_SCHED
4521 * Do not allow realtime tasks into groups that have no runtime
4524 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4525 task_group(p)->rt_bandwidth.rt_runtime == 0)
4529 retval = security_task_setscheduler(p, policy, param);
4535 * make sure no PI-waiters arrive (or leave) while we are
4536 * changing the priority of the task:
4538 raw_spin_lock_irqsave(&p->pi_lock, flags);
4540 * To be able to change p->policy safely, the apropriate
4541 * runqueue lock must be held.
4543 rq = __task_rq_lock(p);
4544 /* recheck policy now with rq lock held */
4545 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4546 policy = oldpolicy = -1;
4547 __task_rq_unlock(rq);
4548 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4551 update_rq_clock(rq);
4552 on_rq = p->se.on_rq;
4553 running = task_current(rq, p);
4555 deactivate_task(rq, p, 0);
4557 p->sched_class->put_prev_task(rq, p);
4559 p->sched_reset_on_fork = reset_on_fork;
4562 prev_class = p->sched_class;
4563 __setscheduler(rq, p, policy, param->sched_priority);
4566 p->sched_class->set_curr_task(rq);
4568 activate_task(rq, p, 0);
4570 check_class_changed(rq, p, prev_class, oldprio, running);
4572 __task_rq_unlock(rq);
4573 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4575 rt_mutex_adjust_pi(p);
4581 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4582 * @p: the task in question.
4583 * @policy: new policy.
4584 * @param: structure containing the new RT priority.
4586 * NOTE that the task may be already dead.
4588 int sched_setscheduler(struct task_struct *p, int policy,
4589 struct sched_param *param)
4591 return __sched_setscheduler(p, policy, param, true);
4593 EXPORT_SYMBOL_GPL(sched_setscheduler);
4596 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4597 * @p: the task in question.
4598 * @policy: new policy.
4599 * @param: structure containing the new RT priority.
4601 * Just like sched_setscheduler, only don't bother checking if the
4602 * current context has permission. For example, this is needed in
4603 * stop_machine(): we create temporary high priority worker threads,
4604 * but our caller might not have that capability.
4606 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4607 struct sched_param *param)
4609 return __sched_setscheduler(p, policy, param, false);
4613 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4615 struct sched_param lparam;
4616 struct task_struct *p;
4619 if (!param || pid < 0)
4621 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4626 p = find_process_by_pid(pid);
4628 retval = sched_setscheduler(p, policy, &lparam);
4635 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4636 * @pid: the pid in question.
4637 * @policy: new policy.
4638 * @param: structure containing the new RT priority.
4640 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4641 struct sched_param __user *, param)
4643 /* negative values for policy are not valid */
4647 return do_sched_setscheduler(pid, policy, param);
4651 * sys_sched_setparam - set/change the RT priority of a thread
4652 * @pid: the pid in question.
4653 * @param: structure containing the new RT priority.
4655 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4657 return do_sched_setscheduler(pid, -1, param);
4661 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4662 * @pid: the pid in question.
4664 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4666 struct task_struct *p;
4674 p = find_process_by_pid(pid);
4676 retval = security_task_getscheduler(p);
4679 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4686 * sys_sched_getparam - get the RT priority of a thread
4687 * @pid: the pid in question.
4688 * @param: structure containing the RT priority.
4690 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4692 struct sched_param lp;
4693 struct task_struct *p;
4696 if (!param || pid < 0)
4700 p = find_process_by_pid(pid);
4705 retval = security_task_getscheduler(p);
4709 lp.sched_priority = p->rt_priority;
4713 * This one might sleep, we cannot do it with a spinlock held ...
4715 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4724 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4726 cpumask_var_t cpus_allowed, new_mask;
4727 struct task_struct *p;
4733 p = find_process_by_pid(pid);
4740 /* Prevent p going away */
4744 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4748 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4750 goto out_free_cpus_allowed;
4753 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
4756 retval = security_task_setscheduler(p, 0, NULL);
4760 cpuset_cpus_allowed(p, cpus_allowed);
4761 cpumask_and(new_mask, in_mask, cpus_allowed);
4763 retval = set_cpus_allowed_ptr(p, new_mask);
4766 cpuset_cpus_allowed(p, cpus_allowed);
4767 if (!cpumask_subset(new_mask, cpus_allowed)) {
4769 * We must have raced with a concurrent cpuset
4770 * update. Just reset the cpus_allowed to the
4771 * cpuset's cpus_allowed
4773 cpumask_copy(new_mask, cpus_allowed);
4778 free_cpumask_var(new_mask);
4779 out_free_cpus_allowed:
4780 free_cpumask_var(cpus_allowed);
4787 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4788 struct cpumask *new_mask)
4790 if (len < cpumask_size())
4791 cpumask_clear(new_mask);
4792 else if (len > cpumask_size())
4793 len = cpumask_size();
4795 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4799 * sys_sched_setaffinity - set the cpu affinity of a process
4800 * @pid: pid of the process
4801 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4802 * @user_mask_ptr: user-space pointer to the new cpu mask
4804 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4805 unsigned long __user *, user_mask_ptr)
4807 cpumask_var_t new_mask;
4810 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4813 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4815 retval = sched_setaffinity(pid, new_mask);
4816 free_cpumask_var(new_mask);
4820 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4822 struct task_struct *p;
4823 unsigned long flags;
4831 p = find_process_by_pid(pid);
4835 retval = security_task_getscheduler(p);
4839 rq = task_rq_lock(p, &flags);
4840 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4841 task_rq_unlock(rq, &flags);
4851 * sys_sched_getaffinity - get the cpu affinity of a process
4852 * @pid: pid of the process
4853 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4854 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4856 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4857 unsigned long __user *, user_mask_ptr)
4862 if (len < cpumask_size())
4865 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4868 ret = sched_getaffinity(pid, mask);
4870 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
4873 ret = cpumask_size();
4875 free_cpumask_var(mask);
4881 * sys_sched_yield - yield the current processor to other threads.
4883 * This function yields the current CPU to other tasks. If there are no
4884 * other threads running on this CPU then this function will return.
4886 SYSCALL_DEFINE0(sched_yield)
4888 struct rq *rq = this_rq_lock();
4890 schedstat_inc(rq, yld_count);
4891 current->sched_class->yield_task(rq);
4894 * Since we are going to call schedule() anyway, there's
4895 * no need to preempt or enable interrupts:
4897 __release(rq->lock);
4898 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4899 do_raw_spin_unlock(&rq->lock);
4900 preempt_enable_no_resched();
4907 static inline int should_resched(void)
4909 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4912 static void __cond_resched(void)
4914 add_preempt_count(PREEMPT_ACTIVE);
4916 sub_preempt_count(PREEMPT_ACTIVE);
4919 int __sched _cond_resched(void)
4921 if (should_resched()) {
4927 EXPORT_SYMBOL(_cond_resched);
4930 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4931 * call schedule, and on return reacquire the lock.
4933 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4934 * operations here to prevent schedule() from being called twice (once via
4935 * spin_unlock(), once by hand).
4937 int __cond_resched_lock(spinlock_t *lock)
4939 int resched = should_resched();
4942 lockdep_assert_held(lock);
4944 if (spin_needbreak(lock) || resched) {
4955 EXPORT_SYMBOL(__cond_resched_lock);
4957 int __sched __cond_resched_softirq(void)
4959 BUG_ON(!in_softirq());
4961 if (should_resched()) {
4969 EXPORT_SYMBOL(__cond_resched_softirq);
4972 * yield - yield the current processor to other threads.
4974 * This is a shortcut for kernel-space yielding - it marks the
4975 * thread runnable and calls sys_sched_yield().
4977 void __sched yield(void)
4979 set_current_state(TASK_RUNNING);
4982 EXPORT_SYMBOL(yield);
4985 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4986 * that process accounting knows that this is a task in IO wait state.
4988 void __sched io_schedule(void)
4990 struct rq *rq = raw_rq();
4992 delayacct_blkio_start();
4993 atomic_inc(&rq->nr_iowait);
4994 current->in_iowait = 1;
4996 current->in_iowait = 0;
4997 atomic_dec(&rq->nr_iowait);
4998 delayacct_blkio_end();
5000 EXPORT_SYMBOL(io_schedule);
5002 long __sched io_schedule_timeout(long timeout)
5004 struct rq *rq = raw_rq();
5007 delayacct_blkio_start();
5008 atomic_inc(&rq->nr_iowait);
5009 current->in_iowait = 1;
5010 ret = schedule_timeout(timeout);
5011 current->in_iowait = 0;
5012 atomic_dec(&rq->nr_iowait);
5013 delayacct_blkio_end();
5018 * sys_sched_get_priority_max - return maximum RT priority.
5019 * @policy: scheduling class.
5021 * this syscall returns the maximum rt_priority that can be used
5022 * by a given scheduling class.
5024 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5031 ret = MAX_USER_RT_PRIO-1;
5043 * sys_sched_get_priority_min - return minimum RT priority.
5044 * @policy: scheduling class.
5046 * this syscall returns the minimum rt_priority that can be used
5047 * by a given scheduling class.
5049 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5067 * sys_sched_rr_get_interval - return the default timeslice of a process.
5068 * @pid: pid of the process.
5069 * @interval: userspace pointer to the timeslice value.
5071 * this syscall writes the default timeslice value of a given process
5072 * into the user-space timespec buffer. A value of '0' means infinity.
5074 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5075 struct timespec __user *, interval)
5077 struct task_struct *p;
5078 unsigned int time_slice;
5079 unsigned long flags;
5089 p = find_process_by_pid(pid);
5093 retval = security_task_getscheduler(p);
5097 rq = task_rq_lock(p, &flags);
5098 time_slice = p->sched_class->get_rr_interval(rq, p);
5099 task_rq_unlock(rq, &flags);
5102 jiffies_to_timespec(time_slice, &t);
5103 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5111 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5113 void sched_show_task(struct task_struct *p)
5115 unsigned long free = 0;
5118 state = p->state ? __ffs(p->state) + 1 : 0;
5119 printk(KERN_INFO "%-13.13s %c", p->comm,
5120 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5121 #if BITS_PER_LONG == 32
5122 if (state == TASK_RUNNING)
5123 printk(KERN_CONT " running ");
5125 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5127 if (state == TASK_RUNNING)
5128 printk(KERN_CONT " running task ");
5130 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5132 #ifdef CONFIG_DEBUG_STACK_USAGE
5133 free = stack_not_used(p);
5135 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5136 task_pid_nr(p), task_pid_nr(p->real_parent),
5137 (unsigned long)task_thread_info(p)->flags);
5139 show_stack(p, NULL);
5142 void show_state_filter(unsigned long state_filter)
5144 struct task_struct *g, *p;
5146 #if BITS_PER_LONG == 32
5148 " task PC stack pid father\n");
5151 " task PC stack pid father\n");
5153 read_lock(&tasklist_lock);
5154 do_each_thread(g, p) {
5156 * reset the NMI-timeout, listing all files on a slow
5157 * console might take alot of time:
5159 touch_nmi_watchdog();
5160 if (!state_filter || (p->state & state_filter))
5162 } while_each_thread(g, p);
5164 touch_all_softlockup_watchdogs();
5166 #ifdef CONFIG_SCHED_DEBUG
5167 sysrq_sched_debug_show();
5169 read_unlock(&tasklist_lock);
5171 * Only show locks if all tasks are dumped:
5174 debug_show_all_locks();
5177 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5179 idle->sched_class = &idle_sched_class;
5183 * init_idle - set up an idle thread for a given CPU
5184 * @idle: task in question
5185 * @cpu: cpu the idle task belongs to
5187 * NOTE: this function does not set the idle thread's NEED_RESCHED
5188 * flag, to make booting more robust.
5190 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5192 struct rq *rq = cpu_rq(cpu);
5193 unsigned long flags;
5195 raw_spin_lock_irqsave(&rq->lock, flags);
5198 idle->state = TASK_RUNNING;
5199 idle->se.exec_start = sched_clock();
5201 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
5202 __set_task_cpu(idle, cpu);
5204 rq->curr = rq->idle = idle;
5205 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5208 raw_spin_unlock_irqrestore(&rq->lock, flags);
5210 /* Set the preempt count _outside_ the spinlocks! */
5211 #if defined(CONFIG_PREEMPT)
5212 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5214 task_thread_info(idle)->preempt_count = 0;
5217 * The idle tasks have their own, simple scheduling class:
5219 idle->sched_class = &idle_sched_class;
5220 ftrace_graph_init_task(idle);
5224 * In a system that switches off the HZ timer nohz_cpu_mask
5225 * indicates which cpus entered this state. This is used
5226 * in the rcu update to wait only for active cpus. For system
5227 * which do not switch off the HZ timer nohz_cpu_mask should
5228 * always be CPU_BITS_NONE.
5230 cpumask_var_t nohz_cpu_mask;
5233 * Increase the granularity value when there are more CPUs,
5234 * because with more CPUs the 'effective latency' as visible
5235 * to users decreases. But the relationship is not linear,
5236 * so pick a second-best guess by going with the log2 of the
5239 * This idea comes from the SD scheduler of Con Kolivas:
5241 static int get_update_sysctl_factor(void)
5243 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5244 unsigned int factor;
5246 switch (sysctl_sched_tunable_scaling) {
5247 case SCHED_TUNABLESCALING_NONE:
5250 case SCHED_TUNABLESCALING_LINEAR:
5253 case SCHED_TUNABLESCALING_LOG:
5255 factor = 1 + ilog2(cpus);
5262 static void update_sysctl(void)
5264 unsigned int factor = get_update_sysctl_factor();
5266 #define SET_SYSCTL(name) \
5267 (sysctl_##name = (factor) * normalized_sysctl_##name)
5268 SET_SYSCTL(sched_min_granularity);
5269 SET_SYSCTL(sched_latency);
5270 SET_SYSCTL(sched_wakeup_granularity);
5271 SET_SYSCTL(sched_shares_ratelimit);
5275 static inline void sched_init_granularity(void)
5282 * This is how migration works:
5284 * 1) we queue a struct migration_req structure in the source CPU's
5285 * runqueue and wake up that CPU's migration thread.
5286 * 2) we down() the locked semaphore => thread blocks.
5287 * 3) migration thread wakes up (implicitly it forces the migrated
5288 * thread off the CPU)
5289 * 4) it gets the migration request and checks whether the migrated
5290 * task is still in the wrong runqueue.
5291 * 5) if it's in the wrong runqueue then the migration thread removes
5292 * it and puts it into the right queue.
5293 * 6) migration thread up()s the semaphore.
5294 * 7) we wake up and the migration is done.
5298 * Change a given task's CPU affinity. Migrate the thread to a
5299 * proper CPU and schedule it away if the CPU it's executing on
5300 * is removed from the allowed bitmask.
5302 * NOTE: the caller must have a valid reference to the task, the
5303 * task must not exit() & deallocate itself prematurely. The
5304 * call is not atomic; no spinlocks may be held.
5306 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5308 struct migration_req req;
5309 unsigned long flags;
5313 rq = task_rq_lock(p, &flags);
5315 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5320 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5321 !cpumask_equal(&p->cpus_allowed, new_mask))) {
5326 if (p->sched_class->set_cpus_allowed)
5327 p->sched_class->set_cpus_allowed(p, new_mask);
5329 cpumask_copy(&p->cpus_allowed, new_mask);
5330 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5333 /* Can the task run on the task's current CPU? If so, we're done */
5334 if (cpumask_test_cpu(task_cpu(p), new_mask))
5337 if (migrate_task(p, cpumask_any_and(cpu_active_mask, new_mask), &req)) {
5338 /* Need help from migration thread: drop lock and wait. */
5339 struct task_struct *mt = rq->migration_thread;
5341 get_task_struct(mt);
5342 task_rq_unlock(rq, &flags);
5343 wake_up_process(rq->migration_thread);
5344 put_task_struct(mt);
5345 wait_for_completion(&req.done);
5346 tlb_migrate_finish(p->mm);
5350 task_rq_unlock(rq, &flags);
5354 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5357 * Move (not current) task off this cpu, onto dest cpu. We're doing
5358 * this because either it can't run here any more (set_cpus_allowed()
5359 * away from this CPU, or CPU going down), or because we're
5360 * attempting to rebalance this task on exec (sched_exec).
5362 * So we race with normal scheduler movements, but that's OK, as long
5363 * as the task is no longer on this CPU.
5365 * Returns non-zero if task was successfully migrated.
5367 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5369 struct rq *rq_dest, *rq_src;
5372 if (unlikely(!cpu_active(dest_cpu)))
5375 rq_src = cpu_rq(src_cpu);
5376 rq_dest = cpu_rq(dest_cpu);
5378 double_rq_lock(rq_src, rq_dest);
5379 /* Already moved. */
5380 if (task_cpu(p) != src_cpu)
5382 /* Affinity changed (again). */
5383 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
5387 * If we're not on a rq, the next wake-up will ensure we're
5391 deactivate_task(rq_src, p, 0);
5392 set_task_cpu(p, dest_cpu);
5393 activate_task(rq_dest, p, 0);
5394 check_preempt_curr(rq_dest, p, 0);
5399 double_rq_unlock(rq_src, rq_dest);
5403 #define RCU_MIGRATION_IDLE 0
5404 #define RCU_MIGRATION_NEED_QS 1
5405 #define RCU_MIGRATION_GOT_QS 2
5406 #define RCU_MIGRATION_MUST_SYNC 3
5409 * migration_thread - this is a highprio system thread that performs
5410 * thread migration by bumping thread off CPU then 'pushing' onto
5413 static int migration_thread(void *data)
5416 int cpu = (long)data;
5420 BUG_ON(rq->migration_thread != current);
5422 set_current_state(TASK_INTERRUPTIBLE);
5423 while (!kthread_should_stop()) {
5424 struct migration_req *req;
5425 struct list_head *head;
5427 raw_spin_lock_irq(&rq->lock);
5429 if (cpu_is_offline(cpu)) {
5430 raw_spin_unlock_irq(&rq->lock);
5434 if (rq->active_balance) {
5435 active_load_balance(rq, cpu);
5436 rq->active_balance = 0;
5439 head = &rq->migration_queue;
5441 if (list_empty(head)) {
5442 raw_spin_unlock_irq(&rq->lock);
5444 set_current_state(TASK_INTERRUPTIBLE);
5447 req = list_entry(head->next, struct migration_req, list);
5448 list_del_init(head->next);
5450 if (req->task != NULL) {
5451 raw_spin_unlock(&rq->lock);
5452 __migrate_task(req->task, cpu, req->dest_cpu);
5453 } else if (likely(cpu == (badcpu = smp_processor_id()))) {
5454 req->dest_cpu = RCU_MIGRATION_GOT_QS;
5455 raw_spin_unlock(&rq->lock);
5457 req->dest_cpu = RCU_MIGRATION_MUST_SYNC;
5458 raw_spin_unlock(&rq->lock);
5459 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu, cpu);
5463 complete(&req->done);
5465 __set_current_state(TASK_RUNNING);
5470 #ifdef CONFIG_HOTPLUG_CPU
5472 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5476 local_irq_disable();
5477 ret = __migrate_task(p, src_cpu, dest_cpu);
5483 * Figure out where task on dead CPU should go, use force if necessary.
5485 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5490 dest_cpu = select_fallback_rq(dead_cpu, p);
5492 /* It can have affinity changed while we were choosing. */
5493 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
5498 * While a dead CPU has no uninterruptible tasks queued at this point,
5499 * it might still have a nonzero ->nr_uninterruptible counter, because
5500 * for performance reasons the counter is not stricly tracking tasks to
5501 * their home CPUs. So we just add the counter to another CPU's counter,
5502 * to keep the global sum constant after CPU-down:
5504 static void migrate_nr_uninterruptible(struct rq *rq_src)
5506 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5507 unsigned long flags;
5509 local_irq_save(flags);
5510 double_rq_lock(rq_src, rq_dest);
5511 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5512 rq_src->nr_uninterruptible = 0;
5513 double_rq_unlock(rq_src, rq_dest);
5514 local_irq_restore(flags);
5517 /* Run through task list and migrate tasks from the dead cpu. */
5518 static void migrate_live_tasks(int src_cpu)
5520 struct task_struct *p, *t;
5522 read_lock(&tasklist_lock);
5524 do_each_thread(t, p) {
5528 if (task_cpu(p) == src_cpu)
5529 move_task_off_dead_cpu(src_cpu, p);
5530 } while_each_thread(t, p);
5532 read_unlock(&tasklist_lock);
5536 * Schedules idle task to be the next runnable task on current CPU.
5537 * It does so by boosting its priority to highest possible.
5538 * Used by CPU offline code.
5540 void sched_idle_next(void)
5542 int this_cpu = smp_processor_id();
5543 struct rq *rq = cpu_rq(this_cpu);
5544 struct task_struct *p = rq->idle;
5545 unsigned long flags;
5547 /* cpu has to be offline */
5548 BUG_ON(cpu_online(this_cpu));
5551 * Strictly not necessary since rest of the CPUs are stopped by now
5552 * and interrupts disabled on the current cpu.
5554 raw_spin_lock_irqsave(&rq->lock, flags);
5556 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5558 update_rq_clock(rq);
5559 activate_task(rq, p, 0);
5561 raw_spin_unlock_irqrestore(&rq->lock, flags);
5565 * Ensures that the idle task is using init_mm right before its cpu goes
5568 void idle_task_exit(void)
5570 struct mm_struct *mm = current->active_mm;
5572 BUG_ON(cpu_online(smp_processor_id()));
5575 switch_mm(mm, &init_mm, current);
5579 /* called under rq->lock with disabled interrupts */
5580 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5582 struct rq *rq = cpu_rq(dead_cpu);
5584 /* Must be exiting, otherwise would be on tasklist. */
5585 BUG_ON(!p->exit_state);
5587 /* Cannot have done final schedule yet: would have vanished. */
5588 BUG_ON(p->state == TASK_DEAD);
5593 * Drop lock around migration; if someone else moves it,
5594 * that's OK. No task can be added to this CPU, so iteration is
5597 raw_spin_unlock_irq(&rq->lock);
5598 move_task_off_dead_cpu(dead_cpu, p);
5599 raw_spin_lock_irq(&rq->lock);
5604 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5605 static void migrate_dead_tasks(unsigned int dead_cpu)
5607 struct rq *rq = cpu_rq(dead_cpu);
5608 struct task_struct *next;
5611 if (!rq->nr_running)
5613 update_rq_clock(rq);
5614 next = pick_next_task(rq);
5617 next->sched_class->put_prev_task(rq, next);
5618 migrate_dead(dead_cpu, next);
5624 * remove the tasks which were accounted by rq from calc_load_tasks.
5626 static void calc_global_load_remove(struct rq *rq)
5628 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5629 rq->calc_load_active = 0;
5631 #endif /* CONFIG_HOTPLUG_CPU */
5633 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5635 static struct ctl_table sd_ctl_dir[] = {
5637 .procname = "sched_domain",
5643 static struct ctl_table sd_ctl_root[] = {
5645 .procname = "kernel",
5647 .child = sd_ctl_dir,
5652 static struct ctl_table *sd_alloc_ctl_entry(int n)
5654 struct ctl_table *entry =
5655 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5660 static void sd_free_ctl_entry(struct ctl_table **tablep)
5662 struct ctl_table *entry;
5665 * In the intermediate directories, both the child directory and
5666 * procname are dynamically allocated and could fail but the mode
5667 * will always be set. In the lowest directory the names are
5668 * static strings and all have proc handlers.
5670 for (entry = *tablep; entry->mode; entry++) {
5672 sd_free_ctl_entry(&entry->child);
5673 if (entry->proc_handler == NULL)
5674 kfree(entry->procname);
5682 set_table_entry(struct ctl_table *entry,
5683 const char *procname, void *data, int maxlen,
5684 mode_t mode, proc_handler *proc_handler)
5686 entry->procname = procname;
5688 entry->maxlen = maxlen;
5690 entry->proc_handler = proc_handler;
5693 static struct ctl_table *
5694 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5696 struct ctl_table *table = sd_alloc_ctl_entry(13);
5701 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5702 sizeof(long), 0644, proc_doulongvec_minmax);
5703 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5704 sizeof(long), 0644, proc_doulongvec_minmax);
5705 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5706 sizeof(int), 0644, proc_dointvec_minmax);
5707 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5708 sizeof(int), 0644, proc_dointvec_minmax);
5709 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5710 sizeof(int), 0644, proc_dointvec_minmax);
5711 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5712 sizeof(int), 0644, proc_dointvec_minmax);
5713 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5714 sizeof(int), 0644, proc_dointvec_minmax);
5715 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5716 sizeof(int), 0644, proc_dointvec_minmax);
5717 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5718 sizeof(int), 0644, proc_dointvec_minmax);
5719 set_table_entry(&table[9], "cache_nice_tries",
5720 &sd->cache_nice_tries,
5721 sizeof(int), 0644, proc_dointvec_minmax);
5722 set_table_entry(&table[10], "flags", &sd->flags,
5723 sizeof(int), 0644, proc_dointvec_minmax);
5724 set_table_entry(&table[11], "name", sd->name,
5725 CORENAME_MAX_SIZE, 0444, proc_dostring);
5726 /* &table[12] is terminator */
5731 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5733 struct ctl_table *entry, *table;
5734 struct sched_domain *sd;
5735 int domain_num = 0, i;
5738 for_each_domain(cpu, sd)
5740 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5745 for_each_domain(cpu, sd) {
5746 snprintf(buf, 32, "domain%d", i);
5747 entry->procname = kstrdup(buf, GFP_KERNEL);
5749 entry->child = sd_alloc_ctl_domain_table(sd);
5756 static struct ctl_table_header *sd_sysctl_header;
5757 static void register_sched_domain_sysctl(void)
5759 int i, cpu_num = num_possible_cpus();
5760 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5763 WARN_ON(sd_ctl_dir[0].child);
5764 sd_ctl_dir[0].child = entry;
5769 for_each_possible_cpu(i) {
5770 snprintf(buf, 32, "cpu%d", i);
5771 entry->procname = kstrdup(buf, GFP_KERNEL);
5773 entry->child = sd_alloc_ctl_cpu_table(i);
5777 WARN_ON(sd_sysctl_header);
5778 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5781 /* may be called multiple times per register */
5782 static void unregister_sched_domain_sysctl(void)
5784 if (sd_sysctl_header)
5785 unregister_sysctl_table(sd_sysctl_header);
5786 sd_sysctl_header = NULL;
5787 if (sd_ctl_dir[0].child)
5788 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5791 static void register_sched_domain_sysctl(void)
5794 static void unregister_sched_domain_sysctl(void)
5799 static void set_rq_online(struct rq *rq)
5802 const struct sched_class *class;
5804 cpumask_set_cpu(rq->cpu, rq->rd->online);
5807 for_each_class(class) {
5808 if (class->rq_online)
5809 class->rq_online(rq);
5814 static void set_rq_offline(struct rq *rq)
5817 const struct sched_class *class;
5819 for_each_class(class) {
5820 if (class->rq_offline)
5821 class->rq_offline(rq);
5824 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5830 * migration_call - callback that gets triggered when a CPU is added.
5831 * Here we can start up the necessary migration thread for the new CPU.
5833 static int __cpuinit
5834 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5836 struct task_struct *p;
5837 int cpu = (long)hcpu;
5838 unsigned long flags;
5843 case CPU_UP_PREPARE:
5844 case CPU_UP_PREPARE_FROZEN:
5845 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
5848 kthread_bind(p, cpu);
5849 /* Must be high prio: stop_machine expects to yield to it. */
5850 rq = task_rq_lock(p, &flags);
5851 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5852 task_rq_unlock(rq, &flags);
5854 cpu_rq(cpu)->migration_thread = p;
5855 rq->calc_load_update = calc_load_update;
5859 case CPU_ONLINE_FROZEN:
5860 /* Strictly unnecessary, as first user will wake it. */
5861 wake_up_process(cpu_rq(cpu)->migration_thread);
5863 /* Update our root-domain */
5865 raw_spin_lock_irqsave(&rq->lock, flags);
5867 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5871 raw_spin_unlock_irqrestore(&rq->lock, flags);
5874 #ifdef CONFIG_HOTPLUG_CPU
5875 case CPU_UP_CANCELED:
5876 case CPU_UP_CANCELED_FROZEN:
5877 if (!cpu_rq(cpu)->migration_thread)
5879 /* Unbind it from offline cpu so it can run. Fall thru. */
5880 kthread_bind(cpu_rq(cpu)->migration_thread,
5881 cpumask_any(cpu_online_mask));
5882 kthread_stop(cpu_rq(cpu)->migration_thread);
5883 put_task_struct(cpu_rq(cpu)->migration_thread);
5884 cpu_rq(cpu)->migration_thread = NULL;
5888 case CPU_DEAD_FROZEN:
5889 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
5890 migrate_live_tasks(cpu);
5892 kthread_stop(rq->migration_thread);
5893 put_task_struct(rq->migration_thread);
5894 rq->migration_thread = NULL;
5895 /* Idle task back to normal (off runqueue, low prio) */
5896 raw_spin_lock_irq(&rq->lock);
5897 update_rq_clock(rq);
5898 deactivate_task(rq, rq->idle, 0);
5899 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5900 rq->idle->sched_class = &idle_sched_class;
5901 migrate_dead_tasks(cpu);
5902 raw_spin_unlock_irq(&rq->lock);
5904 migrate_nr_uninterruptible(rq);
5905 BUG_ON(rq->nr_running != 0);
5906 calc_global_load_remove(rq);
5908 * No need to migrate the tasks: it was best-effort if
5909 * they didn't take sched_hotcpu_mutex. Just wake up
5912 raw_spin_lock_irq(&rq->lock);
5913 while (!list_empty(&rq->migration_queue)) {
5914 struct migration_req *req;
5916 req = list_entry(rq->migration_queue.next,
5917 struct migration_req, list);
5918 list_del_init(&req->list);
5919 raw_spin_unlock_irq(&rq->lock);
5920 complete(&req->done);
5921 raw_spin_lock_irq(&rq->lock);
5923 raw_spin_unlock_irq(&rq->lock);
5927 case CPU_DYING_FROZEN:
5928 /* Update our root-domain */
5930 raw_spin_lock_irqsave(&rq->lock, flags);
5932 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5935 raw_spin_unlock_irqrestore(&rq->lock, flags);
5943 * Register at high priority so that task migration (migrate_all_tasks)
5944 * happens before everything else. This has to be lower priority than
5945 * the notifier in the perf_event subsystem, though.
5947 static struct notifier_block __cpuinitdata migration_notifier = {
5948 .notifier_call = migration_call,
5952 static int __init migration_init(void)
5954 void *cpu = (void *)(long)smp_processor_id();
5957 /* Start one for the boot CPU: */
5958 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5959 BUG_ON(err == NOTIFY_BAD);
5960 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5961 register_cpu_notifier(&migration_notifier);
5965 early_initcall(migration_init);
5970 #ifdef CONFIG_SCHED_DEBUG
5972 static __read_mostly int sched_domain_debug_enabled;
5974 static int __init sched_domain_debug_setup(char *str)
5976 sched_domain_debug_enabled = 1;
5980 early_param("sched_debug", sched_domain_debug_setup);
5982 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5983 struct cpumask *groupmask)
5985 struct sched_group *group = sd->groups;
5988 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5989 cpumask_clear(groupmask);
5991 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5993 if (!(sd->flags & SD_LOAD_BALANCE)) {
5994 printk("does not load-balance\n");
5996 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6001 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6003 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6004 printk(KERN_ERR "ERROR: domain->span does not contain "
6007 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6008 printk(KERN_ERR "ERROR: domain->groups does not contain"
6012 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6016 printk(KERN_ERR "ERROR: group is NULL\n");
6020 if (!group->cpu_power) {
6021 printk(KERN_CONT "\n");
6022 printk(KERN_ERR "ERROR: domain->cpu_power not "
6027 if (!cpumask_weight(sched_group_cpus(group))) {
6028 printk(KERN_CONT "\n");
6029 printk(KERN_ERR "ERROR: empty group\n");
6033 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6034 printk(KERN_CONT "\n");
6035 printk(KERN_ERR "ERROR: repeated CPUs\n");
6039 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6041 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6043 printk(KERN_CONT " %s", str);
6044 if (group->cpu_power != SCHED_LOAD_SCALE) {
6045 printk(KERN_CONT " (cpu_power = %d)",
6049 group = group->next;
6050 } while (group != sd->groups);
6051 printk(KERN_CONT "\n");
6053 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6054 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6057 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6058 printk(KERN_ERR "ERROR: parent span is not a superset "
6059 "of domain->span\n");
6063 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6065 cpumask_var_t groupmask;
6068 if (!sched_domain_debug_enabled)
6072 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6076 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6078 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6079 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6084 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6091 free_cpumask_var(groupmask);
6093 #else /* !CONFIG_SCHED_DEBUG */
6094 # define sched_domain_debug(sd, cpu) do { } while (0)
6095 #endif /* CONFIG_SCHED_DEBUG */
6097 static int sd_degenerate(struct sched_domain *sd)
6099 if (cpumask_weight(sched_domain_span(sd)) == 1)
6102 /* Following flags need at least 2 groups */
6103 if (sd->flags & (SD_LOAD_BALANCE |
6104 SD_BALANCE_NEWIDLE |
6108 SD_SHARE_PKG_RESOURCES)) {
6109 if (sd->groups != sd->groups->next)
6113 /* Following flags don't use groups */
6114 if (sd->flags & (SD_WAKE_AFFINE))
6121 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6123 unsigned long cflags = sd->flags, pflags = parent->flags;
6125 if (sd_degenerate(parent))
6128 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6131 /* Flags needing groups don't count if only 1 group in parent */
6132 if (parent->groups == parent->groups->next) {
6133 pflags &= ~(SD_LOAD_BALANCE |
6134 SD_BALANCE_NEWIDLE |
6138 SD_SHARE_PKG_RESOURCES);
6139 if (nr_node_ids == 1)
6140 pflags &= ~SD_SERIALIZE;
6142 if (~cflags & pflags)
6148 static void free_rootdomain(struct root_domain *rd)
6150 synchronize_sched();
6152 cpupri_cleanup(&rd->cpupri);
6154 free_cpumask_var(rd->rto_mask);
6155 free_cpumask_var(rd->online);
6156 free_cpumask_var(rd->span);
6160 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6162 struct root_domain *old_rd = NULL;
6163 unsigned long flags;
6165 raw_spin_lock_irqsave(&rq->lock, flags);
6170 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6173 cpumask_clear_cpu(rq->cpu, old_rd->span);
6176 * If we dont want to free the old_rt yet then
6177 * set old_rd to NULL to skip the freeing later
6180 if (!atomic_dec_and_test(&old_rd->refcount))
6184 atomic_inc(&rd->refcount);
6187 cpumask_set_cpu(rq->cpu, rd->span);
6188 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6191 raw_spin_unlock_irqrestore(&rq->lock, flags);
6194 free_rootdomain(old_rd);
6197 static int init_rootdomain(struct root_domain *rd, bool bootmem)
6199 gfp_t gfp = GFP_KERNEL;
6201 memset(rd, 0, sizeof(*rd));
6206 if (!alloc_cpumask_var(&rd->span, gfp))
6208 if (!alloc_cpumask_var(&rd->online, gfp))
6210 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
6213 if (cpupri_init(&rd->cpupri, bootmem) != 0)
6218 free_cpumask_var(rd->rto_mask);
6220 free_cpumask_var(rd->online);
6222 free_cpumask_var(rd->span);
6227 static void init_defrootdomain(void)
6229 init_rootdomain(&def_root_domain, true);
6231 atomic_set(&def_root_domain.refcount, 1);
6234 static struct root_domain *alloc_rootdomain(void)
6236 struct root_domain *rd;
6238 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6242 if (init_rootdomain(rd, false) != 0) {
6251 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6252 * hold the hotplug lock.
6255 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6257 struct rq *rq = cpu_rq(cpu);
6258 struct sched_domain *tmp;
6260 /* Remove the sched domains which do not contribute to scheduling. */
6261 for (tmp = sd; tmp; ) {
6262 struct sched_domain *parent = tmp->parent;
6266 if (sd_parent_degenerate(tmp, parent)) {
6267 tmp->parent = parent->parent;
6269 parent->parent->child = tmp;
6274 if (sd && sd_degenerate(sd)) {
6280 sched_domain_debug(sd, cpu);
6282 rq_attach_root(rq, rd);
6283 rcu_assign_pointer(rq->sd, sd);
6286 /* cpus with isolated domains */
6287 static cpumask_var_t cpu_isolated_map;
6289 /* Setup the mask of cpus configured for isolated domains */
6290 static int __init isolated_cpu_setup(char *str)
6292 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6293 cpulist_parse(str, cpu_isolated_map);
6297 __setup("isolcpus=", isolated_cpu_setup);
6300 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6301 * to a function which identifies what group(along with sched group) a CPU
6302 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6303 * (due to the fact that we keep track of groups covered with a struct cpumask).
6305 * init_sched_build_groups will build a circular linked list of the groups
6306 * covered by the given span, and will set each group's ->cpumask correctly,
6307 * and ->cpu_power to 0.
6310 init_sched_build_groups(const struct cpumask *span,
6311 const struct cpumask *cpu_map,
6312 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
6313 struct sched_group **sg,
6314 struct cpumask *tmpmask),
6315 struct cpumask *covered, struct cpumask *tmpmask)
6317 struct sched_group *first = NULL, *last = NULL;
6320 cpumask_clear(covered);
6322 for_each_cpu(i, span) {
6323 struct sched_group *sg;
6324 int group = group_fn(i, cpu_map, &sg, tmpmask);
6327 if (cpumask_test_cpu(i, covered))
6330 cpumask_clear(sched_group_cpus(sg));
6333 for_each_cpu(j, span) {
6334 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6337 cpumask_set_cpu(j, covered);
6338 cpumask_set_cpu(j, sched_group_cpus(sg));
6349 #define SD_NODES_PER_DOMAIN 16
6354 * find_next_best_node - find the next node to include in a sched_domain
6355 * @node: node whose sched_domain we're building
6356 * @used_nodes: nodes already in the sched_domain
6358 * Find the next node to include in a given scheduling domain. Simply
6359 * finds the closest node not already in the @used_nodes map.
6361 * Should use nodemask_t.
6363 static int find_next_best_node(int node, nodemask_t *used_nodes)
6365 int i, n, val, min_val, best_node = 0;
6369 for (i = 0; i < nr_node_ids; i++) {
6370 /* Start at @node */
6371 n = (node + i) % nr_node_ids;
6373 if (!nr_cpus_node(n))
6376 /* Skip already used nodes */
6377 if (node_isset(n, *used_nodes))
6380 /* Simple min distance search */
6381 val = node_distance(node, n);
6383 if (val < min_val) {
6389 node_set(best_node, *used_nodes);
6394 * sched_domain_node_span - get a cpumask for a node's sched_domain
6395 * @node: node whose cpumask we're constructing
6396 * @span: resulting cpumask
6398 * Given a node, construct a good cpumask for its sched_domain to span. It
6399 * should be one that prevents unnecessary balancing, but also spreads tasks
6402 static void sched_domain_node_span(int node, struct cpumask *span)
6404 nodemask_t used_nodes;
6407 cpumask_clear(span);
6408 nodes_clear(used_nodes);
6410 cpumask_or(span, span, cpumask_of_node(node));
6411 node_set(node, used_nodes);
6413 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6414 int next_node = find_next_best_node(node, &used_nodes);
6416 cpumask_or(span, span, cpumask_of_node(next_node));
6419 #endif /* CONFIG_NUMA */
6421 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6424 * The cpus mask in sched_group and sched_domain hangs off the end.
6426 * ( See the the comments in include/linux/sched.h:struct sched_group
6427 * and struct sched_domain. )
6429 struct static_sched_group {
6430 struct sched_group sg;
6431 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
6434 struct static_sched_domain {
6435 struct sched_domain sd;
6436 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
6442 cpumask_var_t domainspan;
6443 cpumask_var_t covered;
6444 cpumask_var_t notcovered;
6446 cpumask_var_t nodemask;
6447 cpumask_var_t this_sibling_map;
6448 cpumask_var_t this_core_map;
6449 cpumask_var_t send_covered;
6450 cpumask_var_t tmpmask;
6451 struct sched_group **sched_group_nodes;
6452 struct root_domain *rd;
6456 sa_sched_groups = 0,
6461 sa_this_sibling_map,
6463 sa_sched_group_nodes,
6473 * SMT sched-domains:
6475 #ifdef CONFIG_SCHED_SMT
6476 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
6477 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
6480 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
6481 struct sched_group **sg, struct cpumask *unused)
6484 *sg = &per_cpu(sched_groups, cpu).sg;
6487 #endif /* CONFIG_SCHED_SMT */
6490 * multi-core sched-domains:
6492 #ifdef CONFIG_SCHED_MC
6493 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
6494 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
6495 #endif /* CONFIG_SCHED_MC */
6497 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6499 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6500 struct sched_group **sg, struct cpumask *mask)
6504 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6505 group = cpumask_first(mask);
6507 *sg = &per_cpu(sched_group_core, group).sg;
6510 #elif defined(CONFIG_SCHED_MC)
6512 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6513 struct sched_group **sg, struct cpumask *unused)
6516 *sg = &per_cpu(sched_group_core, cpu).sg;
6521 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
6522 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
6525 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
6526 struct sched_group **sg, struct cpumask *mask)
6529 #ifdef CONFIG_SCHED_MC
6530 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6531 group = cpumask_first(mask);
6532 #elif defined(CONFIG_SCHED_SMT)
6533 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6534 group = cpumask_first(mask);
6539 *sg = &per_cpu(sched_group_phys, group).sg;
6545 * The init_sched_build_groups can't handle what we want to do with node
6546 * groups, so roll our own. Now each node has its own list of groups which
6547 * gets dynamically allocated.
6549 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
6550 static struct sched_group ***sched_group_nodes_bycpu;
6552 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
6553 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
6555 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
6556 struct sched_group **sg,
6557 struct cpumask *nodemask)
6561 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
6562 group = cpumask_first(nodemask);
6565 *sg = &per_cpu(sched_group_allnodes, group).sg;
6569 static void init_numa_sched_groups_power(struct sched_group *group_head)
6571 struct sched_group *sg = group_head;
6577 for_each_cpu(j, sched_group_cpus(sg)) {
6578 struct sched_domain *sd;
6580 sd = &per_cpu(phys_domains, j).sd;
6581 if (j != group_first_cpu(sd->groups)) {
6583 * Only add "power" once for each
6589 sg->cpu_power += sd->groups->cpu_power;
6592 } while (sg != group_head);
6595 static int build_numa_sched_groups(struct s_data *d,
6596 const struct cpumask *cpu_map, int num)
6598 struct sched_domain *sd;
6599 struct sched_group *sg, *prev;
6602 cpumask_clear(d->covered);
6603 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
6604 if (cpumask_empty(d->nodemask)) {
6605 d->sched_group_nodes[num] = NULL;
6609 sched_domain_node_span(num, d->domainspan);
6610 cpumask_and(d->domainspan, d->domainspan, cpu_map);
6612 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6615 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
6619 d->sched_group_nodes[num] = sg;
6621 for_each_cpu(j, d->nodemask) {
6622 sd = &per_cpu(node_domains, j).sd;
6627 cpumask_copy(sched_group_cpus(sg), d->nodemask);
6629 cpumask_or(d->covered, d->covered, d->nodemask);
6632 for (j = 0; j < nr_node_ids; j++) {
6633 n = (num + j) % nr_node_ids;
6634 cpumask_complement(d->notcovered, d->covered);
6635 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
6636 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
6637 if (cpumask_empty(d->tmpmask))
6639 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
6640 if (cpumask_empty(d->tmpmask))
6642 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6646 "Can not alloc domain group for node %d\n", j);
6650 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
6651 sg->next = prev->next;
6652 cpumask_or(d->covered, d->covered, d->tmpmask);
6659 #endif /* CONFIG_NUMA */
6662 /* Free memory allocated for various sched_group structures */
6663 static void free_sched_groups(const struct cpumask *cpu_map,
6664 struct cpumask *nodemask)
6668 for_each_cpu(cpu, cpu_map) {
6669 struct sched_group **sched_group_nodes
6670 = sched_group_nodes_bycpu[cpu];
6672 if (!sched_group_nodes)
6675 for (i = 0; i < nr_node_ids; i++) {
6676 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6678 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
6679 if (cpumask_empty(nodemask))
6689 if (oldsg != sched_group_nodes[i])
6692 kfree(sched_group_nodes);
6693 sched_group_nodes_bycpu[cpu] = NULL;
6696 #else /* !CONFIG_NUMA */
6697 static void free_sched_groups(const struct cpumask *cpu_map,
6698 struct cpumask *nodemask)
6701 #endif /* CONFIG_NUMA */
6704 * Initialize sched groups cpu_power.
6706 * cpu_power indicates the capacity of sched group, which is used while
6707 * distributing the load between different sched groups in a sched domain.
6708 * Typically cpu_power for all the groups in a sched domain will be same unless
6709 * there are asymmetries in the topology. If there are asymmetries, group
6710 * having more cpu_power will pickup more load compared to the group having
6713 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6715 struct sched_domain *child;
6716 struct sched_group *group;
6720 WARN_ON(!sd || !sd->groups);
6722 if (cpu != group_first_cpu(sd->groups))
6727 sd->groups->cpu_power = 0;
6730 power = SCHED_LOAD_SCALE;
6731 weight = cpumask_weight(sched_domain_span(sd));
6733 * SMT siblings share the power of a single core.
6734 * Usually multiple threads get a better yield out of
6735 * that one core than a single thread would have,
6736 * reflect that in sd->smt_gain.
6738 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
6739 power *= sd->smt_gain;
6741 power >>= SCHED_LOAD_SHIFT;
6743 sd->groups->cpu_power += power;
6748 * Add cpu_power of each child group to this groups cpu_power.
6750 group = child->groups;
6752 sd->groups->cpu_power += group->cpu_power;
6753 group = group->next;
6754 } while (group != child->groups);
6758 * Initializers for schedule domains
6759 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6762 #ifdef CONFIG_SCHED_DEBUG
6763 # define SD_INIT_NAME(sd, type) sd->name = #type
6765 # define SD_INIT_NAME(sd, type) do { } while (0)
6768 #define SD_INIT(sd, type) sd_init_##type(sd)
6770 #define SD_INIT_FUNC(type) \
6771 static noinline void sd_init_##type(struct sched_domain *sd) \
6773 memset(sd, 0, sizeof(*sd)); \
6774 *sd = SD_##type##_INIT; \
6775 sd->level = SD_LV_##type; \
6776 SD_INIT_NAME(sd, type); \
6781 SD_INIT_FUNC(ALLNODES)
6784 #ifdef CONFIG_SCHED_SMT
6785 SD_INIT_FUNC(SIBLING)
6787 #ifdef CONFIG_SCHED_MC
6791 static int default_relax_domain_level = -1;
6793 static int __init setup_relax_domain_level(char *str)
6797 val = simple_strtoul(str, NULL, 0);
6798 if (val < SD_LV_MAX)
6799 default_relax_domain_level = val;
6803 __setup("relax_domain_level=", setup_relax_domain_level);
6805 static void set_domain_attribute(struct sched_domain *sd,
6806 struct sched_domain_attr *attr)
6810 if (!attr || attr->relax_domain_level < 0) {
6811 if (default_relax_domain_level < 0)
6814 request = default_relax_domain_level;
6816 request = attr->relax_domain_level;
6817 if (request < sd->level) {
6818 /* turn off idle balance on this domain */
6819 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6821 /* turn on idle balance on this domain */
6822 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6826 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6827 const struct cpumask *cpu_map)
6830 case sa_sched_groups:
6831 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
6832 d->sched_group_nodes = NULL;
6834 free_rootdomain(d->rd); /* fall through */
6836 free_cpumask_var(d->tmpmask); /* fall through */
6837 case sa_send_covered:
6838 free_cpumask_var(d->send_covered); /* fall through */
6839 case sa_this_core_map:
6840 free_cpumask_var(d->this_core_map); /* fall through */
6841 case sa_this_sibling_map:
6842 free_cpumask_var(d->this_sibling_map); /* fall through */
6844 free_cpumask_var(d->nodemask); /* fall through */
6845 case sa_sched_group_nodes:
6847 kfree(d->sched_group_nodes); /* fall through */
6849 free_cpumask_var(d->notcovered); /* fall through */
6851 free_cpumask_var(d->covered); /* fall through */
6853 free_cpumask_var(d->domainspan); /* fall through */
6860 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6861 const struct cpumask *cpu_map)
6864 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
6866 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
6867 return sa_domainspan;
6868 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
6870 /* Allocate the per-node list of sched groups */
6871 d->sched_group_nodes = kcalloc(nr_node_ids,
6872 sizeof(struct sched_group *), GFP_KERNEL);
6873 if (!d->sched_group_nodes) {
6874 printk(KERN_WARNING "Can not alloc sched group node list\n");
6875 return sa_notcovered;
6877 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
6879 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
6880 return sa_sched_group_nodes;
6881 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
6883 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
6884 return sa_this_sibling_map;
6885 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
6886 return sa_this_core_map;
6887 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
6888 return sa_send_covered;
6889 d->rd = alloc_rootdomain();
6891 printk(KERN_WARNING "Cannot alloc root domain\n");
6894 return sa_rootdomain;
6897 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
6898 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
6900 struct sched_domain *sd = NULL;
6902 struct sched_domain *parent;
6905 if (cpumask_weight(cpu_map) >
6906 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
6907 sd = &per_cpu(allnodes_domains, i).sd;
6908 SD_INIT(sd, ALLNODES);
6909 set_domain_attribute(sd, attr);
6910 cpumask_copy(sched_domain_span(sd), cpu_map);
6911 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
6916 sd = &per_cpu(node_domains, i).sd;
6918 set_domain_attribute(sd, attr);
6919 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
6920 sd->parent = parent;
6923 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
6928 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
6929 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6930 struct sched_domain *parent, int i)
6932 struct sched_domain *sd;
6933 sd = &per_cpu(phys_domains, i).sd;
6935 set_domain_attribute(sd, attr);
6936 cpumask_copy(sched_domain_span(sd), d->nodemask);
6937 sd->parent = parent;
6940 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
6944 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
6945 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6946 struct sched_domain *parent, int i)
6948 struct sched_domain *sd = parent;
6949 #ifdef CONFIG_SCHED_MC
6950 sd = &per_cpu(core_domains, i).sd;
6952 set_domain_attribute(sd, attr);
6953 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
6954 sd->parent = parent;
6956 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
6961 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
6962 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6963 struct sched_domain *parent, int i)
6965 struct sched_domain *sd = parent;
6966 #ifdef CONFIG_SCHED_SMT
6967 sd = &per_cpu(cpu_domains, i).sd;
6968 SD_INIT(sd, SIBLING);
6969 set_domain_attribute(sd, attr);
6970 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
6971 sd->parent = parent;
6973 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
6978 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
6979 const struct cpumask *cpu_map, int cpu)
6982 #ifdef CONFIG_SCHED_SMT
6983 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
6984 cpumask_and(d->this_sibling_map, cpu_map,
6985 topology_thread_cpumask(cpu));
6986 if (cpu == cpumask_first(d->this_sibling_map))
6987 init_sched_build_groups(d->this_sibling_map, cpu_map,
6989 d->send_covered, d->tmpmask);
6992 #ifdef CONFIG_SCHED_MC
6993 case SD_LV_MC: /* set up multi-core groups */
6994 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
6995 if (cpu == cpumask_first(d->this_core_map))
6996 init_sched_build_groups(d->this_core_map, cpu_map,
6998 d->send_covered, d->tmpmask);
7001 case SD_LV_CPU: /* set up physical groups */
7002 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
7003 if (!cpumask_empty(d->nodemask))
7004 init_sched_build_groups(d->nodemask, cpu_map,
7006 d->send_covered, d->tmpmask);
7009 case SD_LV_ALLNODES:
7010 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
7011 d->send_covered, d->tmpmask);
7020 * Build sched domains for a given set of cpus and attach the sched domains
7021 * to the individual cpus
7023 static int __build_sched_domains(const struct cpumask *cpu_map,
7024 struct sched_domain_attr *attr)
7026 enum s_alloc alloc_state = sa_none;
7028 struct sched_domain *sd;
7034 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7035 if (alloc_state != sa_rootdomain)
7037 alloc_state = sa_sched_groups;
7040 * Set up domains for cpus specified by the cpu_map.
7042 for_each_cpu(i, cpu_map) {
7043 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
7046 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
7047 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
7048 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
7049 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
7052 for_each_cpu(i, cpu_map) {
7053 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
7054 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
7057 /* Set up physical groups */
7058 for (i = 0; i < nr_node_ids; i++)
7059 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
7062 /* Set up node groups */
7064 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
7066 for (i = 0; i < nr_node_ids; i++)
7067 if (build_numa_sched_groups(&d, cpu_map, i))
7071 /* Calculate CPU power for physical packages and nodes */
7072 #ifdef CONFIG_SCHED_SMT
7073 for_each_cpu(i, cpu_map) {
7074 sd = &per_cpu(cpu_domains, i).sd;
7075 init_sched_groups_power(i, sd);
7078 #ifdef CONFIG_SCHED_MC
7079 for_each_cpu(i, cpu_map) {
7080 sd = &per_cpu(core_domains, i).sd;
7081 init_sched_groups_power(i, sd);
7085 for_each_cpu(i, cpu_map) {
7086 sd = &per_cpu(phys_domains, i).sd;
7087 init_sched_groups_power(i, sd);
7091 for (i = 0; i < nr_node_ids; i++)
7092 init_numa_sched_groups_power(d.sched_group_nodes[i]);
7094 if (d.sd_allnodes) {
7095 struct sched_group *sg;
7097 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7099 init_numa_sched_groups_power(sg);
7103 /* Attach the domains */
7104 for_each_cpu(i, cpu_map) {
7105 #ifdef CONFIG_SCHED_SMT
7106 sd = &per_cpu(cpu_domains, i).sd;
7107 #elif defined(CONFIG_SCHED_MC)
7108 sd = &per_cpu(core_domains, i).sd;
7110 sd = &per_cpu(phys_domains, i).sd;
7112 cpu_attach_domain(sd, d.rd, i);
7115 d.sched_group_nodes = NULL; /* don't free this we still need it */
7116 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
7120 __free_domain_allocs(&d, alloc_state, cpu_map);
7124 static int build_sched_domains(const struct cpumask *cpu_map)
7126 return __build_sched_domains(cpu_map, NULL);
7129 static cpumask_var_t *doms_cur; /* current sched domains */
7130 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7131 static struct sched_domain_attr *dattr_cur;
7132 /* attribues of custom domains in 'doms_cur' */
7135 * Special case: If a kmalloc of a doms_cur partition (array of
7136 * cpumask) fails, then fallback to a single sched domain,
7137 * as determined by the single cpumask fallback_doms.
7139 static cpumask_var_t fallback_doms;
7142 * arch_update_cpu_topology lets virtualized architectures update the
7143 * cpu core maps. It is supposed to return 1 if the topology changed
7144 * or 0 if it stayed the same.
7146 int __attribute__((weak)) arch_update_cpu_topology(void)
7151 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7154 cpumask_var_t *doms;
7156 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7159 for (i = 0; i < ndoms; i++) {
7160 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7161 free_sched_domains(doms, i);
7168 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7171 for (i = 0; i < ndoms; i++)
7172 free_cpumask_var(doms[i]);
7177 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7178 * For now this just excludes isolated cpus, but could be used to
7179 * exclude other special cases in the future.
7181 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7185 arch_update_cpu_topology();
7187 doms_cur = alloc_sched_domains(ndoms_cur);
7189 doms_cur = &fallback_doms;
7190 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7192 err = build_sched_domains(doms_cur[0]);
7193 register_sched_domain_sysctl();
7198 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7199 struct cpumask *tmpmask)
7201 free_sched_groups(cpu_map, tmpmask);
7205 * Detach sched domains from a group of cpus specified in cpu_map
7206 * These cpus will now be attached to the NULL domain
7208 static void detach_destroy_domains(const struct cpumask *cpu_map)
7210 /* Save because hotplug lock held. */
7211 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7214 for_each_cpu(i, cpu_map)
7215 cpu_attach_domain(NULL, &def_root_domain, i);
7216 synchronize_sched();
7217 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7220 /* handle null as "default" */
7221 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7222 struct sched_domain_attr *new, int idx_new)
7224 struct sched_domain_attr tmp;
7231 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7232 new ? (new + idx_new) : &tmp,
7233 sizeof(struct sched_domain_attr));
7237 * Partition sched domains as specified by the 'ndoms_new'
7238 * cpumasks in the array doms_new[] of cpumasks. This compares
7239 * doms_new[] to the current sched domain partitioning, doms_cur[].
7240 * It destroys each deleted domain and builds each new domain.
7242 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7243 * The masks don't intersect (don't overlap.) We should setup one
7244 * sched domain for each mask. CPUs not in any of the cpumasks will
7245 * not be load balanced. If the same cpumask appears both in the
7246 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7249 * The passed in 'doms_new' should be allocated using
7250 * alloc_sched_domains. This routine takes ownership of it and will
7251 * free_sched_domains it when done with it. If the caller failed the
7252 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7253 * and partition_sched_domains() will fallback to the single partition
7254 * 'fallback_doms', it also forces the domains to be rebuilt.
7256 * If doms_new == NULL it will be replaced with cpu_online_mask.
7257 * ndoms_new == 0 is a special case for destroying existing domains,
7258 * and it will not create the default domain.
7260 * Call with hotplug lock held
7262 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7263 struct sched_domain_attr *dattr_new)
7268 mutex_lock(&sched_domains_mutex);
7270 /* always unregister in case we don't destroy any domains */
7271 unregister_sched_domain_sysctl();
7273 /* Let architecture update cpu core mappings. */
7274 new_topology = arch_update_cpu_topology();
7276 n = doms_new ? ndoms_new : 0;
7278 /* Destroy deleted domains */
7279 for (i = 0; i < ndoms_cur; i++) {
7280 for (j = 0; j < n && !new_topology; j++) {
7281 if (cpumask_equal(doms_cur[i], doms_new[j])
7282 && dattrs_equal(dattr_cur, i, dattr_new, j))
7285 /* no match - a current sched domain not in new doms_new[] */
7286 detach_destroy_domains(doms_cur[i]);
7291 if (doms_new == NULL) {
7293 doms_new = &fallback_doms;
7294 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7295 WARN_ON_ONCE(dattr_new);
7298 /* Build new domains */
7299 for (i = 0; i < ndoms_new; i++) {
7300 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7301 if (cpumask_equal(doms_new[i], doms_cur[j])
7302 && dattrs_equal(dattr_new, i, dattr_cur, j))
7305 /* no match - add a new doms_new */
7306 __build_sched_domains(doms_new[i],
7307 dattr_new ? dattr_new + i : NULL);
7312 /* Remember the new sched domains */
7313 if (doms_cur != &fallback_doms)
7314 free_sched_domains(doms_cur, ndoms_cur);
7315 kfree(dattr_cur); /* kfree(NULL) is safe */
7316 doms_cur = doms_new;
7317 dattr_cur = dattr_new;
7318 ndoms_cur = ndoms_new;
7320 register_sched_domain_sysctl();
7322 mutex_unlock(&sched_domains_mutex);
7325 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7326 static void arch_reinit_sched_domains(void)
7330 /* Destroy domains first to force the rebuild */
7331 partition_sched_domains(0, NULL, NULL);
7333 rebuild_sched_domains();
7337 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7339 unsigned int level = 0;
7341 if (sscanf(buf, "%u", &level) != 1)
7345 * level is always be positive so don't check for
7346 * level < POWERSAVINGS_BALANCE_NONE which is 0
7347 * What happens on 0 or 1 byte write,
7348 * need to check for count as well?
7351 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7355 sched_smt_power_savings = level;
7357 sched_mc_power_savings = level;
7359 arch_reinit_sched_domains();
7364 #ifdef CONFIG_SCHED_MC
7365 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7366 struct sysdev_class_attribute *attr,
7369 return sprintf(page, "%u\n", sched_mc_power_savings);
7371 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7372 struct sysdev_class_attribute *attr,
7373 const char *buf, size_t count)
7375 return sched_power_savings_store(buf, count, 0);
7377 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7378 sched_mc_power_savings_show,
7379 sched_mc_power_savings_store);
7382 #ifdef CONFIG_SCHED_SMT
7383 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7384 struct sysdev_class_attribute *attr,
7387 return sprintf(page, "%u\n", sched_smt_power_savings);
7389 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7390 struct sysdev_class_attribute *attr,
7391 const char *buf, size_t count)
7393 return sched_power_savings_store(buf, count, 1);
7395 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7396 sched_smt_power_savings_show,
7397 sched_smt_power_savings_store);
7400 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7404 #ifdef CONFIG_SCHED_SMT
7406 err = sysfs_create_file(&cls->kset.kobj,
7407 &attr_sched_smt_power_savings.attr);
7409 #ifdef CONFIG_SCHED_MC
7410 if (!err && mc_capable())
7411 err = sysfs_create_file(&cls->kset.kobj,
7412 &attr_sched_mc_power_savings.attr);
7416 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7418 #ifndef CONFIG_CPUSETS
7420 * Add online and remove offline CPUs from the scheduler domains.
7421 * When cpusets are enabled they take over this function.
7423 static int update_sched_domains(struct notifier_block *nfb,
7424 unsigned long action, void *hcpu)
7428 case CPU_ONLINE_FROZEN:
7429 case CPU_DOWN_PREPARE:
7430 case CPU_DOWN_PREPARE_FROZEN:
7431 case CPU_DOWN_FAILED:
7432 case CPU_DOWN_FAILED_FROZEN:
7433 partition_sched_domains(1, NULL, NULL);
7442 static int update_runtime(struct notifier_block *nfb,
7443 unsigned long action, void *hcpu)
7445 int cpu = (int)(long)hcpu;
7448 case CPU_DOWN_PREPARE:
7449 case CPU_DOWN_PREPARE_FROZEN:
7450 disable_runtime(cpu_rq(cpu));
7453 case CPU_DOWN_FAILED:
7454 case CPU_DOWN_FAILED_FROZEN:
7456 case CPU_ONLINE_FROZEN:
7457 enable_runtime(cpu_rq(cpu));
7465 void __init sched_init_smp(void)
7467 cpumask_var_t non_isolated_cpus;
7469 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7470 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7472 #if defined(CONFIG_NUMA)
7473 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7475 BUG_ON(sched_group_nodes_bycpu == NULL);
7478 mutex_lock(&sched_domains_mutex);
7479 arch_init_sched_domains(cpu_active_mask);
7480 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7481 if (cpumask_empty(non_isolated_cpus))
7482 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7483 mutex_unlock(&sched_domains_mutex);
7486 #ifndef CONFIG_CPUSETS
7487 /* XXX: Theoretical race here - CPU may be hotplugged now */
7488 hotcpu_notifier(update_sched_domains, 0);
7491 /* RT runtime code needs to handle some hotplug events */
7492 hotcpu_notifier(update_runtime, 0);
7496 /* Move init over to a non-isolated CPU */
7497 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7499 sched_init_granularity();
7500 free_cpumask_var(non_isolated_cpus);
7502 init_sched_rt_class();
7505 void __init sched_init_smp(void)
7507 sched_init_granularity();
7509 #endif /* CONFIG_SMP */
7511 const_debug unsigned int sysctl_timer_migration = 1;
7513 int in_sched_functions(unsigned long addr)
7515 return in_lock_functions(addr) ||
7516 (addr >= (unsigned long)__sched_text_start
7517 && addr < (unsigned long)__sched_text_end);
7520 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7522 cfs_rq->tasks_timeline = RB_ROOT;
7523 INIT_LIST_HEAD(&cfs_rq->tasks);
7524 #ifdef CONFIG_FAIR_GROUP_SCHED
7527 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7530 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7532 struct rt_prio_array *array;
7535 array = &rt_rq->active;
7536 for (i = 0; i < MAX_RT_PRIO; i++) {
7537 INIT_LIST_HEAD(array->queue + i);
7538 __clear_bit(i, array->bitmap);
7540 /* delimiter for bitsearch: */
7541 __set_bit(MAX_RT_PRIO, array->bitmap);
7543 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7544 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7546 rt_rq->highest_prio.next = MAX_RT_PRIO;
7550 rt_rq->rt_nr_migratory = 0;
7551 rt_rq->overloaded = 0;
7552 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
7556 rt_rq->rt_throttled = 0;
7557 rt_rq->rt_runtime = 0;
7558 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7560 #ifdef CONFIG_RT_GROUP_SCHED
7561 rt_rq->rt_nr_boosted = 0;
7566 #ifdef CONFIG_FAIR_GROUP_SCHED
7567 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7568 struct sched_entity *se, int cpu, int add,
7569 struct sched_entity *parent)
7571 struct rq *rq = cpu_rq(cpu);
7572 tg->cfs_rq[cpu] = cfs_rq;
7573 init_cfs_rq(cfs_rq, rq);
7576 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7579 /* se could be NULL for init_task_group */
7584 se->cfs_rq = &rq->cfs;
7586 se->cfs_rq = parent->my_q;
7589 se->load.weight = tg->shares;
7590 se->load.inv_weight = 0;
7591 se->parent = parent;
7595 #ifdef CONFIG_RT_GROUP_SCHED
7596 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7597 struct sched_rt_entity *rt_se, int cpu, int add,
7598 struct sched_rt_entity *parent)
7600 struct rq *rq = cpu_rq(cpu);
7602 tg->rt_rq[cpu] = rt_rq;
7603 init_rt_rq(rt_rq, rq);
7605 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7607 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7609 tg->rt_se[cpu] = rt_se;
7614 rt_se->rt_rq = &rq->rt;
7616 rt_se->rt_rq = parent->my_q;
7618 rt_se->my_q = rt_rq;
7619 rt_se->parent = parent;
7620 INIT_LIST_HEAD(&rt_se->run_list);
7624 void __init sched_init(void)
7627 unsigned long alloc_size = 0, ptr;
7629 #ifdef CONFIG_FAIR_GROUP_SCHED
7630 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7632 #ifdef CONFIG_RT_GROUP_SCHED
7633 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7635 #ifdef CONFIG_CPUMASK_OFFSTACK
7636 alloc_size += num_possible_cpus() * cpumask_size();
7639 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7641 #ifdef CONFIG_FAIR_GROUP_SCHED
7642 init_task_group.se = (struct sched_entity **)ptr;
7643 ptr += nr_cpu_ids * sizeof(void **);
7645 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7646 ptr += nr_cpu_ids * sizeof(void **);
7648 #endif /* CONFIG_FAIR_GROUP_SCHED */
7649 #ifdef CONFIG_RT_GROUP_SCHED
7650 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7651 ptr += nr_cpu_ids * sizeof(void **);
7653 init_task_group.rt_rq = (struct rt_rq **)ptr;
7654 ptr += nr_cpu_ids * sizeof(void **);
7656 #endif /* CONFIG_RT_GROUP_SCHED */
7657 #ifdef CONFIG_CPUMASK_OFFSTACK
7658 for_each_possible_cpu(i) {
7659 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
7660 ptr += cpumask_size();
7662 #endif /* CONFIG_CPUMASK_OFFSTACK */
7666 init_defrootdomain();
7669 init_rt_bandwidth(&def_rt_bandwidth,
7670 global_rt_period(), global_rt_runtime());
7672 #ifdef CONFIG_RT_GROUP_SCHED
7673 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7674 global_rt_period(), global_rt_runtime());
7675 #endif /* CONFIG_RT_GROUP_SCHED */
7677 #ifdef CONFIG_CGROUP_SCHED
7678 list_add(&init_task_group.list, &task_groups);
7679 INIT_LIST_HEAD(&init_task_group.children);
7681 #endif /* CONFIG_CGROUP_SCHED */
7683 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
7684 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
7685 __alignof__(unsigned long));
7687 for_each_possible_cpu(i) {
7691 raw_spin_lock_init(&rq->lock);
7693 rq->calc_load_active = 0;
7694 rq->calc_load_update = jiffies + LOAD_FREQ;
7695 init_cfs_rq(&rq->cfs, rq);
7696 init_rt_rq(&rq->rt, rq);
7697 #ifdef CONFIG_FAIR_GROUP_SCHED
7698 init_task_group.shares = init_task_group_load;
7699 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7700 #ifdef CONFIG_CGROUP_SCHED
7702 * How much cpu bandwidth does init_task_group get?
7704 * In case of task-groups formed thr' the cgroup filesystem, it
7705 * gets 100% of the cpu resources in the system. This overall
7706 * system cpu resource is divided among the tasks of
7707 * init_task_group and its child task-groups in a fair manner,
7708 * based on each entity's (task or task-group's) weight
7709 * (se->load.weight).
7711 * In other words, if init_task_group has 10 tasks of weight
7712 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7713 * then A0's share of the cpu resource is:
7715 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7717 * We achieve this by letting init_task_group's tasks sit
7718 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7720 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
7722 #endif /* CONFIG_FAIR_GROUP_SCHED */
7724 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7725 #ifdef CONFIG_RT_GROUP_SCHED
7726 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7727 #ifdef CONFIG_CGROUP_SCHED
7728 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
7732 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7733 rq->cpu_load[j] = 0;
7737 rq->post_schedule = 0;
7738 rq->active_balance = 0;
7739 rq->next_balance = jiffies;
7743 rq->migration_thread = NULL;
7745 rq->avg_idle = 2*sysctl_sched_migration_cost;
7746 INIT_LIST_HEAD(&rq->migration_queue);
7747 rq_attach_root(rq, &def_root_domain);
7750 atomic_set(&rq->nr_iowait, 0);
7753 set_load_weight(&init_task);
7755 #ifdef CONFIG_PREEMPT_NOTIFIERS
7756 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7760 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7763 #ifdef CONFIG_RT_MUTEXES
7764 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
7768 * The boot idle thread does lazy MMU switching as well:
7770 atomic_inc(&init_mm.mm_count);
7771 enter_lazy_tlb(&init_mm, current);
7774 * Make us the idle thread. Technically, schedule() should not be
7775 * called from this thread, however somewhere below it might be,
7776 * but because we are the idle thread, we just pick up running again
7777 * when this runqueue becomes "idle".
7779 init_idle(current, smp_processor_id());
7781 calc_load_update = jiffies + LOAD_FREQ;
7784 * During early bootup we pretend to be a normal task:
7786 current->sched_class = &fair_sched_class;
7788 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
7789 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
7792 zalloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
7793 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
7795 /* May be allocated at isolcpus cmdline parse time */
7796 if (cpu_isolated_map == NULL)
7797 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7802 scheduler_running = 1;
7805 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7806 static inline int preempt_count_equals(int preempt_offset)
7808 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7810 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
7813 void __might_sleep(const char *file, int line, int preempt_offset)
7816 static unsigned long prev_jiffy; /* ratelimiting */
7818 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7819 system_state != SYSTEM_RUNNING || oops_in_progress)
7821 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7823 prev_jiffy = jiffies;
7826 "BUG: sleeping function called from invalid context at %s:%d\n",
7829 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7830 in_atomic(), irqs_disabled(),
7831 current->pid, current->comm);
7833 debug_show_held_locks(current);
7834 if (irqs_disabled())
7835 print_irqtrace_events(current);
7839 EXPORT_SYMBOL(__might_sleep);
7842 #ifdef CONFIG_MAGIC_SYSRQ
7843 static void normalize_task(struct rq *rq, struct task_struct *p)
7847 update_rq_clock(rq);
7848 on_rq = p->se.on_rq;
7850 deactivate_task(rq, p, 0);
7851 __setscheduler(rq, p, SCHED_NORMAL, 0);
7853 activate_task(rq, p, 0);
7854 resched_task(rq->curr);
7858 void normalize_rt_tasks(void)
7860 struct task_struct *g, *p;
7861 unsigned long flags;
7864 read_lock_irqsave(&tasklist_lock, flags);
7865 do_each_thread(g, p) {
7867 * Only normalize user tasks:
7872 p->se.exec_start = 0;
7873 #ifdef CONFIG_SCHEDSTATS
7874 p->se.wait_start = 0;
7875 p->se.sleep_start = 0;
7876 p->se.block_start = 0;
7881 * Renice negative nice level userspace
7884 if (TASK_NICE(p) < 0 && p->mm)
7885 set_user_nice(p, 0);
7889 raw_spin_lock(&p->pi_lock);
7890 rq = __task_rq_lock(p);
7892 normalize_task(rq, p);
7894 __task_rq_unlock(rq);
7895 raw_spin_unlock(&p->pi_lock);
7896 } while_each_thread(g, p);
7898 read_unlock_irqrestore(&tasklist_lock, flags);
7901 #endif /* CONFIG_MAGIC_SYSRQ */
7905 * These functions are only useful for the IA64 MCA handling.
7907 * They can only be called when the whole system has been
7908 * stopped - every CPU needs to be quiescent, and no scheduling
7909 * activity can take place. Using them for anything else would
7910 * be a serious bug, and as a result, they aren't even visible
7911 * under any other configuration.
7915 * curr_task - return the current task for a given cpu.
7916 * @cpu: the processor in question.
7918 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7920 struct task_struct *curr_task(int cpu)
7922 return cpu_curr(cpu);
7926 * set_curr_task - set the current task for a given cpu.
7927 * @cpu: the processor in question.
7928 * @p: the task pointer to set.
7930 * Description: This function must only be used when non-maskable interrupts
7931 * are serviced on a separate stack. It allows the architecture to switch the
7932 * notion of the current task on a cpu in a non-blocking manner. This function
7933 * must be called with all CPU's synchronized, and interrupts disabled, the
7934 * and caller must save the original value of the current task (see
7935 * curr_task() above) and restore that value before reenabling interrupts and
7936 * re-starting the system.
7938 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7940 void set_curr_task(int cpu, struct task_struct *p)
7947 #ifdef CONFIG_FAIR_GROUP_SCHED
7948 static void free_fair_sched_group(struct task_group *tg)
7952 for_each_possible_cpu(i) {
7954 kfree(tg->cfs_rq[i]);
7964 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7966 struct cfs_rq *cfs_rq;
7967 struct sched_entity *se;
7971 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7974 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7978 tg->shares = NICE_0_LOAD;
7980 for_each_possible_cpu(i) {
7983 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
7984 GFP_KERNEL, cpu_to_node(i));
7988 se = kzalloc_node(sizeof(struct sched_entity),
7989 GFP_KERNEL, cpu_to_node(i));
7993 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
8004 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8006 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8007 &cpu_rq(cpu)->leaf_cfs_rq_list);
8010 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8012 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8014 #else /* !CONFG_FAIR_GROUP_SCHED */
8015 static inline void free_fair_sched_group(struct task_group *tg)
8020 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8025 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8029 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8032 #endif /* CONFIG_FAIR_GROUP_SCHED */
8034 #ifdef CONFIG_RT_GROUP_SCHED
8035 static void free_rt_sched_group(struct task_group *tg)
8039 destroy_rt_bandwidth(&tg->rt_bandwidth);
8041 for_each_possible_cpu(i) {
8043 kfree(tg->rt_rq[i]);
8045 kfree(tg->rt_se[i]);
8053 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8055 struct rt_rq *rt_rq;
8056 struct sched_rt_entity *rt_se;
8060 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8063 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8067 init_rt_bandwidth(&tg->rt_bandwidth,
8068 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8070 for_each_possible_cpu(i) {
8073 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8074 GFP_KERNEL, cpu_to_node(i));
8078 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8079 GFP_KERNEL, cpu_to_node(i));
8083 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
8094 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8096 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8097 &cpu_rq(cpu)->leaf_rt_rq_list);
8100 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8102 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8104 #else /* !CONFIG_RT_GROUP_SCHED */
8105 static inline void free_rt_sched_group(struct task_group *tg)
8110 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8115 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8119 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8122 #endif /* CONFIG_RT_GROUP_SCHED */
8124 #ifdef CONFIG_CGROUP_SCHED
8125 static void free_sched_group(struct task_group *tg)
8127 free_fair_sched_group(tg);
8128 free_rt_sched_group(tg);
8132 /* allocate runqueue etc for a new task group */
8133 struct task_group *sched_create_group(struct task_group *parent)
8135 struct task_group *tg;
8136 unsigned long flags;
8139 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8141 return ERR_PTR(-ENOMEM);
8143 if (!alloc_fair_sched_group(tg, parent))
8146 if (!alloc_rt_sched_group(tg, parent))
8149 spin_lock_irqsave(&task_group_lock, flags);
8150 for_each_possible_cpu(i) {
8151 register_fair_sched_group(tg, i);
8152 register_rt_sched_group(tg, i);
8154 list_add_rcu(&tg->list, &task_groups);
8156 WARN_ON(!parent); /* root should already exist */
8158 tg->parent = parent;
8159 INIT_LIST_HEAD(&tg->children);
8160 list_add_rcu(&tg->siblings, &parent->children);
8161 spin_unlock_irqrestore(&task_group_lock, flags);
8166 free_sched_group(tg);
8167 return ERR_PTR(-ENOMEM);
8170 /* rcu callback to free various structures associated with a task group */
8171 static void free_sched_group_rcu(struct rcu_head *rhp)
8173 /* now it should be safe to free those cfs_rqs */
8174 free_sched_group(container_of(rhp, struct task_group, rcu));
8177 /* Destroy runqueue etc associated with a task group */
8178 void sched_destroy_group(struct task_group *tg)
8180 unsigned long flags;
8183 spin_lock_irqsave(&task_group_lock, flags);
8184 for_each_possible_cpu(i) {
8185 unregister_fair_sched_group(tg, i);
8186 unregister_rt_sched_group(tg, i);
8188 list_del_rcu(&tg->list);
8189 list_del_rcu(&tg->siblings);
8190 spin_unlock_irqrestore(&task_group_lock, flags);
8192 /* wait for possible concurrent references to cfs_rqs complete */
8193 call_rcu(&tg->rcu, free_sched_group_rcu);
8196 /* change task's runqueue when it moves between groups.
8197 * The caller of this function should have put the task in its new group
8198 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8199 * reflect its new group.
8201 void sched_move_task(struct task_struct *tsk)
8204 unsigned long flags;
8207 rq = task_rq_lock(tsk, &flags);
8209 update_rq_clock(rq);
8211 running = task_current(rq, tsk);
8212 on_rq = tsk->se.on_rq;
8215 dequeue_task(rq, tsk, 0);
8216 if (unlikely(running))
8217 tsk->sched_class->put_prev_task(rq, tsk);
8219 set_task_rq(tsk, task_cpu(tsk));
8221 #ifdef CONFIG_FAIR_GROUP_SCHED
8222 if (tsk->sched_class->moved_group)
8223 tsk->sched_class->moved_group(tsk, on_rq);
8226 if (unlikely(running))
8227 tsk->sched_class->set_curr_task(rq);
8229 enqueue_task(rq, tsk, 0, false);
8231 task_rq_unlock(rq, &flags);
8233 #endif /* CONFIG_CGROUP_SCHED */
8235 #ifdef CONFIG_FAIR_GROUP_SCHED
8236 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8238 struct cfs_rq *cfs_rq = se->cfs_rq;
8243 dequeue_entity(cfs_rq, se, 0);
8245 se->load.weight = shares;
8246 se->load.inv_weight = 0;
8249 enqueue_entity(cfs_rq, se, 0);
8252 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8254 struct cfs_rq *cfs_rq = se->cfs_rq;
8255 struct rq *rq = cfs_rq->rq;
8256 unsigned long flags;
8258 raw_spin_lock_irqsave(&rq->lock, flags);
8259 __set_se_shares(se, shares);
8260 raw_spin_unlock_irqrestore(&rq->lock, flags);
8263 static DEFINE_MUTEX(shares_mutex);
8265 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8268 unsigned long flags;
8271 * We can't change the weight of the root cgroup.
8276 if (shares < MIN_SHARES)
8277 shares = MIN_SHARES;
8278 else if (shares > MAX_SHARES)
8279 shares = MAX_SHARES;
8281 mutex_lock(&shares_mutex);
8282 if (tg->shares == shares)
8285 spin_lock_irqsave(&task_group_lock, flags);
8286 for_each_possible_cpu(i)
8287 unregister_fair_sched_group(tg, i);
8288 list_del_rcu(&tg->siblings);
8289 spin_unlock_irqrestore(&task_group_lock, flags);
8291 /* wait for any ongoing reference to this group to finish */
8292 synchronize_sched();
8295 * Now we are free to modify the group's share on each cpu
8296 * w/o tripping rebalance_share or load_balance_fair.
8298 tg->shares = shares;
8299 for_each_possible_cpu(i) {
8303 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8304 set_se_shares(tg->se[i], shares);
8308 * Enable load balance activity on this group, by inserting it back on
8309 * each cpu's rq->leaf_cfs_rq_list.
8311 spin_lock_irqsave(&task_group_lock, flags);
8312 for_each_possible_cpu(i)
8313 register_fair_sched_group(tg, i);
8314 list_add_rcu(&tg->siblings, &tg->parent->children);
8315 spin_unlock_irqrestore(&task_group_lock, flags);
8317 mutex_unlock(&shares_mutex);
8321 unsigned long sched_group_shares(struct task_group *tg)
8327 #ifdef CONFIG_RT_GROUP_SCHED
8329 * Ensure that the real time constraints are schedulable.
8331 static DEFINE_MUTEX(rt_constraints_mutex);
8333 static unsigned long to_ratio(u64 period, u64 runtime)
8335 if (runtime == RUNTIME_INF)
8338 return div64_u64(runtime << 20, period);
8341 /* Must be called with tasklist_lock held */
8342 static inline int tg_has_rt_tasks(struct task_group *tg)
8344 struct task_struct *g, *p;
8346 do_each_thread(g, p) {
8347 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8349 } while_each_thread(g, p);
8354 struct rt_schedulable_data {
8355 struct task_group *tg;
8360 static int tg_schedulable(struct task_group *tg, void *data)
8362 struct rt_schedulable_data *d = data;
8363 struct task_group *child;
8364 unsigned long total, sum = 0;
8365 u64 period, runtime;
8367 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8368 runtime = tg->rt_bandwidth.rt_runtime;
8371 period = d->rt_period;
8372 runtime = d->rt_runtime;
8376 * Cannot have more runtime than the period.
8378 if (runtime > period && runtime != RUNTIME_INF)
8382 * Ensure we don't starve existing RT tasks.
8384 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8387 total = to_ratio(period, runtime);
8390 * Nobody can have more than the global setting allows.
8392 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8396 * The sum of our children's runtime should not exceed our own.
8398 list_for_each_entry_rcu(child, &tg->children, siblings) {
8399 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8400 runtime = child->rt_bandwidth.rt_runtime;
8402 if (child == d->tg) {
8403 period = d->rt_period;
8404 runtime = d->rt_runtime;
8407 sum += to_ratio(period, runtime);
8416 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8418 struct rt_schedulable_data data = {
8420 .rt_period = period,
8421 .rt_runtime = runtime,
8424 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8427 static int tg_set_bandwidth(struct task_group *tg,
8428 u64 rt_period, u64 rt_runtime)
8432 mutex_lock(&rt_constraints_mutex);
8433 read_lock(&tasklist_lock);
8434 err = __rt_schedulable(tg, rt_period, rt_runtime);
8438 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8439 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8440 tg->rt_bandwidth.rt_runtime = rt_runtime;
8442 for_each_possible_cpu(i) {
8443 struct rt_rq *rt_rq = tg->rt_rq[i];
8445 raw_spin_lock(&rt_rq->rt_runtime_lock);
8446 rt_rq->rt_runtime = rt_runtime;
8447 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8449 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8451 read_unlock(&tasklist_lock);
8452 mutex_unlock(&rt_constraints_mutex);
8457 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8459 u64 rt_runtime, rt_period;
8461 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8462 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8463 if (rt_runtime_us < 0)
8464 rt_runtime = RUNTIME_INF;
8466 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8469 long sched_group_rt_runtime(struct task_group *tg)
8473 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8476 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8477 do_div(rt_runtime_us, NSEC_PER_USEC);
8478 return rt_runtime_us;
8481 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8483 u64 rt_runtime, rt_period;
8485 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8486 rt_runtime = tg->rt_bandwidth.rt_runtime;
8491 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8494 long sched_group_rt_period(struct task_group *tg)
8498 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8499 do_div(rt_period_us, NSEC_PER_USEC);
8500 return rt_period_us;
8503 static int sched_rt_global_constraints(void)
8505 u64 runtime, period;
8508 if (sysctl_sched_rt_period <= 0)
8511 runtime = global_rt_runtime();
8512 period = global_rt_period();
8515 * Sanity check on the sysctl variables.
8517 if (runtime > period && runtime != RUNTIME_INF)
8520 mutex_lock(&rt_constraints_mutex);
8521 read_lock(&tasklist_lock);
8522 ret = __rt_schedulable(NULL, 0, 0);
8523 read_unlock(&tasklist_lock);
8524 mutex_unlock(&rt_constraints_mutex);
8529 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8531 /* Don't accept realtime tasks when there is no way for them to run */
8532 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8538 #else /* !CONFIG_RT_GROUP_SCHED */
8539 static int sched_rt_global_constraints(void)
8541 unsigned long flags;
8544 if (sysctl_sched_rt_period <= 0)
8548 * There's always some RT tasks in the root group
8549 * -- migration, kstopmachine etc..
8551 if (sysctl_sched_rt_runtime == 0)
8554 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8555 for_each_possible_cpu(i) {
8556 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8558 raw_spin_lock(&rt_rq->rt_runtime_lock);
8559 rt_rq->rt_runtime = global_rt_runtime();
8560 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8562 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8566 #endif /* CONFIG_RT_GROUP_SCHED */
8568 int sched_rt_handler(struct ctl_table *table, int write,
8569 void __user *buffer, size_t *lenp,
8573 int old_period, old_runtime;
8574 static DEFINE_MUTEX(mutex);
8577 old_period = sysctl_sched_rt_period;
8578 old_runtime = sysctl_sched_rt_runtime;
8580 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8582 if (!ret && write) {
8583 ret = sched_rt_global_constraints();
8585 sysctl_sched_rt_period = old_period;
8586 sysctl_sched_rt_runtime = old_runtime;
8588 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8589 def_rt_bandwidth.rt_period =
8590 ns_to_ktime(global_rt_period());
8593 mutex_unlock(&mutex);
8598 #ifdef CONFIG_CGROUP_SCHED
8600 /* return corresponding task_group object of a cgroup */
8601 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8603 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8604 struct task_group, css);
8607 static struct cgroup_subsys_state *
8608 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8610 struct task_group *tg, *parent;
8612 if (!cgrp->parent) {
8613 /* This is early initialization for the top cgroup */
8614 return &init_task_group.css;
8617 parent = cgroup_tg(cgrp->parent);
8618 tg = sched_create_group(parent);
8620 return ERR_PTR(-ENOMEM);
8626 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8628 struct task_group *tg = cgroup_tg(cgrp);
8630 sched_destroy_group(tg);
8634 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8636 #ifdef CONFIG_RT_GROUP_SCHED
8637 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
8640 /* We don't support RT-tasks being in separate groups */
8641 if (tsk->sched_class != &fair_sched_class)
8648 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8649 struct task_struct *tsk, bool threadgroup)
8651 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
8655 struct task_struct *c;
8657 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8658 retval = cpu_cgroup_can_attach_task(cgrp, c);
8670 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8671 struct cgroup *old_cont, struct task_struct *tsk,
8674 sched_move_task(tsk);
8676 struct task_struct *c;
8678 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8685 #ifdef CONFIG_FAIR_GROUP_SCHED
8686 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8689 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8692 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8694 struct task_group *tg = cgroup_tg(cgrp);
8696 return (u64) tg->shares;
8698 #endif /* CONFIG_FAIR_GROUP_SCHED */
8700 #ifdef CONFIG_RT_GROUP_SCHED
8701 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8704 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8707 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8709 return sched_group_rt_runtime(cgroup_tg(cgrp));
8712 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8715 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8718 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8720 return sched_group_rt_period(cgroup_tg(cgrp));
8722 #endif /* CONFIG_RT_GROUP_SCHED */
8724 static struct cftype cpu_files[] = {
8725 #ifdef CONFIG_FAIR_GROUP_SCHED
8728 .read_u64 = cpu_shares_read_u64,
8729 .write_u64 = cpu_shares_write_u64,
8732 #ifdef CONFIG_RT_GROUP_SCHED
8734 .name = "rt_runtime_us",
8735 .read_s64 = cpu_rt_runtime_read,
8736 .write_s64 = cpu_rt_runtime_write,
8739 .name = "rt_period_us",
8740 .read_u64 = cpu_rt_period_read_uint,
8741 .write_u64 = cpu_rt_period_write_uint,
8746 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8748 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8751 struct cgroup_subsys cpu_cgroup_subsys = {
8753 .create = cpu_cgroup_create,
8754 .destroy = cpu_cgroup_destroy,
8755 .can_attach = cpu_cgroup_can_attach,
8756 .attach = cpu_cgroup_attach,
8757 .populate = cpu_cgroup_populate,
8758 .subsys_id = cpu_cgroup_subsys_id,
8762 #endif /* CONFIG_CGROUP_SCHED */
8764 #ifdef CONFIG_CGROUP_CPUACCT
8767 * CPU accounting code for task groups.
8769 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8770 * (balbir@in.ibm.com).
8773 /* track cpu usage of a group of tasks and its child groups */
8775 struct cgroup_subsys_state css;
8776 /* cpuusage holds pointer to a u64-type object on every cpu */
8777 u64 __percpu *cpuusage;
8778 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
8779 struct cpuacct *parent;
8782 struct cgroup_subsys cpuacct_subsys;
8784 /* return cpu accounting group corresponding to this container */
8785 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8787 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8788 struct cpuacct, css);
8791 /* return cpu accounting group to which this task belongs */
8792 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8794 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8795 struct cpuacct, css);
8798 /* create a new cpu accounting group */
8799 static struct cgroup_subsys_state *cpuacct_create(
8800 struct cgroup_subsys *ss, struct cgroup *cgrp)
8802 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8808 ca->cpuusage = alloc_percpu(u64);
8812 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8813 if (percpu_counter_init(&ca->cpustat[i], 0))
8814 goto out_free_counters;
8817 ca->parent = cgroup_ca(cgrp->parent);
8823 percpu_counter_destroy(&ca->cpustat[i]);
8824 free_percpu(ca->cpuusage);
8828 return ERR_PTR(-ENOMEM);
8831 /* destroy an existing cpu accounting group */
8833 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8835 struct cpuacct *ca = cgroup_ca(cgrp);
8838 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8839 percpu_counter_destroy(&ca->cpustat[i]);
8840 free_percpu(ca->cpuusage);
8844 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
8846 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8849 #ifndef CONFIG_64BIT
8851 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8853 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8855 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8863 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
8865 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8867 #ifndef CONFIG_64BIT
8869 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8871 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8873 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8879 /* return total cpu usage (in nanoseconds) of a group */
8880 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8882 struct cpuacct *ca = cgroup_ca(cgrp);
8883 u64 totalcpuusage = 0;
8886 for_each_present_cpu(i)
8887 totalcpuusage += cpuacct_cpuusage_read(ca, i);
8889 return totalcpuusage;
8892 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8895 struct cpuacct *ca = cgroup_ca(cgrp);
8904 for_each_present_cpu(i)
8905 cpuacct_cpuusage_write(ca, i, 0);
8911 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
8914 struct cpuacct *ca = cgroup_ca(cgroup);
8918 for_each_present_cpu(i) {
8919 percpu = cpuacct_cpuusage_read(ca, i);
8920 seq_printf(m, "%llu ", (unsigned long long) percpu);
8922 seq_printf(m, "\n");
8926 static const char *cpuacct_stat_desc[] = {
8927 [CPUACCT_STAT_USER] = "user",
8928 [CPUACCT_STAT_SYSTEM] = "system",
8931 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
8932 struct cgroup_map_cb *cb)
8934 struct cpuacct *ca = cgroup_ca(cgrp);
8937 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
8938 s64 val = percpu_counter_read(&ca->cpustat[i]);
8939 val = cputime64_to_clock_t(val);
8940 cb->fill(cb, cpuacct_stat_desc[i], val);
8945 static struct cftype files[] = {
8948 .read_u64 = cpuusage_read,
8949 .write_u64 = cpuusage_write,
8952 .name = "usage_percpu",
8953 .read_seq_string = cpuacct_percpu_seq_read,
8957 .read_map = cpuacct_stats_show,
8961 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
8963 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
8967 * charge this task's execution time to its accounting group.
8969 * called with rq->lock held.
8971 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8976 if (unlikely(!cpuacct_subsys.active))
8979 cpu = task_cpu(tsk);
8985 for (; ca; ca = ca->parent) {
8986 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8987 *cpuusage += cputime;
8994 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
8995 * in cputime_t units. As a result, cpuacct_update_stats calls
8996 * percpu_counter_add with values large enough to always overflow the
8997 * per cpu batch limit causing bad SMP scalability.
8999 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9000 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9001 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9004 #define CPUACCT_BATCH \
9005 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9007 #define CPUACCT_BATCH 0
9011 * Charge the system/user time to the task's accounting group.
9013 static void cpuacct_update_stats(struct task_struct *tsk,
9014 enum cpuacct_stat_index idx, cputime_t val)
9017 int batch = CPUACCT_BATCH;
9019 if (unlikely(!cpuacct_subsys.active))
9026 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9032 struct cgroup_subsys cpuacct_subsys = {
9034 .create = cpuacct_create,
9035 .destroy = cpuacct_destroy,
9036 .populate = cpuacct_populate,
9037 .subsys_id = cpuacct_subsys_id,
9039 #endif /* CONFIG_CGROUP_CPUACCT */
9043 int rcu_expedited_torture_stats(char *page)
9047 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
9049 void synchronize_sched_expedited(void)
9052 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9054 #else /* #ifndef CONFIG_SMP */
9056 static DEFINE_PER_CPU(struct migration_req, rcu_migration_req);
9057 static DEFINE_MUTEX(rcu_sched_expedited_mutex);
9059 #define RCU_EXPEDITED_STATE_POST -2
9060 #define RCU_EXPEDITED_STATE_IDLE -1
9062 static int rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
9064 int rcu_expedited_torture_stats(char *page)
9069 cnt += sprintf(&page[cnt], "state: %d /", rcu_expedited_state);
9070 for_each_online_cpu(cpu) {
9071 cnt += sprintf(&page[cnt], " %d:%d",
9072 cpu, per_cpu(rcu_migration_req, cpu).dest_cpu);
9074 cnt += sprintf(&page[cnt], "\n");
9077 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
9079 static long synchronize_sched_expedited_count;
9082 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
9083 * approach to force grace period to end quickly. This consumes
9084 * significant time on all CPUs, and is thus not recommended for
9085 * any sort of common-case code.
9087 * Note that it is illegal to call this function while holding any
9088 * lock that is acquired by a CPU-hotplug notifier. Failing to
9089 * observe this restriction will result in deadlock.
9091 void synchronize_sched_expedited(void)
9094 unsigned long flags;
9095 bool need_full_sync = 0;
9097 struct migration_req *req;
9101 smp_mb(); /* ensure prior mod happens before capturing snap. */
9102 snap = ACCESS_ONCE(synchronize_sched_expedited_count) + 1;
9104 while (!mutex_trylock(&rcu_sched_expedited_mutex)) {
9106 if (trycount++ < 10)
9107 udelay(trycount * num_online_cpus());
9109 synchronize_sched();
9112 if (ACCESS_ONCE(synchronize_sched_expedited_count) - snap > 0) {
9113 smp_mb(); /* ensure test happens before caller kfree */
9118 rcu_expedited_state = RCU_EXPEDITED_STATE_POST;
9119 for_each_online_cpu(cpu) {
9121 req = &per_cpu(rcu_migration_req, cpu);
9122 init_completion(&req->done);
9124 req->dest_cpu = RCU_MIGRATION_NEED_QS;
9125 raw_spin_lock_irqsave(&rq->lock, flags);
9126 list_add(&req->list, &rq->migration_queue);
9127 raw_spin_unlock_irqrestore(&rq->lock, flags);
9128 wake_up_process(rq->migration_thread);
9130 for_each_online_cpu(cpu) {
9131 rcu_expedited_state = cpu;
9132 req = &per_cpu(rcu_migration_req, cpu);
9134 wait_for_completion(&req->done);
9135 raw_spin_lock_irqsave(&rq->lock, flags);
9136 if (unlikely(req->dest_cpu == RCU_MIGRATION_MUST_SYNC))
9138 req->dest_cpu = RCU_MIGRATION_IDLE;
9139 raw_spin_unlock_irqrestore(&rq->lock, flags);
9141 rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
9142 synchronize_sched_expedited_count++;
9143 mutex_unlock(&rcu_sched_expedited_mutex);
9146 synchronize_sched();
9148 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9150 #endif /* #else #ifndef CONFIG_SMP */