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)
606 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
607 * See detach_destroy_domains: synchronize_sched for details.
609 * The domain tree of any CPU may only be accessed from within
610 * preempt-disabled sections.
612 #define for_each_domain(cpu, __sd) \
613 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
615 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
616 #define this_rq() (&__get_cpu_var(runqueues))
617 #define task_rq(p) cpu_rq(task_cpu(p))
618 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
619 #define raw_rq() (&__raw_get_cpu_var(runqueues))
621 inline void update_rq_clock(struct rq *rq)
623 rq->clock = sched_clock_cpu(cpu_of(rq));
627 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
629 #ifdef CONFIG_SCHED_DEBUG
630 # define const_debug __read_mostly
632 # define const_debug static const
637 * @cpu: the processor in question.
639 * Returns true if the current cpu runqueue is locked.
640 * This interface allows printk to be called with the runqueue lock
641 * held and know whether or not it is OK to wake up the klogd.
643 int runqueue_is_locked(int cpu)
645 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
649 * Debugging: various feature bits
652 #define SCHED_FEAT(name, enabled) \
653 __SCHED_FEAT_##name ,
656 #include "sched_features.h"
661 #define SCHED_FEAT(name, enabled) \
662 (1UL << __SCHED_FEAT_##name) * enabled |
664 const_debug unsigned int sysctl_sched_features =
665 #include "sched_features.h"
670 #ifdef CONFIG_SCHED_DEBUG
671 #define SCHED_FEAT(name, enabled) \
674 static __read_mostly char *sched_feat_names[] = {
675 #include "sched_features.h"
681 static int sched_feat_show(struct seq_file *m, void *v)
685 for (i = 0; sched_feat_names[i]; i++) {
686 if (!(sysctl_sched_features & (1UL << i)))
688 seq_printf(m, "%s ", sched_feat_names[i]);
696 sched_feat_write(struct file *filp, const char __user *ubuf,
697 size_t cnt, loff_t *ppos)
707 if (copy_from_user(&buf, ubuf, cnt))
712 if (strncmp(buf, "NO_", 3) == 0) {
717 for (i = 0; sched_feat_names[i]; i++) {
718 int len = strlen(sched_feat_names[i]);
720 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
722 sysctl_sched_features &= ~(1UL << i);
724 sysctl_sched_features |= (1UL << i);
729 if (!sched_feat_names[i])
737 static int sched_feat_open(struct inode *inode, struct file *filp)
739 return single_open(filp, sched_feat_show, NULL);
742 static const struct file_operations sched_feat_fops = {
743 .open = sched_feat_open,
744 .write = sched_feat_write,
747 .release = single_release,
750 static __init int sched_init_debug(void)
752 debugfs_create_file("sched_features", 0644, NULL, NULL,
757 late_initcall(sched_init_debug);
761 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
764 * Number of tasks to iterate in a single balance run.
765 * Limited because this is done with IRQs disabled.
767 const_debug unsigned int sysctl_sched_nr_migrate = 32;
770 * ratelimit for updating the group shares.
773 unsigned int sysctl_sched_shares_ratelimit = 250000;
774 unsigned int normalized_sysctl_sched_shares_ratelimit = 250000;
777 * Inject some fuzzyness into changing the per-cpu group shares
778 * this avoids remote rq-locks at the expense of fairness.
781 unsigned int sysctl_sched_shares_thresh = 4;
784 * period over which we average the RT time consumption, measured
789 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
792 * period over which we measure -rt task cpu usage in us.
795 unsigned int sysctl_sched_rt_period = 1000000;
797 static __read_mostly int scheduler_running;
800 * part of the period that we allow rt tasks to run in us.
803 int sysctl_sched_rt_runtime = 950000;
805 static inline u64 global_rt_period(void)
807 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
810 static inline u64 global_rt_runtime(void)
812 if (sysctl_sched_rt_runtime < 0)
815 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
818 #ifndef prepare_arch_switch
819 # define prepare_arch_switch(next) do { } while (0)
821 #ifndef finish_arch_switch
822 # define finish_arch_switch(prev) do { } while (0)
825 static inline int task_current(struct rq *rq, struct task_struct *p)
827 return rq->curr == p;
830 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
831 static inline int task_running(struct rq *rq, struct task_struct *p)
833 return task_current(rq, p);
836 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
840 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
842 #ifdef CONFIG_DEBUG_SPINLOCK
843 /* this is a valid case when another task releases the spinlock */
844 rq->lock.owner = current;
847 * If we are tracking spinlock dependencies then we have to
848 * fix up the runqueue lock - which gets 'carried over' from
851 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
853 raw_spin_unlock_irq(&rq->lock);
856 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
857 static inline int task_running(struct rq *rq, struct task_struct *p)
862 return task_current(rq, p);
866 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
870 * We can optimise this out completely for !SMP, because the
871 * SMP rebalancing from interrupt is the only thing that cares
876 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
877 raw_spin_unlock_irq(&rq->lock);
879 raw_spin_unlock(&rq->lock);
883 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
887 * After ->oncpu is cleared, the task can be moved to a different CPU.
888 * We must ensure this doesn't happen until the switch is completely
894 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
898 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
901 * Check whether the task is waking, we use this to synchronize against
902 * ttwu() so that task_cpu() reports a stable number.
904 * We need to make an exception for PF_STARTING tasks because the fork
905 * path might require task_rq_lock() to work, eg. it can call
906 * set_cpus_allowed_ptr() from the cpuset clone_ns code.
908 static inline int task_is_waking(struct task_struct *p)
910 return unlikely((p->state == TASK_WAKING) && !(p->flags & PF_STARTING));
914 * __task_rq_lock - lock the runqueue a given task resides on.
915 * Must be called interrupts disabled.
917 static inline struct rq *__task_rq_lock(struct task_struct *p)
923 while (task_is_waking(p))
926 raw_spin_lock(&rq->lock);
927 if (likely(rq == task_rq(p) && !task_is_waking(p)))
929 raw_spin_unlock(&rq->lock);
934 * task_rq_lock - lock the runqueue a given task resides on and disable
935 * interrupts. Note the ordering: we can safely lookup the task_rq without
936 * explicitly disabling preemption.
938 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
944 while (task_is_waking(p))
946 local_irq_save(*flags);
948 raw_spin_lock(&rq->lock);
949 if (likely(rq == task_rq(p) && !task_is_waking(p)))
951 raw_spin_unlock_irqrestore(&rq->lock, *flags);
955 void task_rq_unlock_wait(struct task_struct *p)
957 struct rq *rq = task_rq(p);
959 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
960 raw_spin_unlock_wait(&rq->lock);
963 static void __task_rq_unlock(struct rq *rq)
966 raw_spin_unlock(&rq->lock);
969 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
972 raw_spin_unlock_irqrestore(&rq->lock, *flags);
976 * this_rq_lock - lock this runqueue and disable interrupts.
978 static struct rq *this_rq_lock(void)
985 raw_spin_lock(&rq->lock);
990 #ifdef CONFIG_SCHED_HRTICK
992 * Use HR-timers to deliver accurate preemption points.
994 * Its all a bit involved since we cannot program an hrt while holding the
995 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
998 * When we get rescheduled we reprogram the hrtick_timer outside of the
1004 * - enabled by features
1005 * - hrtimer is actually high res
1007 static inline int hrtick_enabled(struct rq *rq)
1009 if (!sched_feat(HRTICK))
1011 if (!cpu_active(cpu_of(rq)))
1013 return hrtimer_is_hres_active(&rq->hrtick_timer);
1016 static void hrtick_clear(struct rq *rq)
1018 if (hrtimer_active(&rq->hrtick_timer))
1019 hrtimer_cancel(&rq->hrtick_timer);
1023 * High-resolution timer tick.
1024 * Runs from hardirq context with interrupts disabled.
1026 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1028 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1030 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1032 raw_spin_lock(&rq->lock);
1033 update_rq_clock(rq);
1034 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1035 raw_spin_unlock(&rq->lock);
1037 return HRTIMER_NORESTART;
1042 * called from hardirq (IPI) context
1044 static void __hrtick_start(void *arg)
1046 struct rq *rq = arg;
1048 raw_spin_lock(&rq->lock);
1049 hrtimer_restart(&rq->hrtick_timer);
1050 rq->hrtick_csd_pending = 0;
1051 raw_spin_unlock(&rq->lock);
1055 * Called to set the hrtick timer state.
1057 * called with rq->lock held and irqs disabled
1059 static void hrtick_start(struct rq *rq, u64 delay)
1061 struct hrtimer *timer = &rq->hrtick_timer;
1062 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1064 hrtimer_set_expires(timer, time);
1066 if (rq == this_rq()) {
1067 hrtimer_restart(timer);
1068 } else if (!rq->hrtick_csd_pending) {
1069 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1070 rq->hrtick_csd_pending = 1;
1075 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1077 int cpu = (int)(long)hcpu;
1080 case CPU_UP_CANCELED:
1081 case CPU_UP_CANCELED_FROZEN:
1082 case CPU_DOWN_PREPARE:
1083 case CPU_DOWN_PREPARE_FROZEN:
1085 case CPU_DEAD_FROZEN:
1086 hrtick_clear(cpu_rq(cpu));
1093 static __init void init_hrtick(void)
1095 hotcpu_notifier(hotplug_hrtick, 0);
1099 * Called to set the hrtick timer state.
1101 * called with rq->lock held and irqs disabled
1103 static void hrtick_start(struct rq *rq, u64 delay)
1105 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1106 HRTIMER_MODE_REL_PINNED, 0);
1109 static inline void init_hrtick(void)
1112 #endif /* CONFIG_SMP */
1114 static void init_rq_hrtick(struct rq *rq)
1117 rq->hrtick_csd_pending = 0;
1119 rq->hrtick_csd.flags = 0;
1120 rq->hrtick_csd.func = __hrtick_start;
1121 rq->hrtick_csd.info = rq;
1124 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1125 rq->hrtick_timer.function = hrtick;
1127 #else /* CONFIG_SCHED_HRTICK */
1128 static inline void hrtick_clear(struct rq *rq)
1132 static inline void init_rq_hrtick(struct rq *rq)
1136 static inline void init_hrtick(void)
1139 #endif /* CONFIG_SCHED_HRTICK */
1142 * resched_task - mark a task 'to be rescheduled now'.
1144 * On UP this means the setting of the need_resched flag, on SMP it
1145 * might also involve a cross-CPU call to trigger the scheduler on
1150 #ifndef tsk_is_polling
1151 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1154 static void resched_task(struct task_struct *p)
1158 assert_raw_spin_locked(&task_rq(p)->lock);
1160 if (test_tsk_need_resched(p))
1163 set_tsk_need_resched(p);
1166 if (cpu == smp_processor_id())
1169 /* NEED_RESCHED must be visible before we test polling */
1171 if (!tsk_is_polling(p))
1172 smp_send_reschedule(cpu);
1175 static void resched_cpu(int cpu)
1177 struct rq *rq = cpu_rq(cpu);
1178 unsigned long flags;
1180 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1182 resched_task(cpu_curr(cpu));
1183 raw_spin_unlock_irqrestore(&rq->lock, flags);
1188 * When add_timer_on() enqueues a timer into the timer wheel of an
1189 * idle CPU then this timer might expire before the next timer event
1190 * which is scheduled to wake up that CPU. In case of a completely
1191 * idle system the next event might even be infinite time into the
1192 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1193 * leaves the inner idle loop so the newly added timer is taken into
1194 * account when the CPU goes back to idle and evaluates the timer
1195 * wheel for the next timer event.
1197 void wake_up_idle_cpu(int cpu)
1199 struct rq *rq = cpu_rq(cpu);
1201 if (cpu == smp_processor_id())
1205 * This is safe, as this function is called with the timer
1206 * wheel base lock of (cpu) held. When the CPU is on the way
1207 * to idle and has not yet set rq->curr to idle then it will
1208 * be serialized on the timer wheel base lock and take the new
1209 * timer into account automatically.
1211 if (rq->curr != rq->idle)
1215 * We can set TIF_RESCHED on the idle task of the other CPU
1216 * lockless. The worst case is that the other CPU runs the
1217 * idle task through an additional NOOP schedule()
1219 set_tsk_need_resched(rq->idle);
1221 /* NEED_RESCHED must be visible before we test polling */
1223 if (!tsk_is_polling(rq->idle))
1224 smp_send_reschedule(cpu);
1226 #endif /* CONFIG_NO_HZ */
1228 static u64 sched_avg_period(void)
1230 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1233 static void sched_avg_update(struct rq *rq)
1235 s64 period = sched_avg_period();
1237 while ((s64)(rq->clock - rq->age_stamp) > period) {
1238 rq->age_stamp += period;
1243 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1245 rq->rt_avg += rt_delta;
1246 sched_avg_update(rq);
1249 #else /* !CONFIG_SMP */
1250 static void resched_task(struct task_struct *p)
1252 assert_raw_spin_locked(&task_rq(p)->lock);
1253 set_tsk_need_resched(p);
1256 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1259 #endif /* CONFIG_SMP */
1261 #if BITS_PER_LONG == 32
1262 # define WMULT_CONST (~0UL)
1264 # define WMULT_CONST (1UL << 32)
1267 #define WMULT_SHIFT 32
1270 * Shift right and round:
1272 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1275 * delta *= weight / lw
1277 static unsigned long
1278 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1279 struct load_weight *lw)
1283 if (!lw->inv_weight) {
1284 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1287 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1291 tmp = (u64)delta_exec * weight;
1293 * Check whether we'd overflow the 64-bit multiplication:
1295 if (unlikely(tmp > WMULT_CONST))
1296 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1299 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1301 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1304 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1310 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1317 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1318 * of tasks with abnormal "nice" values across CPUs the contribution that
1319 * each task makes to its run queue's load is weighted according to its
1320 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1321 * scaled version of the new time slice allocation that they receive on time
1325 #define WEIGHT_IDLEPRIO 3
1326 #define WMULT_IDLEPRIO 1431655765
1329 * Nice levels are multiplicative, with a gentle 10% change for every
1330 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1331 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1332 * that remained on nice 0.
1334 * The "10% effect" is relative and cumulative: from _any_ nice level,
1335 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1336 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1337 * If a task goes up by ~10% and another task goes down by ~10% then
1338 * the relative distance between them is ~25%.)
1340 static const int prio_to_weight[40] = {
1341 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1342 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1343 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1344 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1345 /* 0 */ 1024, 820, 655, 526, 423,
1346 /* 5 */ 335, 272, 215, 172, 137,
1347 /* 10 */ 110, 87, 70, 56, 45,
1348 /* 15 */ 36, 29, 23, 18, 15,
1352 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1354 * In cases where the weight does not change often, we can use the
1355 * precalculated inverse to speed up arithmetics by turning divisions
1356 * into multiplications:
1358 static const u32 prio_to_wmult[40] = {
1359 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1360 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1361 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1362 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1363 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1364 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1365 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1366 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1369 /* Time spent by the tasks of the cpu accounting group executing in ... */
1370 enum cpuacct_stat_index {
1371 CPUACCT_STAT_USER, /* ... user mode */
1372 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1374 CPUACCT_STAT_NSTATS,
1377 #ifdef CONFIG_CGROUP_CPUACCT
1378 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1379 static void cpuacct_update_stats(struct task_struct *tsk,
1380 enum cpuacct_stat_index idx, cputime_t val);
1382 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1383 static inline void cpuacct_update_stats(struct task_struct *tsk,
1384 enum cpuacct_stat_index idx, cputime_t val) {}
1387 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1389 update_load_add(&rq->load, load);
1392 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1394 update_load_sub(&rq->load, load);
1397 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1398 typedef int (*tg_visitor)(struct task_group *, void *);
1401 * Iterate the full tree, calling @down when first entering a node and @up when
1402 * leaving it for the final time.
1404 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1406 struct task_group *parent, *child;
1410 parent = &root_task_group;
1412 ret = (*down)(parent, data);
1415 list_for_each_entry_rcu(child, &parent->children, siblings) {
1422 ret = (*up)(parent, data);
1427 parent = parent->parent;
1436 static int tg_nop(struct task_group *tg, void *data)
1443 /* Used instead of source_load when we know the type == 0 */
1444 static unsigned long weighted_cpuload(const int cpu)
1446 return cpu_rq(cpu)->load.weight;
1450 * Return a low guess at the load of a migration-source cpu weighted
1451 * according to the scheduling class and "nice" value.
1453 * We want to under-estimate the load of migration sources, to
1454 * balance conservatively.
1456 static unsigned long source_load(int cpu, int type)
1458 struct rq *rq = cpu_rq(cpu);
1459 unsigned long total = weighted_cpuload(cpu);
1461 if (type == 0 || !sched_feat(LB_BIAS))
1464 return min(rq->cpu_load[type-1], total);
1468 * Return a high guess at the load of a migration-target cpu weighted
1469 * according to the scheduling class and "nice" value.
1471 static unsigned long target_load(int cpu, int type)
1473 struct rq *rq = cpu_rq(cpu);
1474 unsigned long total = weighted_cpuload(cpu);
1476 if (type == 0 || !sched_feat(LB_BIAS))
1479 return max(rq->cpu_load[type-1], total);
1482 static struct sched_group *group_of(int cpu)
1484 struct sched_domain *sd = rcu_dereference(cpu_rq(cpu)->sd);
1492 static unsigned long power_of(int cpu)
1494 struct sched_group *group = group_of(cpu);
1497 return SCHED_LOAD_SCALE;
1499 return group->cpu_power;
1502 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1504 static unsigned long cpu_avg_load_per_task(int cpu)
1506 struct rq *rq = cpu_rq(cpu);
1507 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1510 rq->avg_load_per_task = rq->load.weight / nr_running;
1512 rq->avg_load_per_task = 0;
1514 return rq->avg_load_per_task;
1517 #ifdef CONFIG_FAIR_GROUP_SCHED
1519 static __read_mostly unsigned long *update_shares_data;
1521 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1524 * Calculate and set the cpu's group shares.
1526 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1527 unsigned long sd_shares,
1528 unsigned long sd_rq_weight,
1529 unsigned long *usd_rq_weight)
1531 unsigned long shares, rq_weight;
1534 rq_weight = usd_rq_weight[cpu];
1537 rq_weight = NICE_0_LOAD;
1541 * \Sum_j shares_j * rq_weight_i
1542 * shares_i = -----------------------------
1543 * \Sum_j rq_weight_j
1545 shares = (sd_shares * rq_weight) / sd_rq_weight;
1546 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1548 if (abs(shares - tg->se[cpu]->load.weight) >
1549 sysctl_sched_shares_thresh) {
1550 struct rq *rq = cpu_rq(cpu);
1551 unsigned long flags;
1553 raw_spin_lock_irqsave(&rq->lock, flags);
1554 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1555 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1556 __set_se_shares(tg->se[cpu], shares);
1557 raw_spin_unlock_irqrestore(&rq->lock, flags);
1562 * Re-compute the task group their per cpu shares over the given domain.
1563 * This needs to be done in a bottom-up fashion because the rq weight of a
1564 * parent group depends on the shares of its child groups.
1566 static int tg_shares_up(struct task_group *tg, void *data)
1568 unsigned long weight, rq_weight = 0, sum_weight = 0, shares = 0;
1569 unsigned long *usd_rq_weight;
1570 struct sched_domain *sd = data;
1571 unsigned long flags;
1577 local_irq_save(flags);
1578 usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id());
1580 for_each_cpu(i, sched_domain_span(sd)) {
1581 weight = tg->cfs_rq[i]->load.weight;
1582 usd_rq_weight[i] = weight;
1584 rq_weight += weight;
1586 * If there are currently no tasks on the cpu pretend there
1587 * is one of average load so that when a new task gets to
1588 * run here it will not get delayed by group starvation.
1591 weight = NICE_0_LOAD;
1593 sum_weight += weight;
1594 shares += tg->cfs_rq[i]->shares;
1598 rq_weight = sum_weight;
1600 if ((!shares && rq_weight) || shares > tg->shares)
1601 shares = tg->shares;
1603 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1604 shares = tg->shares;
1606 for_each_cpu(i, sched_domain_span(sd))
1607 update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight);
1609 local_irq_restore(flags);
1615 * Compute the cpu's hierarchical load factor for each task group.
1616 * This needs to be done in a top-down fashion because the load of a child
1617 * group is a fraction of its parents load.
1619 static int tg_load_down(struct task_group *tg, void *data)
1622 long cpu = (long)data;
1625 load = cpu_rq(cpu)->load.weight;
1627 load = tg->parent->cfs_rq[cpu]->h_load;
1628 load *= tg->cfs_rq[cpu]->shares;
1629 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1632 tg->cfs_rq[cpu]->h_load = load;
1637 static void update_shares(struct sched_domain *sd)
1642 if (root_task_group_empty())
1645 now = cpu_clock(raw_smp_processor_id());
1646 elapsed = now - sd->last_update;
1648 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1649 sd->last_update = now;
1650 walk_tg_tree(tg_nop, tg_shares_up, sd);
1654 static void update_h_load(long cpu)
1656 if (root_task_group_empty())
1659 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1664 static inline void update_shares(struct sched_domain *sd)
1670 #ifdef CONFIG_PREEMPT
1672 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1675 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1676 * way at the expense of forcing extra atomic operations in all
1677 * invocations. This assures that the double_lock is acquired using the
1678 * same underlying policy as the spinlock_t on this architecture, which
1679 * reduces latency compared to the unfair variant below. However, it
1680 * also adds more overhead and therefore may reduce throughput.
1682 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1683 __releases(this_rq->lock)
1684 __acquires(busiest->lock)
1685 __acquires(this_rq->lock)
1687 raw_spin_unlock(&this_rq->lock);
1688 double_rq_lock(this_rq, busiest);
1695 * Unfair double_lock_balance: Optimizes throughput at the expense of
1696 * latency by eliminating extra atomic operations when the locks are
1697 * already in proper order on entry. This favors lower cpu-ids and will
1698 * grant the double lock to lower cpus over higher ids under contention,
1699 * regardless of entry order into the function.
1701 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1702 __releases(this_rq->lock)
1703 __acquires(busiest->lock)
1704 __acquires(this_rq->lock)
1708 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1709 if (busiest < this_rq) {
1710 raw_spin_unlock(&this_rq->lock);
1711 raw_spin_lock(&busiest->lock);
1712 raw_spin_lock_nested(&this_rq->lock,
1713 SINGLE_DEPTH_NESTING);
1716 raw_spin_lock_nested(&busiest->lock,
1717 SINGLE_DEPTH_NESTING);
1722 #endif /* CONFIG_PREEMPT */
1725 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1727 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1729 if (unlikely(!irqs_disabled())) {
1730 /* printk() doesn't work good under rq->lock */
1731 raw_spin_unlock(&this_rq->lock);
1735 return _double_lock_balance(this_rq, busiest);
1738 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1739 __releases(busiest->lock)
1741 raw_spin_unlock(&busiest->lock);
1742 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1746 * double_rq_lock - safely lock two runqueues
1748 * Note this does not disable interrupts like task_rq_lock,
1749 * you need to do so manually before calling.
1751 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1752 __acquires(rq1->lock)
1753 __acquires(rq2->lock)
1755 BUG_ON(!irqs_disabled());
1757 raw_spin_lock(&rq1->lock);
1758 __acquire(rq2->lock); /* Fake it out ;) */
1761 raw_spin_lock(&rq1->lock);
1762 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1764 raw_spin_lock(&rq2->lock);
1765 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1768 update_rq_clock(rq1);
1769 update_rq_clock(rq2);
1773 * double_rq_unlock - safely unlock two runqueues
1775 * Note this does not restore interrupts like task_rq_unlock,
1776 * you need to do so manually after calling.
1778 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1779 __releases(rq1->lock)
1780 __releases(rq2->lock)
1782 raw_spin_unlock(&rq1->lock);
1784 raw_spin_unlock(&rq2->lock);
1786 __release(rq2->lock);
1791 #ifdef CONFIG_FAIR_GROUP_SCHED
1792 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1795 cfs_rq->shares = shares;
1800 static void calc_load_account_active(struct rq *this_rq);
1801 static void update_sysctl(void);
1802 static int get_update_sysctl_factor(void);
1804 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1806 set_task_rq(p, cpu);
1809 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1810 * successfuly executed on another CPU. We must ensure that updates of
1811 * per-task data have been completed by this moment.
1814 task_thread_info(p)->cpu = cpu;
1818 static const struct sched_class rt_sched_class;
1820 #define sched_class_highest (&rt_sched_class)
1821 #define for_each_class(class) \
1822 for (class = sched_class_highest; class; class = class->next)
1824 #include "sched_stats.h"
1826 static void inc_nr_running(struct rq *rq)
1831 static void dec_nr_running(struct rq *rq)
1836 static void set_load_weight(struct task_struct *p)
1838 if (task_has_rt_policy(p)) {
1839 p->se.load.weight = prio_to_weight[0] * 2;
1840 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1845 * SCHED_IDLE tasks get minimal weight:
1847 if (p->policy == SCHED_IDLE) {
1848 p->se.load.weight = WEIGHT_IDLEPRIO;
1849 p->se.load.inv_weight = WMULT_IDLEPRIO;
1853 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1854 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1857 static void update_avg(u64 *avg, u64 sample)
1859 s64 diff = sample - *avg;
1864 enqueue_task(struct rq *rq, struct task_struct *p, int wakeup, bool head)
1867 p->se.start_runtime = p->se.sum_exec_runtime;
1869 sched_info_queued(p);
1870 p->sched_class->enqueue_task(rq, p, wakeup, head);
1874 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1877 if (p->se.last_wakeup) {
1878 update_avg(&p->se.avg_overlap,
1879 p->se.sum_exec_runtime - p->se.last_wakeup);
1880 p->se.last_wakeup = 0;
1882 update_avg(&p->se.avg_wakeup,
1883 sysctl_sched_wakeup_granularity);
1887 sched_info_dequeued(p);
1888 p->sched_class->dequeue_task(rq, p, sleep);
1893 * activate_task - move a task to the runqueue.
1895 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1897 if (task_contributes_to_load(p))
1898 rq->nr_uninterruptible--;
1900 enqueue_task(rq, p, wakeup, false);
1905 * deactivate_task - remove a task from the runqueue.
1907 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1909 if (task_contributes_to_load(p))
1910 rq->nr_uninterruptible++;
1912 dequeue_task(rq, p, sleep);
1916 #include "sched_idletask.c"
1917 #include "sched_fair.c"
1918 #include "sched_rt.c"
1919 #ifdef CONFIG_SCHED_DEBUG
1920 # include "sched_debug.c"
1924 * __normal_prio - return the priority that is based on the static prio
1926 static inline int __normal_prio(struct task_struct *p)
1928 return p->static_prio;
1932 * Calculate the expected normal priority: i.e. priority
1933 * without taking RT-inheritance into account. Might be
1934 * boosted by interactivity modifiers. Changes upon fork,
1935 * setprio syscalls, and whenever the interactivity
1936 * estimator recalculates.
1938 static inline int normal_prio(struct task_struct *p)
1942 if (task_has_rt_policy(p))
1943 prio = MAX_RT_PRIO-1 - p->rt_priority;
1945 prio = __normal_prio(p);
1950 * Calculate the current priority, i.e. the priority
1951 * taken into account by the scheduler. This value might
1952 * be boosted by RT tasks, or might be boosted by
1953 * interactivity modifiers. Will be RT if the task got
1954 * RT-boosted. If not then it returns p->normal_prio.
1956 static int effective_prio(struct task_struct *p)
1958 p->normal_prio = normal_prio(p);
1960 * If we are RT tasks or we were boosted to RT priority,
1961 * keep the priority unchanged. Otherwise, update priority
1962 * to the normal priority:
1964 if (!rt_prio(p->prio))
1965 return p->normal_prio;
1970 * task_curr - is this task currently executing on a CPU?
1971 * @p: the task in question.
1973 inline int task_curr(const struct task_struct *p)
1975 return cpu_curr(task_cpu(p)) == p;
1978 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1979 const struct sched_class *prev_class,
1980 int oldprio, int running)
1982 if (prev_class != p->sched_class) {
1983 if (prev_class->switched_from)
1984 prev_class->switched_from(rq, p, running);
1985 p->sched_class->switched_to(rq, p, running);
1987 p->sched_class->prio_changed(rq, p, oldprio, running);
1992 * Is this task likely cache-hot:
1995 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1999 if (p->sched_class != &fair_sched_class)
2003 * Buddy candidates are cache hot:
2005 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2006 (&p->se == cfs_rq_of(&p->se)->next ||
2007 &p->se == cfs_rq_of(&p->se)->last))
2010 if (sysctl_sched_migration_cost == -1)
2012 if (sysctl_sched_migration_cost == 0)
2015 delta = now - p->se.exec_start;
2017 return delta < (s64)sysctl_sched_migration_cost;
2020 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2022 #ifdef CONFIG_SCHED_DEBUG
2024 * We should never call set_task_cpu() on a blocked task,
2025 * ttwu() will sort out the placement.
2027 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2028 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2031 trace_sched_migrate_task(p, new_cpu);
2033 if (task_cpu(p) != new_cpu) {
2034 p->se.nr_migrations++;
2035 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2038 __set_task_cpu(p, new_cpu);
2041 struct migration_req {
2042 struct list_head list;
2044 struct task_struct *task;
2047 struct completion done;
2051 * The task's runqueue lock must be held.
2052 * Returns true if you have to wait for migration thread.
2055 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2057 struct rq *rq = task_rq(p);
2060 * If the task is not on a runqueue (and not running), then
2061 * the next wake-up will properly place the task.
2063 if (!p->se.on_rq && !task_running(rq, p))
2066 init_completion(&req->done);
2068 req->dest_cpu = dest_cpu;
2069 list_add(&req->list, &rq->migration_queue);
2075 * wait_task_context_switch - wait for a thread to complete at least one
2078 * @p must not be current.
2080 void wait_task_context_switch(struct task_struct *p)
2082 unsigned long nvcsw, nivcsw, flags;
2090 * The runqueue is assigned before the actual context
2091 * switch. We need to take the runqueue lock.
2093 * We could check initially without the lock but it is
2094 * very likely that we need to take the lock in every
2097 rq = task_rq_lock(p, &flags);
2098 running = task_running(rq, p);
2099 task_rq_unlock(rq, &flags);
2101 if (likely(!running))
2104 * The switch count is incremented before the actual
2105 * context switch. We thus wait for two switches to be
2106 * sure at least one completed.
2108 if ((p->nvcsw - nvcsw) > 1)
2110 if ((p->nivcsw - nivcsw) > 1)
2118 * wait_task_inactive - wait for a thread to unschedule.
2120 * If @match_state is nonzero, it's the @p->state value just checked and
2121 * not expected to change. If it changes, i.e. @p might have woken up,
2122 * then return zero. When we succeed in waiting for @p to be off its CPU,
2123 * we return a positive number (its total switch count). If a second call
2124 * a short while later returns the same number, the caller can be sure that
2125 * @p has remained unscheduled the whole time.
2127 * The caller must ensure that the task *will* unschedule sometime soon,
2128 * else this function might spin for a *long* time. This function can't
2129 * be called with interrupts off, or it may introduce deadlock with
2130 * smp_call_function() if an IPI is sent by the same process we are
2131 * waiting to become inactive.
2133 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2135 unsigned long flags;
2142 * We do the initial early heuristics without holding
2143 * any task-queue locks at all. We'll only try to get
2144 * the runqueue lock when things look like they will
2150 * If the task is actively running on another CPU
2151 * still, just relax and busy-wait without holding
2154 * NOTE! Since we don't hold any locks, it's not
2155 * even sure that "rq" stays as the right runqueue!
2156 * But we don't care, since "task_running()" will
2157 * return false if the runqueue has changed and p
2158 * is actually now running somewhere else!
2160 while (task_running(rq, p)) {
2161 if (match_state && unlikely(p->state != match_state))
2167 * Ok, time to look more closely! We need the rq
2168 * lock now, to be *sure*. If we're wrong, we'll
2169 * just go back and repeat.
2171 rq = task_rq_lock(p, &flags);
2172 trace_sched_wait_task(rq, p);
2173 running = task_running(rq, p);
2174 on_rq = p->se.on_rq;
2176 if (!match_state || p->state == match_state)
2177 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2178 task_rq_unlock(rq, &flags);
2181 * If it changed from the expected state, bail out now.
2183 if (unlikely(!ncsw))
2187 * Was it really running after all now that we
2188 * checked with the proper locks actually held?
2190 * Oops. Go back and try again..
2192 if (unlikely(running)) {
2198 * It's not enough that it's not actively running,
2199 * it must be off the runqueue _entirely_, and not
2202 * So if it was still runnable (but just not actively
2203 * running right now), it's preempted, and we should
2204 * yield - it could be a while.
2206 if (unlikely(on_rq)) {
2207 schedule_timeout_uninterruptible(1);
2212 * Ahh, all good. It wasn't running, and it wasn't
2213 * runnable, which means that it will never become
2214 * running in the future either. We're all done!
2223 * kick_process - kick a running thread to enter/exit the kernel
2224 * @p: the to-be-kicked thread
2226 * Cause a process which is running on another CPU to enter
2227 * kernel-mode, without any delay. (to get signals handled.)
2229 * NOTE: this function doesnt have to take the runqueue lock,
2230 * because all it wants to ensure is that the remote task enters
2231 * the kernel. If the IPI races and the task has been migrated
2232 * to another CPU then no harm is done and the purpose has been
2235 void kick_process(struct task_struct *p)
2241 if ((cpu != smp_processor_id()) && task_curr(p))
2242 smp_send_reschedule(cpu);
2245 EXPORT_SYMBOL_GPL(kick_process);
2246 #endif /* CONFIG_SMP */
2249 * task_oncpu_function_call - call a function on the cpu on which a task runs
2250 * @p: the task to evaluate
2251 * @func: the function to be called
2252 * @info: the function call argument
2254 * Calls the function @func when the task is currently running. This might
2255 * be on the current CPU, which just calls the function directly
2257 void task_oncpu_function_call(struct task_struct *p,
2258 void (*func) (void *info), void *info)
2265 smp_call_function_single(cpu, func, info, 1);
2270 static int select_fallback_rq(int cpu, struct task_struct *p)
2273 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2275 /* Look for allowed, online CPU in same node. */
2276 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2277 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2280 /* Any allowed, online CPU? */
2281 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2282 if (dest_cpu < nr_cpu_ids)
2285 /* No more Mr. Nice Guy. */
2286 if (dest_cpu >= nr_cpu_ids) {
2288 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
2290 dest_cpu = cpumask_any_and(cpu_active_mask, &p->cpus_allowed);
2293 * Don't tell them about moving exiting tasks or
2294 * kernel threads (both mm NULL), since they never
2297 if (p->mm && printk_ratelimit()) {
2298 printk(KERN_INFO "process %d (%s) no "
2299 "longer affine to cpu%d\n",
2300 task_pid_nr(p), p->comm, cpu);
2308 * Gets called from 3 sites (exec, fork, wakeup), since it is called without
2309 * holding rq->lock we need to ensure ->cpus_allowed is stable, this is done
2312 * exec: is unstable, retry loop
2313 * fork & wake-up: serialize ->cpus_allowed against TASK_WAKING
2316 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2318 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2321 * In order not to call set_task_cpu() on a blocking task we need
2322 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2325 * Since this is common to all placement strategies, this lives here.
2327 * [ this allows ->select_task() to simply return task_cpu(p) and
2328 * not worry about this generic constraint ]
2330 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2332 cpu = select_fallback_rq(task_cpu(p), p);
2339 * try_to_wake_up - wake up a thread
2340 * @p: the to-be-woken-up thread
2341 * @state: the mask of task states that can be woken
2342 * @sync: do a synchronous wakeup?
2344 * Put it on the run-queue if it's not already there. The "current"
2345 * thread is always on the run-queue (except when the actual
2346 * re-schedule is in progress), and as such you're allowed to do
2347 * the simpler "current->state = TASK_RUNNING" to mark yourself
2348 * runnable without the overhead of this.
2350 * returns failure only if the task is already active.
2352 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2355 int cpu, orig_cpu, this_cpu, success = 0;
2356 unsigned long flags;
2357 struct rq *rq, *orig_rq;
2359 if (!sched_feat(SYNC_WAKEUPS))
2360 wake_flags &= ~WF_SYNC;
2362 this_cpu = get_cpu();
2365 rq = orig_rq = task_rq_lock(p, &flags);
2366 update_rq_clock(rq);
2367 if (!(p->state & state))
2377 if (unlikely(task_running(rq, p)))
2381 * In order to handle concurrent wakeups and release the rq->lock
2382 * we put the task in TASK_WAKING state.
2384 * First fix up the nr_uninterruptible count:
2386 if (task_contributes_to_load(p))
2387 rq->nr_uninterruptible--;
2388 p->state = TASK_WAKING;
2390 if (p->sched_class->task_waking)
2391 p->sched_class->task_waking(rq, p);
2393 __task_rq_unlock(rq);
2395 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2396 if (cpu != orig_cpu) {
2398 * Since we migrate the task without holding any rq->lock,
2399 * we need to be careful with task_rq_lock(), since that
2400 * might end up locking an invalid rq.
2402 set_task_cpu(p, cpu);
2406 raw_spin_lock(&rq->lock);
2407 update_rq_clock(rq);
2410 * We migrated the task without holding either rq->lock, however
2411 * since the task is not on the task list itself, nobody else
2412 * will try and migrate the task, hence the rq should match the
2413 * cpu we just moved it to.
2415 WARN_ON(task_cpu(p) != cpu);
2416 WARN_ON(p->state != TASK_WAKING);
2418 #ifdef CONFIG_SCHEDSTATS
2419 schedstat_inc(rq, ttwu_count);
2420 if (cpu == this_cpu)
2421 schedstat_inc(rq, ttwu_local);
2423 struct sched_domain *sd;
2424 for_each_domain(this_cpu, sd) {
2425 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2426 schedstat_inc(sd, ttwu_wake_remote);
2431 #endif /* CONFIG_SCHEDSTATS */
2434 #endif /* CONFIG_SMP */
2435 schedstat_inc(p, se.nr_wakeups);
2436 if (wake_flags & WF_SYNC)
2437 schedstat_inc(p, se.nr_wakeups_sync);
2438 if (orig_cpu != cpu)
2439 schedstat_inc(p, se.nr_wakeups_migrate);
2440 if (cpu == this_cpu)
2441 schedstat_inc(p, se.nr_wakeups_local);
2443 schedstat_inc(p, se.nr_wakeups_remote);
2444 activate_task(rq, p, 1);
2448 * Only attribute actual wakeups done by this task.
2450 if (!in_interrupt()) {
2451 struct sched_entity *se = ¤t->se;
2452 u64 sample = se->sum_exec_runtime;
2454 if (se->last_wakeup)
2455 sample -= se->last_wakeup;
2457 sample -= se->start_runtime;
2458 update_avg(&se->avg_wakeup, sample);
2460 se->last_wakeup = se->sum_exec_runtime;
2464 trace_sched_wakeup(rq, p, success);
2465 check_preempt_curr(rq, p, wake_flags);
2467 p->state = TASK_RUNNING;
2469 if (p->sched_class->task_woken)
2470 p->sched_class->task_woken(rq, p);
2472 if (unlikely(rq->idle_stamp)) {
2473 u64 delta = rq->clock - rq->idle_stamp;
2474 u64 max = 2*sysctl_sched_migration_cost;
2479 update_avg(&rq->avg_idle, delta);
2484 task_rq_unlock(rq, &flags);
2491 * wake_up_process - Wake up a specific process
2492 * @p: The process to be woken up.
2494 * Attempt to wake up the nominated process and move it to the set of runnable
2495 * processes. Returns 1 if the process was woken up, 0 if it was already
2498 * It may be assumed that this function implies a write memory barrier before
2499 * changing the task state if and only if any tasks are woken up.
2501 int wake_up_process(struct task_struct *p)
2503 return try_to_wake_up(p, TASK_ALL, 0);
2505 EXPORT_SYMBOL(wake_up_process);
2507 int wake_up_state(struct task_struct *p, unsigned int state)
2509 return try_to_wake_up(p, state, 0);
2513 * Perform scheduler related setup for a newly forked process p.
2514 * p is forked by current.
2516 * __sched_fork() is basic setup used by init_idle() too:
2518 static void __sched_fork(struct task_struct *p)
2520 p->se.exec_start = 0;
2521 p->se.sum_exec_runtime = 0;
2522 p->se.prev_sum_exec_runtime = 0;
2523 p->se.nr_migrations = 0;
2524 p->se.last_wakeup = 0;
2525 p->se.avg_overlap = 0;
2526 p->se.start_runtime = 0;
2527 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2529 #ifdef CONFIG_SCHEDSTATS
2530 p->se.wait_start = 0;
2532 p->se.wait_count = 0;
2535 p->se.sleep_start = 0;
2536 p->se.sleep_max = 0;
2537 p->se.sum_sleep_runtime = 0;
2539 p->se.block_start = 0;
2540 p->se.block_max = 0;
2542 p->se.slice_max = 0;
2544 p->se.nr_migrations_cold = 0;
2545 p->se.nr_failed_migrations_affine = 0;
2546 p->se.nr_failed_migrations_running = 0;
2547 p->se.nr_failed_migrations_hot = 0;
2548 p->se.nr_forced_migrations = 0;
2550 p->se.nr_wakeups = 0;
2551 p->se.nr_wakeups_sync = 0;
2552 p->se.nr_wakeups_migrate = 0;
2553 p->se.nr_wakeups_local = 0;
2554 p->se.nr_wakeups_remote = 0;
2555 p->se.nr_wakeups_affine = 0;
2556 p->se.nr_wakeups_affine_attempts = 0;
2557 p->se.nr_wakeups_passive = 0;
2558 p->se.nr_wakeups_idle = 0;
2562 INIT_LIST_HEAD(&p->rt.run_list);
2564 INIT_LIST_HEAD(&p->se.group_node);
2566 #ifdef CONFIG_PREEMPT_NOTIFIERS
2567 INIT_HLIST_HEAD(&p->preempt_notifiers);
2572 * fork()/clone()-time setup:
2574 void sched_fork(struct task_struct *p, int clone_flags)
2576 int cpu = get_cpu();
2580 * We mark the process as waking here. This guarantees that
2581 * nobody will actually run it, and a signal or other external
2582 * event cannot wake it up and insert it on the runqueue either.
2584 p->state = TASK_WAKING;
2587 * Revert to default priority/policy on fork if requested.
2589 if (unlikely(p->sched_reset_on_fork)) {
2590 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2591 p->policy = SCHED_NORMAL;
2592 p->normal_prio = p->static_prio;
2595 if (PRIO_TO_NICE(p->static_prio) < 0) {
2596 p->static_prio = NICE_TO_PRIO(0);
2597 p->normal_prio = p->static_prio;
2602 * We don't need the reset flag anymore after the fork. It has
2603 * fulfilled its duty:
2605 p->sched_reset_on_fork = 0;
2609 * Make sure we do not leak PI boosting priority to the child.
2611 p->prio = current->normal_prio;
2613 if (!rt_prio(p->prio))
2614 p->sched_class = &fair_sched_class;
2616 if (p->sched_class->task_fork)
2617 p->sched_class->task_fork(p);
2619 set_task_cpu(p, cpu);
2621 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2622 if (likely(sched_info_on()))
2623 memset(&p->sched_info, 0, sizeof(p->sched_info));
2625 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2628 #ifdef CONFIG_PREEMPT
2629 /* Want to start with kernel preemption disabled. */
2630 task_thread_info(p)->preempt_count = 1;
2632 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2638 * wake_up_new_task - wake up a newly created task for the first time.
2640 * This function will do some initial scheduler statistics housekeeping
2641 * that must be done for every newly created context, then puts the task
2642 * on the runqueue and wakes it.
2644 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2646 unsigned long flags;
2648 int cpu = get_cpu();
2652 * Fork balancing, do it here and not earlier because:
2653 * - cpus_allowed can change in the fork path
2654 * - any previously selected cpu might disappear through hotplug
2656 * We still have TASK_WAKING but PF_STARTING is gone now, meaning
2657 * ->cpus_allowed is stable, we have preemption disabled, meaning
2658 * cpu_online_mask is stable.
2660 cpu = select_task_rq(p, SD_BALANCE_FORK, 0);
2661 set_task_cpu(p, cpu);
2665 * Since the task is not on the rq and we still have TASK_WAKING set
2666 * nobody else will migrate this task.
2669 raw_spin_lock_irqsave(&rq->lock, flags);
2671 BUG_ON(p->state != TASK_WAKING);
2672 p->state = TASK_RUNNING;
2673 update_rq_clock(rq);
2674 activate_task(rq, p, 0);
2675 trace_sched_wakeup_new(rq, p, 1);
2676 check_preempt_curr(rq, p, WF_FORK);
2678 if (p->sched_class->task_woken)
2679 p->sched_class->task_woken(rq, p);
2681 task_rq_unlock(rq, &flags);
2685 #ifdef CONFIG_PREEMPT_NOTIFIERS
2688 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2689 * @notifier: notifier struct to register
2691 void preempt_notifier_register(struct preempt_notifier *notifier)
2693 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2695 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2698 * preempt_notifier_unregister - no longer interested in preemption notifications
2699 * @notifier: notifier struct to unregister
2701 * This is safe to call from within a preemption notifier.
2703 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2705 hlist_del(¬ifier->link);
2707 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2709 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2711 struct preempt_notifier *notifier;
2712 struct hlist_node *node;
2714 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2715 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2719 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2720 struct task_struct *next)
2722 struct preempt_notifier *notifier;
2723 struct hlist_node *node;
2725 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2726 notifier->ops->sched_out(notifier, next);
2729 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2731 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2736 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2737 struct task_struct *next)
2741 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2744 * prepare_task_switch - prepare to switch tasks
2745 * @rq: the runqueue preparing to switch
2746 * @prev: the current task that is being switched out
2747 * @next: the task we are going to switch to.
2749 * This is called with the rq lock held and interrupts off. It must
2750 * be paired with a subsequent finish_task_switch after the context
2753 * prepare_task_switch sets up locking and calls architecture specific
2757 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2758 struct task_struct *next)
2760 fire_sched_out_preempt_notifiers(prev, next);
2761 prepare_lock_switch(rq, next);
2762 prepare_arch_switch(next);
2766 * finish_task_switch - clean up after a task-switch
2767 * @rq: runqueue associated with task-switch
2768 * @prev: the thread we just switched away from.
2770 * finish_task_switch must be called after the context switch, paired
2771 * with a prepare_task_switch call before the context switch.
2772 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2773 * and do any other architecture-specific cleanup actions.
2775 * Note that we may have delayed dropping an mm in context_switch(). If
2776 * so, we finish that here outside of the runqueue lock. (Doing it
2777 * with the lock held can cause deadlocks; see schedule() for
2780 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2781 __releases(rq->lock)
2783 struct mm_struct *mm = rq->prev_mm;
2789 * A task struct has one reference for the use as "current".
2790 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2791 * schedule one last time. The schedule call will never return, and
2792 * the scheduled task must drop that reference.
2793 * The test for TASK_DEAD must occur while the runqueue locks are
2794 * still held, otherwise prev could be scheduled on another cpu, die
2795 * there before we look at prev->state, and then the reference would
2797 * Manfred Spraul <manfred@colorfullife.com>
2799 prev_state = prev->state;
2800 finish_arch_switch(prev);
2801 perf_event_task_sched_in(current, cpu_of(rq));
2802 finish_lock_switch(rq, prev);
2804 fire_sched_in_preempt_notifiers(current);
2807 if (unlikely(prev_state == TASK_DEAD)) {
2809 * Remove function-return probe instances associated with this
2810 * task and put them back on the free list.
2812 kprobe_flush_task(prev);
2813 put_task_struct(prev);
2819 /* assumes rq->lock is held */
2820 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2822 if (prev->sched_class->pre_schedule)
2823 prev->sched_class->pre_schedule(rq, prev);
2826 /* rq->lock is NOT held, but preemption is disabled */
2827 static inline void post_schedule(struct rq *rq)
2829 if (rq->post_schedule) {
2830 unsigned long flags;
2832 raw_spin_lock_irqsave(&rq->lock, flags);
2833 if (rq->curr->sched_class->post_schedule)
2834 rq->curr->sched_class->post_schedule(rq);
2835 raw_spin_unlock_irqrestore(&rq->lock, flags);
2837 rq->post_schedule = 0;
2843 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2847 static inline void post_schedule(struct rq *rq)
2854 * schedule_tail - first thing a freshly forked thread must call.
2855 * @prev: the thread we just switched away from.
2857 asmlinkage void schedule_tail(struct task_struct *prev)
2858 __releases(rq->lock)
2860 struct rq *rq = this_rq();
2862 finish_task_switch(rq, prev);
2865 * FIXME: do we need to worry about rq being invalidated by the
2870 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2871 /* In this case, finish_task_switch does not reenable preemption */
2874 if (current->set_child_tid)
2875 put_user(task_pid_vnr(current), current->set_child_tid);
2879 * context_switch - switch to the new MM and the new
2880 * thread's register state.
2883 context_switch(struct rq *rq, struct task_struct *prev,
2884 struct task_struct *next)
2886 struct mm_struct *mm, *oldmm;
2888 prepare_task_switch(rq, prev, next);
2889 trace_sched_switch(rq, prev, next);
2891 oldmm = prev->active_mm;
2893 * For paravirt, this is coupled with an exit in switch_to to
2894 * combine the page table reload and the switch backend into
2897 arch_start_context_switch(prev);
2900 next->active_mm = oldmm;
2901 atomic_inc(&oldmm->mm_count);
2902 enter_lazy_tlb(oldmm, next);
2904 switch_mm(oldmm, mm, next);
2906 if (likely(!prev->mm)) {
2907 prev->active_mm = NULL;
2908 rq->prev_mm = oldmm;
2911 * Since the runqueue lock will be released by the next
2912 * task (which is an invalid locking op but in the case
2913 * of the scheduler it's an obvious special-case), so we
2914 * do an early lockdep release here:
2916 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2917 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2920 /* Here we just switch the register state and the stack. */
2921 switch_to(prev, next, prev);
2925 * this_rq must be evaluated again because prev may have moved
2926 * CPUs since it called schedule(), thus the 'rq' on its stack
2927 * frame will be invalid.
2929 finish_task_switch(this_rq(), prev);
2933 * nr_running, nr_uninterruptible and nr_context_switches:
2935 * externally visible scheduler statistics: current number of runnable
2936 * threads, current number of uninterruptible-sleeping threads, total
2937 * number of context switches performed since bootup.
2939 unsigned long nr_running(void)
2941 unsigned long i, sum = 0;
2943 for_each_online_cpu(i)
2944 sum += cpu_rq(i)->nr_running;
2949 unsigned long nr_uninterruptible(void)
2951 unsigned long i, sum = 0;
2953 for_each_possible_cpu(i)
2954 sum += cpu_rq(i)->nr_uninterruptible;
2957 * Since we read the counters lockless, it might be slightly
2958 * inaccurate. Do not allow it to go below zero though:
2960 if (unlikely((long)sum < 0))
2966 unsigned long long nr_context_switches(void)
2969 unsigned long long sum = 0;
2971 for_each_possible_cpu(i)
2972 sum += cpu_rq(i)->nr_switches;
2977 unsigned long nr_iowait(void)
2979 unsigned long i, sum = 0;
2981 for_each_possible_cpu(i)
2982 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2987 unsigned long nr_iowait_cpu(void)
2989 struct rq *this = this_rq();
2990 return atomic_read(&this->nr_iowait);
2993 unsigned long this_cpu_load(void)
2995 struct rq *this = this_rq();
2996 return this->cpu_load[0];
3000 /* Variables and functions for calc_load */
3001 static atomic_long_t calc_load_tasks;
3002 static unsigned long calc_load_update;
3003 unsigned long avenrun[3];
3004 EXPORT_SYMBOL(avenrun);
3007 * get_avenrun - get the load average array
3008 * @loads: pointer to dest load array
3009 * @offset: offset to add
3010 * @shift: shift count to shift the result left
3012 * These values are estimates at best, so no need for locking.
3014 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3016 loads[0] = (avenrun[0] + offset) << shift;
3017 loads[1] = (avenrun[1] + offset) << shift;
3018 loads[2] = (avenrun[2] + offset) << shift;
3021 static unsigned long
3022 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3025 load += active * (FIXED_1 - exp);
3026 return load >> FSHIFT;
3030 * calc_load - update the avenrun load estimates 10 ticks after the
3031 * CPUs have updated calc_load_tasks.
3033 void calc_global_load(void)
3035 unsigned long upd = calc_load_update + 10;
3038 if (time_before(jiffies, upd))
3041 active = atomic_long_read(&calc_load_tasks);
3042 active = active > 0 ? active * FIXED_1 : 0;
3044 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3045 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3046 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3048 calc_load_update += LOAD_FREQ;
3052 * Either called from update_cpu_load() or from a cpu going idle
3054 static void calc_load_account_active(struct rq *this_rq)
3056 long nr_active, delta;
3058 nr_active = this_rq->nr_running;
3059 nr_active += (long) this_rq->nr_uninterruptible;
3061 if (nr_active != this_rq->calc_load_active) {
3062 delta = nr_active - this_rq->calc_load_active;
3063 this_rq->calc_load_active = nr_active;
3064 atomic_long_add(delta, &calc_load_tasks);
3069 * Update rq->cpu_load[] statistics. This function is usually called every
3070 * scheduler tick (TICK_NSEC).
3072 static void update_cpu_load(struct rq *this_rq)
3074 unsigned long this_load = this_rq->load.weight;
3077 this_rq->nr_load_updates++;
3079 /* Update our load: */
3080 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3081 unsigned long old_load, new_load;
3083 /* scale is effectively 1 << i now, and >> i divides by scale */
3085 old_load = this_rq->cpu_load[i];
3086 new_load = this_load;
3088 * Round up the averaging division if load is increasing. This
3089 * prevents us from getting stuck on 9 if the load is 10, for
3092 if (new_load > old_load)
3093 new_load += scale-1;
3094 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3097 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3098 this_rq->calc_load_update += LOAD_FREQ;
3099 calc_load_account_active(this_rq);
3106 * sched_exec - execve() is a valuable balancing opportunity, because at
3107 * this point the task has the smallest effective memory and cache footprint.
3109 void sched_exec(void)
3111 struct task_struct *p = current;
3112 struct migration_req req;
3113 int dest_cpu, this_cpu;
3114 unsigned long flags;
3118 this_cpu = get_cpu();
3119 dest_cpu = select_task_rq(p, SD_BALANCE_EXEC, 0);
3120 if (dest_cpu == this_cpu) {
3125 rq = task_rq_lock(p, &flags);
3129 * select_task_rq() can race against ->cpus_allowed
3131 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
3132 || unlikely(!cpu_active(dest_cpu))) {
3133 task_rq_unlock(rq, &flags);
3137 /* force the process onto the specified CPU */
3138 if (migrate_task(p, dest_cpu, &req)) {
3139 /* Need to wait for migration thread (might exit: take ref). */
3140 struct task_struct *mt = rq->migration_thread;
3142 get_task_struct(mt);
3143 task_rq_unlock(rq, &flags);
3144 wake_up_process(mt);
3145 put_task_struct(mt);
3146 wait_for_completion(&req.done);
3150 task_rq_unlock(rq, &flags);
3155 DEFINE_PER_CPU(struct kernel_stat, kstat);
3157 EXPORT_PER_CPU_SYMBOL(kstat);
3160 * Return any ns on the sched_clock that have not yet been accounted in
3161 * @p in case that task is currently running.
3163 * Called with task_rq_lock() held on @rq.
3165 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3169 if (task_current(rq, p)) {
3170 update_rq_clock(rq);
3171 ns = rq->clock - p->se.exec_start;
3179 unsigned long long task_delta_exec(struct task_struct *p)
3181 unsigned long flags;
3185 rq = task_rq_lock(p, &flags);
3186 ns = do_task_delta_exec(p, rq);
3187 task_rq_unlock(rq, &flags);
3193 * Return accounted runtime for the task.
3194 * In case the task is currently running, return the runtime plus current's
3195 * pending runtime that have not been accounted yet.
3197 unsigned long long task_sched_runtime(struct task_struct *p)
3199 unsigned long flags;
3203 rq = task_rq_lock(p, &flags);
3204 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3205 task_rq_unlock(rq, &flags);
3211 * Return sum_exec_runtime for the thread group.
3212 * In case the task is currently running, return the sum plus current's
3213 * pending runtime that have not been accounted yet.
3215 * Note that the thread group might have other running tasks as well,
3216 * so the return value not includes other pending runtime that other
3217 * running tasks might have.
3219 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3221 struct task_cputime totals;
3222 unsigned long flags;
3226 rq = task_rq_lock(p, &flags);
3227 thread_group_cputime(p, &totals);
3228 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3229 task_rq_unlock(rq, &flags);
3235 * Account user cpu time to a process.
3236 * @p: the process that the cpu time gets accounted to
3237 * @cputime: the cpu time spent in user space since the last update
3238 * @cputime_scaled: cputime scaled by cpu frequency
3240 void account_user_time(struct task_struct *p, cputime_t cputime,
3241 cputime_t cputime_scaled)
3243 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3246 /* Add user time to process. */
3247 p->utime = cputime_add(p->utime, cputime);
3248 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3249 account_group_user_time(p, cputime);
3251 /* Add user time to cpustat. */
3252 tmp = cputime_to_cputime64(cputime);
3253 if (TASK_NICE(p) > 0)
3254 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3256 cpustat->user = cputime64_add(cpustat->user, tmp);
3258 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3259 /* Account for user time used */
3260 acct_update_integrals(p);
3264 * Account guest cpu time to a process.
3265 * @p: the process that the cpu time gets accounted to
3266 * @cputime: the cpu time spent in virtual machine since the last update
3267 * @cputime_scaled: cputime scaled by cpu frequency
3269 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3270 cputime_t cputime_scaled)
3273 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3275 tmp = cputime_to_cputime64(cputime);
3277 /* Add guest time to process. */
3278 p->utime = cputime_add(p->utime, cputime);
3279 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3280 account_group_user_time(p, cputime);
3281 p->gtime = cputime_add(p->gtime, cputime);
3283 /* Add guest time to cpustat. */
3284 if (TASK_NICE(p) > 0) {
3285 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3286 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3288 cpustat->user = cputime64_add(cpustat->user, tmp);
3289 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3294 * Account system cpu time to a process.
3295 * @p: the process that the cpu time gets accounted to
3296 * @hardirq_offset: the offset to subtract from hardirq_count()
3297 * @cputime: the cpu time spent in kernel space since the last update
3298 * @cputime_scaled: cputime scaled by cpu frequency
3300 void account_system_time(struct task_struct *p, int hardirq_offset,
3301 cputime_t cputime, cputime_t cputime_scaled)
3303 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3306 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3307 account_guest_time(p, cputime, cputime_scaled);
3311 /* Add system time to process. */
3312 p->stime = cputime_add(p->stime, cputime);
3313 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3314 account_group_system_time(p, cputime);
3316 /* Add system time to cpustat. */
3317 tmp = cputime_to_cputime64(cputime);
3318 if (hardirq_count() - hardirq_offset)
3319 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3320 else if (softirq_count())
3321 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3323 cpustat->system = cputime64_add(cpustat->system, tmp);
3325 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3327 /* Account for system time used */
3328 acct_update_integrals(p);
3332 * Account for involuntary wait time.
3333 * @steal: the cpu time spent in involuntary wait
3335 void account_steal_time(cputime_t cputime)
3337 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3338 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3340 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3344 * Account for idle time.
3345 * @cputime: the cpu time spent in idle wait
3347 void account_idle_time(cputime_t cputime)
3349 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3350 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3351 struct rq *rq = this_rq();
3353 if (atomic_read(&rq->nr_iowait) > 0)
3354 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3356 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3359 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3362 * Account a single tick of cpu time.
3363 * @p: the process that the cpu time gets accounted to
3364 * @user_tick: indicates if the tick is a user or a system tick
3366 void account_process_tick(struct task_struct *p, int user_tick)
3368 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3369 struct rq *rq = this_rq();
3372 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3373 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3374 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3377 account_idle_time(cputime_one_jiffy);
3381 * Account multiple ticks of steal time.
3382 * @p: the process from which the cpu time has been stolen
3383 * @ticks: number of stolen ticks
3385 void account_steal_ticks(unsigned long ticks)
3387 account_steal_time(jiffies_to_cputime(ticks));
3391 * Account multiple ticks of idle time.
3392 * @ticks: number of stolen ticks
3394 void account_idle_ticks(unsigned long ticks)
3396 account_idle_time(jiffies_to_cputime(ticks));
3402 * Use precise platform statistics if available:
3404 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3405 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3411 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3413 struct task_cputime cputime;
3415 thread_group_cputime(p, &cputime);
3417 *ut = cputime.utime;
3418 *st = cputime.stime;
3422 #ifndef nsecs_to_cputime
3423 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3426 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3428 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
3431 * Use CFS's precise accounting:
3433 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3438 temp = (u64)(rtime * utime);
3439 do_div(temp, total);
3440 utime = (cputime_t)temp;
3445 * Compare with previous values, to keep monotonicity:
3447 p->prev_utime = max(p->prev_utime, utime);
3448 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
3450 *ut = p->prev_utime;
3451 *st = p->prev_stime;
3455 * Must be called with siglock held.
3457 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3459 struct signal_struct *sig = p->signal;
3460 struct task_cputime cputime;
3461 cputime_t rtime, utime, total;
3463 thread_group_cputime(p, &cputime);
3465 total = cputime_add(cputime.utime, cputime.stime);
3466 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3471 temp = (u64)(rtime * cputime.utime);
3472 do_div(temp, total);
3473 utime = (cputime_t)temp;
3477 sig->prev_utime = max(sig->prev_utime, utime);
3478 sig->prev_stime = max(sig->prev_stime,
3479 cputime_sub(rtime, sig->prev_utime));
3481 *ut = sig->prev_utime;
3482 *st = sig->prev_stime;
3487 * This function gets called by the timer code, with HZ frequency.
3488 * We call it with interrupts disabled.
3490 * It also gets called by the fork code, when changing the parent's
3493 void scheduler_tick(void)
3495 int cpu = smp_processor_id();
3496 struct rq *rq = cpu_rq(cpu);
3497 struct task_struct *curr = rq->curr;
3501 raw_spin_lock(&rq->lock);
3502 update_rq_clock(rq);
3503 update_cpu_load(rq);
3504 curr->sched_class->task_tick(rq, curr, 0);
3505 raw_spin_unlock(&rq->lock);
3507 perf_event_task_tick(curr, cpu);
3510 rq->idle_at_tick = idle_cpu(cpu);
3511 trigger_load_balance(rq, cpu);
3515 notrace unsigned long get_parent_ip(unsigned long addr)
3517 if (in_lock_functions(addr)) {
3518 addr = CALLER_ADDR2;
3519 if (in_lock_functions(addr))
3520 addr = CALLER_ADDR3;
3525 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3526 defined(CONFIG_PREEMPT_TRACER))
3528 void __kprobes add_preempt_count(int val)
3530 #ifdef CONFIG_DEBUG_PREEMPT
3534 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3537 preempt_count() += val;
3538 #ifdef CONFIG_DEBUG_PREEMPT
3540 * Spinlock count overflowing soon?
3542 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3545 if (preempt_count() == val)
3546 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3548 EXPORT_SYMBOL(add_preempt_count);
3550 void __kprobes sub_preempt_count(int val)
3552 #ifdef CONFIG_DEBUG_PREEMPT
3556 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3559 * Is the spinlock portion underflowing?
3561 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3562 !(preempt_count() & PREEMPT_MASK)))
3566 if (preempt_count() == val)
3567 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3568 preempt_count() -= val;
3570 EXPORT_SYMBOL(sub_preempt_count);
3575 * Print scheduling while atomic bug:
3577 static noinline void __schedule_bug(struct task_struct *prev)
3579 struct pt_regs *regs = get_irq_regs();
3581 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3582 prev->comm, prev->pid, preempt_count());
3584 debug_show_held_locks(prev);
3586 if (irqs_disabled())
3587 print_irqtrace_events(prev);
3596 * Various schedule()-time debugging checks and statistics:
3598 static inline void schedule_debug(struct task_struct *prev)
3601 * Test if we are atomic. Since do_exit() needs to call into
3602 * schedule() atomically, we ignore that path for now.
3603 * Otherwise, whine if we are scheduling when we should not be.
3605 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3606 __schedule_bug(prev);
3608 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3610 schedstat_inc(this_rq(), sched_count);
3611 #ifdef CONFIG_SCHEDSTATS
3612 if (unlikely(prev->lock_depth >= 0)) {
3613 schedstat_inc(this_rq(), bkl_count);
3614 schedstat_inc(prev, sched_info.bkl_count);
3619 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3621 if (prev->state == TASK_RUNNING) {
3622 u64 runtime = prev->se.sum_exec_runtime;
3624 runtime -= prev->se.prev_sum_exec_runtime;
3625 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
3628 * In order to avoid avg_overlap growing stale when we are
3629 * indeed overlapping and hence not getting put to sleep, grow
3630 * the avg_overlap on preemption.
3632 * We use the average preemption runtime because that
3633 * correlates to the amount of cache footprint a task can
3636 update_avg(&prev->se.avg_overlap, runtime);
3638 prev->sched_class->put_prev_task(rq, prev);
3642 * Pick up the highest-prio task:
3644 static inline struct task_struct *
3645 pick_next_task(struct rq *rq)
3647 const struct sched_class *class;
3648 struct task_struct *p;
3651 * Optimization: we know that if all tasks are in
3652 * the fair class we can call that function directly:
3654 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3655 p = fair_sched_class.pick_next_task(rq);
3660 class = sched_class_highest;
3662 p = class->pick_next_task(rq);
3666 * Will never be NULL as the idle class always
3667 * returns a non-NULL p:
3669 class = class->next;
3674 * schedule() is the main scheduler function.
3676 asmlinkage void __sched schedule(void)
3678 struct task_struct *prev, *next;
3679 unsigned long *switch_count;
3685 cpu = smp_processor_id();
3689 switch_count = &prev->nivcsw;
3691 release_kernel_lock(prev);
3692 need_resched_nonpreemptible:
3694 schedule_debug(prev);
3696 if (sched_feat(HRTICK))
3699 raw_spin_lock_irq(&rq->lock);
3700 update_rq_clock(rq);
3701 clear_tsk_need_resched(prev);
3703 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3704 if (unlikely(signal_pending_state(prev->state, prev)))
3705 prev->state = TASK_RUNNING;
3707 deactivate_task(rq, prev, 1);
3708 switch_count = &prev->nvcsw;
3711 pre_schedule(rq, prev);
3713 if (unlikely(!rq->nr_running))
3714 idle_balance(cpu, rq);
3716 put_prev_task(rq, prev);
3717 next = pick_next_task(rq);
3719 if (likely(prev != next)) {
3720 sched_info_switch(prev, next);
3721 perf_event_task_sched_out(prev, next, cpu);
3727 context_switch(rq, prev, next); /* unlocks the rq */
3729 * the context switch might have flipped the stack from under
3730 * us, hence refresh the local variables.
3732 cpu = smp_processor_id();
3735 raw_spin_unlock_irq(&rq->lock);
3739 if (unlikely(reacquire_kernel_lock(current) < 0)) {
3741 switch_count = &prev->nivcsw;
3742 goto need_resched_nonpreemptible;
3745 preempt_enable_no_resched();
3749 EXPORT_SYMBOL(schedule);
3751 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3753 * Look out! "owner" is an entirely speculative pointer
3754 * access and not reliable.
3756 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
3761 if (!sched_feat(OWNER_SPIN))
3764 #ifdef CONFIG_DEBUG_PAGEALLOC
3766 * Need to access the cpu field knowing that
3767 * DEBUG_PAGEALLOC could have unmapped it if
3768 * the mutex owner just released it and exited.
3770 if (probe_kernel_address(&owner->cpu, cpu))
3777 * Even if the access succeeded (likely case),
3778 * the cpu field may no longer be valid.
3780 if (cpu >= nr_cpumask_bits)
3784 * We need to validate that we can do a
3785 * get_cpu() and that we have the percpu area.
3787 if (!cpu_online(cpu))
3794 * Owner changed, break to re-assess state.
3796 if (lock->owner != owner)
3800 * Is that owner really running on that cpu?
3802 if (task_thread_info(rq->curr) != owner || need_resched())
3812 #ifdef CONFIG_PREEMPT
3814 * this is the entry point to schedule() from in-kernel preemption
3815 * off of preempt_enable. Kernel preemptions off return from interrupt
3816 * occur there and call schedule directly.
3818 asmlinkage void __sched preempt_schedule(void)
3820 struct thread_info *ti = current_thread_info();
3823 * If there is a non-zero preempt_count or interrupts are disabled,
3824 * we do not want to preempt the current task. Just return..
3826 if (likely(ti->preempt_count || irqs_disabled()))
3830 add_preempt_count(PREEMPT_ACTIVE);
3832 sub_preempt_count(PREEMPT_ACTIVE);
3835 * Check again in case we missed a preemption opportunity
3836 * between schedule and now.
3839 } while (need_resched());
3841 EXPORT_SYMBOL(preempt_schedule);
3844 * this is the entry point to schedule() from kernel preemption
3845 * off of irq context.
3846 * Note, that this is called and return with irqs disabled. This will
3847 * protect us against recursive calling from irq.
3849 asmlinkage void __sched preempt_schedule_irq(void)
3851 struct thread_info *ti = current_thread_info();
3853 /* Catch callers which need to be fixed */
3854 BUG_ON(ti->preempt_count || !irqs_disabled());
3857 add_preempt_count(PREEMPT_ACTIVE);
3860 local_irq_disable();
3861 sub_preempt_count(PREEMPT_ACTIVE);
3864 * Check again in case we missed a preemption opportunity
3865 * between schedule and now.
3868 } while (need_resched());
3871 #endif /* CONFIG_PREEMPT */
3873 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3876 return try_to_wake_up(curr->private, mode, wake_flags);
3878 EXPORT_SYMBOL(default_wake_function);
3881 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3882 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3883 * number) then we wake all the non-exclusive tasks and one exclusive task.
3885 * There are circumstances in which we can try to wake a task which has already
3886 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3887 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3889 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3890 int nr_exclusive, int wake_flags, void *key)
3892 wait_queue_t *curr, *next;
3894 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3895 unsigned flags = curr->flags;
3897 if (curr->func(curr, mode, wake_flags, key) &&
3898 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3904 * __wake_up - wake up threads blocked on a waitqueue.
3906 * @mode: which threads
3907 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3908 * @key: is directly passed to the wakeup function
3910 * It may be assumed that this function implies a write memory barrier before
3911 * changing the task state if and only if any tasks are woken up.
3913 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3914 int nr_exclusive, void *key)
3916 unsigned long flags;
3918 spin_lock_irqsave(&q->lock, flags);
3919 __wake_up_common(q, mode, nr_exclusive, 0, key);
3920 spin_unlock_irqrestore(&q->lock, flags);
3922 EXPORT_SYMBOL(__wake_up);
3925 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3927 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3929 __wake_up_common(q, mode, 1, 0, NULL);
3932 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3934 __wake_up_common(q, mode, 1, 0, key);
3938 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3940 * @mode: which threads
3941 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3942 * @key: opaque value to be passed to wakeup targets
3944 * The sync wakeup differs that the waker knows that it will schedule
3945 * away soon, so while the target thread will be woken up, it will not
3946 * be migrated to another CPU - ie. the two threads are 'synchronized'
3947 * with each other. This can prevent needless bouncing between CPUs.
3949 * On UP it can prevent extra preemption.
3951 * It may be assumed that this function implies a write memory barrier before
3952 * changing the task state if and only if any tasks are woken up.
3954 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3955 int nr_exclusive, void *key)
3957 unsigned long flags;
3958 int wake_flags = WF_SYNC;
3963 if (unlikely(!nr_exclusive))
3966 spin_lock_irqsave(&q->lock, flags);
3967 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3968 spin_unlock_irqrestore(&q->lock, flags);
3970 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3973 * __wake_up_sync - see __wake_up_sync_key()
3975 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3977 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
3979 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3982 * complete: - signals a single thread waiting on this completion
3983 * @x: holds the state of this particular completion
3985 * This will wake up a single thread waiting on this completion. Threads will be
3986 * awakened in the same order in which they were queued.
3988 * See also complete_all(), wait_for_completion() and related routines.
3990 * It may be assumed that this function implies a write memory barrier before
3991 * changing the task state if and only if any tasks are woken up.
3993 void complete(struct completion *x)
3995 unsigned long flags;
3997 spin_lock_irqsave(&x->wait.lock, flags);
3999 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4000 spin_unlock_irqrestore(&x->wait.lock, flags);
4002 EXPORT_SYMBOL(complete);
4005 * complete_all: - signals all threads waiting on this completion
4006 * @x: holds the state of this particular completion
4008 * This will wake up all threads waiting on this particular completion event.
4010 * It may be assumed that this function implies a write memory barrier before
4011 * changing the task state if and only if any tasks are woken up.
4013 void complete_all(struct completion *x)
4015 unsigned long flags;
4017 spin_lock_irqsave(&x->wait.lock, flags);
4018 x->done += UINT_MAX/2;
4019 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4020 spin_unlock_irqrestore(&x->wait.lock, flags);
4022 EXPORT_SYMBOL(complete_all);
4024 static inline long __sched
4025 do_wait_for_common(struct completion *x, long timeout, int state)
4028 DECLARE_WAITQUEUE(wait, current);
4030 wait.flags |= WQ_FLAG_EXCLUSIVE;
4031 __add_wait_queue_tail(&x->wait, &wait);
4033 if (signal_pending_state(state, current)) {
4034 timeout = -ERESTARTSYS;
4037 __set_current_state(state);
4038 spin_unlock_irq(&x->wait.lock);
4039 timeout = schedule_timeout(timeout);
4040 spin_lock_irq(&x->wait.lock);
4041 } while (!x->done && timeout);
4042 __remove_wait_queue(&x->wait, &wait);
4047 return timeout ?: 1;
4051 wait_for_common(struct completion *x, long timeout, int state)
4055 spin_lock_irq(&x->wait.lock);
4056 timeout = do_wait_for_common(x, timeout, state);
4057 spin_unlock_irq(&x->wait.lock);
4062 * wait_for_completion: - waits for completion of a task
4063 * @x: holds the state of this particular completion
4065 * This waits to be signaled for completion of a specific task. It is NOT
4066 * interruptible and there is no timeout.
4068 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4069 * and interrupt capability. Also see complete().
4071 void __sched wait_for_completion(struct completion *x)
4073 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4075 EXPORT_SYMBOL(wait_for_completion);
4078 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
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. The timeout is in jiffies. It is not
4086 unsigned long __sched
4087 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4089 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4091 EXPORT_SYMBOL(wait_for_completion_timeout);
4094 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4095 * @x: holds the state of this particular completion
4097 * This waits for completion of a specific task to be signaled. It is
4100 int __sched wait_for_completion_interruptible(struct completion *x)
4102 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4103 if (t == -ERESTARTSYS)
4107 EXPORT_SYMBOL(wait_for_completion_interruptible);
4110 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4111 * @x: holds the state of this particular completion
4112 * @timeout: timeout value in jiffies
4114 * This waits for either a completion of a specific task to be signaled or for a
4115 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4117 unsigned long __sched
4118 wait_for_completion_interruptible_timeout(struct completion *x,
4119 unsigned long timeout)
4121 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4123 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4126 * wait_for_completion_killable: - waits for completion of a task (killable)
4127 * @x: holds the state of this particular completion
4129 * This waits to be signaled for completion of a specific task. It can be
4130 * interrupted by a kill signal.
4132 int __sched wait_for_completion_killable(struct completion *x)
4134 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4135 if (t == -ERESTARTSYS)
4139 EXPORT_SYMBOL(wait_for_completion_killable);
4142 * try_wait_for_completion - try to decrement a completion without blocking
4143 * @x: completion structure
4145 * Returns: 0 if a decrement cannot be done without blocking
4146 * 1 if a decrement succeeded.
4148 * If a completion is being used as a counting completion,
4149 * attempt to decrement the counter without blocking. This
4150 * enables us to avoid waiting if the resource the completion
4151 * is protecting is not available.
4153 bool try_wait_for_completion(struct completion *x)
4155 unsigned long flags;
4158 spin_lock_irqsave(&x->wait.lock, flags);
4163 spin_unlock_irqrestore(&x->wait.lock, flags);
4166 EXPORT_SYMBOL(try_wait_for_completion);
4169 * completion_done - Test to see if a completion has any waiters
4170 * @x: completion structure
4172 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4173 * 1 if there are no waiters.
4176 bool completion_done(struct completion *x)
4178 unsigned long flags;
4181 spin_lock_irqsave(&x->wait.lock, flags);
4184 spin_unlock_irqrestore(&x->wait.lock, flags);
4187 EXPORT_SYMBOL(completion_done);
4190 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4192 unsigned long flags;
4195 init_waitqueue_entry(&wait, current);
4197 __set_current_state(state);
4199 spin_lock_irqsave(&q->lock, flags);
4200 __add_wait_queue(q, &wait);
4201 spin_unlock(&q->lock);
4202 timeout = schedule_timeout(timeout);
4203 spin_lock_irq(&q->lock);
4204 __remove_wait_queue(q, &wait);
4205 spin_unlock_irqrestore(&q->lock, flags);
4210 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4212 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4214 EXPORT_SYMBOL(interruptible_sleep_on);
4217 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4219 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4221 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4223 void __sched sleep_on(wait_queue_head_t *q)
4225 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4227 EXPORT_SYMBOL(sleep_on);
4229 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4231 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4233 EXPORT_SYMBOL(sleep_on_timeout);
4235 #ifdef CONFIG_RT_MUTEXES
4238 * rt_mutex_setprio - set the current priority of a task
4240 * @prio: prio value (kernel-internal form)
4242 * This function changes the 'effective' priority of a task. It does
4243 * not touch ->normal_prio like __setscheduler().
4245 * Used by the rt_mutex code to implement priority inheritance logic.
4247 void rt_mutex_setprio(struct task_struct *p, int prio)
4249 unsigned long flags;
4250 int oldprio, on_rq, running;
4252 const struct sched_class *prev_class = p->sched_class;
4254 BUG_ON(prio < 0 || prio > MAX_PRIO);
4256 rq = task_rq_lock(p, &flags);
4257 update_rq_clock(rq);
4260 on_rq = p->se.on_rq;
4261 running = task_current(rq, p);
4263 dequeue_task(rq, p, 0);
4265 p->sched_class->put_prev_task(rq, p);
4268 p->sched_class = &rt_sched_class;
4270 p->sched_class = &fair_sched_class;
4275 p->sched_class->set_curr_task(rq);
4277 enqueue_task(rq, p, 0, oldprio < prio);
4279 check_class_changed(rq, p, prev_class, oldprio, running);
4281 task_rq_unlock(rq, &flags);
4286 void set_user_nice(struct task_struct *p, long nice)
4288 int old_prio, delta, on_rq;
4289 unsigned long flags;
4292 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4295 * We have to be careful, if called from sys_setpriority(),
4296 * the task might be in the middle of scheduling on another CPU.
4298 rq = task_rq_lock(p, &flags);
4299 update_rq_clock(rq);
4301 * The RT priorities are set via sched_setscheduler(), but we still
4302 * allow the 'normal' nice value to be set - but as expected
4303 * it wont have any effect on scheduling until the task is
4304 * SCHED_FIFO/SCHED_RR:
4306 if (task_has_rt_policy(p)) {
4307 p->static_prio = NICE_TO_PRIO(nice);
4310 on_rq = p->se.on_rq;
4312 dequeue_task(rq, p, 0);
4314 p->static_prio = NICE_TO_PRIO(nice);
4317 p->prio = effective_prio(p);
4318 delta = p->prio - old_prio;
4321 enqueue_task(rq, p, 0, false);
4323 * If the task increased its priority or is running and
4324 * lowered its priority, then reschedule its CPU:
4326 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4327 resched_task(rq->curr);
4330 task_rq_unlock(rq, &flags);
4332 EXPORT_SYMBOL(set_user_nice);
4335 * can_nice - check if a task can reduce its nice value
4339 int can_nice(const struct task_struct *p, const int nice)
4341 /* convert nice value [19,-20] to rlimit style value [1,40] */
4342 int nice_rlim = 20 - nice;
4344 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4345 capable(CAP_SYS_NICE));
4348 #ifdef __ARCH_WANT_SYS_NICE
4351 * sys_nice - change the priority of the current process.
4352 * @increment: priority increment
4354 * sys_setpriority is a more generic, but much slower function that
4355 * does similar things.
4357 SYSCALL_DEFINE1(nice, int, increment)
4362 * Setpriority might change our priority at the same moment.
4363 * We don't have to worry. Conceptually one call occurs first
4364 * and we have a single winner.
4366 if (increment < -40)
4371 nice = TASK_NICE(current) + increment;
4377 if (increment < 0 && !can_nice(current, nice))
4380 retval = security_task_setnice(current, nice);
4384 set_user_nice(current, nice);
4391 * task_prio - return the priority value of a given task.
4392 * @p: the task in question.
4394 * This is the priority value as seen by users in /proc.
4395 * RT tasks are offset by -200. Normal tasks are centered
4396 * around 0, value goes from -16 to +15.
4398 int task_prio(const struct task_struct *p)
4400 return p->prio - MAX_RT_PRIO;
4404 * task_nice - return the nice value of a given task.
4405 * @p: the task in question.
4407 int task_nice(const struct task_struct *p)
4409 return TASK_NICE(p);
4411 EXPORT_SYMBOL(task_nice);
4414 * idle_cpu - is a given cpu idle currently?
4415 * @cpu: the processor in question.
4417 int idle_cpu(int cpu)
4419 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4423 * idle_task - return the idle task for a given cpu.
4424 * @cpu: the processor in question.
4426 struct task_struct *idle_task(int cpu)
4428 return cpu_rq(cpu)->idle;
4432 * find_process_by_pid - find a process with a matching PID value.
4433 * @pid: the pid in question.
4435 static struct task_struct *find_process_by_pid(pid_t pid)
4437 return pid ? find_task_by_vpid(pid) : current;
4440 /* Actually do priority change: must hold rq lock. */
4442 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4444 BUG_ON(p->se.on_rq);
4447 p->rt_priority = prio;
4448 p->normal_prio = normal_prio(p);
4449 /* we are holding p->pi_lock already */
4450 p->prio = rt_mutex_getprio(p);
4451 if (rt_prio(p->prio))
4452 p->sched_class = &rt_sched_class;
4454 p->sched_class = &fair_sched_class;
4459 * check the target process has a UID that matches the current process's
4461 static bool check_same_owner(struct task_struct *p)
4463 const struct cred *cred = current_cred(), *pcred;
4467 pcred = __task_cred(p);
4468 match = (cred->euid == pcred->euid ||
4469 cred->euid == pcred->uid);
4474 static int __sched_setscheduler(struct task_struct *p, int policy,
4475 struct sched_param *param, bool user)
4477 int retval, oldprio, oldpolicy = -1, on_rq, running;
4478 unsigned long flags;
4479 const struct sched_class *prev_class = p->sched_class;
4483 /* may grab non-irq protected spin_locks */
4484 BUG_ON(in_interrupt());
4486 /* double check policy once rq lock held */
4488 reset_on_fork = p->sched_reset_on_fork;
4489 policy = oldpolicy = p->policy;
4491 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4492 policy &= ~SCHED_RESET_ON_FORK;
4494 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4495 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4496 policy != SCHED_IDLE)
4501 * Valid priorities for SCHED_FIFO and SCHED_RR are
4502 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4503 * SCHED_BATCH and SCHED_IDLE is 0.
4505 if (param->sched_priority < 0 ||
4506 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4507 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4509 if (rt_policy(policy) != (param->sched_priority != 0))
4513 * Allow unprivileged RT tasks to decrease priority:
4515 if (user && !capable(CAP_SYS_NICE)) {
4516 if (rt_policy(policy)) {
4517 unsigned long rlim_rtprio;
4519 if (!lock_task_sighand(p, &flags))
4521 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4522 unlock_task_sighand(p, &flags);
4524 /* can't set/change the rt policy */
4525 if (policy != p->policy && !rlim_rtprio)
4528 /* can't increase priority */
4529 if (param->sched_priority > p->rt_priority &&
4530 param->sched_priority > rlim_rtprio)
4534 * Like positive nice levels, dont allow tasks to
4535 * move out of SCHED_IDLE either:
4537 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4540 /* can't change other user's priorities */
4541 if (!check_same_owner(p))
4544 /* Normal users shall not reset the sched_reset_on_fork flag */
4545 if (p->sched_reset_on_fork && !reset_on_fork)
4550 #ifdef CONFIG_RT_GROUP_SCHED
4552 * Do not allow realtime tasks into groups that have no runtime
4555 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4556 task_group(p)->rt_bandwidth.rt_runtime == 0)
4560 retval = security_task_setscheduler(p, policy, param);
4566 * make sure no PI-waiters arrive (or leave) while we are
4567 * changing the priority of the task:
4569 raw_spin_lock_irqsave(&p->pi_lock, flags);
4571 * To be able to change p->policy safely, the apropriate
4572 * runqueue lock must be held.
4574 rq = __task_rq_lock(p);
4575 /* recheck policy now with rq lock held */
4576 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4577 policy = oldpolicy = -1;
4578 __task_rq_unlock(rq);
4579 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4582 update_rq_clock(rq);
4583 on_rq = p->se.on_rq;
4584 running = task_current(rq, p);
4586 deactivate_task(rq, p, 0);
4588 p->sched_class->put_prev_task(rq, p);
4590 p->sched_reset_on_fork = reset_on_fork;
4593 __setscheduler(rq, p, policy, param->sched_priority);
4596 p->sched_class->set_curr_task(rq);
4598 activate_task(rq, p, 0);
4600 check_class_changed(rq, p, prev_class, oldprio, running);
4602 __task_rq_unlock(rq);
4603 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4605 rt_mutex_adjust_pi(p);
4611 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4612 * @p: the task in question.
4613 * @policy: new policy.
4614 * @param: structure containing the new RT priority.
4616 * NOTE that the task may be already dead.
4618 int sched_setscheduler(struct task_struct *p, int policy,
4619 struct sched_param *param)
4621 return __sched_setscheduler(p, policy, param, true);
4623 EXPORT_SYMBOL_GPL(sched_setscheduler);
4626 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4627 * @p: the task in question.
4628 * @policy: new policy.
4629 * @param: structure containing the new RT priority.
4631 * Just like sched_setscheduler, only don't bother checking if the
4632 * current context has permission. For example, this is needed in
4633 * stop_machine(): we create temporary high priority worker threads,
4634 * but our caller might not have that capability.
4636 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4637 struct sched_param *param)
4639 return __sched_setscheduler(p, policy, param, false);
4643 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4645 struct sched_param lparam;
4646 struct task_struct *p;
4649 if (!param || pid < 0)
4651 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4656 p = find_process_by_pid(pid);
4658 retval = sched_setscheduler(p, policy, &lparam);
4665 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4666 * @pid: the pid in question.
4667 * @policy: new policy.
4668 * @param: structure containing the new RT priority.
4670 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4671 struct sched_param __user *, param)
4673 /* negative values for policy are not valid */
4677 return do_sched_setscheduler(pid, policy, param);
4681 * sys_sched_setparam - set/change the RT priority of a thread
4682 * @pid: the pid in question.
4683 * @param: structure containing the new RT priority.
4685 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4687 return do_sched_setscheduler(pid, -1, param);
4691 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4692 * @pid: the pid in question.
4694 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4696 struct task_struct *p;
4704 p = find_process_by_pid(pid);
4706 retval = security_task_getscheduler(p);
4709 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4716 * sys_sched_getparam - get the RT priority of a thread
4717 * @pid: the pid in question.
4718 * @param: structure containing the RT priority.
4720 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4722 struct sched_param lp;
4723 struct task_struct *p;
4726 if (!param || pid < 0)
4730 p = find_process_by_pid(pid);
4735 retval = security_task_getscheduler(p);
4739 lp.sched_priority = p->rt_priority;
4743 * This one might sleep, we cannot do it with a spinlock held ...
4745 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4754 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4756 cpumask_var_t cpus_allowed, new_mask;
4757 struct task_struct *p;
4763 p = find_process_by_pid(pid);
4770 /* Prevent p going away */
4774 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4778 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4780 goto out_free_cpus_allowed;
4783 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
4786 retval = security_task_setscheduler(p, 0, NULL);
4790 cpuset_cpus_allowed(p, cpus_allowed);
4791 cpumask_and(new_mask, in_mask, cpus_allowed);
4793 retval = set_cpus_allowed_ptr(p, new_mask);
4796 cpuset_cpus_allowed(p, cpus_allowed);
4797 if (!cpumask_subset(new_mask, cpus_allowed)) {
4799 * We must have raced with a concurrent cpuset
4800 * update. Just reset the cpus_allowed to the
4801 * cpuset's cpus_allowed
4803 cpumask_copy(new_mask, cpus_allowed);
4808 free_cpumask_var(new_mask);
4809 out_free_cpus_allowed:
4810 free_cpumask_var(cpus_allowed);
4817 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4818 struct cpumask *new_mask)
4820 if (len < cpumask_size())
4821 cpumask_clear(new_mask);
4822 else if (len > cpumask_size())
4823 len = cpumask_size();
4825 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4829 * sys_sched_setaffinity - set the cpu affinity of a process
4830 * @pid: pid of the process
4831 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4832 * @user_mask_ptr: user-space pointer to the new cpu mask
4834 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4835 unsigned long __user *, user_mask_ptr)
4837 cpumask_var_t new_mask;
4840 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4843 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4845 retval = sched_setaffinity(pid, new_mask);
4846 free_cpumask_var(new_mask);
4850 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4852 struct task_struct *p;
4853 unsigned long flags;
4861 p = find_process_by_pid(pid);
4865 retval = security_task_getscheduler(p);
4869 rq = task_rq_lock(p, &flags);
4870 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4871 task_rq_unlock(rq, &flags);
4881 * sys_sched_getaffinity - get the cpu affinity of a process
4882 * @pid: pid of the process
4883 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4884 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4886 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4887 unsigned long __user *, user_mask_ptr)
4892 if (len < cpumask_size())
4895 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4898 ret = sched_getaffinity(pid, mask);
4900 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
4903 ret = cpumask_size();
4905 free_cpumask_var(mask);
4911 * sys_sched_yield - yield the current processor to other threads.
4913 * This function yields the current CPU to other tasks. If there are no
4914 * other threads running on this CPU then this function will return.
4916 SYSCALL_DEFINE0(sched_yield)
4918 struct rq *rq = this_rq_lock();
4920 schedstat_inc(rq, yld_count);
4921 current->sched_class->yield_task(rq);
4924 * Since we are going to call schedule() anyway, there's
4925 * no need to preempt or enable interrupts:
4927 __release(rq->lock);
4928 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4929 do_raw_spin_unlock(&rq->lock);
4930 preempt_enable_no_resched();
4937 static inline int should_resched(void)
4939 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4942 static void __cond_resched(void)
4944 add_preempt_count(PREEMPT_ACTIVE);
4946 sub_preempt_count(PREEMPT_ACTIVE);
4949 int __sched _cond_resched(void)
4951 if (should_resched()) {
4957 EXPORT_SYMBOL(_cond_resched);
4960 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4961 * call schedule, and on return reacquire the lock.
4963 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4964 * operations here to prevent schedule() from being called twice (once via
4965 * spin_unlock(), once by hand).
4967 int __cond_resched_lock(spinlock_t *lock)
4969 int resched = should_resched();
4972 lockdep_assert_held(lock);
4974 if (spin_needbreak(lock) || resched) {
4985 EXPORT_SYMBOL(__cond_resched_lock);
4987 int __sched __cond_resched_softirq(void)
4989 BUG_ON(!in_softirq());
4991 if (should_resched()) {
4999 EXPORT_SYMBOL(__cond_resched_softirq);
5002 * yield - yield the current processor to other threads.
5004 * This is a shortcut for kernel-space yielding - it marks the
5005 * thread runnable and calls sys_sched_yield().
5007 void __sched yield(void)
5009 set_current_state(TASK_RUNNING);
5012 EXPORT_SYMBOL(yield);
5015 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5016 * that process accounting knows that this is a task in IO wait state.
5018 void __sched io_schedule(void)
5020 struct rq *rq = raw_rq();
5022 delayacct_blkio_start();
5023 atomic_inc(&rq->nr_iowait);
5024 current->in_iowait = 1;
5026 current->in_iowait = 0;
5027 atomic_dec(&rq->nr_iowait);
5028 delayacct_blkio_end();
5030 EXPORT_SYMBOL(io_schedule);
5032 long __sched io_schedule_timeout(long timeout)
5034 struct rq *rq = raw_rq();
5037 delayacct_blkio_start();
5038 atomic_inc(&rq->nr_iowait);
5039 current->in_iowait = 1;
5040 ret = schedule_timeout(timeout);
5041 current->in_iowait = 0;
5042 atomic_dec(&rq->nr_iowait);
5043 delayacct_blkio_end();
5048 * sys_sched_get_priority_max - return maximum RT priority.
5049 * @policy: scheduling class.
5051 * this syscall returns the maximum rt_priority that can be used
5052 * by a given scheduling class.
5054 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5061 ret = MAX_USER_RT_PRIO-1;
5073 * sys_sched_get_priority_min - return minimum RT priority.
5074 * @policy: scheduling class.
5076 * this syscall returns the minimum rt_priority that can be used
5077 * by a given scheduling class.
5079 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5097 * sys_sched_rr_get_interval - return the default timeslice of a process.
5098 * @pid: pid of the process.
5099 * @interval: userspace pointer to the timeslice value.
5101 * this syscall writes the default timeslice value of a given process
5102 * into the user-space timespec buffer. A value of '0' means infinity.
5104 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5105 struct timespec __user *, interval)
5107 struct task_struct *p;
5108 unsigned int time_slice;
5109 unsigned long flags;
5119 p = find_process_by_pid(pid);
5123 retval = security_task_getscheduler(p);
5127 rq = task_rq_lock(p, &flags);
5128 time_slice = p->sched_class->get_rr_interval(rq, p);
5129 task_rq_unlock(rq, &flags);
5132 jiffies_to_timespec(time_slice, &t);
5133 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5141 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5143 void sched_show_task(struct task_struct *p)
5145 unsigned long free = 0;
5148 state = p->state ? __ffs(p->state) + 1 : 0;
5149 printk(KERN_INFO "%-13.13s %c", p->comm,
5150 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5151 #if BITS_PER_LONG == 32
5152 if (state == TASK_RUNNING)
5153 printk(KERN_CONT " running ");
5155 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5157 if (state == TASK_RUNNING)
5158 printk(KERN_CONT " running task ");
5160 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5162 #ifdef CONFIG_DEBUG_STACK_USAGE
5163 free = stack_not_used(p);
5165 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5166 task_pid_nr(p), task_pid_nr(p->real_parent),
5167 (unsigned long)task_thread_info(p)->flags);
5169 show_stack(p, NULL);
5172 void show_state_filter(unsigned long state_filter)
5174 struct task_struct *g, *p;
5176 #if BITS_PER_LONG == 32
5178 " task PC stack pid father\n");
5181 " task PC stack pid father\n");
5183 read_lock(&tasklist_lock);
5184 do_each_thread(g, p) {
5186 * reset the NMI-timeout, listing all files on a slow
5187 * console might take alot of time:
5189 touch_nmi_watchdog();
5190 if (!state_filter || (p->state & state_filter))
5192 } while_each_thread(g, p);
5194 touch_all_softlockup_watchdogs();
5196 #ifdef CONFIG_SCHED_DEBUG
5197 sysrq_sched_debug_show();
5199 read_unlock(&tasklist_lock);
5201 * Only show locks if all tasks are dumped:
5204 debug_show_all_locks();
5207 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5209 idle->sched_class = &idle_sched_class;
5213 * init_idle - set up an idle thread for a given CPU
5214 * @idle: task in question
5215 * @cpu: cpu the idle task belongs to
5217 * NOTE: this function does not set the idle thread's NEED_RESCHED
5218 * flag, to make booting more robust.
5220 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5222 struct rq *rq = cpu_rq(cpu);
5223 unsigned long flags;
5225 raw_spin_lock_irqsave(&rq->lock, flags);
5228 idle->state = TASK_RUNNING;
5229 idle->se.exec_start = sched_clock();
5231 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
5232 __set_task_cpu(idle, cpu);
5234 rq->curr = rq->idle = idle;
5235 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5238 raw_spin_unlock_irqrestore(&rq->lock, flags);
5240 /* Set the preempt count _outside_ the spinlocks! */
5241 #if defined(CONFIG_PREEMPT)
5242 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5244 task_thread_info(idle)->preempt_count = 0;
5247 * The idle tasks have their own, simple scheduling class:
5249 idle->sched_class = &idle_sched_class;
5250 ftrace_graph_init_task(idle);
5254 * In a system that switches off the HZ timer nohz_cpu_mask
5255 * indicates which cpus entered this state. This is used
5256 * in the rcu update to wait only for active cpus. For system
5257 * which do not switch off the HZ timer nohz_cpu_mask should
5258 * always be CPU_BITS_NONE.
5260 cpumask_var_t nohz_cpu_mask;
5263 * Increase the granularity value when there are more CPUs,
5264 * because with more CPUs the 'effective latency' as visible
5265 * to users decreases. But the relationship is not linear,
5266 * so pick a second-best guess by going with the log2 of the
5269 * This idea comes from the SD scheduler of Con Kolivas:
5271 static int get_update_sysctl_factor(void)
5273 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5274 unsigned int factor;
5276 switch (sysctl_sched_tunable_scaling) {
5277 case SCHED_TUNABLESCALING_NONE:
5280 case SCHED_TUNABLESCALING_LINEAR:
5283 case SCHED_TUNABLESCALING_LOG:
5285 factor = 1 + ilog2(cpus);
5292 static void update_sysctl(void)
5294 unsigned int factor = get_update_sysctl_factor();
5296 #define SET_SYSCTL(name) \
5297 (sysctl_##name = (factor) * normalized_sysctl_##name)
5298 SET_SYSCTL(sched_min_granularity);
5299 SET_SYSCTL(sched_latency);
5300 SET_SYSCTL(sched_wakeup_granularity);
5301 SET_SYSCTL(sched_shares_ratelimit);
5305 static inline void sched_init_granularity(void)
5312 * This is how migration works:
5314 * 1) we queue a struct migration_req structure in the source CPU's
5315 * runqueue and wake up that CPU's migration thread.
5316 * 2) we down() the locked semaphore => thread blocks.
5317 * 3) migration thread wakes up (implicitly it forces the migrated
5318 * thread off the CPU)
5319 * 4) it gets the migration request and checks whether the migrated
5320 * task is still in the wrong runqueue.
5321 * 5) if it's in the wrong runqueue then the migration thread removes
5322 * it and puts it into the right queue.
5323 * 6) migration thread up()s the semaphore.
5324 * 7) we wake up and the migration is done.
5328 * Change a given task's CPU affinity. Migrate the thread to a
5329 * proper CPU and schedule it away if the CPU it's executing on
5330 * is removed from the allowed bitmask.
5332 * NOTE: the caller must have a valid reference to the task, the
5333 * task must not exit() & deallocate itself prematurely. The
5334 * call is not atomic; no spinlocks may be held.
5336 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5338 struct migration_req req;
5339 unsigned long flags;
5343 rq = task_rq_lock(p, &flags);
5345 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5350 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5351 !cpumask_equal(&p->cpus_allowed, new_mask))) {
5356 if (p->sched_class->set_cpus_allowed)
5357 p->sched_class->set_cpus_allowed(p, new_mask);
5359 cpumask_copy(&p->cpus_allowed, new_mask);
5360 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5363 /* Can the task run on the task's current CPU? If so, we're done */
5364 if (cpumask_test_cpu(task_cpu(p), new_mask))
5367 if (migrate_task(p, cpumask_any_and(cpu_active_mask, new_mask), &req)) {
5368 /* Need help from migration thread: drop lock and wait. */
5369 struct task_struct *mt = rq->migration_thread;
5371 get_task_struct(mt);
5372 task_rq_unlock(rq, &flags);
5373 wake_up_process(rq->migration_thread);
5374 put_task_struct(mt);
5375 wait_for_completion(&req.done);
5376 tlb_migrate_finish(p->mm);
5380 task_rq_unlock(rq, &flags);
5384 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5387 * Move (not current) task off this cpu, onto dest cpu. We're doing
5388 * this because either it can't run here any more (set_cpus_allowed()
5389 * away from this CPU, or CPU going down), or because we're
5390 * attempting to rebalance this task on exec (sched_exec).
5392 * So we race with normal scheduler movements, but that's OK, as long
5393 * as the task is no longer on this CPU.
5395 * Returns non-zero if task was successfully migrated.
5397 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5399 struct rq *rq_dest, *rq_src;
5402 if (unlikely(!cpu_active(dest_cpu)))
5405 rq_src = cpu_rq(src_cpu);
5406 rq_dest = cpu_rq(dest_cpu);
5408 double_rq_lock(rq_src, rq_dest);
5409 /* Already moved. */
5410 if (task_cpu(p) != src_cpu)
5412 /* Affinity changed (again). */
5413 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
5417 * If we're not on a rq, the next wake-up will ensure we're
5421 deactivate_task(rq_src, p, 0);
5422 set_task_cpu(p, dest_cpu);
5423 activate_task(rq_dest, p, 0);
5424 check_preempt_curr(rq_dest, p, 0);
5429 double_rq_unlock(rq_src, rq_dest);
5433 #define RCU_MIGRATION_IDLE 0
5434 #define RCU_MIGRATION_NEED_QS 1
5435 #define RCU_MIGRATION_GOT_QS 2
5436 #define RCU_MIGRATION_MUST_SYNC 3
5439 * migration_thread - this is a highprio system thread that performs
5440 * thread migration by bumping thread off CPU then 'pushing' onto
5443 static int migration_thread(void *data)
5446 int cpu = (long)data;
5450 BUG_ON(rq->migration_thread != current);
5452 set_current_state(TASK_INTERRUPTIBLE);
5453 while (!kthread_should_stop()) {
5454 struct migration_req *req;
5455 struct list_head *head;
5457 raw_spin_lock_irq(&rq->lock);
5459 if (cpu_is_offline(cpu)) {
5460 raw_spin_unlock_irq(&rq->lock);
5464 if (rq->active_balance) {
5465 active_load_balance(rq, cpu);
5466 rq->active_balance = 0;
5469 head = &rq->migration_queue;
5471 if (list_empty(head)) {
5472 raw_spin_unlock_irq(&rq->lock);
5474 set_current_state(TASK_INTERRUPTIBLE);
5477 req = list_entry(head->next, struct migration_req, list);
5478 list_del_init(head->next);
5480 if (req->task != NULL) {
5481 raw_spin_unlock(&rq->lock);
5482 __migrate_task(req->task, cpu, req->dest_cpu);
5483 } else if (likely(cpu == (badcpu = smp_processor_id()))) {
5484 req->dest_cpu = RCU_MIGRATION_GOT_QS;
5485 raw_spin_unlock(&rq->lock);
5487 req->dest_cpu = RCU_MIGRATION_MUST_SYNC;
5488 raw_spin_unlock(&rq->lock);
5489 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu, cpu);
5493 complete(&req->done);
5495 __set_current_state(TASK_RUNNING);
5500 #ifdef CONFIG_HOTPLUG_CPU
5502 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5506 local_irq_disable();
5507 ret = __migrate_task(p, src_cpu, dest_cpu);
5513 * Figure out where task on dead CPU should go, use force if necessary.
5515 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5520 dest_cpu = select_fallback_rq(dead_cpu, p);
5522 /* It can have affinity changed while we were choosing. */
5523 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
5528 * While a dead CPU has no uninterruptible tasks queued at this point,
5529 * it might still have a nonzero ->nr_uninterruptible counter, because
5530 * for performance reasons the counter is not stricly tracking tasks to
5531 * their home CPUs. So we just add the counter to another CPU's counter,
5532 * to keep the global sum constant after CPU-down:
5534 static void migrate_nr_uninterruptible(struct rq *rq_src)
5536 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5537 unsigned long flags;
5539 local_irq_save(flags);
5540 double_rq_lock(rq_src, rq_dest);
5541 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5542 rq_src->nr_uninterruptible = 0;
5543 double_rq_unlock(rq_src, rq_dest);
5544 local_irq_restore(flags);
5547 /* Run through task list and migrate tasks from the dead cpu. */
5548 static void migrate_live_tasks(int src_cpu)
5550 struct task_struct *p, *t;
5552 read_lock(&tasklist_lock);
5554 do_each_thread(t, p) {
5558 if (task_cpu(p) == src_cpu)
5559 move_task_off_dead_cpu(src_cpu, p);
5560 } while_each_thread(t, p);
5562 read_unlock(&tasklist_lock);
5566 * Schedules idle task to be the next runnable task on current CPU.
5567 * It does so by boosting its priority to highest possible.
5568 * Used by CPU offline code.
5570 void sched_idle_next(void)
5572 int this_cpu = smp_processor_id();
5573 struct rq *rq = cpu_rq(this_cpu);
5574 struct task_struct *p = rq->idle;
5575 unsigned long flags;
5577 /* cpu has to be offline */
5578 BUG_ON(cpu_online(this_cpu));
5581 * Strictly not necessary since rest of the CPUs are stopped by now
5582 * and interrupts disabled on the current cpu.
5584 raw_spin_lock_irqsave(&rq->lock, flags);
5586 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5588 update_rq_clock(rq);
5589 activate_task(rq, p, 0);
5591 raw_spin_unlock_irqrestore(&rq->lock, flags);
5595 * Ensures that the idle task is using init_mm right before its cpu goes
5598 void idle_task_exit(void)
5600 struct mm_struct *mm = current->active_mm;
5602 BUG_ON(cpu_online(smp_processor_id()));
5605 switch_mm(mm, &init_mm, current);
5609 /* called under rq->lock with disabled interrupts */
5610 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5612 struct rq *rq = cpu_rq(dead_cpu);
5614 /* Must be exiting, otherwise would be on tasklist. */
5615 BUG_ON(!p->exit_state);
5617 /* Cannot have done final schedule yet: would have vanished. */
5618 BUG_ON(p->state == TASK_DEAD);
5623 * Drop lock around migration; if someone else moves it,
5624 * that's OK. No task can be added to this CPU, so iteration is
5627 raw_spin_unlock_irq(&rq->lock);
5628 move_task_off_dead_cpu(dead_cpu, p);
5629 raw_spin_lock_irq(&rq->lock);
5634 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5635 static void migrate_dead_tasks(unsigned int dead_cpu)
5637 struct rq *rq = cpu_rq(dead_cpu);
5638 struct task_struct *next;
5641 if (!rq->nr_running)
5643 update_rq_clock(rq);
5644 next = pick_next_task(rq);
5647 next->sched_class->put_prev_task(rq, next);
5648 migrate_dead(dead_cpu, next);
5654 * remove the tasks which were accounted by rq from calc_load_tasks.
5656 static void calc_global_load_remove(struct rq *rq)
5658 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5659 rq->calc_load_active = 0;
5661 #endif /* CONFIG_HOTPLUG_CPU */
5663 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5665 static struct ctl_table sd_ctl_dir[] = {
5667 .procname = "sched_domain",
5673 static struct ctl_table sd_ctl_root[] = {
5675 .procname = "kernel",
5677 .child = sd_ctl_dir,
5682 static struct ctl_table *sd_alloc_ctl_entry(int n)
5684 struct ctl_table *entry =
5685 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5690 static void sd_free_ctl_entry(struct ctl_table **tablep)
5692 struct ctl_table *entry;
5695 * In the intermediate directories, both the child directory and
5696 * procname are dynamically allocated and could fail but the mode
5697 * will always be set. In the lowest directory the names are
5698 * static strings and all have proc handlers.
5700 for (entry = *tablep; entry->mode; entry++) {
5702 sd_free_ctl_entry(&entry->child);
5703 if (entry->proc_handler == NULL)
5704 kfree(entry->procname);
5712 set_table_entry(struct ctl_table *entry,
5713 const char *procname, void *data, int maxlen,
5714 mode_t mode, proc_handler *proc_handler)
5716 entry->procname = procname;
5718 entry->maxlen = maxlen;
5720 entry->proc_handler = proc_handler;
5723 static struct ctl_table *
5724 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5726 struct ctl_table *table = sd_alloc_ctl_entry(13);
5731 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5732 sizeof(long), 0644, proc_doulongvec_minmax);
5733 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5734 sizeof(long), 0644, proc_doulongvec_minmax);
5735 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5736 sizeof(int), 0644, proc_dointvec_minmax);
5737 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5738 sizeof(int), 0644, proc_dointvec_minmax);
5739 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5740 sizeof(int), 0644, proc_dointvec_minmax);
5741 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5742 sizeof(int), 0644, proc_dointvec_minmax);
5743 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5744 sizeof(int), 0644, proc_dointvec_minmax);
5745 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5746 sizeof(int), 0644, proc_dointvec_minmax);
5747 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5748 sizeof(int), 0644, proc_dointvec_minmax);
5749 set_table_entry(&table[9], "cache_nice_tries",
5750 &sd->cache_nice_tries,
5751 sizeof(int), 0644, proc_dointvec_minmax);
5752 set_table_entry(&table[10], "flags", &sd->flags,
5753 sizeof(int), 0644, proc_dointvec_minmax);
5754 set_table_entry(&table[11], "name", sd->name,
5755 CORENAME_MAX_SIZE, 0444, proc_dostring);
5756 /* &table[12] is terminator */
5761 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5763 struct ctl_table *entry, *table;
5764 struct sched_domain *sd;
5765 int domain_num = 0, i;
5768 for_each_domain(cpu, sd)
5770 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5775 for_each_domain(cpu, sd) {
5776 snprintf(buf, 32, "domain%d", i);
5777 entry->procname = kstrdup(buf, GFP_KERNEL);
5779 entry->child = sd_alloc_ctl_domain_table(sd);
5786 static struct ctl_table_header *sd_sysctl_header;
5787 static void register_sched_domain_sysctl(void)
5789 int i, cpu_num = num_possible_cpus();
5790 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5793 WARN_ON(sd_ctl_dir[0].child);
5794 sd_ctl_dir[0].child = entry;
5799 for_each_possible_cpu(i) {
5800 snprintf(buf, 32, "cpu%d", i);
5801 entry->procname = kstrdup(buf, GFP_KERNEL);
5803 entry->child = sd_alloc_ctl_cpu_table(i);
5807 WARN_ON(sd_sysctl_header);
5808 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5811 /* may be called multiple times per register */
5812 static void unregister_sched_domain_sysctl(void)
5814 if (sd_sysctl_header)
5815 unregister_sysctl_table(sd_sysctl_header);
5816 sd_sysctl_header = NULL;
5817 if (sd_ctl_dir[0].child)
5818 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5821 static void register_sched_domain_sysctl(void)
5824 static void unregister_sched_domain_sysctl(void)
5829 static void set_rq_online(struct rq *rq)
5832 const struct sched_class *class;
5834 cpumask_set_cpu(rq->cpu, rq->rd->online);
5837 for_each_class(class) {
5838 if (class->rq_online)
5839 class->rq_online(rq);
5844 static void set_rq_offline(struct rq *rq)
5847 const struct sched_class *class;
5849 for_each_class(class) {
5850 if (class->rq_offline)
5851 class->rq_offline(rq);
5854 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5860 * migration_call - callback that gets triggered when a CPU is added.
5861 * Here we can start up the necessary migration thread for the new CPU.
5863 static int __cpuinit
5864 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5866 struct task_struct *p;
5867 int cpu = (long)hcpu;
5868 unsigned long flags;
5873 case CPU_UP_PREPARE:
5874 case CPU_UP_PREPARE_FROZEN:
5875 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
5878 kthread_bind(p, cpu);
5879 /* Must be high prio: stop_machine expects to yield to it. */
5880 rq = task_rq_lock(p, &flags);
5881 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5882 task_rq_unlock(rq, &flags);
5884 cpu_rq(cpu)->migration_thread = p;
5885 rq->calc_load_update = calc_load_update;
5889 case CPU_ONLINE_FROZEN:
5890 /* Strictly unnecessary, as first user will wake it. */
5891 wake_up_process(cpu_rq(cpu)->migration_thread);
5893 /* Update our root-domain */
5895 raw_spin_lock_irqsave(&rq->lock, flags);
5897 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5901 raw_spin_unlock_irqrestore(&rq->lock, flags);
5904 #ifdef CONFIG_HOTPLUG_CPU
5905 case CPU_UP_CANCELED:
5906 case CPU_UP_CANCELED_FROZEN:
5907 if (!cpu_rq(cpu)->migration_thread)
5909 /* Unbind it from offline cpu so it can run. Fall thru. */
5910 kthread_bind(cpu_rq(cpu)->migration_thread,
5911 cpumask_any(cpu_online_mask));
5912 kthread_stop(cpu_rq(cpu)->migration_thread);
5913 put_task_struct(cpu_rq(cpu)->migration_thread);
5914 cpu_rq(cpu)->migration_thread = NULL;
5918 case CPU_DEAD_FROZEN:
5919 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
5920 migrate_live_tasks(cpu);
5922 kthread_stop(rq->migration_thread);
5923 put_task_struct(rq->migration_thread);
5924 rq->migration_thread = NULL;
5925 /* Idle task back to normal (off runqueue, low prio) */
5926 raw_spin_lock_irq(&rq->lock);
5927 update_rq_clock(rq);
5928 deactivate_task(rq, rq->idle, 0);
5929 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5930 rq->idle->sched_class = &idle_sched_class;
5931 migrate_dead_tasks(cpu);
5932 raw_spin_unlock_irq(&rq->lock);
5934 migrate_nr_uninterruptible(rq);
5935 BUG_ON(rq->nr_running != 0);
5936 calc_global_load_remove(rq);
5938 * No need to migrate the tasks: it was best-effort if
5939 * they didn't take sched_hotcpu_mutex. Just wake up
5942 raw_spin_lock_irq(&rq->lock);
5943 while (!list_empty(&rq->migration_queue)) {
5944 struct migration_req *req;
5946 req = list_entry(rq->migration_queue.next,
5947 struct migration_req, list);
5948 list_del_init(&req->list);
5949 raw_spin_unlock_irq(&rq->lock);
5950 complete(&req->done);
5951 raw_spin_lock_irq(&rq->lock);
5953 raw_spin_unlock_irq(&rq->lock);
5957 case CPU_DYING_FROZEN:
5958 /* Update our root-domain */
5960 raw_spin_lock_irqsave(&rq->lock, flags);
5962 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5965 raw_spin_unlock_irqrestore(&rq->lock, flags);
5973 * Register at high priority so that task migration (migrate_all_tasks)
5974 * happens before everything else. This has to be lower priority than
5975 * the notifier in the perf_event subsystem, though.
5977 static struct notifier_block __cpuinitdata migration_notifier = {
5978 .notifier_call = migration_call,
5982 static int __init migration_init(void)
5984 void *cpu = (void *)(long)smp_processor_id();
5987 /* Start one for the boot CPU: */
5988 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5989 BUG_ON(err == NOTIFY_BAD);
5990 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5991 register_cpu_notifier(&migration_notifier);
5995 early_initcall(migration_init);
6000 #ifdef CONFIG_SCHED_DEBUG
6002 static __read_mostly int sched_domain_debug_enabled;
6004 static int __init sched_domain_debug_setup(char *str)
6006 sched_domain_debug_enabled = 1;
6010 early_param("sched_debug", sched_domain_debug_setup);
6012 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6013 struct cpumask *groupmask)
6015 struct sched_group *group = sd->groups;
6018 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6019 cpumask_clear(groupmask);
6021 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6023 if (!(sd->flags & SD_LOAD_BALANCE)) {
6024 printk("does not load-balance\n");
6026 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6031 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6033 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6034 printk(KERN_ERR "ERROR: domain->span does not contain "
6037 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6038 printk(KERN_ERR "ERROR: domain->groups does not contain"
6042 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6046 printk(KERN_ERR "ERROR: group is NULL\n");
6050 if (!group->cpu_power) {
6051 printk(KERN_CONT "\n");
6052 printk(KERN_ERR "ERROR: domain->cpu_power not "
6057 if (!cpumask_weight(sched_group_cpus(group))) {
6058 printk(KERN_CONT "\n");
6059 printk(KERN_ERR "ERROR: empty group\n");
6063 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6064 printk(KERN_CONT "\n");
6065 printk(KERN_ERR "ERROR: repeated CPUs\n");
6069 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6071 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6073 printk(KERN_CONT " %s", str);
6074 if (group->cpu_power != SCHED_LOAD_SCALE) {
6075 printk(KERN_CONT " (cpu_power = %d)",
6079 group = group->next;
6080 } while (group != sd->groups);
6081 printk(KERN_CONT "\n");
6083 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6084 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6087 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6088 printk(KERN_ERR "ERROR: parent span is not a superset "
6089 "of domain->span\n");
6093 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6095 cpumask_var_t groupmask;
6098 if (!sched_domain_debug_enabled)
6102 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6106 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6108 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6109 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6114 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6121 free_cpumask_var(groupmask);
6123 #else /* !CONFIG_SCHED_DEBUG */
6124 # define sched_domain_debug(sd, cpu) do { } while (0)
6125 #endif /* CONFIG_SCHED_DEBUG */
6127 static int sd_degenerate(struct sched_domain *sd)
6129 if (cpumask_weight(sched_domain_span(sd)) == 1)
6132 /* Following flags need at least 2 groups */
6133 if (sd->flags & (SD_LOAD_BALANCE |
6134 SD_BALANCE_NEWIDLE |
6138 SD_SHARE_PKG_RESOURCES)) {
6139 if (sd->groups != sd->groups->next)
6143 /* Following flags don't use groups */
6144 if (sd->flags & (SD_WAKE_AFFINE))
6151 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6153 unsigned long cflags = sd->flags, pflags = parent->flags;
6155 if (sd_degenerate(parent))
6158 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6161 /* Flags needing groups don't count if only 1 group in parent */
6162 if (parent->groups == parent->groups->next) {
6163 pflags &= ~(SD_LOAD_BALANCE |
6164 SD_BALANCE_NEWIDLE |
6168 SD_SHARE_PKG_RESOURCES);
6169 if (nr_node_ids == 1)
6170 pflags &= ~SD_SERIALIZE;
6172 if (~cflags & pflags)
6178 static void free_rootdomain(struct root_domain *rd)
6180 synchronize_sched();
6182 cpupri_cleanup(&rd->cpupri);
6184 free_cpumask_var(rd->rto_mask);
6185 free_cpumask_var(rd->online);
6186 free_cpumask_var(rd->span);
6190 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6192 struct root_domain *old_rd = NULL;
6193 unsigned long flags;
6195 raw_spin_lock_irqsave(&rq->lock, flags);
6200 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6203 cpumask_clear_cpu(rq->cpu, old_rd->span);
6206 * If we dont want to free the old_rt yet then
6207 * set old_rd to NULL to skip the freeing later
6210 if (!atomic_dec_and_test(&old_rd->refcount))
6214 atomic_inc(&rd->refcount);
6217 cpumask_set_cpu(rq->cpu, rd->span);
6218 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6221 raw_spin_unlock_irqrestore(&rq->lock, flags);
6224 free_rootdomain(old_rd);
6227 static int init_rootdomain(struct root_domain *rd, bool bootmem)
6229 gfp_t gfp = GFP_KERNEL;
6231 memset(rd, 0, sizeof(*rd));
6236 if (!alloc_cpumask_var(&rd->span, gfp))
6238 if (!alloc_cpumask_var(&rd->online, gfp))
6240 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
6243 if (cpupri_init(&rd->cpupri, bootmem) != 0)
6248 free_cpumask_var(rd->rto_mask);
6250 free_cpumask_var(rd->online);
6252 free_cpumask_var(rd->span);
6257 static void init_defrootdomain(void)
6259 init_rootdomain(&def_root_domain, true);
6261 atomic_set(&def_root_domain.refcount, 1);
6264 static struct root_domain *alloc_rootdomain(void)
6266 struct root_domain *rd;
6268 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6272 if (init_rootdomain(rd, false) != 0) {
6281 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6282 * hold the hotplug lock.
6285 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6287 struct rq *rq = cpu_rq(cpu);
6288 struct sched_domain *tmp;
6290 /* Remove the sched domains which do not contribute to scheduling. */
6291 for (tmp = sd; tmp; ) {
6292 struct sched_domain *parent = tmp->parent;
6296 if (sd_parent_degenerate(tmp, parent)) {
6297 tmp->parent = parent->parent;
6299 parent->parent->child = tmp;
6304 if (sd && sd_degenerate(sd)) {
6310 sched_domain_debug(sd, cpu);
6312 rq_attach_root(rq, rd);
6313 rcu_assign_pointer(rq->sd, sd);
6316 /* cpus with isolated domains */
6317 static cpumask_var_t cpu_isolated_map;
6319 /* Setup the mask of cpus configured for isolated domains */
6320 static int __init isolated_cpu_setup(char *str)
6322 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6323 cpulist_parse(str, cpu_isolated_map);
6327 __setup("isolcpus=", isolated_cpu_setup);
6330 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6331 * to a function which identifies what group(along with sched group) a CPU
6332 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6333 * (due to the fact that we keep track of groups covered with a struct cpumask).
6335 * init_sched_build_groups will build a circular linked list of the groups
6336 * covered by the given span, and will set each group's ->cpumask correctly,
6337 * and ->cpu_power to 0.
6340 init_sched_build_groups(const struct cpumask *span,
6341 const struct cpumask *cpu_map,
6342 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
6343 struct sched_group **sg,
6344 struct cpumask *tmpmask),
6345 struct cpumask *covered, struct cpumask *tmpmask)
6347 struct sched_group *first = NULL, *last = NULL;
6350 cpumask_clear(covered);
6352 for_each_cpu(i, span) {
6353 struct sched_group *sg;
6354 int group = group_fn(i, cpu_map, &sg, tmpmask);
6357 if (cpumask_test_cpu(i, covered))
6360 cpumask_clear(sched_group_cpus(sg));
6363 for_each_cpu(j, span) {
6364 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6367 cpumask_set_cpu(j, covered);
6368 cpumask_set_cpu(j, sched_group_cpus(sg));
6379 #define SD_NODES_PER_DOMAIN 16
6384 * find_next_best_node - find the next node to include in a sched_domain
6385 * @node: node whose sched_domain we're building
6386 * @used_nodes: nodes already in the sched_domain
6388 * Find the next node to include in a given scheduling domain. Simply
6389 * finds the closest node not already in the @used_nodes map.
6391 * Should use nodemask_t.
6393 static int find_next_best_node(int node, nodemask_t *used_nodes)
6395 int i, n, val, min_val, best_node = 0;
6399 for (i = 0; i < nr_node_ids; i++) {
6400 /* Start at @node */
6401 n = (node + i) % nr_node_ids;
6403 if (!nr_cpus_node(n))
6406 /* Skip already used nodes */
6407 if (node_isset(n, *used_nodes))
6410 /* Simple min distance search */
6411 val = node_distance(node, n);
6413 if (val < min_val) {
6419 node_set(best_node, *used_nodes);
6424 * sched_domain_node_span - get a cpumask for a node's sched_domain
6425 * @node: node whose cpumask we're constructing
6426 * @span: resulting cpumask
6428 * Given a node, construct a good cpumask for its sched_domain to span. It
6429 * should be one that prevents unnecessary balancing, but also spreads tasks
6432 static void sched_domain_node_span(int node, struct cpumask *span)
6434 nodemask_t used_nodes;
6437 cpumask_clear(span);
6438 nodes_clear(used_nodes);
6440 cpumask_or(span, span, cpumask_of_node(node));
6441 node_set(node, used_nodes);
6443 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6444 int next_node = find_next_best_node(node, &used_nodes);
6446 cpumask_or(span, span, cpumask_of_node(next_node));
6449 #endif /* CONFIG_NUMA */
6451 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6454 * The cpus mask in sched_group and sched_domain hangs off the end.
6456 * ( See the the comments in include/linux/sched.h:struct sched_group
6457 * and struct sched_domain. )
6459 struct static_sched_group {
6460 struct sched_group sg;
6461 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
6464 struct static_sched_domain {
6465 struct sched_domain sd;
6466 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
6472 cpumask_var_t domainspan;
6473 cpumask_var_t covered;
6474 cpumask_var_t notcovered;
6476 cpumask_var_t nodemask;
6477 cpumask_var_t this_sibling_map;
6478 cpumask_var_t this_core_map;
6479 cpumask_var_t send_covered;
6480 cpumask_var_t tmpmask;
6481 struct sched_group **sched_group_nodes;
6482 struct root_domain *rd;
6486 sa_sched_groups = 0,
6491 sa_this_sibling_map,
6493 sa_sched_group_nodes,
6503 * SMT sched-domains:
6505 #ifdef CONFIG_SCHED_SMT
6506 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
6507 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
6510 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
6511 struct sched_group **sg, struct cpumask *unused)
6514 *sg = &per_cpu(sched_groups, cpu).sg;
6517 #endif /* CONFIG_SCHED_SMT */
6520 * multi-core sched-domains:
6522 #ifdef CONFIG_SCHED_MC
6523 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
6524 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
6525 #endif /* CONFIG_SCHED_MC */
6527 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6529 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6530 struct sched_group **sg, struct cpumask *mask)
6534 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6535 group = cpumask_first(mask);
6537 *sg = &per_cpu(sched_group_core, group).sg;
6540 #elif defined(CONFIG_SCHED_MC)
6542 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6543 struct sched_group **sg, struct cpumask *unused)
6546 *sg = &per_cpu(sched_group_core, cpu).sg;
6551 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
6552 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
6555 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
6556 struct sched_group **sg, struct cpumask *mask)
6559 #ifdef CONFIG_SCHED_MC
6560 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6561 group = cpumask_first(mask);
6562 #elif defined(CONFIG_SCHED_SMT)
6563 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6564 group = cpumask_first(mask);
6569 *sg = &per_cpu(sched_group_phys, group).sg;
6575 * The init_sched_build_groups can't handle what we want to do with node
6576 * groups, so roll our own. Now each node has its own list of groups which
6577 * gets dynamically allocated.
6579 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
6580 static struct sched_group ***sched_group_nodes_bycpu;
6582 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
6583 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
6585 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
6586 struct sched_group **sg,
6587 struct cpumask *nodemask)
6591 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
6592 group = cpumask_first(nodemask);
6595 *sg = &per_cpu(sched_group_allnodes, group).sg;
6599 static void init_numa_sched_groups_power(struct sched_group *group_head)
6601 struct sched_group *sg = group_head;
6607 for_each_cpu(j, sched_group_cpus(sg)) {
6608 struct sched_domain *sd;
6610 sd = &per_cpu(phys_domains, j).sd;
6611 if (j != group_first_cpu(sd->groups)) {
6613 * Only add "power" once for each
6619 sg->cpu_power += sd->groups->cpu_power;
6622 } while (sg != group_head);
6625 static int build_numa_sched_groups(struct s_data *d,
6626 const struct cpumask *cpu_map, int num)
6628 struct sched_domain *sd;
6629 struct sched_group *sg, *prev;
6632 cpumask_clear(d->covered);
6633 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
6634 if (cpumask_empty(d->nodemask)) {
6635 d->sched_group_nodes[num] = NULL;
6639 sched_domain_node_span(num, d->domainspan);
6640 cpumask_and(d->domainspan, d->domainspan, cpu_map);
6642 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6645 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
6649 d->sched_group_nodes[num] = sg;
6651 for_each_cpu(j, d->nodemask) {
6652 sd = &per_cpu(node_domains, j).sd;
6657 cpumask_copy(sched_group_cpus(sg), d->nodemask);
6659 cpumask_or(d->covered, d->covered, d->nodemask);
6662 for (j = 0; j < nr_node_ids; j++) {
6663 n = (num + j) % nr_node_ids;
6664 cpumask_complement(d->notcovered, d->covered);
6665 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
6666 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
6667 if (cpumask_empty(d->tmpmask))
6669 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
6670 if (cpumask_empty(d->tmpmask))
6672 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6676 "Can not alloc domain group for node %d\n", j);
6680 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
6681 sg->next = prev->next;
6682 cpumask_or(d->covered, d->covered, d->tmpmask);
6689 #endif /* CONFIG_NUMA */
6692 /* Free memory allocated for various sched_group structures */
6693 static void free_sched_groups(const struct cpumask *cpu_map,
6694 struct cpumask *nodemask)
6698 for_each_cpu(cpu, cpu_map) {
6699 struct sched_group **sched_group_nodes
6700 = sched_group_nodes_bycpu[cpu];
6702 if (!sched_group_nodes)
6705 for (i = 0; i < nr_node_ids; i++) {
6706 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6708 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
6709 if (cpumask_empty(nodemask))
6719 if (oldsg != sched_group_nodes[i])
6722 kfree(sched_group_nodes);
6723 sched_group_nodes_bycpu[cpu] = NULL;
6726 #else /* !CONFIG_NUMA */
6727 static void free_sched_groups(const struct cpumask *cpu_map,
6728 struct cpumask *nodemask)
6731 #endif /* CONFIG_NUMA */
6734 * Initialize sched groups cpu_power.
6736 * cpu_power indicates the capacity of sched group, which is used while
6737 * distributing the load between different sched groups in a sched domain.
6738 * Typically cpu_power for all the groups in a sched domain will be same unless
6739 * there are asymmetries in the topology. If there are asymmetries, group
6740 * having more cpu_power will pickup more load compared to the group having
6743 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6745 struct sched_domain *child;
6746 struct sched_group *group;
6750 WARN_ON(!sd || !sd->groups);
6752 if (cpu != group_first_cpu(sd->groups))
6757 sd->groups->cpu_power = 0;
6760 power = SCHED_LOAD_SCALE;
6761 weight = cpumask_weight(sched_domain_span(sd));
6763 * SMT siblings share the power of a single core.
6764 * Usually multiple threads get a better yield out of
6765 * that one core than a single thread would have,
6766 * reflect that in sd->smt_gain.
6768 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
6769 power *= sd->smt_gain;
6771 power >>= SCHED_LOAD_SHIFT;
6773 sd->groups->cpu_power += power;
6778 * Add cpu_power of each child group to this groups cpu_power.
6780 group = child->groups;
6782 sd->groups->cpu_power += group->cpu_power;
6783 group = group->next;
6784 } while (group != child->groups);
6788 * Initializers for schedule domains
6789 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6792 #ifdef CONFIG_SCHED_DEBUG
6793 # define SD_INIT_NAME(sd, type) sd->name = #type
6795 # define SD_INIT_NAME(sd, type) do { } while (0)
6798 #define SD_INIT(sd, type) sd_init_##type(sd)
6800 #define SD_INIT_FUNC(type) \
6801 static noinline void sd_init_##type(struct sched_domain *sd) \
6803 memset(sd, 0, sizeof(*sd)); \
6804 *sd = SD_##type##_INIT; \
6805 sd->level = SD_LV_##type; \
6806 SD_INIT_NAME(sd, type); \
6811 SD_INIT_FUNC(ALLNODES)
6814 #ifdef CONFIG_SCHED_SMT
6815 SD_INIT_FUNC(SIBLING)
6817 #ifdef CONFIG_SCHED_MC
6821 static int default_relax_domain_level = -1;
6823 static int __init setup_relax_domain_level(char *str)
6827 val = simple_strtoul(str, NULL, 0);
6828 if (val < SD_LV_MAX)
6829 default_relax_domain_level = val;
6833 __setup("relax_domain_level=", setup_relax_domain_level);
6835 static void set_domain_attribute(struct sched_domain *sd,
6836 struct sched_domain_attr *attr)
6840 if (!attr || attr->relax_domain_level < 0) {
6841 if (default_relax_domain_level < 0)
6844 request = default_relax_domain_level;
6846 request = attr->relax_domain_level;
6847 if (request < sd->level) {
6848 /* turn off idle balance on this domain */
6849 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6851 /* turn on idle balance on this domain */
6852 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6856 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6857 const struct cpumask *cpu_map)
6860 case sa_sched_groups:
6861 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
6862 d->sched_group_nodes = NULL;
6864 free_rootdomain(d->rd); /* fall through */
6866 free_cpumask_var(d->tmpmask); /* fall through */
6867 case sa_send_covered:
6868 free_cpumask_var(d->send_covered); /* fall through */
6869 case sa_this_core_map:
6870 free_cpumask_var(d->this_core_map); /* fall through */
6871 case sa_this_sibling_map:
6872 free_cpumask_var(d->this_sibling_map); /* fall through */
6874 free_cpumask_var(d->nodemask); /* fall through */
6875 case sa_sched_group_nodes:
6877 kfree(d->sched_group_nodes); /* fall through */
6879 free_cpumask_var(d->notcovered); /* fall through */
6881 free_cpumask_var(d->covered); /* fall through */
6883 free_cpumask_var(d->domainspan); /* fall through */
6890 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6891 const struct cpumask *cpu_map)
6894 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
6896 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
6897 return sa_domainspan;
6898 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
6900 /* Allocate the per-node list of sched groups */
6901 d->sched_group_nodes = kcalloc(nr_node_ids,
6902 sizeof(struct sched_group *), GFP_KERNEL);
6903 if (!d->sched_group_nodes) {
6904 printk(KERN_WARNING "Can not alloc sched group node list\n");
6905 return sa_notcovered;
6907 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
6909 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
6910 return sa_sched_group_nodes;
6911 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
6913 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
6914 return sa_this_sibling_map;
6915 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
6916 return sa_this_core_map;
6917 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
6918 return sa_send_covered;
6919 d->rd = alloc_rootdomain();
6921 printk(KERN_WARNING "Cannot alloc root domain\n");
6924 return sa_rootdomain;
6927 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
6928 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
6930 struct sched_domain *sd = NULL;
6932 struct sched_domain *parent;
6935 if (cpumask_weight(cpu_map) >
6936 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
6937 sd = &per_cpu(allnodes_domains, i).sd;
6938 SD_INIT(sd, ALLNODES);
6939 set_domain_attribute(sd, attr);
6940 cpumask_copy(sched_domain_span(sd), cpu_map);
6941 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
6946 sd = &per_cpu(node_domains, i).sd;
6948 set_domain_attribute(sd, attr);
6949 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
6950 sd->parent = parent;
6953 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
6958 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
6959 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6960 struct sched_domain *parent, int i)
6962 struct sched_domain *sd;
6963 sd = &per_cpu(phys_domains, i).sd;
6965 set_domain_attribute(sd, attr);
6966 cpumask_copy(sched_domain_span(sd), d->nodemask);
6967 sd->parent = parent;
6970 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
6974 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
6975 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6976 struct sched_domain *parent, int i)
6978 struct sched_domain *sd = parent;
6979 #ifdef CONFIG_SCHED_MC
6980 sd = &per_cpu(core_domains, i).sd;
6982 set_domain_attribute(sd, attr);
6983 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
6984 sd->parent = parent;
6986 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
6991 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
6992 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6993 struct sched_domain *parent, int i)
6995 struct sched_domain *sd = parent;
6996 #ifdef CONFIG_SCHED_SMT
6997 sd = &per_cpu(cpu_domains, i).sd;
6998 SD_INIT(sd, SIBLING);
6999 set_domain_attribute(sd, attr);
7000 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
7001 sd->parent = parent;
7003 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
7008 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
7009 const struct cpumask *cpu_map, int cpu)
7012 #ifdef CONFIG_SCHED_SMT
7013 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
7014 cpumask_and(d->this_sibling_map, cpu_map,
7015 topology_thread_cpumask(cpu));
7016 if (cpu == cpumask_first(d->this_sibling_map))
7017 init_sched_build_groups(d->this_sibling_map, cpu_map,
7019 d->send_covered, d->tmpmask);
7022 #ifdef CONFIG_SCHED_MC
7023 case SD_LV_MC: /* set up multi-core groups */
7024 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
7025 if (cpu == cpumask_first(d->this_core_map))
7026 init_sched_build_groups(d->this_core_map, cpu_map,
7028 d->send_covered, d->tmpmask);
7031 case SD_LV_CPU: /* set up physical groups */
7032 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
7033 if (!cpumask_empty(d->nodemask))
7034 init_sched_build_groups(d->nodemask, cpu_map,
7036 d->send_covered, d->tmpmask);
7039 case SD_LV_ALLNODES:
7040 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
7041 d->send_covered, d->tmpmask);
7050 * Build sched domains for a given set of cpus and attach the sched domains
7051 * to the individual cpus
7053 static int __build_sched_domains(const struct cpumask *cpu_map,
7054 struct sched_domain_attr *attr)
7056 enum s_alloc alloc_state = sa_none;
7058 struct sched_domain *sd;
7064 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7065 if (alloc_state != sa_rootdomain)
7067 alloc_state = sa_sched_groups;
7070 * Set up domains for cpus specified by the cpu_map.
7072 for_each_cpu(i, cpu_map) {
7073 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
7076 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
7077 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
7078 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
7079 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
7082 for_each_cpu(i, cpu_map) {
7083 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
7084 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
7087 /* Set up physical groups */
7088 for (i = 0; i < nr_node_ids; i++)
7089 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
7092 /* Set up node groups */
7094 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
7096 for (i = 0; i < nr_node_ids; i++)
7097 if (build_numa_sched_groups(&d, cpu_map, i))
7101 /* Calculate CPU power for physical packages and nodes */
7102 #ifdef CONFIG_SCHED_SMT
7103 for_each_cpu(i, cpu_map) {
7104 sd = &per_cpu(cpu_domains, i).sd;
7105 init_sched_groups_power(i, sd);
7108 #ifdef CONFIG_SCHED_MC
7109 for_each_cpu(i, cpu_map) {
7110 sd = &per_cpu(core_domains, i).sd;
7111 init_sched_groups_power(i, sd);
7115 for_each_cpu(i, cpu_map) {
7116 sd = &per_cpu(phys_domains, i).sd;
7117 init_sched_groups_power(i, sd);
7121 for (i = 0; i < nr_node_ids; i++)
7122 init_numa_sched_groups_power(d.sched_group_nodes[i]);
7124 if (d.sd_allnodes) {
7125 struct sched_group *sg;
7127 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7129 init_numa_sched_groups_power(sg);
7133 /* Attach the domains */
7134 for_each_cpu(i, cpu_map) {
7135 #ifdef CONFIG_SCHED_SMT
7136 sd = &per_cpu(cpu_domains, i).sd;
7137 #elif defined(CONFIG_SCHED_MC)
7138 sd = &per_cpu(core_domains, i).sd;
7140 sd = &per_cpu(phys_domains, i).sd;
7142 cpu_attach_domain(sd, d.rd, i);
7145 d.sched_group_nodes = NULL; /* don't free this we still need it */
7146 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
7150 __free_domain_allocs(&d, alloc_state, cpu_map);
7154 static int build_sched_domains(const struct cpumask *cpu_map)
7156 return __build_sched_domains(cpu_map, NULL);
7159 static cpumask_var_t *doms_cur; /* current sched domains */
7160 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7161 static struct sched_domain_attr *dattr_cur;
7162 /* attribues of custom domains in 'doms_cur' */
7165 * Special case: If a kmalloc of a doms_cur partition (array of
7166 * cpumask) fails, then fallback to a single sched domain,
7167 * as determined by the single cpumask fallback_doms.
7169 static cpumask_var_t fallback_doms;
7172 * arch_update_cpu_topology lets virtualized architectures update the
7173 * cpu core maps. It is supposed to return 1 if the topology changed
7174 * or 0 if it stayed the same.
7176 int __attribute__((weak)) arch_update_cpu_topology(void)
7181 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7184 cpumask_var_t *doms;
7186 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7189 for (i = 0; i < ndoms; i++) {
7190 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7191 free_sched_domains(doms, i);
7198 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7201 for (i = 0; i < ndoms; i++)
7202 free_cpumask_var(doms[i]);
7207 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7208 * For now this just excludes isolated cpus, but could be used to
7209 * exclude other special cases in the future.
7211 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7215 arch_update_cpu_topology();
7217 doms_cur = alloc_sched_domains(ndoms_cur);
7219 doms_cur = &fallback_doms;
7220 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7222 err = build_sched_domains(doms_cur[0]);
7223 register_sched_domain_sysctl();
7228 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7229 struct cpumask *tmpmask)
7231 free_sched_groups(cpu_map, tmpmask);
7235 * Detach sched domains from a group of cpus specified in cpu_map
7236 * These cpus will now be attached to the NULL domain
7238 static void detach_destroy_domains(const struct cpumask *cpu_map)
7240 /* Save because hotplug lock held. */
7241 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7244 for_each_cpu(i, cpu_map)
7245 cpu_attach_domain(NULL, &def_root_domain, i);
7246 synchronize_sched();
7247 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7250 /* handle null as "default" */
7251 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7252 struct sched_domain_attr *new, int idx_new)
7254 struct sched_domain_attr tmp;
7261 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7262 new ? (new + idx_new) : &tmp,
7263 sizeof(struct sched_domain_attr));
7267 * Partition sched domains as specified by the 'ndoms_new'
7268 * cpumasks in the array doms_new[] of cpumasks. This compares
7269 * doms_new[] to the current sched domain partitioning, doms_cur[].
7270 * It destroys each deleted domain and builds each new domain.
7272 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7273 * The masks don't intersect (don't overlap.) We should setup one
7274 * sched domain for each mask. CPUs not in any of the cpumasks will
7275 * not be load balanced. If the same cpumask appears both in the
7276 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7279 * The passed in 'doms_new' should be allocated using
7280 * alloc_sched_domains. This routine takes ownership of it and will
7281 * free_sched_domains it when done with it. If the caller failed the
7282 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7283 * and partition_sched_domains() will fallback to the single partition
7284 * 'fallback_doms', it also forces the domains to be rebuilt.
7286 * If doms_new == NULL it will be replaced with cpu_online_mask.
7287 * ndoms_new == 0 is a special case for destroying existing domains,
7288 * and it will not create the default domain.
7290 * Call with hotplug lock held
7292 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7293 struct sched_domain_attr *dattr_new)
7298 mutex_lock(&sched_domains_mutex);
7300 /* always unregister in case we don't destroy any domains */
7301 unregister_sched_domain_sysctl();
7303 /* Let architecture update cpu core mappings. */
7304 new_topology = arch_update_cpu_topology();
7306 n = doms_new ? ndoms_new : 0;
7308 /* Destroy deleted domains */
7309 for (i = 0; i < ndoms_cur; i++) {
7310 for (j = 0; j < n && !new_topology; j++) {
7311 if (cpumask_equal(doms_cur[i], doms_new[j])
7312 && dattrs_equal(dattr_cur, i, dattr_new, j))
7315 /* no match - a current sched domain not in new doms_new[] */
7316 detach_destroy_domains(doms_cur[i]);
7321 if (doms_new == NULL) {
7323 doms_new = &fallback_doms;
7324 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7325 WARN_ON_ONCE(dattr_new);
7328 /* Build new domains */
7329 for (i = 0; i < ndoms_new; i++) {
7330 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7331 if (cpumask_equal(doms_new[i], doms_cur[j])
7332 && dattrs_equal(dattr_new, i, dattr_cur, j))
7335 /* no match - add a new doms_new */
7336 __build_sched_domains(doms_new[i],
7337 dattr_new ? dattr_new + i : NULL);
7342 /* Remember the new sched domains */
7343 if (doms_cur != &fallback_doms)
7344 free_sched_domains(doms_cur, ndoms_cur);
7345 kfree(dattr_cur); /* kfree(NULL) is safe */
7346 doms_cur = doms_new;
7347 dattr_cur = dattr_new;
7348 ndoms_cur = ndoms_new;
7350 register_sched_domain_sysctl();
7352 mutex_unlock(&sched_domains_mutex);
7355 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7356 static void arch_reinit_sched_domains(void)
7360 /* Destroy domains first to force the rebuild */
7361 partition_sched_domains(0, NULL, NULL);
7363 rebuild_sched_domains();
7367 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7369 unsigned int level = 0;
7371 if (sscanf(buf, "%u", &level) != 1)
7375 * level is always be positive so don't check for
7376 * level < POWERSAVINGS_BALANCE_NONE which is 0
7377 * What happens on 0 or 1 byte write,
7378 * need to check for count as well?
7381 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7385 sched_smt_power_savings = level;
7387 sched_mc_power_savings = level;
7389 arch_reinit_sched_domains();
7394 #ifdef CONFIG_SCHED_MC
7395 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7398 return sprintf(page, "%u\n", sched_mc_power_savings);
7400 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7401 const char *buf, size_t count)
7403 return sched_power_savings_store(buf, count, 0);
7405 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7406 sched_mc_power_savings_show,
7407 sched_mc_power_savings_store);
7410 #ifdef CONFIG_SCHED_SMT
7411 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7414 return sprintf(page, "%u\n", sched_smt_power_savings);
7416 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7417 const char *buf, size_t count)
7419 return sched_power_savings_store(buf, count, 1);
7421 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7422 sched_smt_power_savings_show,
7423 sched_smt_power_savings_store);
7426 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7430 #ifdef CONFIG_SCHED_SMT
7432 err = sysfs_create_file(&cls->kset.kobj,
7433 &attr_sched_smt_power_savings.attr);
7435 #ifdef CONFIG_SCHED_MC
7436 if (!err && mc_capable())
7437 err = sysfs_create_file(&cls->kset.kobj,
7438 &attr_sched_mc_power_savings.attr);
7442 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7444 #ifndef CONFIG_CPUSETS
7446 * Add online and remove offline CPUs from the scheduler domains.
7447 * When cpusets are enabled they take over this function.
7449 static int update_sched_domains(struct notifier_block *nfb,
7450 unsigned long action, void *hcpu)
7454 case CPU_ONLINE_FROZEN:
7455 case CPU_DOWN_PREPARE:
7456 case CPU_DOWN_PREPARE_FROZEN:
7457 case CPU_DOWN_FAILED:
7458 case CPU_DOWN_FAILED_FROZEN:
7459 partition_sched_domains(1, NULL, NULL);
7468 static int update_runtime(struct notifier_block *nfb,
7469 unsigned long action, void *hcpu)
7471 int cpu = (int)(long)hcpu;
7474 case CPU_DOWN_PREPARE:
7475 case CPU_DOWN_PREPARE_FROZEN:
7476 disable_runtime(cpu_rq(cpu));
7479 case CPU_DOWN_FAILED:
7480 case CPU_DOWN_FAILED_FROZEN:
7482 case CPU_ONLINE_FROZEN:
7483 enable_runtime(cpu_rq(cpu));
7491 void __init sched_init_smp(void)
7493 cpumask_var_t non_isolated_cpus;
7495 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7496 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7498 #if defined(CONFIG_NUMA)
7499 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7501 BUG_ON(sched_group_nodes_bycpu == NULL);
7504 mutex_lock(&sched_domains_mutex);
7505 arch_init_sched_domains(cpu_active_mask);
7506 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7507 if (cpumask_empty(non_isolated_cpus))
7508 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7509 mutex_unlock(&sched_domains_mutex);
7512 #ifndef CONFIG_CPUSETS
7513 /* XXX: Theoretical race here - CPU may be hotplugged now */
7514 hotcpu_notifier(update_sched_domains, 0);
7517 /* RT runtime code needs to handle some hotplug events */
7518 hotcpu_notifier(update_runtime, 0);
7522 /* Move init over to a non-isolated CPU */
7523 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7525 sched_init_granularity();
7526 free_cpumask_var(non_isolated_cpus);
7528 init_sched_rt_class();
7531 void __init sched_init_smp(void)
7533 sched_init_granularity();
7535 #endif /* CONFIG_SMP */
7537 const_debug unsigned int sysctl_timer_migration = 1;
7539 int in_sched_functions(unsigned long addr)
7541 return in_lock_functions(addr) ||
7542 (addr >= (unsigned long)__sched_text_start
7543 && addr < (unsigned long)__sched_text_end);
7546 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7548 cfs_rq->tasks_timeline = RB_ROOT;
7549 INIT_LIST_HEAD(&cfs_rq->tasks);
7550 #ifdef CONFIG_FAIR_GROUP_SCHED
7553 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7556 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7558 struct rt_prio_array *array;
7561 array = &rt_rq->active;
7562 for (i = 0; i < MAX_RT_PRIO; i++) {
7563 INIT_LIST_HEAD(array->queue + i);
7564 __clear_bit(i, array->bitmap);
7566 /* delimiter for bitsearch: */
7567 __set_bit(MAX_RT_PRIO, array->bitmap);
7569 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7570 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7572 rt_rq->highest_prio.next = MAX_RT_PRIO;
7576 rt_rq->rt_nr_migratory = 0;
7577 rt_rq->overloaded = 0;
7578 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
7582 rt_rq->rt_throttled = 0;
7583 rt_rq->rt_runtime = 0;
7584 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7586 #ifdef CONFIG_RT_GROUP_SCHED
7587 rt_rq->rt_nr_boosted = 0;
7592 #ifdef CONFIG_FAIR_GROUP_SCHED
7593 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7594 struct sched_entity *se, int cpu, int add,
7595 struct sched_entity *parent)
7597 struct rq *rq = cpu_rq(cpu);
7598 tg->cfs_rq[cpu] = cfs_rq;
7599 init_cfs_rq(cfs_rq, rq);
7602 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7605 /* se could be NULL for init_task_group */
7610 se->cfs_rq = &rq->cfs;
7612 se->cfs_rq = parent->my_q;
7615 se->load.weight = tg->shares;
7616 se->load.inv_weight = 0;
7617 se->parent = parent;
7621 #ifdef CONFIG_RT_GROUP_SCHED
7622 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7623 struct sched_rt_entity *rt_se, int cpu, int add,
7624 struct sched_rt_entity *parent)
7626 struct rq *rq = cpu_rq(cpu);
7628 tg->rt_rq[cpu] = rt_rq;
7629 init_rt_rq(rt_rq, rq);
7631 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7633 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7635 tg->rt_se[cpu] = rt_se;
7640 rt_se->rt_rq = &rq->rt;
7642 rt_se->rt_rq = parent->my_q;
7644 rt_se->my_q = rt_rq;
7645 rt_se->parent = parent;
7646 INIT_LIST_HEAD(&rt_se->run_list);
7650 void __init sched_init(void)
7653 unsigned long alloc_size = 0, ptr;
7655 #ifdef CONFIG_FAIR_GROUP_SCHED
7656 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7658 #ifdef CONFIG_RT_GROUP_SCHED
7659 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7661 #ifdef CONFIG_CPUMASK_OFFSTACK
7662 alloc_size += num_possible_cpus() * cpumask_size();
7665 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7667 #ifdef CONFIG_FAIR_GROUP_SCHED
7668 init_task_group.se = (struct sched_entity **)ptr;
7669 ptr += nr_cpu_ids * sizeof(void **);
7671 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7672 ptr += nr_cpu_ids * sizeof(void **);
7674 #endif /* CONFIG_FAIR_GROUP_SCHED */
7675 #ifdef CONFIG_RT_GROUP_SCHED
7676 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7677 ptr += nr_cpu_ids * sizeof(void **);
7679 init_task_group.rt_rq = (struct rt_rq **)ptr;
7680 ptr += nr_cpu_ids * sizeof(void **);
7682 #endif /* CONFIG_RT_GROUP_SCHED */
7683 #ifdef CONFIG_CPUMASK_OFFSTACK
7684 for_each_possible_cpu(i) {
7685 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
7686 ptr += cpumask_size();
7688 #endif /* CONFIG_CPUMASK_OFFSTACK */
7692 init_defrootdomain();
7695 init_rt_bandwidth(&def_rt_bandwidth,
7696 global_rt_period(), global_rt_runtime());
7698 #ifdef CONFIG_RT_GROUP_SCHED
7699 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7700 global_rt_period(), global_rt_runtime());
7701 #endif /* CONFIG_RT_GROUP_SCHED */
7703 #ifdef CONFIG_CGROUP_SCHED
7704 list_add(&init_task_group.list, &task_groups);
7705 INIT_LIST_HEAD(&init_task_group.children);
7707 #endif /* CONFIG_CGROUP_SCHED */
7709 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
7710 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
7711 __alignof__(unsigned long));
7713 for_each_possible_cpu(i) {
7717 raw_spin_lock_init(&rq->lock);
7719 rq->calc_load_active = 0;
7720 rq->calc_load_update = jiffies + LOAD_FREQ;
7721 init_cfs_rq(&rq->cfs, rq);
7722 init_rt_rq(&rq->rt, rq);
7723 #ifdef CONFIG_FAIR_GROUP_SCHED
7724 init_task_group.shares = init_task_group_load;
7725 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7726 #ifdef CONFIG_CGROUP_SCHED
7728 * How much cpu bandwidth does init_task_group get?
7730 * In case of task-groups formed thr' the cgroup filesystem, it
7731 * gets 100% of the cpu resources in the system. This overall
7732 * system cpu resource is divided among the tasks of
7733 * init_task_group and its child task-groups in a fair manner,
7734 * based on each entity's (task or task-group's) weight
7735 * (se->load.weight).
7737 * In other words, if init_task_group has 10 tasks of weight
7738 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7739 * then A0's share of the cpu resource is:
7741 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7743 * We achieve this by letting init_task_group's tasks sit
7744 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7746 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
7748 #endif /* CONFIG_FAIR_GROUP_SCHED */
7750 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7751 #ifdef CONFIG_RT_GROUP_SCHED
7752 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7753 #ifdef CONFIG_CGROUP_SCHED
7754 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
7758 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7759 rq->cpu_load[j] = 0;
7763 rq->post_schedule = 0;
7764 rq->active_balance = 0;
7765 rq->next_balance = jiffies;
7769 rq->migration_thread = NULL;
7771 rq->avg_idle = 2*sysctl_sched_migration_cost;
7772 INIT_LIST_HEAD(&rq->migration_queue);
7773 rq_attach_root(rq, &def_root_domain);
7776 atomic_set(&rq->nr_iowait, 0);
7779 set_load_weight(&init_task);
7781 #ifdef CONFIG_PREEMPT_NOTIFIERS
7782 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7786 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7789 #ifdef CONFIG_RT_MUTEXES
7790 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
7794 * The boot idle thread does lazy MMU switching as well:
7796 atomic_inc(&init_mm.mm_count);
7797 enter_lazy_tlb(&init_mm, current);
7800 * Make us the idle thread. Technically, schedule() should not be
7801 * called from this thread, however somewhere below it might be,
7802 * but because we are the idle thread, we just pick up running again
7803 * when this runqueue becomes "idle".
7805 init_idle(current, smp_processor_id());
7807 calc_load_update = jiffies + LOAD_FREQ;
7810 * During early bootup we pretend to be a normal task:
7812 current->sched_class = &fair_sched_class;
7814 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
7815 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
7818 zalloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
7819 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
7821 /* May be allocated at isolcpus cmdline parse time */
7822 if (cpu_isolated_map == NULL)
7823 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7828 scheduler_running = 1;
7831 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7832 static inline int preempt_count_equals(int preempt_offset)
7834 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7836 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
7839 void __might_sleep(const char *file, int line, int preempt_offset)
7842 static unsigned long prev_jiffy; /* ratelimiting */
7844 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7845 system_state != SYSTEM_RUNNING || oops_in_progress)
7847 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7849 prev_jiffy = jiffies;
7852 "BUG: sleeping function called from invalid context at %s:%d\n",
7855 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7856 in_atomic(), irqs_disabled(),
7857 current->pid, current->comm);
7859 debug_show_held_locks(current);
7860 if (irqs_disabled())
7861 print_irqtrace_events(current);
7865 EXPORT_SYMBOL(__might_sleep);
7868 #ifdef CONFIG_MAGIC_SYSRQ
7869 static void normalize_task(struct rq *rq, struct task_struct *p)
7873 update_rq_clock(rq);
7874 on_rq = p->se.on_rq;
7876 deactivate_task(rq, p, 0);
7877 __setscheduler(rq, p, SCHED_NORMAL, 0);
7879 activate_task(rq, p, 0);
7880 resched_task(rq->curr);
7884 void normalize_rt_tasks(void)
7886 struct task_struct *g, *p;
7887 unsigned long flags;
7890 read_lock_irqsave(&tasklist_lock, flags);
7891 do_each_thread(g, p) {
7893 * Only normalize user tasks:
7898 p->se.exec_start = 0;
7899 #ifdef CONFIG_SCHEDSTATS
7900 p->se.wait_start = 0;
7901 p->se.sleep_start = 0;
7902 p->se.block_start = 0;
7907 * Renice negative nice level userspace
7910 if (TASK_NICE(p) < 0 && p->mm)
7911 set_user_nice(p, 0);
7915 raw_spin_lock(&p->pi_lock);
7916 rq = __task_rq_lock(p);
7918 normalize_task(rq, p);
7920 __task_rq_unlock(rq);
7921 raw_spin_unlock(&p->pi_lock);
7922 } while_each_thread(g, p);
7924 read_unlock_irqrestore(&tasklist_lock, flags);
7927 #endif /* CONFIG_MAGIC_SYSRQ */
7931 * These functions are only useful for the IA64 MCA handling.
7933 * They can only be called when the whole system has been
7934 * stopped - every CPU needs to be quiescent, and no scheduling
7935 * activity can take place. Using them for anything else would
7936 * be a serious bug, and as a result, they aren't even visible
7937 * under any other configuration.
7941 * curr_task - return the current task for a given cpu.
7942 * @cpu: the processor in question.
7944 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7946 struct task_struct *curr_task(int cpu)
7948 return cpu_curr(cpu);
7952 * set_curr_task - set the current task for a given cpu.
7953 * @cpu: the processor in question.
7954 * @p: the task pointer to set.
7956 * Description: This function must only be used when non-maskable interrupts
7957 * are serviced on a separate stack. It allows the architecture to switch the
7958 * notion of the current task on a cpu in a non-blocking manner. This function
7959 * must be called with all CPU's synchronized, and interrupts disabled, the
7960 * and caller must save the original value of the current task (see
7961 * curr_task() above) and restore that value before reenabling interrupts and
7962 * re-starting the system.
7964 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7966 void set_curr_task(int cpu, struct task_struct *p)
7973 #ifdef CONFIG_FAIR_GROUP_SCHED
7974 static void free_fair_sched_group(struct task_group *tg)
7978 for_each_possible_cpu(i) {
7980 kfree(tg->cfs_rq[i]);
7990 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7992 struct cfs_rq *cfs_rq;
7993 struct sched_entity *se;
7997 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8000 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8004 tg->shares = NICE_0_LOAD;
8006 for_each_possible_cpu(i) {
8009 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8010 GFP_KERNEL, cpu_to_node(i));
8014 se = kzalloc_node(sizeof(struct sched_entity),
8015 GFP_KERNEL, cpu_to_node(i));
8019 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
8030 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8032 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8033 &cpu_rq(cpu)->leaf_cfs_rq_list);
8036 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8038 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8040 #else /* !CONFG_FAIR_GROUP_SCHED */
8041 static inline void free_fair_sched_group(struct task_group *tg)
8046 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8051 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8055 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8058 #endif /* CONFIG_FAIR_GROUP_SCHED */
8060 #ifdef CONFIG_RT_GROUP_SCHED
8061 static void free_rt_sched_group(struct task_group *tg)
8065 destroy_rt_bandwidth(&tg->rt_bandwidth);
8067 for_each_possible_cpu(i) {
8069 kfree(tg->rt_rq[i]);
8071 kfree(tg->rt_se[i]);
8079 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8081 struct rt_rq *rt_rq;
8082 struct sched_rt_entity *rt_se;
8086 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8089 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8093 init_rt_bandwidth(&tg->rt_bandwidth,
8094 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8096 for_each_possible_cpu(i) {
8099 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8100 GFP_KERNEL, cpu_to_node(i));
8104 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8105 GFP_KERNEL, cpu_to_node(i));
8109 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
8120 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8122 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8123 &cpu_rq(cpu)->leaf_rt_rq_list);
8126 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8128 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8130 #else /* !CONFIG_RT_GROUP_SCHED */
8131 static inline void free_rt_sched_group(struct task_group *tg)
8136 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8141 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8145 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8148 #endif /* CONFIG_RT_GROUP_SCHED */
8150 #ifdef CONFIG_CGROUP_SCHED
8151 static void free_sched_group(struct task_group *tg)
8153 free_fair_sched_group(tg);
8154 free_rt_sched_group(tg);
8158 /* allocate runqueue etc for a new task group */
8159 struct task_group *sched_create_group(struct task_group *parent)
8161 struct task_group *tg;
8162 unsigned long flags;
8165 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8167 return ERR_PTR(-ENOMEM);
8169 if (!alloc_fair_sched_group(tg, parent))
8172 if (!alloc_rt_sched_group(tg, parent))
8175 spin_lock_irqsave(&task_group_lock, flags);
8176 for_each_possible_cpu(i) {
8177 register_fair_sched_group(tg, i);
8178 register_rt_sched_group(tg, i);
8180 list_add_rcu(&tg->list, &task_groups);
8182 WARN_ON(!parent); /* root should already exist */
8184 tg->parent = parent;
8185 INIT_LIST_HEAD(&tg->children);
8186 list_add_rcu(&tg->siblings, &parent->children);
8187 spin_unlock_irqrestore(&task_group_lock, flags);
8192 free_sched_group(tg);
8193 return ERR_PTR(-ENOMEM);
8196 /* rcu callback to free various structures associated with a task group */
8197 static void free_sched_group_rcu(struct rcu_head *rhp)
8199 /* now it should be safe to free those cfs_rqs */
8200 free_sched_group(container_of(rhp, struct task_group, rcu));
8203 /* Destroy runqueue etc associated with a task group */
8204 void sched_destroy_group(struct task_group *tg)
8206 unsigned long flags;
8209 spin_lock_irqsave(&task_group_lock, flags);
8210 for_each_possible_cpu(i) {
8211 unregister_fair_sched_group(tg, i);
8212 unregister_rt_sched_group(tg, i);
8214 list_del_rcu(&tg->list);
8215 list_del_rcu(&tg->siblings);
8216 spin_unlock_irqrestore(&task_group_lock, flags);
8218 /* wait for possible concurrent references to cfs_rqs complete */
8219 call_rcu(&tg->rcu, free_sched_group_rcu);
8222 /* change task's runqueue when it moves between groups.
8223 * The caller of this function should have put the task in its new group
8224 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8225 * reflect its new group.
8227 void sched_move_task(struct task_struct *tsk)
8230 unsigned long flags;
8233 rq = task_rq_lock(tsk, &flags);
8235 update_rq_clock(rq);
8237 running = task_current(rq, tsk);
8238 on_rq = tsk->se.on_rq;
8241 dequeue_task(rq, tsk, 0);
8242 if (unlikely(running))
8243 tsk->sched_class->put_prev_task(rq, tsk);
8245 set_task_rq(tsk, task_cpu(tsk));
8247 #ifdef CONFIG_FAIR_GROUP_SCHED
8248 if (tsk->sched_class->moved_group)
8249 tsk->sched_class->moved_group(tsk, on_rq);
8252 if (unlikely(running))
8253 tsk->sched_class->set_curr_task(rq);
8255 enqueue_task(rq, tsk, 0, false);
8257 task_rq_unlock(rq, &flags);
8259 #endif /* CONFIG_CGROUP_SCHED */
8261 #ifdef CONFIG_FAIR_GROUP_SCHED
8262 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8264 struct cfs_rq *cfs_rq = se->cfs_rq;
8269 dequeue_entity(cfs_rq, se, 0);
8271 se->load.weight = shares;
8272 se->load.inv_weight = 0;
8275 enqueue_entity(cfs_rq, se, 0);
8278 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8280 struct cfs_rq *cfs_rq = se->cfs_rq;
8281 struct rq *rq = cfs_rq->rq;
8282 unsigned long flags;
8284 raw_spin_lock_irqsave(&rq->lock, flags);
8285 __set_se_shares(se, shares);
8286 raw_spin_unlock_irqrestore(&rq->lock, flags);
8289 static DEFINE_MUTEX(shares_mutex);
8291 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8294 unsigned long flags;
8297 * We can't change the weight of the root cgroup.
8302 if (shares < MIN_SHARES)
8303 shares = MIN_SHARES;
8304 else if (shares > MAX_SHARES)
8305 shares = MAX_SHARES;
8307 mutex_lock(&shares_mutex);
8308 if (tg->shares == shares)
8311 spin_lock_irqsave(&task_group_lock, flags);
8312 for_each_possible_cpu(i)
8313 unregister_fair_sched_group(tg, i);
8314 list_del_rcu(&tg->siblings);
8315 spin_unlock_irqrestore(&task_group_lock, flags);
8317 /* wait for any ongoing reference to this group to finish */
8318 synchronize_sched();
8321 * Now we are free to modify the group's share on each cpu
8322 * w/o tripping rebalance_share or load_balance_fair.
8324 tg->shares = shares;
8325 for_each_possible_cpu(i) {
8329 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8330 set_se_shares(tg->se[i], shares);
8334 * Enable load balance activity on this group, by inserting it back on
8335 * each cpu's rq->leaf_cfs_rq_list.
8337 spin_lock_irqsave(&task_group_lock, flags);
8338 for_each_possible_cpu(i)
8339 register_fair_sched_group(tg, i);
8340 list_add_rcu(&tg->siblings, &tg->parent->children);
8341 spin_unlock_irqrestore(&task_group_lock, flags);
8343 mutex_unlock(&shares_mutex);
8347 unsigned long sched_group_shares(struct task_group *tg)
8353 #ifdef CONFIG_RT_GROUP_SCHED
8355 * Ensure that the real time constraints are schedulable.
8357 static DEFINE_MUTEX(rt_constraints_mutex);
8359 static unsigned long to_ratio(u64 period, u64 runtime)
8361 if (runtime == RUNTIME_INF)
8364 return div64_u64(runtime << 20, period);
8367 /* Must be called with tasklist_lock held */
8368 static inline int tg_has_rt_tasks(struct task_group *tg)
8370 struct task_struct *g, *p;
8372 do_each_thread(g, p) {
8373 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8375 } while_each_thread(g, p);
8380 struct rt_schedulable_data {
8381 struct task_group *tg;
8386 static int tg_schedulable(struct task_group *tg, void *data)
8388 struct rt_schedulable_data *d = data;
8389 struct task_group *child;
8390 unsigned long total, sum = 0;
8391 u64 period, runtime;
8393 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8394 runtime = tg->rt_bandwidth.rt_runtime;
8397 period = d->rt_period;
8398 runtime = d->rt_runtime;
8402 * Cannot have more runtime than the period.
8404 if (runtime > period && runtime != RUNTIME_INF)
8408 * Ensure we don't starve existing RT tasks.
8410 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8413 total = to_ratio(period, runtime);
8416 * Nobody can have more than the global setting allows.
8418 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8422 * The sum of our children's runtime should not exceed our own.
8424 list_for_each_entry_rcu(child, &tg->children, siblings) {
8425 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8426 runtime = child->rt_bandwidth.rt_runtime;
8428 if (child == d->tg) {
8429 period = d->rt_period;
8430 runtime = d->rt_runtime;
8433 sum += to_ratio(period, runtime);
8442 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8444 struct rt_schedulable_data data = {
8446 .rt_period = period,
8447 .rt_runtime = runtime,
8450 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8453 static int tg_set_bandwidth(struct task_group *tg,
8454 u64 rt_period, u64 rt_runtime)
8458 mutex_lock(&rt_constraints_mutex);
8459 read_lock(&tasklist_lock);
8460 err = __rt_schedulable(tg, rt_period, rt_runtime);
8464 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8465 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8466 tg->rt_bandwidth.rt_runtime = rt_runtime;
8468 for_each_possible_cpu(i) {
8469 struct rt_rq *rt_rq = tg->rt_rq[i];
8471 raw_spin_lock(&rt_rq->rt_runtime_lock);
8472 rt_rq->rt_runtime = rt_runtime;
8473 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8475 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8477 read_unlock(&tasklist_lock);
8478 mutex_unlock(&rt_constraints_mutex);
8483 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8485 u64 rt_runtime, rt_period;
8487 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8488 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8489 if (rt_runtime_us < 0)
8490 rt_runtime = RUNTIME_INF;
8492 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8495 long sched_group_rt_runtime(struct task_group *tg)
8499 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8502 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8503 do_div(rt_runtime_us, NSEC_PER_USEC);
8504 return rt_runtime_us;
8507 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8509 u64 rt_runtime, rt_period;
8511 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8512 rt_runtime = tg->rt_bandwidth.rt_runtime;
8517 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8520 long sched_group_rt_period(struct task_group *tg)
8524 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8525 do_div(rt_period_us, NSEC_PER_USEC);
8526 return rt_period_us;
8529 static int sched_rt_global_constraints(void)
8531 u64 runtime, period;
8534 if (sysctl_sched_rt_period <= 0)
8537 runtime = global_rt_runtime();
8538 period = global_rt_period();
8541 * Sanity check on the sysctl variables.
8543 if (runtime > period && runtime != RUNTIME_INF)
8546 mutex_lock(&rt_constraints_mutex);
8547 read_lock(&tasklist_lock);
8548 ret = __rt_schedulable(NULL, 0, 0);
8549 read_unlock(&tasklist_lock);
8550 mutex_unlock(&rt_constraints_mutex);
8555 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8557 /* Don't accept realtime tasks when there is no way for them to run */
8558 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8564 #else /* !CONFIG_RT_GROUP_SCHED */
8565 static int sched_rt_global_constraints(void)
8567 unsigned long flags;
8570 if (sysctl_sched_rt_period <= 0)
8574 * There's always some RT tasks in the root group
8575 * -- migration, kstopmachine etc..
8577 if (sysctl_sched_rt_runtime == 0)
8580 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8581 for_each_possible_cpu(i) {
8582 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8584 raw_spin_lock(&rt_rq->rt_runtime_lock);
8585 rt_rq->rt_runtime = global_rt_runtime();
8586 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8588 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8592 #endif /* CONFIG_RT_GROUP_SCHED */
8594 int sched_rt_handler(struct ctl_table *table, int write,
8595 void __user *buffer, size_t *lenp,
8599 int old_period, old_runtime;
8600 static DEFINE_MUTEX(mutex);
8603 old_period = sysctl_sched_rt_period;
8604 old_runtime = sysctl_sched_rt_runtime;
8606 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8608 if (!ret && write) {
8609 ret = sched_rt_global_constraints();
8611 sysctl_sched_rt_period = old_period;
8612 sysctl_sched_rt_runtime = old_runtime;
8614 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8615 def_rt_bandwidth.rt_period =
8616 ns_to_ktime(global_rt_period());
8619 mutex_unlock(&mutex);
8624 #ifdef CONFIG_CGROUP_SCHED
8626 /* return corresponding task_group object of a cgroup */
8627 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8629 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8630 struct task_group, css);
8633 static struct cgroup_subsys_state *
8634 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8636 struct task_group *tg, *parent;
8638 if (!cgrp->parent) {
8639 /* This is early initialization for the top cgroup */
8640 return &init_task_group.css;
8643 parent = cgroup_tg(cgrp->parent);
8644 tg = sched_create_group(parent);
8646 return ERR_PTR(-ENOMEM);
8652 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8654 struct task_group *tg = cgroup_tg(cgrp);
8656 sched_destroy_group(tg);
8660 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8662 #ifdef CONFIG_RT_GROUP_SCHED
8663 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
8666 /* We don't support RT-tasks being in separate groups */
8667 if (tsk->sched_class != &fair_sched_class)
8674 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8675 struct task_struct *tsk, bool threadgroup)
8677 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
8681 struct task_struct *c;
8683 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8684 retval = cpu_cgroup_can_attach_task(cgrp, c);
8696 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8697 struct cgroup *old_cont, struct task_struct *tsk,
8700 sched_move_task(tsk);
8702 struct task_struct *c;
8704 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8711 #ifdef CONFIG_FAIR_GROUP_SCHED
8712 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8715 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8718 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8720 struct task_group *tg = cgroup_tg(cgrp);
8722 return (u64) tg->shares;
8724 #endif /* CONFIG_FAIR_GROUP_SCHED */
8726 #ifdef CONFIG_RT_GROUP_SCHED
8727 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8730 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8733 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8735 return sched_group_rt_runtime(cgroup_tg(cgrp));
8738 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8741 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8744 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8746 return sched_group_rt_period(cgroup_tg(cgrp));
8748 #endif /* CONFIG_RT_GROUP_SCHED */
8750 static struct cftype cpu_files[] = {
8751 #ifdef CONFIG_FAIR_GROUP_SCHED
8754 .read_u64 = cpu_shares_read_u64,
8755 .write_u64 = cpu_shares_write_u64,
8758 #ifdef CONFIG_RT_GROUP_SCHED
8760 .name = "rt_runtime_us",
8761 .read_s64 = cpu_rt_runtime_read,
8762 .write_s64 = cpu_rt_runtime_write,
8765 .name = "rt_period_us",
8766 .read_u64 = cpu_rt_period_read_uint,
8767 .write_u64 = cpu_rt_period_write_uint,
8772 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8774 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8777 struct cgroup_subsys cpu_cgroup_subsys = {
8779 .create = cpu_cgroup_create,
8780 .destroy = cpu_cgroup_destroy,
8781 .can_attach = cpu_cgroup_can_attach,
8782 .attach = cpu_cgroup_attach,
8783 .populate = cpu_cgroup_populate,
8784 .subsys_id = cpu_cgroup_subsys_id,
8788 #endif /* CONFIG_CGROUP_SCHED */
8790 #ifdef CONFIG_CGROUP_CPUACCT
8793 * CPU accounting code for task groups.
8795 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8796 * (balbir@in.ibm.com).
8799 /* track cpu usage of a group of tasks and its child groups */
8801 struct cgroup_subsys_state css;
8802 /* cpuusage holds pointer to a u64-type object on every cpu */
8804 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
8805 struct cpuacct *parent;
8808 struct cgroup_subsys cpuacct_subsys;
8810 /* return cpu accounting group corresponding to this container */
8811 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8813 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8814 struct cpuacct, css);
8817 /* return cpu accounting group to which this task belongs */
8818 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8820 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8821 struct cpuacct, css);
8824 /* create a new cpu accounting group */
8825 static struct cgroup_subsys_state *cpuacct_create(
8826 struct cgroup_subsys *ss, struct cgroup *cgrp)
8828 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8834 ca->cpuusage = alloc_percpu(u64);
8838 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8839 if (percpu_counter_init(&ca->cpustat[i], 0))
8840 goto out_free_counters;
8843 ca->parent = cgroup_ca(cgrp->parent);
8849 percpu_counter_destroy(&ca->cpustat[i]);
8850 free_percpu(ca->cpuusage);
8854 return ERR_PTR(-ENOMEM);
8857 /* destroy an existing cpu accounting group */
8859 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8861 struct cpuacct *ca = cgroup_ca(cgrp);
8864 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8865 percpu_counter_destroy(&ca->cpustat[i]);
8866 free_percpu(ca->cpuusage);
8870 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
8872 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8875 #ifndef CONFIG_64BIT
8877 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8879 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8881 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8889 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
8891 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8893 #ifndef CONFIG_64BIT
8895 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8897 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8899 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8905 /* return total cpu usage (in nanoseconds) of a group */
8906 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8908 struct cpuacct *ca = cgroup_ca(cgrp);
8909 u64 totalcpuusage = 0;
8912 for_each_present_cpu(i)
8913 totalcpuusage += cpuacct_cpuusage_read(ca, i);
8915 return totalcpuusage;
8918 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8921 struct cpuacct *ca = cgroup_ca(cgrp);
8930 for_each_present_cpu(i)
8931 cpuacct_cpuusage_write(ca, i, 0);
8937 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
8940 struct cpuacct *ca = cgroup_ca(cgroup);
8944 for_each_present_cpu(i) {
8945 percpu = cpuacct_cpuusage_read(ca, i);
8946 seq_printf(m, "%llu ", (unsigned long long) percpu);
8948 seq_printf(m, "\n");
8952 static const char *cpuacct_stat_desc[] = {
8953 [CPUACCT_STAT_USER] = "user",
8954 [CPUACCT_STAT_SYSTEM] = "system",
8957 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
8958 struct cgroup_map_cb *cb)
8960 struct cpuacct *ca = cgroup_ca(cgrp);
8963 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
8964 s64 val = percpu_counter_read(&ca->cpustat[i]);
8965 val = cputime64_to_clock_t(val);
8966 cb->fill(cb, cpuacct_stat_desc[i], val);
8971 static struct cftype files[] = {
8974 .read_u64 = cpuusage_read,
8975 .write_u64 = cpuusage_write,
8978 .name = "usage_percpu",
8979 .read_seq_string = cpuacct_percpu_seq_read,
8983 .read_map = cpuacct_stats_show,
8987 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
8989 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
8993 * charge this task's execution time to its accounting group.
8995 * called with rq->lock held.
8997 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9002 if (unlikely(!cpuacct_subsys.active))
9005 cpu = task_cpu(tsk);
9011 for (; ca; ca = ca->parent) {
9012 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9013 *cpuusage += cputime;
9020 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9021 * in cputime_t units. As a result, cpuacct_update_stats calls
9022 * percpu_counter_add with values large enough to always overflow the
9023 * per cpu batch limit causing bad SMP scalability.
9025 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9026 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9027 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9030 #define CPUACCT_BATCH \
9031 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9033 #define CPUACCT_BATCH 0
9037 * Charge the system/user time to the task's accounting group.
9039 static void cpuacct_update_stats(struct task_struct *tsk,
9040 enum cpuacct_stat_index idx, cputime_t val)
9043 int batch = CPUACCT_BATCH;
9045 if (unlikely(!cpuacct_subsys.active))
9052 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9058 struct cgroup_subsys cpuacct_subsys = {
9060 .create = cpuacct_create,
9061 .destroy = cpuacct_destroy,
9062 .populate = cpuacct_populate,
9063 .subsys_id = cpuacct_subsys_id,
9065 #endif /* CONFIG_CGROUP_CPUACCT */
9069 int rcu_expedited_torture_stats(char *page)
9073 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
9075 void synchronize_sched_expedited(void)
9078 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9080 #else /* #ifndef CONFIG_SMP */
9082 static DEFINE_PER_CPU(struct migration_req, rcu_migration_req);
9083 static DEFINE_MUTEX(rcu_sched_expedited_mutex);
9085 #define RCU_EXPEDITED_STATE_POST -2
9086 #define RCU_EXPEDITED_STATE_IDLE -1
9088 static int rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
9090 int rcu_expedited_torture_stats(char *page)
9095 cnt += sprintf(&page[cnt], "state: %d /", rcu_expedited_state);
9096 for_each_online_cpu(cpu) {
9097 cnt += sprintf(&page[cnt], " %d:%d",
9098 cpu, per_cpu(rcu_migration_req, cpu).dest_cpu);
9100 cnt += sprintf(&page[cnt], "\n");
9103 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
9105 static long synchronize_sched_expedited_count;
9108 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
9109 * approach to force grace period to end quickly. This consumes
9110 * significant time on all CPUs, and is thus not recommended for
9111 * any sort of common-case code.
9113 * Note that it is illegal to call this function while holding any
9114 * lock that is acquired by a CPU-hotplug notifier. Failing to
9115 * observe this restriction will result in deadlock.
9117 void synchronize_sched_expedited(void)
9120 unsigned long flags;
9121 bool need_full_sync = 0;
9123 struct migration_req *req;
9127 smp_mb(); /* ensure prior mod happens before capturing snap. */
9128 snap = ACCESS_ONCE(synchronize_sched_expedited_count) + 1;
9130 while (!mutex_trylock(&rcu_sched_expedited_mutex)) {
9132 if (trycount++ < 10)
9133 udelay(trycount * num_online_cpus());
9135 synchronize_sched();
9138 if (ACCESS_ONCE(synchronize_sched_expedited_count) - snap > 0) {
9139 smp_mb(); /* ensure test happens before caller kfree */
9144 rcu_expedited_state = RCU_EXPEDITED_STATE_POST;
9145 for_each_online_cpu(cpu) {
9147 req = &per_cpu(rcu_migration_req, cpu);
9148 init_completion(&req->done);
9150 req->dest_cpu = RCU_MIGRATION_NEED_QS;
9151 raw_spin_lock_irqsave(&rq->lock, flags);
9152 list_add(&req->list, &rq->migration_queue);
9153 raw_spin_unlock_irqrestore(&rq->lock, flags);
9154 wake_up_process(rq->migration_thread);
9156 for_each_online_cpu(cpu) {
9157 rcu_expedited_state = cpu;
9158 req = &per_cpu(rcu_migration_req, cpu);
9160 wait_for_completion(&req->done);
9161 raw_spin_lock_irqsave(&rq->lock, flags);
9162 if (unlikely(req->dest_cpu == RCU_MIGRATION_MUST_SYNC))
9164 req->dest_cpu = RCU_MIGRATION_IDLE;
9165 raw_spin_unlock_irqrestore(&rq->lock, flags);
9167 rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
9168 synchronize_sched_expedited_count++;
9169 mutex_unlock(&rcu_sched_expedited_mutex);
9172 synchronize_sched();
9174 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9176 #endif /* #else #ifndef CONFIG_SMP */