4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/perf_event.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/proc_fs.h>
59 #include <linux/seq_file.h>
60 #include <linux/stop_machine.h>
61 #include <linux/sysctl.h>
62 #include <linux/syscalls.h>
63 #include <linux/times.h>
64 #include <linux/tsacct_kern.h>
65 #include <linux/kprobes.h>
66 #include <linux/delayacct.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
73 #include <linux/ftrace.h>
74 #include <linux/slab.h>
77 #include <asm/irq_regs.h>
79 #include "sched_cpupri.h"
81 #define CREATE_TRACE_POINTS
82 #include <trace/events/sched.h>
85 * Convert user-nice values [ -20 ... 0 ... 19 ]
86 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
89 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
90 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
91 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
94 * 'User priority' is the nice value converted to something we
95 * can work with better when scaling various scheduler parameters,
96 * it's a [ 0 ... 39 ] range.
98 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
99 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
100 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
103 * Helpers for converting nanosecond timing to jiffy resolution
105 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
107 #define NICE_0_LOAD SCHED_LOAD_SCALE
108 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
111 * These are the 'tuning knobs' of the scheduler:
113 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
114 * Timeslices get refilled after they expire.
116 #define DEF_TIMESLICE (100 * HZ / 1000)
119 * single value that denotes runtime == period, ie unlimited time.
121 #define RUNTIME_INF ((u64)~0ULL)
123 static inline int rt_policy(int policy)
125 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
130 static inline int task_has_rt_policy(struct task_struct *p)
132 return rt_policy(p->policy);
136 * This is the priority-queue data structure of the RT scheduling class:
138 struct rt_prio_array {
139 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
140 struct list_head queue[MAX_RT_PRIO];
143 struct rt_bandwidth {
144 /* nests inside the rq lock: */
145 raw_spinlock_t rt_runtime_lock;
148 struct hrtimer rt_period_timer;
151 static struct rt_bandwidth def_rt_bandwidth;
153 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
155 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
157 struct rt_bandwidth *rt_b =
158 container_of(timer, struct rt_bandwidth, rt_period_timer);
164 now = hrtimer_cb_get_time(timer);
165 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
170 idle = do_sched_rt_period_timer(rt_b, overrun);
173 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
177 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
179 rt_b->rt_period = ns_to_ktime(period);
180 rt_b->rt_runtime = runtime;
182 raw_spin_lock_init(&rt_b->rt_runtime_lock);
184 hrtimer_init(&rt_b->rt_period_timer,
185 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
186 rt_b->rt_period_timer.function = sched_rt_period_timer;
189 static inline int rt_bandwidth_enabled(void)
191 return sysctl_sched_rt_runtime >= 0;
194 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
198 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
201 if (hrtimer_active(&rt_b->rt_period_timer))
204 raw_spin_lock(&rt_b->rt_runtime_lock);
209 if (hrtimer_active(&rt_b->rt_period_timer))
212 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
213 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
215 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
216 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
217 delta = ktime_to_ns(ktime_sub(hard, soft));
218 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
219 HRTIMER_MODE_ABS_PINNED, 0);
221 raw_spin_unlock(&rt_b->rt_runtime_lock);
224 #ifdef CONFIG_RT_GROUP_SCHED
225 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
227 hrtimer_cancel(&rt_b->rt_period_timer);
232 * sched_domains_mutex serializes calls to arch_init_sched_domains,
233 * detach_destroy_domains and partition_sched_domains.
235 static DEFINE_MUTEX(sched_domains_mutex);
237 #ifdef CONFIG_CGROUP_SCHED
239 #include <linux/cgroup.h>
243 static LIST_HEAD(task_groups);
245 /* task group related information */
247 struct cgroup_subsys_state css;
249 #ifdef CONFIG_FAIR_GROUP_SCHED
250 /* schedulable entities of this group on each cpu */
251 struct sched_entity **se;
252 /* runqueue "owned" by this group on each cpu */
253 struct cfs_rq **cfs_rq;
254 unsigned long shares;
257 #ifdef CONFIG_RT_GROUP_SCHED
258 struct sched_rt_entity **rt_se;
259 struct rt_rq **rt_rq;
261 struct rt_bandwidth rt_bandwidth;
265 struct list_head list;
267 struct task_group *parent;
268 struct list_head siblings;
269 struct list_head children;
272 #define root_task_group init_task_group
274 /* task_group_lock serializes add/remove of task groups and also changes to
275 * a task group's cpu shares.
277 static DEFINE_SPINLOCK(task_group_lock);
279 #ifdef CONFIG_FAIR_GROUP_SCHED
282 static int root_task_group_empty(void)
284 return list_empty(&root_task_group.children);
288 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
291 * A weight of 0 or 1 can cause arithmetics problems.
292 * A weight of a cfs_rq is the sum of weights of which entities
293 * are queued on this cfs_rq, so a weight of a entity should not be
294 * too large, so as the shares value of a task group.
295 * (The default weight is 1024 - so there's no practical
296 * limitation from this.)
299 #define MAX_SHARES (1UL << 18)
301 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
304 /* Default task group.
305 * Every task in system belong to this group at bootup.
307 struct task_group init_task_group;
309 #endif /* CONFIG_CGROUP_SCHED */
311 /* CFS-related fields in a runqueue */
313 struct load_weight load;
314 unsigned long nr_running;
319 struct rb_root tasks_timeline;
320 struct rb_node *rb_leftmost;
322 struct list_head tasks;
323 struct list_head *balance_iterator;
326 * 'curr' points to currently running entity on this cfs_rq.
327 * It is set to NULL otherwise (i.e when none are currently running).
329 struct sched_entity *curr, *next, *last;
331 unsigned int nr_spread_over;
333 #ifdef CONFIG_FAIR_GROUP_SCHED
334 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
337 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
338 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
339 * (like users, containers etc.)
341 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
342 * list is used during load balance.
344 struct list_head leaf_cfs_rq_list;
345 struct task_group *tg; /* group that "owns" this runqueue */
349 * the part of load.weight contributed by tasks
351 unsigned long task_weight;
354 * h_load = weight * f(tg)
356 * Where f(tg) is the recursive weight fraction assigned to
359 unsigned long h_load;
362 * this cpu's part of tg->shares
364 unsigned long shares;
367 * load.weight at the time we set shares
369 unsigned long rq_weight;
374 /* Real-Time classes' related field in a runqueue: */
376 struct rt_prio_array active;
377 unsigned long rt_nr_running;
378 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
380 int curr; /* highest queued rt task prio */
382 int next; /* next highest */
387 unsigned long rt_nr_migratory;
388 unsigned long rt_nr_total;
390 struct plist_head pushable_tasks;
395 /* Nests inside the rq lock: */
396 raw_spinlock_t rt_runtime_lock;
398 #ifdef CONFIG_RT_GROUP_SCHED
399 unsigned long rt_nr_boosted;
402 struct list_head leaf_rt_rq_list;
403 struct task_group *tg;
410 * We add the notion of a root-domain which will be used to define per-domain
411 * variables. Each exclusive cpuset essentially defines an island domain by
412 * fully partitioning the member cpus from any other cpuset. Whenever a new
413 * exclusive cpuset is created, we also create and attach a new root-domain
420 cpumask_var_t online;
423 * The "RT overload" flag: it gets set if a CPU has more than
424 * one runnable RT task.
426 cpumask_var_t rto_mask;
429 struct cpupri cpupri;
434 * By default the system creates a single root-domain with all cpus as
435 * members (mimicking the global state we have today).
437 static struct root_domain def_root_domain;
442 * This is the main, per-CPU runqueue data structure.
444 * Locking rule: those places that want to lock multiple runqueues
445 * (such as the load balancing or the thread migration code), lock
446 * acquire operations must be ordered by ascending &runqueue.
453 * nr_running and cpu_load should be in the same cacheline because
454 * remote CPUs use both these fields when doing load calculation.
456 unsigned long nr_running;
457 #define CPU_LOAD_IDX_MAX 5
458 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
461 unsigned char in_nohz_recently;
463 unsigned int skip_clock_update;
465 /* capture load from *all* tasks on this cpu: */
466 struct load_weight load;
467 unsigned long nr_load_updates;
473 #ifdef CONFIG_FAIR_GROUP_SCHED
474 /* list of leaf cfs_rq on this cpu: */
475 struct list_head leaf_cfs_rq_list;
477 #ifdef CONFIG_RT_GROUP_SCHED
478 struct list_head leaf_rt_rq_list;
482 * This is part of a global counter where only the total sum
483 * over all CPUs matters. A task can increase this counter on
484 * one CPU and if it got migrated afterwards it may decrease
485 * it on another CPU. Always updated under the runqueue lock:
487 unsigned long nr_uninterruptible;
489 struct task_struct *curr, *idle;
490 unsigned long next_balance;
491 struct mm_struct *prev_mm;
498 struct root_domain *rd;
499 struct sched_domain *sd;
501 unsigned long cpu_power;
503 unsigned char idle_at_tick;
504 /* For active balancing */
508 struct cpu_stop_work active_balance_work;
509 /* cpu of this runqueue: */
513 unsigned long avg_load_per_task;
521 /* calc_load related fields */
522 unsigned long calc_load_update;
523 long calc_load_active;
525 #ifdef CONFIG_SCHED_HRTICK
527 int hrtick_csd_pending;
528 struct call_single_data hrtick_csd;
530 struct hrtimer hrtick_timer;
533 #ifdef CONFIG_SCHEDSTATS
535 struct sched_info rq_sched_info;
536 unsigned long long rq_cpu_time;
537 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
539 /* sys_sched_yield() stats */
540 unsigned int yld_count;
542 /* schedule() stats */
543 unsigned int sched_switch;
544 unsigned int sched_count;
545 unsigned int sched_goidle;
547 /* try_to_wake_up() stats */
548 unsigned int ttwu_count;
549 unsigned int ttwu_local;
552 unsigned int bkl_count;
556 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
559 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
561 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
564 * A queue event has occurred, and we're going to schedule. In
565 * this case, we can save a useless back to back clock update.
567 if (test_tsk_need_resched(p))
568 rq->skip_clock_update = 1;
571 static inline int cpu_of(struct rq *rq)
580 #define rcu_dereference_check_sched_domain(p) \
581 rcu_dereference_check((p), \
582 rcu_read_lock_sched_held() || \
583 lockdep_is_held(&sched_domains_mutex))
586 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
587 * See detach_destroy_domains: synchronize_sched for details.
589 * The domain tree of any CPU may only be accessed from within
590 * preempt-disabled sections.
592 #define for_each_domain(cpu, __sd) \
593 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
595 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
596 #define this_rq() (&__get_cpu_var(runqueues))
597 #define task_rq(p) cpu_rq(task_cpu(p))
598 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
599 #define raw_rq() (&__raw_get_cpu_var(runqueues))
601 #ifdef CONFIG_CGROUP_SCHED
604 * Return the group to which this tasks belongs.
606 * We use task_subsys_state_check() and extend the RCU verification
607 * with lockdep_is_held(&task_rq(p)->lock) because cpu_cgroup_attach()
608 * holds that lock for each task it moves into the cgroup. Therefore
609 * by holding that lock, we pin the task to the current cgroup.
611 static inline struct task_group *task_group(struct task_struct *p)
613 struct cgroup_subsys_state *css;
615 css = task_subsys_state_check(p, cpu_cgroup_subsys_id,
616 lockdep_is_held(&task_rq(p)->lock));
617 return container_of(css, struct task_group, css);
620 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
621 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
623 #ifdef CONFIG_FAIR_GROUP_SCHED
624 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
625 p->se.parent = task_group(p)->se[cpu];
628 #ifdef CONFIG_RT_GROUP_SCHED
629 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
630 p->rt.parent = task_group(p)->rt_se[cpu];
634 #else /* CONFIG_CGROUP_SCHED */
636 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
637 static inline struct task_group *task_group(struct task_struct *p)
642 #endif /* CONFIG_CGROUP_SCHED */
644 inline void update_rq_clock(struct rq *rq)
646 if (!rq->skip_clock_update)
647 rq->clock = sched_clock_cpu(cpu_of(rq));
651 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
653 #ifdef CONFIG_SCHED_DEBUG
654 # define const_debug __read_mostly
656 # define const_debug static const
661 * @cpu: the processor in question.
663 * Returns true if the current cpu runqueue is locked.
664 * This interface allows printk to be called with the runqueue lock
665 * held and know whether or not it is OK to wake up the klogd.
667 int runqueue_is_locked(int cpu)
669 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
673 * Debugging: various feature bits
676 #define SCHED_FEAT(name, enabled) \
677 __SCHED_FEAT_##name ,
680 #include "sched_features.h"
685 #define SCHED_FEAT(name, enabled) \
686 (1UL << __SCHED_FEAT_##name) * enabled |
688 const_debug unsigned int sysctl_sched_features =
689 #include "sched_features.h"
694 #ifdef CONFIG_SCHED_DEBUG
695 #define SCHED_FEAT(name, enabled) \
698 static __read_mostly char *sched_feat_names[] = {
699 #include "sched_features.h"
705 static int sched_feat_show(struct seq_file *m, void *v)
709 for (i = 0; sched_feat_names[i]; i++) {
710 if (!(sysctl_sched_features & (1UL << i)))
712 seq_printf(m, "%s ", sched_feat_names[i]);
720 sched_feat_write(struct file *filp, const char __user *ubuf,
721 size_t cnt, loff_t *ppos)
731 if (copy_from_user(&buf, ubuf, cnt))
737 if (strncmp(buf, "NO_", 3) == 0) {
742 for (i = 0; sched_feat_names[i]; i++) {
743 if (strcmp(cmp, sched_feat_names[i]) == 0) {
745 sysctl_sched_features &= ~(1UL << i);
747 sysctl_sched_features |= (1UL << i);
752 if (!sched_feat_names[i])
760 static int sched_feat_open(struct inode *inode, struct file *filp)
762 return single_open(filp, sched_feat_show, NULL);
765 static const struct file_operations sched_feat_fops = {
766 .open = sched_feat_open,
767 .write = sched_feat_write,
770 .release = single_release,
773 static __init int sched_init_debug(void)
775 debugfs_create_file("sched_features", 0644, NULL, NULL,
780 late_initcall(sched_init_debug);
784 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
787 * Number of tasks to iterate in a single balance run.
788 * Limited because this is done with IRQs disabled.
790 const_debug unsigned int sysctl_sched_nr_migrate = 32;
793 * ratelimit for updating the group shares.
796 unsigned int sysctl_sched_shares_ratelimit = 250000;
797 unsigned int normalized_sysctl_sched_shares_ratelimit = 250000;
800 * Inject some fuzzyness into changing the per-cpu group shares
801 * this avoids remote rq-locks at the expense of fairness.
804 unsigned int sysctl_sched_shares_thresh = 4;
807 * period over which we average the RT time consumption, measured
812 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
815 * period over which we measure -rt task cpu usage in us.
818 unsigned int sysctl_sched_rt_period = 1000000;
820 static __read_mostly int scheduler_running;
823 * part of the period that we allow rt tasks to run in us.
826 int sysctl_sched_rt_runtime = 950000;
828 static inline u64 global_rt_period(void)
830 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
833 static inline u64 global_rt_runtime(void)
835 if (sysctl_sched_rt_runtime < 0)
838 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
841 #ifndef prepare_arch_switch
842 # define prepare_arch_switch(next) do { } while (0)
844 #ifndef finish_arch_switch
845 # define finish_arch_switch(prev) do { } while (0)
848 static inline int task_current(struct rq *rq, struct task_struct *p)
850 return rq->curr == p;
853 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
854 static inline int task_running(struct rq *rq, struct task_struct *p)
856 return task_current(rq, p);
859 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
863 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
865 #ifdef CONFIG_DEBUG_SPINLOCK
866 /* this is a valid case when another task releases the spinlock */
867 rq->lock.owner = current;
870 * If we are tracking spinlock dependencies then we have to
871 * fix up the runqueue lock - which gets 'carried over' from
874 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
876 raw_spin_unlock_irq(&rq->lock);
879 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
880 static inline int task_running(struct rq *rq, struct task_struct *p)
885 return task_current(rq, p);
889 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
893 * We can optimise this out completely for !SMP, because the
894 * SMP rebalancing from interrupt is the only thing that cares
899 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
900 raw_spin_unlock_irq(&rq->lock);
902 raw_spin_unlock(&rq->lock);
906 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
910 * After ->oncpu is cleared, the task can be moved to a different CPU.
911 * We must ensure this doesn't happen until the switch is completely
917 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
921 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
924 * Check whether the task is waking, we use this to synchronize ->cpus_allowed
927 static inline int task_is_waking(struct task_struct *p)
929 return unlikely(p->state == TASK_WAKING);
933 * __task_rq_lock - lock the runqueue a given task resides on.
934 * Must be called interrupts disabled.
936 static inline struct rq *__task_rq_lock(struct task_struct *p)
943 raw_spin_lock(&rq->lock);
944 if (likely(rq == task_rq(p)))
946 raw_spin_unlock(&rq->lock);
951 * task_rq_lock - lock the runqueue a given task resides on and disable
952 * interrupts. Note the ordering: we can safely lookup the task_rq without
953 * explicitly disabling preemption.
955 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
961 local_irq_save(*flags);
963 raw_spin_lock(&rq->lock);
964 if (likely(rq == task_rq(p)))
966 raw_spin_unlock_irqrestore(&rq->lock, *flags);
970 static void __task_rq_unlock(struct rq *rq)
973 raw_spin_unlock(&rq->lock);
976 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
979 raw_spin_unlock_irqrestore(&rq->lock, *flags);
983 * this_rq_lock - lock this runqueue and disable interrupts.
985 static struct rq *this_rq_lock(void)
992 raw_spin_lock(&rq->lock);
997 #ifdef CONFIG_SCHED_HRTICK
999 * Use HR-timers to deliver accurate preemption points.
1001 * Its all a bit involved since we cannot program an hrt while holding the
1002 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1005 * When we get rescheduled we reprogram the hrtick_timer outside of the
1011 * - enabled by features
1012 * - hrtimer is actually high res
1014 static inline int hrtick_enabled(struct rq *rq)
1016 if (!sched_feat(HRTICK))
1018 if (!cpu_active(cpu_of(rq)))
1020 return hrtimer_is_hres_active(&rq->hrtick_timer);
1023 static void hrtick_clear(struct rq *rq)
1025 if (hrtimer_active(&rq->hrtick_timer))
1026 hrtimer_cancel(&rq->hrtick_timer);
1030 * High-resolution timer tick.
1031 * Runs from hardirq context with interrupts disabled.
1033 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1035 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1037 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1039 raw_spin_lock(&rq->lock);
1040 update_rq_clock(rq);
1041 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1042 raw_spin_unlock(&rq->lock);
1044 return HRTIMER_NORESTART;
1049 * called from hardirq (IPI) context
1051 static void __hrtick_start(void *arg)
1053 struct rq *rq = arg;
1055 raw_spin_lock(&rq->lock);
1056 hrtimer_restart(&rq->hrtick_timer);
1057 rq->hrtick_csd_pending = 0;
1058 raw_spin_unlock(&rq->lock);
1062 * Called to set the hrtick timer state.
1064 * called with rq->lock held and irqs disabled
1066 static void hrtick_start(struct rq *rq, u64 delay)
1068 struct hrtimer *timer = &rq->hrtick_timer;
1069 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1071 hrtimer_set_expires(timer, time);
1073 if (rq == this_rq()) {
1074 hrtimer_restart(timer);
1075 } else if (!rq->hrtick_csd_pending) {
1076 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1077 rq->hrtick_csd_pending = 1;
1082 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1084 int cpu = (int)(long)hcpu;
1087 case CPU_UP_CANCELED:
1088 case CPU_UP_CANCELED_FROZEN:
1089 case CPU_DOWN_PREPARE:
1090 case CPU_DOWN_PREPARE_FROZEN:
1092 case CPU_DEAD_FROZEN:
1093 hrtick_clear(cpu_rq(cpu));
1100 static __init void init_hrtick(void)
1102 hotcpu_notifier(hotplug_hrtick, 0);
1106 * Called to set the hrtick timer state.
1108 * called with rq->lock held and irqs disabled
1110 static void hrtick_start(struct rq *rq, u64 delay)
1112 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1113 HRTIMER_MODE_REL_PINNED, 0);
1116 static inline void init_hrtick(void)
1119 #endif /* CONFIG_SMP */
1121 static void init_rq_hrtick(struct rq *rq)
1124 rq->hrtick_csd_pending = 0;
1126 rq->hrtick_csd.flags = 0;
1127 rq->hrtick_csd.func = __hrtick_start;
1128 rq->hrtick_csd.info = rq;
1131 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1132 rq->hrtick_timer.function = hrtick;
1134 #else /* CONFIG_SCHED_HRTICK */
1135 static inline void hrtick_clear(struct rq *rq)
1139 static inline void init_rq_hrtick(struct rq *rq)
1143 static inline void init_hrtick(void)
1146 #endif /* CONFIG_SCHED_HRTICK */
1149 * resched_task - mark a task 'to be rescheduled now'.
1151 * On UP this means the setting of the need_resched flag, on SMP it
1152 * might also involve a cross-CPU call to trigger the scheduler on
1157 #ifndef tsk_is_polling
1158 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1161 static void resched_task(struct task_struct *p)
1165 assert_raw_spin_locked(&task_rq(p)->lock);
1167 if (test_tsk_need_resched(p))
1170 set_tsk_need_resched(p);
1173 if (cpu == smp_processor_id())
1176 /* NEED_RESCHED must be visible before we test polling */
1178 if (!tsk_is_polling(p))
1179 smp_send_reschedule(cpu);
1182 static void resched_cpu(int cpu)
1184 struct rq *rq = cpu_rq(cpu);
1185 unsigned long flags;
1187 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1189 resched_task(cpu_curr(cpu));
1190 raw_spin_unlock_irqrestore(&rq->lock, flags);
1195 * When add_timer_on() enqueues a timer into the timer wheel of an
1196 * idle CPU then this timer might expire before the next timer event
1197 * which is scheduled to wake up that CPU. In case of a completely
1198 * idle system the next event might even be infinite time into the
1199 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1200 * leaves the inner idle loop so the newly added timer is taken into
1201 * account when the CPU goes back to idle and evaluates the timer
1202 * wheel for the next timer event.
1204 void wake_up_idle_cpu(int cpu)
1206 struct rq *rq = cpu_rq(cpu);
1208 if (cpu == smp_processor_id())
1212 * This is safe, as this function is called with the timer
1213 * wheel base lock of (cpu) held. When the CPU is on the way
1214 * to idle and has not yet set rq->curr to idle then it will
1215 * be serialized on the timer wheel base lock and take the new
1216 * timer into account automatically.
1218 if (rq->curr != rq->idle)
1222 * We can set TIF_RESCHED on the idle task of the other CPU
1223 * lockless. The worst case is that the other CPU runs the
1224 * idle task through an additional NOOP schedule()
1226 set_tsk_need_resched(rq->idle);
1228 /* NEED_RESCHED must be visible before we test polling */
1230 if (!tsk_is_polling(rq->idle))
1231 smp_send_reschedule(cpu);
1234 #endif /* CONFIG_NO_HZ */
1236 static u64 sched_avg_period(void)
1238 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1241 static void sched_avg_update(struct rq *rq)
1243 s64 period = sched_avg_period();
1245 while ((s64)(rq->clock - rq->age_stamp) > period) {
1247 * Inline assembly required to prevent the compiler
1248 * optimising this loop into a divmod call.
1249 * See __iter_div_u64_rem() for another example of this.
1251 asm("" : "+rm" (rq->age_stamp));
1252 rq->age_stamp += period;
1257 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1259 rq->rt_avg += rt_delta;
1260 sched_avg_update(rq);
1263 #else /* !CONFIG_SMP */
1264 static void resched_task(struct task_struct *p)
1266 assert_raw_spin_locked(&task_rq(p)->lock);
1267 set_tsk_need_resched(p);
1270 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1273 #endif /* CONFIG_SMP */
1275 #if BITS_PER_LONG == 32
1276 # define WMULT_CONST (~0UL)
1278 # define WMULT_CONST (1UL << 32)
1281 #define WMULT_SHIFT 32
1284 * Shift right and round:
1286 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1289 * delta *= weight / lw
1291 static unsigned long
1292 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1293 struct load_weight *lw)
1297 if (!lw->inv_weight) {
1298 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1301 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1305 tmp = (u64)delta_exec * weight;
1307 * Check whether we'd overflow the 64-bit multiplication:
1309 if (unlikely(tmp > WMULT_CONST))
1310 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1313 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1315 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1318 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1324 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1331 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1332 * of tasks with abnormal "nice" values across CPUs the contribution that
1333 * each task makes to its run queue's load is weighted according to its
1334 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1335 * scaled version of the new time slice allocation that they receive on time
1339 #define WEIGHT_IDLEPRIO 3
1340 #define WMULT_IDLEPRIO 1431655765
1343 * Nice levels are multiplicative, with a gentle 10% change for every
1344 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1345 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1346 * that remained on nice 0.
1348 * The "10% effect" is relative and cumulative: from _any_ nice level,
1349 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1350 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1351 * If a task goes up by ~10% and another task goes down by ~10% then
1352 * the relative distance between them is ~25%.)
1354 static const int prio_to_weight[40] = {
1355 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1356 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1357 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1358 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1359 /* 0 */ 1024, 820, 655, 526, 423,
1360 /* 5 */ 335, 272, 215, 172, 137,
1361 /* 10 */ 110, 87, 70, 56, 45,
1362 /* 15 */ 36, 29, 23, 18, 15,
1366 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1368 * In cases where the weight does not change often, we can use the
1369 * precalculated inverse to speed up arithmetics by turning divisions
1370 * into multiplications:
1372 static const u32 prio_to_wmult[40] = {
1373 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1374 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1375 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1376 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1377 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1378 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1379 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1380 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1383 /* Time spent by the tasks of the cpu accounting group executing in ... */
1384 enum cpuacct_stat_index {
1385 CPUACCT_STAT_USER, /* ... user mode */
1386 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1388 CPUACCT_STAT_NSTATS,
1391 #ifdef CONFIG_CGROUP_CPUACCT
1392 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1393 static void cpuacct_update_stats(struct task_struct *tsk,
1394 enum cpuacct_stat_index idx, cputime_t val);
1396 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1397 static inline void cpuacct_update_stats(struct task_struct *tsk,
1398 enum cpuacct_stat_index idx, cputime_t val) {}
1401 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1403 update_load_add(&rq->load, load);
1406 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1408 update_load_sub(&rq->load, load);
1411 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1412 typedef int (*tg_visitor)(struct task_group *, void *);
1415 * Iterate the full tree, calling @down when first entering a node and @up when
1416 * leaving it for the final time.
1418 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1420 struct task_group *parent, *child;
1424 parent = &root_task_group;
1426 ret = (*down)(parent, data);
1429 list_for_each_entry_rcu(child, &parent->children, siblings) {
1436 ret = (*up)(parent, data);
1441 parent = parent->parent;
1450 static int tg_nop(struct task_group *tg, void *data)
1457 /* Used instead of source_load when we know the type == 0 */
1458 static unsigned long weighted_cpuload(const int cpu)
1460 return cpu_rq(cpu)->load.weight;
1464 * Return a low guess at the load of a migration-source cpu weighted
1465 * according to the scheduling class and "nice" value.
1467 * We want to under-estimate the load of migration sources, to
1468 * balance conservatively.
1470 static unsigned long source_load(int cpu, int type)
1472 struct rq *rq = cpu_rq(cpu);
1473 unsigned long total = weighted_cpuload(cpu);
1475 if (type == 0 || !sched_feat(LB_BIAS))
1478 return min(rq->cpu_load[type-1], total);
1482 * Return a high guess at the load of a migration-target cpu weighted
1483 * according to the scheduling class and "nice" value.
1485 static unsigned long target_load(int cpu, int type)
1487 struct rq *rq = cpu_rq(cpu);
1488 unsigned long total = weighted_cpuload(cpu);
1490 if (type == 0 || !sched_feat(LB_BIAS))
1493 return max(rq->cpu_load[type-1], total);
1496 static unsigned long power_of(int cpu)
1498 return cpu_rq(cpu)->cpu_power;
1501 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1503 static unsigned long cpu_avg_load_per_task(int cpu)
1505 struct rq *rq = cpu_rq(cpu);
1506 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1509 rq->avg_load_per_task = rq->load.weight / nr_running;
1511 rq->avg_load_per_task = 0;
1513 return rq->avg_load_per_task;
1516 #ifdef CONFIG_FAIR_GROUP_SCHED
1518 static __read_mostly unsigned long __percpu *update_shares_data;
1520 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1523 * Calculate and set the cpu's group shares.
1525 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1526 unsigned long sd_shares,
1527 unsigned long sd_rq_weight,
1528 unsigned long *usd_rq_weight)
1530 unsigned long shares, rq_weight;
1533 rq_weight = usd_rq_weight[cpu];
1536 rq_weight = NICE_0_LOAD;
1540 * \Sum_j shares_j * rq_weight_i
1541 * shares_i = -----------------------------
1542 * \Sum_j rq_weight_j
1544 shares = (sd_shares * rq_weight) / sd_rq_weight;
1545 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1547 if (abs(shares - tg->se[cpu]->load.weight) >
1548 sysctl_sched_shares_thresh) {
1549 struct rq *rq = cpu_rq(cpu);
1550 unsigned long flags;
1552 raw_spin_lock_irqsave(&rq->lock, flags);
1553 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1554 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1555 __set_se_shares(tg->se[cpu], shares);
1556 raw_spin_unlock_irqrestore(&rq->lock, flags);
1561 * Re-compute the task group their per cpu shares over the given domain.
1562 * This needs to be done in a bottom-up fashion because the rq weight of a
1563 * parent group depends on the shares of its child groups.
1565 static int tg_shares_up(struct task_group *tg, void *data)
1567 unsigned long weight, rq_weight = 0, sum_weight = 0, shares = 0;
1568 unsigned long *usd_rq_weight;
1569 struct sched_domain *sd = data;
1570 unsigned long flags;
1576 local_irq_save(flags);
1577 usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id());
1579 for_each_cpu(i, sched_domain_span(sd)) {
1580 weight = tg->cfs_rq[i]->load.weight;
1581 usd_rq_weight[i] = weight;
1583 rq_weight += weight;
1585 * If there are currently no tasks on the cpu pretend there
1586 * is one of average load so that when a new task gets to
1587 * run here it will not get delayed by group starvation.
1590 weight = NICE_0_LOAD;
1592 sum_weight += weight;
1593 shares += tg->cfs_rq[i]->shares;
1597 rq_weight = sum_weight;
1599 if ((!shares && rq_weight) || shares > tg->shares)
1600 shares = tg->shares;
1602 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1603 shares = tg->shares;
1605 for_each_cpu(i, sched_domain_span(sd))
1606 update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight);
1608 local_irq_restore(flags);
1614 * Compute the cpu's hierarchical load factor for each task group.
1615 * This needs to be done in a top-down fashion because the load of a child
1616 * group is a fraction of its parents load.
1618 static int tg_load_down(struct task_group *tg, void *data)
1621 long cpu = (long)data;
1624 load = cpu_rq(cpu)->load.weight;
1626 load = tg->parent->cfs_rq[cpu]->h_load;
1627 load *= tg->cfs_rq[cpu]->shares;
1628 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1631 tg->cfs_rq[cpu]->h_load = load;
1636 static void update_shares(struct sched_domain *sd)
1641 if (root_task_group_empty())
1644 now = cpu_clock(raw_smp_processor_id());
1645 elapsed = now - sd->last_update;
1647 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1648 sd->last_update = now;
1649 walk_tg_tree(tg_nop, tg_shares_up, sd);
1653 static void update_h_load(long cpu)
1655 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1660 static inline void update_shares(struct sched_domain *sd)
1666 #ifdef CONFIG_PREEMPT
1668 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1671 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1672 * way at the expense of forcing extra atomic operations in all
1673 * invocations. This assures that the double_lock is acquired using the
1674 * same underlying policy as the spinlock_t on this architecture, which
1675 * reduces latency compared to the unfair variant below. However, it
1676 * also adds more overhead and therefore may reduce throughput.
1678 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1679 __releases(this_rq->lock)
1680 __acquires(busiest->lock)
1681 __acquires(this_rq->lock)
1683 raw_spin_unlock(&this_rq->lock);
1684 double_rq_lock(this_rq, busiest);
1691 * Unfair double_lock_balance: Optimizes throughput at the expense of
1692 * latency by eliminating extra atomic operations when the locks are
1693 * already in proper order on entry. This favors lower cpu-ids and will
1694 * grant the double lock to lower cpus over higher ids under contention,
1695 * regardless of entry order into the function.
1697 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1698 __releases(this_rq->lock)
1699 __acquires(busiest->lock)
1700 __acquires(this_rq->lock)
1704 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1705 if (busiest < this_rq) {
1706 raw_spin_unlock(&this_rq->lock);
1707 raw_spin_lock(&busiest->lock);
1708 raw_spin_lock_nested(&this_rq->lock,
1709 SINGLE_DEPTH_NESTING);
1712 raw_spin_lock_nested(&busiest->lock,
1713 SINGLE_DEPTH_NESTING);
1718 #endif /* CONFIG_PREEMPT */
1721 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1723 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1725 if (unlikely(!irqs_disabled())) {
1726 /* printk() doesn't work good under rq->lock */
1727 raw_spin_unlock(&this_rq->lock);
1731 return _double_lock_balance(this_rq, busiest);
1734 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1735 __releases(busiest->lock)
1737 raw_spin_unlock(&busiest->lock);
1738 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1742 * double_rq_lock - safely lock two runqueues
1744 * Note this does not disable interrupts like task_rq_lock,
1745 * you need to do so manually before calling.
1747 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1748 __acquires(rq1->lock)
1749 __acquires(rq2->lock)
1751 BUG_ON(!irqs_disabled());
1753 raw_spin_lock(&rq1->lock);
1754 __acquire(rq2->lock); /* Fake it out ;) */
1757 raw_spin_lock(&rq1->lock);
1758 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1760 raw_spin_lock(&rq2->lock);
1761 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1767 * double_rq_unlock - safely unlock two runqueues
1769 * Note this does not restore interrupts like task_rq_unlock,
1770 * you need to do so manually after calling.
1772 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1773 __releases(rq1->lock)
1774 __releases(rq2->lock)
1776 raw_spin_unlock(&rq1->lock);
1778 raw_spin_unlock(&rq2->lock);
1780 __release(rq2->lock);
1785 #ifdef CONFIG_FAIR_GROUP_SCHED
1786 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1789 cfs_rq->shares = shares;
1794 static void calc_load_account_idle(struct rq *this_rq);
1795 static void update_sysctl(void);
1796 static int get_update_sysctl_factor(void);
1798 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1800 set_task_rq(p, cpu);
1803 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1804 * successfuly executed on another CPU. We must ensure that updates of
1805 * per-task data have been completed by this moment.
1808 task_thread_info(p)->cpu = cpu;
1812 static const struct sched_class rt_sched_class;
1814 #define sched_class_highest (&rt_sched_class)
1815 #define for_each_class(class) \
1816 for (class = sched_class_highest; class; class = class->next)
1818 #include "sched_stats.h"
1820 static void inc_nr_running(struct rq *rq)
1825 static void dec_nr_running(struct rq *rq)
1830 static void set_load_weight(struct task_struct *p)
1833 * SCHED_IDLE tasks get minimal weight:
1835 if (p->policy == SCHED_IDLE) {
1836 p->se.load.weight = WEIGHT_IDLEPRIO;
1837 p->se.load.inv_weight = WMULT_IDLEPRIO;
1841 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1842 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1845 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1847 update_rq_clock(rq);
1848 sched_info_queued(p);
1849 p->sched_class->enqueue_task(rq, p, flags);
1853 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1855 update_rq_clock(rq);
1856 sched_info_dequeued(p);
1857 p->sched_class->dequeue_task(rq, p, flags);
1862 * activate_task - move a task to the runqueue.
1864 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1866 if (task_contributes_to_load(p))
1867 rq->nr_uninterruptible--;
1869 enqueue_task(rq, p, flags);
1874 * deactivate_task - remove a task from the runqueue.
1876 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1878 if (task_contributes_to_load(p))
1879 rq->nr_uninterruptible++;
1881 dequeue_task(rq, p, flags);
1885 #include "sched_idletask.c"
1886 #include "sched_fair.c"
1887 #include "sched_rt.c"
1888 #ifdef CONFIG_SCHED_DEBUG
1889 # include "sched_debug.c"
1893 * __normal_prio - return the priority that is based on the static prio
1895 static inline int __normal_prio(struct task_struct *p)
1897 return p->static_prio;
1901 * Calculate the expected normal priority: i.e. priority
1902 * without taking RT-inheritance into account. Might be
1903 * boosted by interactivity modifiers. Changes upon fork,
1904 * setprio syscalls, and whenever the interactivity
1905 * estimator recalculates.
1907 static inline int normal_prio(struct task_struct *p)
1911 if (task_has_rt_policy(p))
1912 prio = MAX_RT_PRIO-1 - p->rt_priority;
1914 prio = __normal_prio(p);
1919 * Calculate the current priority, i.e. the priority
1920 * taken into account by the scheduler. This value might
1921 * be boosted by RT tasks, or might be boosted by
1922 * interactivity modifiers. Will be RT if the task got
1923 * RT-boosted. If not then it returns p->normal_prio.
1925 static int effective_prio(struct task_struct *p)
1927 p->normal_prio = normal_prio(p);
1929 * If we are RT tasks or we were boosted to RT priority,
1930 * keep the priority unchanged. Otherwise, update priority
1931 * to the normal priority:
1933 if (!rt_prio(p->prio))
1934 return p->normal_prio;
1939 * task_curr - is this task currently executing on a CPU?
1940 * @p: the task in question.
1942 inline int task_curr(const struct task_struct *p)
1944 return cpu_curr(task_cpu(p)) == p;
1947 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1948 const struct sched_class *prev_class,
1949 int oldprio, int running)
1951 if (prev_class != p->sched_class) {
1952 if (prev_class->switched_from)
1953 prev_class->switched_from(rq, p, running);
1954 p->sched_class->switched_to(rq, p, running);
1956 p->sched_class->prio_changed(rq, p, oldprio, running);
1961 * Is this task likely cache-hot:
1964 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1968 if (p->sched_class != &fair_sched_class)
1972 * Buddy candidates are cache hot:
1974 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
1975 (&p->se == cfs_rq_of(&p->se)->next ||
1976 &p->se == cfs_rq_of(&p->se)->last))
1979 if (sysctl_sched_migration_cost == -1)
1981 if (sysctl_sched_migration_cost == 0)
1984 delta = now - p->se.exec_start;
1986 return delta < (s64)sysctl_sched_migration_cost;
1989 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1991 #ifdef CONFIG_SCHED_DEBUG
1993 * We should never call set_task_cpu() on a blocked task,
1994 * ttwu() will sort out the placement.
1996 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1997 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2000 trace_sched_migrate_task(p, new_cpu);
2002 if (task_cpu(p) != new_cpu) {
2003 p->se.nr_migrations++;
2004 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2007 __set_task_cpu(p, new_cpu);
2010 struct migration_arg {
2011 struct task_struct *task;
2015 static int migration_cpu_stop(void *data);
2018 * The task's runqueue lock must be held.
2019 * Returns true if you have to wait for migration thread.
2021 static bool migrate_task(struct task_struct *p, int dest_cpu)
2023 struct rq *rq = task_rq(p);
2026 * If the task is not on a runqueue (and not running), then
2027 * the next wake-up will properly place the task.
2029 return p->se.on_rq || task_running(rq, p);
2033 * wait_task_inactive - wait for a thread to unschedule.
2035 * If @match_state is nonzero, it's the @p->state value just checked and
2036 * not expected to change. If it changes, i.e. @p might have woken up,
2037 * then return zero. When we succeed in waiting for @p to be off its CPU,
2038 * we return a positive number (its total switch count). If a second call
2039 * a short while later returns the same number, the caller can be sure that
2040 * @p has remained unscheduled the whole time.
2042 * The caller must ensure that the task *will* unschedule sometime soon,
2043 * else this function might spin for a *long* time. This function can't
2044 * be called with interrupts off, or it may introduce deadlock with
2045 * smp_call_function() if an IPI is sent by the same process we are
2046 * waiting to become inactive.
2048 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2050 unsigned long flags;
2057 * We do the initial early heuristics without holding
2058 * any task-queue locks at all. We'll only try to get
2059 * the runqueue lock when things look like they will
2065 * If the task is actively running on another CPU
2066 * still, just relax and busy-wait without holding
2069 * NOTE! Since we don't hold any locks, it's not
2070 * even sure that "rq" stays as the right runqueue!
2071 * But we don't care, since "task_running()" will
2072 * return false if the runqueue has changed and p
2073 * is actually now running somewhere else!
2075 while (task_running(rq, p)) {
2076 if (match_state && unlikely(p->state != match_state))
2082 * Ok, time to look more closely! We need the rq
2083 * lock now, to be *sure*. If we're wrong, we'll
2084 * just go back and repeat.
2086 rq = task_rq_lock(p, &flags);
2087 trace_sched_wait_task(p);
2088 running = task_running(rq, p);
2089 on_rq = p->se.on_rq;
2091 if (!match_state || p->state == match_state)
2092 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2093 task_rq_unlock(rq, &flags);
2096 * If it changed from the expected state, bail out now.
2098 if (unlikely(!ncsw))
2102 * Was it really running after all now that we
2103 * checked with the proper locks actually held?
2105 * Oops. Go back and try again..
2107 if (unlikely(running)) {
2113 * It's not enough that it's not actively running,
2114 * it must be off the runqueue _entirely_, and not
2117 * So if it was still runnable (but just not actively
2118 * running right now), it's preempted, and we should
2119 * yield - it could be a while.
2121 if (unlikely(on_rq)) {
2122 schedule_timeout_uninterruptible(1);
2127 * Ahh, all good. It wasn't running, and it wasn't
2128 * runnable, which means that it will never become
2129 * running in the future either. We're all done!
2138 * kick_process - kick a running thread to enter/exit the kernel
2139 * @p: the to-be-kicked thread
2141 * Cause a process which is running on another CPU to enter
2142 * kernel-mode, without any delay. (to get signals handled.)
2144 * NOTE: this function doesnt have to take the runqueue lock,
2145 * because all it wants to ensure is that the remote task enters
2146 * the kernel. If the IPI races and the task has been migrated
2147 * to another CPU then no harm is done and the purpose has been
2150 void kick_process(struct task_struct *p)
2156 if ((cpu != smp_processor_id()) && task_curr(p))
2157 smp_send_reschedule(cpu);
2160 EXPORT_SYMBOL_GPL(kick_process);
2161 #endif /* CONFIG_SMP */
2164 * task_oncpu_function_call - call a function on the cpu on which a task runs
2165 * @p: the task to evaluate
2166 * @func: the function to be called
2167 * @info: the function call argument
2169 * Calls the function @func when the task is currently running. This might
2170 * be on the current CPU, which just calls the function directly
2172 void task_oncpu_function_call(struct task_struct *p,
2173 void (*func) (void *info), void *info)
2180 smp_call_function_single(cpu, func, info, 1);
2186 * ->cpus_allowed is protected by either TASK_WAKING or rq->lock held.
2188 static int select_fallback_rq(int cpu, struct task_struct *p)
2191 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2193 /* Look for allowed, online CPU in same node. */
2194 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2195 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2198 /* Any allowed, online CPU? */
2199 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2200 if (dest_cpu < nr_cpu_ids)
2203 /* No more Mr. Nice Guy. */
2204 if (unlikely(dest_cpu >= nr_cpu_ids)) {
2205 dest_cpu = cpuset_cpus_allowed_fallback(p);
2207 * Don't tell them about moving exiting tasks or
2208 * kernel threads (both mm NULL), since they never
2211 if (p->mm && printk_ratelimit()) {
2212 printk(KERN_INFO "process %d (%s) no "
2213 "longer affine to cpu%d\n",
2214 task_pid_nr(p), p->comm, cpu);
2222 * The caller (fork, wakeup) owns TASK_WAKING, ->cpus_allowed is stable.
2225 int select_task_rq(struct rq *rq, struct task_struct *p, int sd_flags, int wake_flags)
2227 int cpu = p->sched_class->select_task_rq(rq, p, sd_flags, wake_flags);
2230 * In order not to call set_task_cpu() on a blocking task we need
2231 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2234 * Since this is common to all placement strategies, this lives here.
2236 * [ this allows ->select_task() to simply return task_cpu(p) and
2237 * not worry about this generic constraint ]
2239 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2241 cpu = select_fallback_rq(task_cpu(p), p);
2246 static void update_avg(u64 *avg, u64 sample)
2248 s64 diff = sample - *avg;
2254 * try_to_wake_up - wake up a thread
2255 * @p: the to-be-woken-up thread
2256 * @state: the mask of task states that can be woken
2257 * @sync: do a synchronous wakeup?
2259 * Put it on the run-queue if it's not already there. The "current"
2260 * thread is always on the run-queue (except when the actual
2261 * re-schedule is in progress), and as such you're allowed to do
2262 * the simpler "current->state = TASK_RUNNING" to mark yourself
2263 * runnable without the overhead of this.
2265 * returns failure only if the task is already active.
2267 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2270 int cpu, orig_cpu, this_cpu, success = 0;
2271 unsigned long flags;
2272 unsigned long en_flags = ENQUEUE_WAKEUP;
2275 this_cpu = get_cpu();
2278 rq = task_rq_lock(p, &flags);
2279 if (!(p->state & state))
2289 if (unlikely(task_running(rq, p)))
2293 * In order to handle concurrent wakeups and release the rq->lock
2294 * we put the task in TASK_WAKING state.
2296 * First fix up the nr_uninterruptible count:
2298 if (task_contributes_to_load(p)) {
2299 if (likely(cpu_online(orig_cpu)))
2300 rq->nr_uninterruptible--;
2302 this_rq()->nr_uninterruptible--;
2304 p->state = TASK_WAKING;
2306 if (p->sched_class->task_waking) {
2307 p->sched_class->task_waking(rq, p);
2308 en_flags |= ENQUEUE_WAKING;
2311 cpu = select_task_rq(rq, p, SD_BALANCE_WAKE, wake_flags);
2312 if (cpu != orig_cpu)
2313 set_task_cpu(p, cpu);
2314 __task_rq_unlock(rq);
2317 raw_spin_lock(&rq->lock);
2320 * We migrated the task without holding either rq->lock, however
2321 * since the task is not on the task list itself, nobody else
2322 * will try and migrate the task, hence the rq should match the
2323 * cpu we just moved it to.
2325 WARN_ON(task_cpu(p) != cpu);
2326 WARN_ON(p->state != TASK_WAKING);
2328 #ifdef CONFIG_SCHEDSTATS
2329 schedstat_inc(rq, ttwu_count);
2330 if (cpu == this_cpu)
2331 schedstat_inc(rq, ttwu_local);
2333 struct sched_domain *sd;
2334 for_each_domain(this_cpu, sd) {
2335 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2336 schedstat_inc(sd, ttwu_wake_remote);
2341 #endif /* CONFIG_SCHEDSTATS */
2344 #endif /* CONFIG_SMP */
2345 schedstat_inc(p, se.statistics.nr_wakeups);
2346 if (wake_flags & WF_SYNC)
2347 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2348 if (orig_cpu != cpu)
2349 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2350 if (cpu == this_cpu)
2351 schedstat_inc(p, se.statistics.nr_wakeups_local);
2353 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2354 activate_task(rq, p, en_flags);
2358 trace_sched_wakeup(p, success);
2359 check_preempt_curr(rq, p, wake_flags);
2361 p->state = TASK_RUNNING;
2363 if (p->sched_class->task_woken)
2364 p->sched_class->task_woken(rq, p);
2366 if (unlikely(rq->idle_stamp)) {
2367 u64 delta = rq->clock - rq->idle_stamp;
2368 u64 max = 2*sysctl_sched_migration_cost;
2373 update_avg(&rq->avg_idle, delta);
2378 task_rq_unlock(rq, &flags);
2385 * wake_up_process - Wake up a specific process
2386 * @p: The process to be woken up.
2388 * Attempt to wake up the nominated process and move it to the set of runnable
2389 * processes. Returns 1 if the process was woken up, 0 if it was already
2392 * It may be assumed that this function implies a write memory barrier before
2393 * changing the task state if and only if any tasks are woken up.
2395 int wake_up_process(struct task_struct *p)
2397 return try_to_wake_up(p, TASK_ALL, 0);
2399 EXPORT_SYMBOL(wake_up_process);
2401 int wake_up_state(struct task_struct *p, unsigned int state)
2403 return try_to_wake_up(p, state, 0);
2407 * Perform scheduler related setup for a newly forked process p.
2408 * p is forked by current.
2410 * __sched_fork() is basic setup used by init_idle() too:
2412 static void __sched_fork(struct task_struct *p)
2414 p->se.exec_start = 0;
2415 p->se.sum_exec_runtime = 0;
2416 p->se.prev_sum_exec_runtime = 0;
2417 p->se.nr_migrations = 0;
2419 #ifdef CONFIG_SCHEDSTATS
2420 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2423 INIT_LIST_HEAD(&p->rt.run_list);
2425 INIT_LIST_HEAD(&p->se.group_node);
2427 #ifdef CONFIG_PREEMPT_NOTIFIERS
2428 INIT_HLIST_HEAD(&p->preempt_notifiers);
2433 * fork()/clone()-time setup:
2435 void sched_fork(struct task_struct *p, int clone_flags)
2437 int cpu = get_cpu();
2441 * We mark the process as running here. This guarantees that
2442 * nobody will actually run it, and a signal or other external
2443 * event cannot wake it up and insert it on the runqueue either.
2445 p->state = TASK_RUNNING;
2448 * Revert to default priority/policy on fork if requested.
2450 if (unlikely(p->sched_reset_on_fork)) {
2451 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2452 p->policy = SCHED_NORMAL;
2453 p->normal_prio = p->static_prio;
2456 if (PRIO_TO_NICE(p->static_prio) < 0) {
2457 p->static_prio = NICE_TO_PRIO(0);
2458 p->normal_prio = p->static_prio;
2463 * We don't need the reset flag anymore after the fork. It has
2464 * fulfilled its duty:
2466 p->sched_reset_on_fork = 0;
2470 * Make sure we do not leak PI boosting priority to the child.
2472 p->prio = current->normal_prio;
2474 if (!rt_prio(p->prio))
2475 p->sched_class = &fair_sched_class;
2477 if (p->sched_class->task_fork)
2478 p->sched_class->task_fork(p);
2481 * The child is not yet in the pid-hash so no cgroup attach races,
2482 * and the cgroup is pinned to this child due to cgroup_fork()
2483 * is ran before sched_fork().
2485 * Silence PROVE_RCU.
2488 set_task_cpu(p, cpu);
2491 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2492 if (likely(sched_info_on()))
2493 memset(&p->sched_info, 0, sizeof(p->sched_info));
2495 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2498 #ifdef CONFIG_PREEMPT
2499 /* Want to start with kernel preemption disabled. */
2500 task_thread_info(p)->preempt_count = 1;
2502 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2508 * wake_up_new_task - wake up a newly created task for the first time.
2510 * This function will do some initial scheduler statistics housekeeping
2511 * that must be done for every newly created context, then puts the task
2512 * on the runqueue and wakes it.
2514 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2516 unsigned long flags;
2518 int cpu __maybe_unused = get_cpu();
2521 rq = task_rq_lock(p, &flags);
2522 p->state = TASK_WAKING;
2525 * Fork balancing, do it here and not earlier because:
2526 * - cpus_allowed can change in the fork path
2527 * - any previously selected cpu might disappear through hotplug
2529 * We set TASK_WAKING so that select_task_rq() can drop rq->lock
2530 * without people poking at ->cpus_allowed.
2532 cpu = select_task_rq(rq, p, SD_BALANCE_FORK, 0);
2533 set_task_cpu(p, cpu);
2535 p->state = TASK_RUNNING;
2536 task_rq_unlock(rq, &flags);
2539 rq = task_rq_lock(p, &flags);
2540 activate_task(rq, p, 0);
2541 trace_sched_wakeup_new(p, 1);
2542 check_preempt_curr(rq, p, WF_FORK);
2544 if (p->sched_class->task_woken)
2545 p->sched_class->task_woken(rq, p);
2547 task_rq_unlock(rq, &flags);
2551 #ifdef CONFIG_PREEMPT_NOTIFIERS
2554 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2555 * @notifier: notifier struct to register
2557 void preempt_notifier_register(struct preempt_notifier *notifier)
2559 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2561 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2564 * preempt_notifier_unregister - no longer interested in preemption notifications
2565 * @notifier: notifier struct to unregister
2567 * This is safe to call from within a preemption notifier.
2569 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2571 hlist_del(¬ifier->link);
2573 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2575 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2577 struct preempt_notifier *notifier;
2578 struct hlist_node *node;
2580 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2581 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2585 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2586 struct task_struct *next)
2588 struct preempt_notifier *notifier;
2589 struct hlist_node *node;
2591 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2592 notifier->ops->sched_out(notifier, next);
2595 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2597 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2602 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2603 struct task_struct *next)
2607 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2610 * prepare_task_switch - prepare to switch tasks
2611 * @rq: the runqueue preparing to switch
2612 * @prev: the current task that is being switched out
2613 * @next: the task we are going to switch to.
2615 * This is called with the rq lock held and interrupts off. It must
2616 * be paired with a subsequent finish_task_switch after the context
2619 * prepare_task_switch sets up locking and calls architecture specific
2623 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2624 struct task_struct *next)
2626 fire_sched_out_preempt_notifiers(prev, next);
2627 prepare_lock_switch(rq, next);
2628 prepare_arch_switch(next);
2632 * finish_task_switch - clean up after a task-switch
2633 * @rq: runqueue associated with task-switch
2634 * @prev: the thread we just switched away from.
2636 * finish_task_switch must be called after the context switch, paired
2637 * with a prepare_task_switch call before the context switch.
2638 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2639 * and do any other architecture-specific cleanup actions.
2641 * Note that we may have delayed dropping an mm in context_switch(). If
2642 * so, we finish that here outside of the runqueue lock. (Doing it
2643 * with the lock held can cause deadlocks; see schedule() for
2646 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2647 __releases(rq->lock)
2649 struct mm_struct *mm = rq->prev_mm;
2655 * A task struct has one reference for the use as "current".
2656 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2657 * schedule one last time. The schedule call will never return, and
2658 * the scheduled task must drop that reference.
2659 * The test for TASK_DEAD must occur while the runqueue locks are
2660 * still held, otherwise prev could be scheduled on another cpu, die
2661 * there before we look at prev->state, and then the reference would
2663 * Manfred Spraul <manfred@colorfullife.com>
2665 prev_state = prev->state;
2666 finish_arch_switch(prev);
2667 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2668 local_irq_disable();
2669 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2670 perf_event_task_sched_in(current);
2671 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2673 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2674 finish_lock_switch(rq, prev);
2676 fire_sched_in_preempt_notifiers(current);
2679 if (unlikely(prev_state == TASK_DEAD)) {
2681 * Remove function-return probe instances associated with this
2682 * task and put them back on the free list.
2684 kprobe_flush_task(prev);
2685 put_task_struct(prev);
2691 /* assumes rq->lock is held */
2692 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2694 if (prev->sched_class->pre_schedule)
2695 prev->sched_class->pre_schedule(rq, prev);
2698 /* rq->lock is NOT held, but preemption is disabled */
2699 static inline void post_schedule(struct rq *rq)
2701 if (rq->post_schedule) {
2702 unsigned long flags;
2704 raw_spin_lock_irqsave(&rq->lock, flags);
2705 if (rq->curr->sched_class->post_schedule)
2706 rq->curr->sched_class->post_schedule(rq);
2707 raw_spin_unlock_irqrestore(&rq->lock, flags);
2709 rq->post_schedule = 0;
2715 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2719 static inline void post_schedule(struct rq *rq)
2726 * schedule_tail - first thing a freshly forked thread must call.
2727 * @prev: the thread we just switched away from.
2729 asmlinkage void schedule_tail(struct task_struct *prev)
2730 __releases(rq->lock)
2732 struct rq *rq = this_rq();
2734 finish_task_switch(rq, prev);
2737 * FIXME: do we need to worry about rq being invalidated by the
2742 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2743 /* In this case, finish_task_switch does not reenable preemption */
2746 if (current->set_child_tid)
2747 put_user(task_pid_vnr(current), current->set_child_tid);
2751 * context_switch - switch to the new MM and the new
2752 * thread's register state.
2755 context_switch(struct rq *rq, struct task_struct *prev,
2756 struct task_struct *next)
2758 struct mm_struct *mm, *oldmm;
2760 prepare_task_switch(rq, prev, next);
2761 trace_sched_switch(prev, next);
2763 oldmm = prev->active_mm;
2765 * For paravirt, this is coupled with an exit in switch_to to
2766 * combine the page table reload and the switch backend into
2769 arch_start_context_switch(prev);
2772 next->active_mm = oldmm;
2773 atomic_inc(&oldmm->mm_count);
2774 enter_lazy_tlb(oldmm, next);
2776 switch_mm(oldmm, mm, next);
2778 if (likely(!prev->mm)) {
2779 prev->active_mm = NULL;
2780 rq->prev_mm = oldmm;
2783 * Since the runqueue lock will be released by the next
2784 * task (which is an invalid locking op but in the case
2785 * of the scheduler it's an obvious special-case), so we
2786 * do an early lockdep release here:
2788 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2789 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2792 /* Here we just switch the register state and the stack. */
2793 switch_to(prev, next, prev);
2797 * this_rq must be evaluated again because prev may have moved
2798 * CPUs since it called schedule(), thus the 'rq' on its stack
2799 * frame will be invalid.
2801 finish_task_switch(this_rq(), prev);
2805 * nr_running, nr_uninterruptible and nr_context_switches:
2807 * externally visible scheduler statistics: current number of runnable
2808 * threads, current number of uninterruptible-sleeping threads, total
2809 * number of context switches performed since bootup.
2811 unsigned long nr_running(void)
2813 unsigned long i, sum = 0;
2815 for_each_online_cpu(i)
2816 sum += cpu_rq(i)->nr_running;
2821 unsigned long nr_uninterruptible(void)
2823 unsigned long i, sum = 0;
2825 for_each_possible_cpu(i)
2826 sum += cpu_rq(i)->nr_uninterruptible;
2829 * Since we read the counters lockless, it might be slightly
2830 * inaccurate. Do not allow it to go below zero though:
2832 if (unlikely((long)sum < 0))
2838 unsigned long long nr_context_switches(void)
2841 unsigned long long sum = 0;
2843 for_each_possible_cpu(i)
2844 sum += cpu_rq(i)->nr_switches;
2849 unsigned long nr_iowait(void)
2851 unsigned long i, sum = 0;
2853 for_each_possible_cpu(i)
2854 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2859 unsigned long nr_iowait_cpu(int cpu)
2861 struct rq *this = cpu_rq(cpu);
2862 return atomic_read(&this->nr_iowait);
2865 unsigned long this_cpu_load(void)
2867 struct rq *this = this_rq();
2868 return this->cpu_load[0];
2872 /* Variables and functions for calc_load */
2873 static atomic_long_t calc_load_tasks;
2874 static unsigned long calc_load_update;
2875 unsigned long avenrun[3];
2876 EXPORT_SYMBOL(avenrun);
2878 static long calc_load_fold_active(struct rq *this_rq)
2880 long nr_active, delta = 0;
2882 nr_active = this_rq->nr_running;
2883 nr_active += (long) this_rq->nr_uninterruptible;
2885 if (nr_active != this_rq->calc_load_active) {
2886 delta = nr_active - this_rq->calc_load_active;
2887 this_rq->calc_load_active = nr_active;
2895 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
2897 * When making the ILB scale, we should try to pull this in as well.
2899 static atomic_long_t calc_load_tasks_idle;
2901 static void calc_load_account_idle(struct rq *this_rq)
2905 delta = calc_load_fold_active(this_rq);
2907 atomic_long_add(delta, &calc_load_tasks_idle);
2910 static long calc_load_fold_idle(void)
2915 * Its got a race, we don't care...
2917 if (atomic_long_read(&calc_load_tasks_idle))
2918 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
2923 static void calc_load_account_idle(struct rq *this_rq)
2927 static inline long calc_load_fold_idle(void)
2934 * get_avenrun - get the load average array
2935 * @loads: pointer to dest load array
2936 * @offset: offset to add
2937 * @shift: shift count to shift the result left
2939 * These values are estimates at best, so no need for locking.
2941 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2943 loads[0] = (avenrun[0] + offset) << shift;
2944 loads[1] = (avenrun[1] + offset) << shift;
2945 loads[2] = (avenrun[2] + offset) << shift;
2948 static unsigned long
2949 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2952 load += active * (FIXED_1 - exp);
2953 return load >> FSHIFT;
2957 * calc_load - update the avenrun load estimates 10 ticks after the
2958 * CPUs have updated calc_load_tasks.
2960 void calc_global_load(void)
2962 unsigned long upd = calc_load_update + 10;
2965 if (time_before(jiffies, upd))
2968 active = atomic_long_read(&calc_load_tasks);
2969 active = active > 0 ? active * FIXED_1 : 0;
2971 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2972 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2973 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2975 calc_load_update += LOAD_FREQ;
2979 * Called from update_cpu_load() to periodically update this CPU's
2982 static void calc_load_account_active(struct rq *this_rq)
2986 if (time_before(jiffies, this_rq->calc_load_update))
2989 delta = calc_load_fold_active(this_rq);
2990 delta += calc_load_fold_idle();
2992 atomic_long_add(delta, &calc_load_tasks);
2994 this_rq->calc_load_update += LOAD_FREQ;
2998 * Update rq->cpu_load[] statistics. This function is usually called every
2999 * scheduler tick (TICK_NSEC).
3001 static void update_cpu_load(struct rq *this_rq)
3003 unsigned long this_load = this_rq->load.weight;
3006 this_rq->nr_load_updates++;
3008 /* Update our load: */
3009 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3010 unsigned long old_load, new_load;
3012 /* scale is effectively 1 << i now, and >> i divides by scale */
3014 old_load = this_rq->cpu_load[i];
3015 new_load = this_load;
3017 * Round up the averaging division if load is increasing. This
3018 * prevents us from getting stuck on 9 if the load is 10, for
3021 if (new_load > old_load)
3022 new_load += scale-1;
3023 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3026 calc_load_account_active(this_rq);
3032 * sched_exec - execve() is a valuable balancing opportunity, because at
3033 * this point the task has the smallest effective memory and cache footprint.
3035 void sched_exec(void)
3037 struct task_struct *p = current;
3038 unsigned long flags;
3042 rq = task_rq_lock(p, &flags);
3043 dest_cpu = p->sched_class->select_task_rq(rq, p, SD_BALANCE_EXEC, 0);
3044 if (dest_cpu == smp_processor_id())
3048 * select_task_rq() can race against ->cpus_allowed
3050 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed) &&
3051 likely(cpu_active(dest_cpu)) && migrate_task(p, dest_cpu)) {
3052 struct migration_arg arg = { p, dest_cpu };
3054 task_rq_unlock(rq, &flags);
3055 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
3059 task_rq_unlock(rq, &flags);
3064 DEFINE_PER_CPU(struct kernel_stat, kstat);
3066 EXPORT_PER_CPU_SYMBOL(kstat);
3069 * Return any ns on the sched_clock that have not yet been accounted in
3070 * @p in case that task is currently running.
3072 * Called with task_rq_lock() held on @rq.
3074 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3078 if (task_current(rq, p)) {
3079 update_rq_clock(rq);
3080 ns = rq->clock - p->se.exec_start;
3088 unsigned long long task_delta_exec(struct task_struct *p)
3090 unsigned long flags;
3094 rq = task_rq_lock(p, &flags);
3095 ns = do_task_delta_exec(p, rq);
3096 task_rq_unlock(rq, &flags);
3102 * Return accounted runtime for the task.
3103 * In case the task is currently running, return the runtime plus current's
3104 * pending runtime that have not been accounted yet.
3106 unsigned long long task_sched_runtime(struct task_struct *p)
3108 unsigned long flags;
3112 rq = task_rq_lock(p, &flags);
3113 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3114 task_rq_unlock(rq, &flags);
3120 * Return sum_exec_runtime for the thread group.
3121 * In case the task is currently running, return the sum plus current's
3122 * pending runtime that have not been accounted yet.
3124 * Note that the thread group might have other running tasks as well,
3125 * so the return value not includes other pending runtime that other
3126 * running tasks might have.
3128 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3130 struct task_cputime totals;
3131 unsigned long flags;
3135 rq = task_rq_lock(p, &flags);
3136 thread_group_cputime(p, &totals);
3137 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3138 task_rq_unlock(rq, &flags);
3144 * Account user cpu time to a process.
3145 * @p: the process that the cpu time gets accounted to
3146 * @cputime: the cpu time spent in user space since the last update
3147 * @cputime_scaled: cputime scaled by cpu frequency
3149 void account_user_time(struct task_struct *p, cputime_t cputime,
3150 cputime_t cputime_scaled)
3152 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3155 /* Add user time to process. */
3156 p->utime = cputime_add(p->utime, cputime);
3157 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3158 account_group_user_time(p, cputime);
3160 /* Add user time to cpustat. */
3161 tmp = cputime_to_cputime64(cputime);
3162 if (TASK_NICE(p) > 0)
3163 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3165 cpustat->user = cputime64_add(cpustat->user, tmp);
3167 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3168 /* Account for user time used */
3169 acct_update_integrals(p);
3173 * Account guest cpu time to a process.
3174 * @p: the process that the cpu time gets accounted to
3175 * @cputime: the cpu time spent in virtual machine since the last update
3176 * @cputime_scaled: cputime scaled by cpu frequency
3178 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3179 cputime_t cputime_scaled)
3182 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3184 tmp = cputime_to_cputime64(cputime);
3186 /* Add guest time to process. */
3187 p->utime = cputime_add(p->utime, cputime);
3188 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3189 account_group_user_time(p, cputime);
3190 p->gtime = cputime_add(p->gtime, cputime);
3192 /* Add guest time to cpustat. */
3193 if (TASK_NICE(p) > 0) {
3194 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3195 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3197 cpustat->user = cputime64_add(cpustat->user, tmp);
3198 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3203 * Account system cpu time to a process.
3204 * @p: the process that the cpu time gets accounted to
3205 * @hardirq_offset: the offset to subtract from hardirq_count()
3206 * @cputime: the cpu time spent in kernel space since the last update
3207 * @cputime_scaled: cputime scaled by cpu frequency
3209 void account_system_time(struct task_struct *p, int hardirq_offset,
3210 cputime_t cputime, cputime_t cputime_scaled)
3212 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3215 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3216 account_guest_time(p, cputime, cputime_scaled);
3220 /* Add system time to process. */
3221 p->stime = cputime_add(p->stime, cputime);
3222 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3223 account_group_system_time(p, cputime);
3225 /* Add system time to cpustat. */
3226 tmp = cputime_to_cputime64(cputime);
3227 if (hardirq_count() - hardirq_offset)
3228 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3229 else if (softirq_count())
3230 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3232 cpustat->system = cputime64_add(cpustat->system, tmp);
3234 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3236 /* Account for system time used */
3237 acct_update_integrals(p);
3241 * Account for involuntary wait time.
3242 * @steal: the cpu time spent in involuntary wait
3244 void account_steal_time(cputime_t cputime)
3246 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3247 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3249 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3253 * Account for idle time.
3254 * @cputime: the cpu time spent in idle wait
3256 void account_idle_time(cputime_t cputime)
3258 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3259 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3260 struct rq *rq = this_rq();
3262 if (atomic_read(&rq->nr_iowait) > 0)
3263 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3265 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3268 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3271 * Account a single tick of cpu time.
3272 * @p: the process that the cpu time gets accounted to
3273 * @user_tick: indicates if the tick is a user or a system tick
3275 void account_process_tick(struct task_struct *p, int user_tick)
3277 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3278 struct rq *rq = this_rq();
3281 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3282 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3283 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3286 account_idle_time(cputime_one_jiffy);
3290 * Account multiple ticks of steal time.
3291 * @p: the process from which the cpu time has been stolen
3292 * @ticks: number of stolen ticks
3294 void account_steal_ticks(unsigned long ticks)
3296 account_steal_time(jiffies_to_cputime(ticks));
3300 * Account multiple ticks of idle time.
3301 * @ticks: number of stolen ticks
3303 void account_idle_ticks(unsigned long ticks)
3305 account_idle_time(jiffies_to_cputime(ticks));
3311 * Use precise platform statistics if available:
3313 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3314 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3320 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3322 struct task_cputime cputime;
3324 thread_group_cputime(p, &cputime);
3326 *ut = cputime.utime;
3327 *st = cputime.stime;
3331 #ifndef nsecs_to_cputime
3332 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3335 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3337 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
3340 * Use CFS's precise accounting:
3342 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3348 do_div(temp, total);
3349 utime = (cputime_t)temp;
3354 * Compare with previous values, to keep monotonicity:
3356 p->prev_utime = max(p->prev_utime, utime);
3357 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
3359 *ut = p->prev_utime;
3360 *st = p->prev_stime;
3364 * Must be called with siglock held.
3366 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3368 struct signal_struct *sig = p->signal;
3369 struct task_cputime cputime;
3370 cputime_t rtime, utime, total;
3372 thread_group_cputime(p, &cputime);
3374 total = cputime_add(cputime.utime, cputime.stime);
3375 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3380 temp *= cputime.utime;
3381 do_div(temp, total);
3382 utime = (cputime_t)temp;
3386 sig->prev_utime = max(sig->prev_utime, utime);
3387 sig->prev_stime = max(sig->prev_stime,
3388 cputime_sub(rtime, sig->prev_utime));
3390 *ut = sig->prev_utime;
3391 *st = sig->prev_stime;
3396 * This function gets called by the timer code, with HZ frequency.
3397 * We call it with interrupts disabled.
3399 * It also gets called by the fork code, when changing the parent's
3402 void scheduler_tick(void)
3404 int cpu = smp_processor_id();
3405 struct rq *rq = cpu_rq(cpu);
3406 struct task_struct *curr = rq->curr;
3410 raw_spin_lock(&rq->lock);
3411 update_rq_clock(rq);
3412 update_cpu_load(rq);
3413 curr->sched_class->task_tick(rq, curr, 0);
3414 raw_spin_unlock(&rq->lock);
3416 perf_event_task_tick(curr);
3419 rq->idle_at_tick = idle_cpu(cpu);
3420 trigger_load_balance(rq, cpu);
3424 notrace unsigned long get_parent_ip(unsigned long addr)
3426 if (in_lock_functions(addr)) {
3427 addr = CALLER_ADDR2;
3428 if (in_lock_functions(addr))
3429 addr = CALLER_ADDR3;
3434 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3435 defined(CONFIG_PREEMPT_TRACER))
3437 void __kprobes add_preempt_count(int val)
3439 #ifdef CONFIG_DEBUG_PREEMPT
3443 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3446 preempt_count() += val;
3447 #ifdef CONFIG_DEBUG_PREEMPT
3449 * Spinlock count overflowing soon?
3451 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3454 if (preempt_count() == val)
3455 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3457 EXPORT_SYMBOL(add_preempt_count);
3459 void __kprobes sub_preempt_count(int val)
3461 #ifdef CONFIG_DEBUG_PREEMPT
3465 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3468 * Is the spinlock portion underflowing?
3470 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3471 !(preempt_count() & PREEMPT_MASK)))
3475 if (preempt_count() == val)
3476 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3477 preempt_count() -= val;
3479 EXPORT_SYMBOL(sub_preempt_count);
3484 * Print scheduling while atomic bug:
3486 static noinline void __schedule_bug(struct task_struct *prev)
3488 struct pt_regs *regs = get_irq_regs();
3490 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3491 prev->comm, prev->pid, preempt_count());
3493 debug_show_held_locks(prev);
3495 if (irqs_disabled())
3496 print_irqtrace_events(prev);
3505 * Various schedule()-time debugging checks and statistics:
3507 static inline void schedule_debug(struct task_struct *prev)
3510 * Test if we are atomic. Since do_exit() needs to call into
3511 * schedule() atomically, we ignore that path for now.
3512 * Otherwise, whine if we are scheduling when we should not be.
3514 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3515 __schedule_bug(prev);
3517 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3519 schedstat_inc(this_rq(), sched_count);
3520 #ifdef CONFIG_SCHEDSTATS
3521 if (unlikely(prev->lock_depth >= 0)) {
3522 schedstat_inc(this_rq(), bkl_count);
3523 schedstat_inc(prev, sched_info.bkl_count);
3528 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3531 update_rq_clock(rq);
3532 rq->skip_clock_update = 0;
3533 prev->sched_class->put_prev_task(rq, prev);
3537 * Pick up the highest-prio task:
3539 static inline struct task_struct *
3540 pick_next_task(struct rq *rq)
3542 const struct sched_class *class;
3543 struct task_struct *p;
3546 * Optimization: we know that if all tasks are in
3547 * the fair class we can call that function directly:
3549 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3550 p = fair_sched_class.pick_next_task(rq);
3555 class = sched_class_highest;
3557 p = class->pick_next_task(rq);
3561 * Will never be NULL as the idle class always
3562 * returns a non-NULL p:
3564 class = class->next;
3569 * schedule() is the main scheduler function.
3571 asmlinkage void __sched schedule(void)
3573 struct task_struct *prev, *next;
3574 unsigned long *switch_count;
3580 cpu = smp_processor_id();
3582 rcu_note_context_switch(cpu);
3584 switch_count = &prev->nivcsw;
3586 release_kernel_lock(prev);
3587 need_resched_nonpreemptible:
3589 schedule_debug(prev);
3591 if (sched_feat(HRTICK))
3594 raw_spin_lock_irq(&rq->lock);
3595 clear_tsk_need_resched(prev);
3597 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3598 if (unlikely(signal_pending_state(prev->state, prev)))
3599 prev->state = TASK_RUNNING;
3601 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3602 switch_count = &prev->nvcsw;
3605 pre_schedule(rq, prev);
3607 if (unlikely(!rq->nr_running))
3608 idle_balance(cpu, rq);
3610 put_prev_task(rq, prev);
3611 next = pick_next_task(rq);
3613 if (likely(prev != next)) {
3614 sched_info_switch(prev, next);
3615 perf_event_task_sched_out(prev, next);
3621 context_switch(rq, prev, next); /* unlocks the rq */
3623 * the context switch might have flipped the stack from under
3624 * us, hence refresh the local variables.
3626 cpu = smp_processor_id();
3629 raw_spin_unlock_irq(&rq->lock);
3633 if (unlikely(reacquire_kernel_lock(current) < 0)) {
3635 switch_count = &prev->nivcsw;
3636 goto need_resched_nonpreemptible;
3639 preempt_enable_no_resched();
3643 EXPORT_SYMBOL(schedule);
3645 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3647 * Look out! "owner" is an entirely speculative pointer
3648 * access and not reliable.
3650 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
3655 if (!sched_feat(OWNER_SPIN))
3658 #ifdef CONFIG_DEBUG_PAGEALLOC
3660 * Need to access the cpu field knowing that
3661 * DEBUG_PAGEALLOC could have unmapped it if
3662 * the mutex owner just released it and exited.
3664 if (probe_kernel_address(&owner->cpu, cpu))
3671 * Even if the access succeeded (likely case),
3672 * the cpu field may no longer be valid.
3674 if (cpu >= nr_cpumask_bits)
3678 * We need to validate that we can do a
3679 * get_cpu() and that we have the percpu area.
3681 if (!cpu_online(cpu))
3688 * Owner changed, break to re-assess state.
3690 if (lock->owner != owner) {
3692 * If the lock has switched to a different owner,
3693 * we likely have heavy contention. Return 0 to quit
3694 * optimistic spinning and not contend further:
3702 * Is that owner really running on that cpu?
3704 if (task_thread_info(rq->curr) != owner || need_resched())
3714 #ifdef CONFIG_PREEMPT
3716 * this is the entry point to schedule() from in-kernel preemption
3717 * off of preempt_enable. Kernel preemptions off return from interrupt
3718 * occur there and call schedule directly.
3720 asmlinkage void __sched preempt_schedule(void)
3722 struct thread_info *ti = current_thread_info();
3725 * If there is a non-zero preempt_count or interrupts are disabled,
3726 * we do not want to preempt the current task. Just return..
3728 if (likely(ti->preempt_count || irqs_disabled()))
3732 add_preempt_count(PREEMPT_ACTIVE);
3734 sub_preempt_count(PREEMPT_ACTIVE);
3737 * Check again in case we missed a preemption opportunity
3738 * between schedule and now.
3741 } while (need_resched());
3743 EXPORT_SYMBOL(preempt_schedule);
3746 * this is the entry point to schedule() from kernel preemption
3747 * off of irq context.
3748 * Note, that this is called and return with irqs disabled. This will
3749 * protect us against recursive calling from irq.
3751 asmlinkage void __sched preempt_schedule_irq(void)
3753 struct thread_info *ti = current_thread_info();
3755 /* Catch callers which need to be fixed */
3756 BUG_ON(ti->preempt_count || !irqs_disabled());
3759 add_preempt_count(PREEMPT_ACTIVE);
3762 local_irq_disable();
3763 sub_preempt_count(PREEMPT_ACTIVE);
3766 * Check again in case we missed a preemption opportunity
3767 * between schedule and now.
3770 } while (need_resched());
3773 #endif /* CONFIG_PREEMPT */
3775 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3778 return try_to_wake_up(curr->private, mode, wake_flags);
3780 EXPORT_SYMBOL(default_wake_function);
3783 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3784 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3785 * number) then we wake all the non-exclusive tasks and one exclusive task.
3787 * There are circumstances in which we can try to wake a task which has already
3788 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3789 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3791 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3792 int nr_exclusive, int wake_flags, void *key)
3794 wait_queue_t *curr, *next;
3796 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3797 unsigned flags = curr->flags;
3799 if (curr->func(curr, mode, wake_flags, key) &&
3800 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3806 * __wake_up - wake up threads blocked on a waitqueue.
3808 * @mode: which threads
3809 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3810 * @key: is directly passed to the wakeup function
3812 * It may be assumed that this function implies a write memory barrier before
3813 * changing the task state if and only if any tasks are woken up.
3815 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3816 int nr_exclusive, void *key)
3818 unsigned long flags;
3820 spin_lock_irqsave(&q->lock, flags);
3821 __wake_up_common(q, mode, nr_exclusive, 0, key);
3822 spin_unlock_irqrestore(&q->lock, flags);
3824 EXPORT_SYMBOL(__wake_up);
3827 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3829 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3831 __wake_up_common(q, mode, 1, 0, NULL);
3833 EXPORT_SYMBOL_GPL(__wake_up_locked);
3835 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3837 __wake_up_common(q, mode, 1, 0, key);
3841 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3843 * @mode: which threads
3844 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3845 * @key: opaque value to be passed to wakeup targets
3847 * The sync wakeup differs that the waker knows that it will schedule
3848 * away soon, so while the target thread will be woken up, it will not
3849 * be migrated to another CPU - ie. the two threads are 'synchronized'
3850 * with each other. This can prevent needless bouncing between CPUs.
3852 * On UP it can prevent extra preemption.
3854 * It may be assumed that this function implies a write memory barrier before
3855 * changing the task state if and only if any tasks are woken up.
3857 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3858 int nr_exclusive, void *key)
3860 unsigned long flags;
3861 int wake_flags = WF_SYNC;
3866 if (unlikely(!nr_exclusive))
3869 spin_lock_irqsave(&q->lock, flags);
3870 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3871 spin_unlock_irqrestore(&q->lock, flags);
3873 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3876 * __wake_up_sync - see __wake_up_sync_key()
3878 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3880 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
3882 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3885 * complete: - signals a single thread waiting on this completion
3886 * @x: holds the state of this particular completion
3888 * This will wake up a single thread waiting on this completion. Threads will be
3889 * awakened in the same order in which they were queued.
3891 * See also complete_all(), wait_for_completion() and related routines.
3893 * It may be assumed that this function implies a write memory barrier before
3894 * changing the task state if and only if any tasks are woken up.
3896 void complete(struct completion *x)
3898 unsigned long flags;
3900 spin_lock_irqsave(&x->wait.lock, flags);
3902 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3903 spin_unlock_irqrestore(&x->wait.lock, flags);
3905 EXPORT_SYMBOL(complete);
3908 * complete_all: - signals all threads waiting on this completion
3909 * @x: holds the state of this particular completion
3911 * This will wake up all threads waiting on this particular completion event.
3913 * It may be assumed that this function implies a write memory barrier before
3914 * changing the task state if and only if any tasks are woken up.
3916 void complete_all(struct completion *x)
3918 unsigned long flags;
3920 spin_lock_irqsave(&x->wait.lock, flags);
3921 x->done += UINT_MAX/2;
3922 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3923 spin_unlock_irqrestore(&x->wait.lock, flags);
3925 EXPORT_SYMBOL(complete_all);
3927 static inline long __sched
3928 do_wait_for_common(struct completion *x, long timeout, int state)
3931 DECLARE_WAITQUEUE(wait, current);
3933 __add_wait_queue_tail_exclusive(&x->wait, &wait);
3935 if (signal_pending_state(state, current)) {
3936 timeout = -ERESTARTSYS;
3939 __set_current_state(state);
3940 spin_unlock_irq(&x->wait.lock);
3941 timeout = schedule_timeout(timeout);
3942 spin_lock_irq(&x->wait.lock);
3943 } while (!x->done && timeout);
3944 __remove_wait_queue(&x->wait, &wait);
3949 return timeout ?: 1;
3953 wait_for_common(struct completion *x, long timeout, int state)
3957 spin_lock_irq(&x->wait.lock);
3958 timeout = do_wait_for_common(x, timeout, state);
3959 spin_unlock_irq(&x->wait.lock);
3964 * wait_for_completion: - waits for completion of a task
3965 * @x: holds the state of this particular completion
3967 * This waits to be signaled for completion of a specific task. It is NOT
3968 * interruptible and there is no timeout.
3970 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3971 * and interrupt capability. Also see complete().
3973 void __sched wait_for_completion(struct completion *x)
3975 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3977 EXPORT_SYMBOL(wait_for_completion);
3980 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3981 * @x: holds the state of this particular completion
3982 * @timeout: timeout value in jiffies
3984 * This waits for either a completion of a specific task to be signaled or for a
3985 * specified timeout to expire. The timeout is in jiffies. It is not
3988 unsigned long __sched
3989 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3991 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3993 EXPORT_SYMBOL(wait_for_completion_timeout);
3996 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3997 * @x: holds the state of this particular completion
3999 * This waits for completion of a specific task to be signaled. It is
4002 int __sched wait_for_completion_interruptible(struct completion *x)
4004 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4005 if (t == -ERESTARTSYS)
4009 EXPORT_SYMBOL(wait_for_completion_interruptible);
4012 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4013 * @x: holds the state of this particular completion
4014 * @timeout: timeout value in jiffies
4016 * This waits for either a completion of a specific task to be signaled or for a
4017 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4019 unsigned long __sched
4020 wait_for_completion_interruptible_timeout(struct completion *x,
4021 unsigned long timeout)
4023 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4025 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4028 * wait_for_completion_killable: - waits for completion of a task (killable)
4029 * @x: holds the state of this particular completion
4031 * This waits to be signaled for completion of a specific task. It can be
4032 * interrupted by a kill signal.
4034 int __sched wait_for_completion_killable(struct completion *x)
4036 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4037 if (t == -ERESTARTSYS)
4041 EXPORT_SYMBOL(wait_for_completion_killable);
4044 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4045 * @x: holds the state of this particular completion
4046 * @timeout: timeout value in jiffies
4048 * This waits for either a completion of a specific task to be
4049 * signaled or for a specified timeout to expire. It can be
4050 * interrupted by a kill signal. The timeout is in jiffies.
4052 unsigned long __sched
4053 wait_for_completion_killable_timeout(struct completion *x,
4054 unsigned long timeout)
4056 return wait_for_common(x, timeout, TASK_KILLABLE);
4058 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
4061 * try_wait_for_completion - try to decrement a completion without blocking
4062 * @x: completion structure
4064 * Returns: 0 if a decrement cannot be done without blocking
4065 * 1 if a decrement succeeded.
4067 * If a completion is being used as a counting completion,
4068 * attempt to decrement the counter without blocking. This
4069 * enables us to avoid waiting if the resource the completion
4070 * is protecting is not available.
4072 bool try_wait_for_completion(struct completion *x)
4074 unsigned long flags;
4077 spin_lock_irqsave(&x->wait.lock, flags);
4082 spin_unlock_irqrestore(&x->wait.lock, flags);
4085 EXPORT_SYMBOL(try_wait_for_completion);
4088 * completion_done - Test to see if a completion has any waiters
4089 * @x: completion structure
4091 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4092 * 1 if there are no waiters.
4095 bool completion_done(struct completion *x)
4097 unsigned long flags;
4100 spin_lock_irqsave(&x->wait.lock, flags);
4103 spin_unlock_irqrestore(&x->wait.lock, flags);
4106 EXPORT_SYMBOL(completion_done);
4109 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4111 unsigned long flags;
4114 init_waitqueue_entry(&wait, current);
4116 __set_current_state(state);
4118 spin_lock_irqsave(&q->lock, flags);
4119 __add_wait_queue(q, &wait);
4120 spin_unlock(&q->lock);
4121 timeout = schedule_timeout(timeout);
4122 spin_lock_irq(&q->lock);
4123 __remove_wait_queue(q, &wait);
4124 spin_unlock_irqrestore(&q->lock, flags);
4129 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4131 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4133 EXPORT_SYMBOL(interruptible_sleep_on);
4136 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4138 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4140 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4142 void __sched sleep_on(wait_queue_head_t *q)
4144 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4146 EXPORT_SYMBOL(sleep_on);
4148 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4150 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4152 EXPORT_SYMBOL(sleep_on_timeout);
4154 #ifdef CONFIG_RT_MUTEXES
4157 * rt_mutex_setprio - set the current priority of a task
4159 * @prio: prio value (kernel-internal form)
4161 * This function changes the 'effective' priority of a task. It does
4162 * not touch ->normal_prio like __setscheduler().
4164 * Used by the rt_mutex code to implement priority inheritance logic.
4166 void rt_mutex_setprio(struct task_struct *p, int prio)
4168 unsigned long flags;
4169 int oldprio, on_rq, running;
4171 const struct sched_class *prev_class;
4173 BUG_ON(prio < 0 || prio > MAX_PRIO);
4175 rq = task_rq_lock(p, &flags);
4178 prev_class = p->sched_class;
4179 on_rq = p->se.on_rq;
4180 running = task_current(rq, p);
4182 dequeue_task(rq, p, 0);
4184 p->sched_class->put_prev_task(rq, p);
4187 p->sched_class = &rt_sched_class;
4189 p->sched_class = &fair_sched_class;
4194 p->sched_class->set_curr_task(rq);
4196 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4198 check_class_changed(rq, p, prev_class, oldprio, running);
4200 task_rq_unlock(rq, &flags);
4205 void set_user_nice(struct task_struct *p, long nice)
4207 int old_prio, delta, on_rq;
4208 unsigned long flags;
4211 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4214 * We have to be careful, if called from sys_setpriority(),
4215 * the task might be in the middle of scheduling on another CPU.
4217 rq = task_rq_lock(p, &flags);
4219 * The RT priorities are set via sched_setscheduler(), but we still
4220 * allow the 'normal' nice value to be set - but as expected
4221 * it wont have any effect on scheduling until the task is
4222 * SCHED_FIFO/SCHED_RR:
4224 if (task_has_rt_policy(p)) {
4225 p->static_prio = NICE_TO_PRIO(nice);
4228 on_rq = p->se.on_rq;
4230 dequeue_task(rq, p, 0);
4232 p->static_prio = NICE_TO_PRIO(nice);
4235 p->prio = effective_prio(p);
4236 delta = p->prio - old_prio;
4239 enqueue_task(rq, p, 0);
4241 * If the task increased its priority or is running and
4242 * lowered its priority, then reschedule its CPU:
4244 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4245 resched_task(rq->curr);
4248 task_rq_unlock(rq, &flags);
4250 EXPORT_SYMBOL(set_user_nice);
4253 * can_nice - check if a task can reduce its nice value
4257 int can_nice(const struct task_struct *p, const int nice)
4259 /* convert nice value [19,-20] to rlimit style value [1,40] */
4260 int nice_rlim = 20 - nice;
4262 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4263 capable(CAP_SYS_NICE));
4266 #ifdef __ARCH_WANT_SYS_NICE
4269 * sys_nice - change the priority of the current process.
4270 * @increment: priority increment
4272 * sys_setpriority is a more generic, but much slower function that
4273 * does similar things.
4275 SYSCALL_DEFINE1(nice, int, increment)
4280 * Setpriority might change our priority at the same moment.
4281 * We don't have to worry. Conceptually one call occurs first
4282 * and we have a single winner.
4284 if (increment < -40)
4289 nice = TASK_NICE(current) + increment;
4295 if (increment < 0 && !can_nice(current, nice))
4298 retval = security_task_setnice(current, nice);
4302 set_user_nice(current, nice);
4309 * task_prio - return the priority value of a given task.
4310 * @p: the task in question.
4312 * This is the priority value as seen by users in /proc.
4313 * RT tasks are offset by -200. Normal tasks are centered
4314 * around 0, value goes from -16 to +15.
4316 int task_prio(const struct task_struct *p)
4318 return p->prio - MAX_RT_PRIO;
4322 * task_nice - return the nice value of a given task.
4323 * @p: the task in question.
4325 int task_nice(const struct task_struct *p)
4327 return TASK_NICE(p);
4329 EXPORT_SYMBOL(task_nice);
4332 * idle_cpu - is a given cpu idle currently?
4333 * @cpu: the processor in question.
4335 int idle_cpu(int cpu)
4337 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4341 * idle_task - return the idle task for a given cpu.
4342 * @cpu: the processor in question.
4344 struct task_struct *idle_task(int cpu)
4346 return cpu_rq(cpu)->idle;
4350 * find_process_by_pid - find a process with a matching PID value.
4351 * @pid: the pid in question.
4353 static struct task_struct *find_process_by_pid(pid_t pid)
4355 return pid ? find_task_by_vpid(pid) : current;
4358 /* Actually do priority change: must hold rq lock. */
4360 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4362 BUG_ON(p->se.on_rq);
4365 p->rt_priority = prio;
4366 p->normal_prio = normal_prio(p);
4367 /* we are holding p->pi_lock already */
4368 p->prio = rt_mutex_getprio(p);
4369 if (rt_prio(p->prio))
4370 p->sched_class = &rt_sched_class;
4372 p->sched_class = &fair_sched_class;
4377 * check the target process has a UID that matches the current process's
4379 static bool check_same_owner(struct task_struct *p)
4381 const struct cred *cred = current_cred(), *pcred;
4385 pcred = __task_cred(p);
4386 match = (cred->euid == pcred->euid ||
4387 cred->euid == pcred->uid);
4392 static int __sched_setscheduler(struct task_struct *p, int policy,
4393 struct sched_param *param, bool user)
4395 int retval, oldprio, oldpolicy = -1, on_rq, running;
4396 unsigned long flags;
4397 const struct sched_class *prev_class;
4401 /* may grab non-irq protected spin_locks */
4402 BUG_ON(in_interrupt());
4404 /* double check policy once rq lock held */
4406 reset_on_fork = p->sched_reset_on_fork;
4407 policy = oldpolicy = p->policy;
4409 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4410 policy &= ~SCHED_RESET_ON_FORK;
4412 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4413 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4414 policy != SCHED_IDLE)
4419 * Valid priorities for SCHED_FIFO and SCHED_RR are
4420 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4421 * SCHED_BATCH and SCHED_IDLE is 0.
4423 if (param->sched_priority < 0 ||
4424 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4425 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4427 if (rt_policy(policy) != (param->sched_priority != 0))
4431 * Allow unprivileged RT tasks to decrease priority:
4433 if (user && !capable(CAP_SYS_NICE)) {
4434 if (rt_policy(policy)) {
4435 unsigned long rlim_rtprio;
4437 if (!lock_task_sighand(p, &flags))
4439 rlim_rtprio = task_rlimit(p, RLIMIT_RTPRIO);
4440 unlock_task_sighand(p, &flags);
4442 /* can't set/change the rt policy */
4443 if (policy != p->policy && !rlim_rtprio)
4446 /* can't increase priority */
4447 if (param->sched_priority > p->rt_priority &&
4448 param->sched_priority > rlim_rtprio)
4452 * Like positive nice levels, dont allow tasks to
4453 * move out of SCHED_IDLE either:
4455 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4458 /* can't change other user's priorities */
4459 if (!check_same_owner(p))
4462 /* Normal users shall not reset the sched_reset_on_fork flag */
4463 if (p->sched_reset_on_fork && !reset_on_fork)
4468 retval = security_task_setscheduler(p, policy, param);
4474 * make sure no PI-waiters arrive (or leave) while we are
4475 * changing the priority of the task:
4477 raw_spin_lock_irqsave(&p->pi_lock, flags);
4479 * To be able to change p->policy safely, the apropriate
4480 * runqueue lock must be held.
4482 rq = __task_rq_lock(p);
4484 #ifdef CONFIG_RT_GROUP_SCHED
4487 * Do not allow realtime tasks into groups that have no runtime
4490 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4491 task_group(p)->rt_bandwidth.rt_runtime == 0) {
4492 __task_rq_unlock(rq);
4493 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4499 /* recheck policy now with rq lock held */
4500 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4501 policy = oldpolicy = -1;
4502 __task_rq_unlock(rq);
4503 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4506 on_rq = p->se.on_rq;
4507 running = task_current(rq, p);
4509 deactivate_task(rq, p, 0);
4511 p->sched_class->put_prev_task(rq, p);
4513 p->sched_reset_on_fork = reset_on_fork;
4516 prev_class = p->sched_class;
4517 __setscheduler(rq, p, policy, param->sched_priority);
4520 p->sched_class->set_curr_task(rq);
4522 activate_task(rq, p, 0);
4524 check_class_changed(rq, p, prev_class, oldprio, running);
4526 __task_rq_unlock(rq);
4527 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4529 rt_mutex_adjust_pi(p);
4535 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4536 * @p: the task in question.
4537 * @policy: new policy.
4538 * @param: structure containing the new RT priority.
4540 * NOTE that the task may be already dead.
4542 int sched_setscheduler(struct task_struct *p, int policy,
4543 struct sched_param *param)
4545 return __sched_setscheduler(p, policy, param, true);
4547 EXPORT_SYMBOL_GPL(sched_setscheduler);
4550 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4551 * @p: the task in question.
4552 * @policy: new policy.
4553 * @param: structure containing the new RT priority.
4555 * Just like sched_setscheduler, only don't bother checking if the
4556 * current context has permission. For example, this is needed in
4557 * stop_machine(): we create temporary high priority worker threads,
4558 * but our caller might not have that capability.
4560 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4561 struct sched_param *param)
4563 return __sched_setscheduler(p, policy, param, false);
4567 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4569 struct sched_param lparam;
4570 struct task_struct *p;
4573 if (!param || pid < 0)
4575 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4580 p = find_process_by_pid(pid);
4582 retval = sched_setscheduler(p, policy, &lparam);
4589 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4590 * @pid: the pid in question.
4591 * @policy: new policy.
4592 * @param: structure containing the new RT priority.
4594 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4595 struct sched_param __user *, param)
4597 /* negative values for policy are not valid */
4601 return do_sched_setscheduler(pid, policy, param);
4605 * sys_sched_setparam - set/change the RT priority of a thread
4606 * @pid: the pid in question.
4607 * @param: structure containing the new RT priority.
4609 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4611 return do_sched_setscheduler(pid, -1, param);
4615 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4616 * @pid: the pid in question.
4618 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4620 struct task_struct *p;
4628 p = find_process_by_pid(pid);
4630 retval = security_task_getscheduler(p);
4633 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4640 * sys_sched_getparam - get the RT priority of a thread
4641 * @pid: the pid in question.
4642 * @param: structure containing the RT priority.
4644 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4646 struct sched_param lp;
4647 struct task_struct *p;
4650 if (!param || pid < 0)
4654 p = find_process_by_pid(pid);
4659 retval = security_task_getscheduler(p);
4663 lp.sched_priority = p->rt_priority;
4667 * This one might sleep, we cannot do it with a spinlock held ...
4669 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4678 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4680 cpumask_var_t cpus_allowed, new_mask;
4681 struct task_struct *p;
4687 p = find_process_by_pid(pid);
4694 /* Prevent p going away */
4698 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4702 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4704 goto out_free_cpus_allowed;
4707 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
4710 retval = security_task_setscheduler(p, 0, NULL);
4714 cpuset_cpus_allowed(p, cpus_allowed);
4715 cpumask_and(new_mask, in_mask, cpus_allowed);
4717 retval = set_cpus_allowed_ptr(p, new_mask);
4720 cpuset_cpus_allowed(p, cpus_allowed);
4721 if (!cpumask_subset(new_mask, cpus_allowed)) {
4723 * We must have raced with a concurrent cpuset
4724 * update. Just reset the cpus_allowed to the
4725 * cpuset's cpus_allowed
4727 cpumask_copy(new_mask, cpus_allowed);
4732 free_cpumask_var(new_mask);
4733 out_free_cpus_allowed:
4734 free_cpumask_var(cpus_allowed);
4741 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4742 struct cpumask *new_mask)
4744 if (len < cpumask_size())
4745 cpumask_clear(new_mask);
4746 else if (len > cpumask_size())
4747 len = cpumask_size();
4749 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4753 * sys_sched_setaffinity - set the cpu affinity of a process
4754 * @pid: pid of the process
4755 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4756 * @user_mask_ptr: user-space pointer to the new cpu mask
4758 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4759 unsigned long __user *, user_mask_ptr)
4761 cpumask_var_t new_mask;
4764 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4767 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4769 retval = sched_setaffinity(pid, new_mask);
4770 free_cpumask_var(new_mask);
4774 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4776 struct task_struct *p;
4777 unsigned long flags;
4785 p = find_process_by_pid(pid);
4789 retval = security_task_getscheduler(p);
4793 rq = task_rq_lock(p, &flags);
4794 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4795 task_rq_unlock(rq, &flags);
4805 * sys_sched_getaffinity - get the cpu affinity of a process
4806 * @pid: pid of the process
4807 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4808 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4810 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4811 unsigned long __user *, user_mask_ptr)
4816 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4818 if (len & (sizeof(unsigned long)-1))
4821 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4824 ret = sched_getaffinity(pid, mask);
4826 size_t retlen = min_t(size_t, len, cpumask_size());
4828 if (copy_to_user(user_mask_ptr, mask, retlen))
4833 free_cpumask_var(mask);
4839 * sys_sched_yield - yield the current processor to other threads.
4841 * This function yields the current CPU to other tasks. If there are no
4842 * other threads running on this CPU then this function will return.
4844 SYSCALL_DEFINE0(sched_yield)
4846 struct rq *rq = this_rq_lock();
4848 schedstat_inc(rq, yld_count);
4849 current->sched_class->yield_task(rq);
4852 * Since we are going to call schedule() anyway, there's
4853 * no need to preempt or enable interrupts:
4855 __release(rq->lock);
4856 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4857 do_raw_spin_unlock(&rq->lock);
4858 preempt_enable_no_resched();
4865 static inline int should_resched(void)
4867 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4870 static void __cond_resched(void)
4872 add_preempt_count(PREEMPT_ACTIVE);
4874 sub_preempt_count(PREEMPT_ACTIVE);
4877 int __sched _cond_resched(void)
4879 if (should_resched()) {
4885 EXPORT_SYMBOL(_cond_resched);
4888 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4889 * call schedule, and on return reacquire the lock.
4891 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4892 * operations here to prevent schedule() from being called twice (once via
4893 * spin_unlock(), once by hand).
4895 int __cond_resched_lock(spinlock_t *lock)
4897 int resched = should_resched();
4900 lockdep_assert_held(lock);
4902 if (spin_needbreak(lock) || resched) {
4913 EXPORT_SYMBOL(__cond_resched_lock);
4915 int __sched __cond_resched_softirq(void)
4917 BUG_ON(!in_softirq());
4919 if (should_resched()) {
4927 EXPORT_SYMBOL(__cond_resched_softirq);
4930 * yield - yield the current processor to other threads.
4932 * This is a shortcut for kernel-space yielding - it marks the
4933 * thread runnable and calls sys_sched_yield().
4935 void __sched yield(void)
4937 set_current_state(TASK_RUNNING);
4940 EXPORT_SYMBOL(yield);
4943 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4944 * that process accounting knows that this is a task in IO wait state.
4946 void __sched io_schedule(void)
4948 struct rq *rq = raw_rq();
4950 delayacct_blkio_start();
4951 atomic_inc(&rq->nr_iowait);
4952 current->in_iowait = 1;
4954 current->in_iowait = 0;
4955 atomic_dec(&rq->nr_iowait);
4956 delayacct_blkio_end();
4958 EXPORT_SYMBOL(io_schedule);
4960 long __sched io_schedule_timeout(long timeout)
4962 struct rq *rq = raw_rq();
4965 delayacct_blkio_start();
4966 atomic_inc(&rq->nr_iowait);
4967 current->in_iowait = 1;
4968 ret = schedule_timeout(timeout);
4969 current->in_iowait = 0;
4970 atomic_dec(&rq->nr_iowait);
4971 delayacct_blkio_end();
4976 * sys_sched_get_priority_max - return maximum RT priority.
4977 * @policy: scheduling class.
4979 * this syscall returns the maximum rt_priority that can be used
4980 * by a given scheduling class.
4982 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4989 ret = MAX_USER_RT_PRIO-1;
5001 * sys_sched_get_priority_min - return minimum RT priority.
5002 * @policy: scheduling class.
5004 * this syscall returns the minimum rt_priority that can be used
5005 * by a given scheduling class.
5007 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5025 * sys_sched_rr_get_interval - return the default timeslice of a process.
5026 * @pid: pid of the process.
5027 * @interval: userspace pointer to the timeslice value.
5029 * this syscall writes the default timeslice value of a given process
5030 * into the user-space timespec buffer. A value of '0' means infinity.
5032 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5033 struct timespec __user *, interval)
5035 struct task_struct *p;
5036 unsigned int time_slice;
5037 unsigned long flags;
5047 p = find_process_by_pid(pid);
5051 retval = security_task_getscheduler(p);
5055 rq = task_rq_lock(p, &flags);
5056 time_slice = p->sched_class->get_rr_interval(rq, p);
5057 task_rq_unlock(rq, &flags);
5060 jiffies_to_timespec(time_slice, &t);
5061 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5069 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5071 void sched_show_task(struct task_struct *p)
5073 unsigned long free = 0;
5076 state = p->state ? __ffs(p->state) + 1 : 0;
5077 printk(KERN_INFO "%-13.13s %c", p->comm,
5078 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5079 #if BITS_PER_LONG == 32
5080 if (state == TASK_RUNNING)
5081 printk(KERN_CONT " running ");
5083 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5085 if (state == TASK_RUNNING)
5086 printk(KERN_CONT " running task ");
5088 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5090 #ifdef CONFIG_DEBUG_STACK_USAGE
5091 free = stack_not_used(p);
5093 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5094 task_pid_nr(p), task_pid_nr(p->real_parent),
5095 (unsigned long)task_thread_info(p)->flags);
5097 show_stack(p, NULL);
5100 void show_state_filter(unsigned long state_filter)
5102 struct task_struct *g, *p;
5104 #if BITS_PER_LONG == 32
5106 " task PC stack pid father\n");
5109 " task PC stack pid father\n");
5111 read_lock(&tasklist_lock);
5112 do_each_thread(g, p) {
5114 * reset the NMI-timeout, listing all files on a slow
5115 * console might take alot of time:
5117 touch_nmi_watchdog();
5118 if (!state_filter || (p->state & state_filter))
5120 } while_each_thread(g, p);
5122 touch_all_softlockup_watchdogs();
5124 #ifdef CONFIG_SCHED_DEBUG
5125 sysrq_sched_debug_show();
5127 read_unlock(&tasklist_lock);
5129 * Only show locks if all tasks are dumped:
5132 debug_show_all_locks();
5135 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5137 idle->sched_class = &idle_sched_class;
5141 * init_idle - set up an idle thread for a given CPU
5142 * @idle: task in question
5143 * @cpu: cpu the idle task belongs to
5145 * NOTE: this function does not set the idle thread's NEED_RESCHED
5146 * flag, to make booting more robust.
5148 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5150 struct rq *rq = cpu_rq(cpu);
5151 unsigned long flags;
5153 raw_spin_lock_irqsave(&rq->lock, flags);
5156 idle->state = TASK_RUNNING;
5157 idle->se.exec_start = sched_clock();
5159 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
5161 * We're having a chicken and egg problem, even though we are
5162 * holding rq->lock, the cpu isn't yet set to this cpu so the
5163 * lockdep check in task_group() will fail.
5165 * Similar case to sched_fork(). / Alternatively we could
5166 * use task_rq_lock() here and obtain the other rq->lock.
5171 __set_task_cpu(idle, cpu);
5174 rq->curr = rq->idle = idle;
5175 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5178 raw_spin_unlock_irqrestore(&rq->lock, flags);
5180 /* Set the preempt count _outside_ the spinlocks! */
5181 #if defined(CONFIG_PREEMPT)
5182 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5184 task_thread_info(idle)->preempt_count = 0;
5187 * The idle tasks have their own, simple scheduling class:
5189 idle->sched_class = &idle_sched_class;
5190 ftrace_graph_init_task(idle);
5194 * In a system that switches off the HZ timer nohz_cpu_mask
5195 * indicates which cpus entered this state. This is used
5196 * in the rcu update to wait only for active cpus. For system
5197 * which do not switch off the HZ timer nohz_cpu_mask should
5198 * always be CPU_BITS_NONE.
5200 cpumask_var_t nohz_cpu_mask;
5203 * Increase the granularity value when there are more CPUs,
5204 * because with more CPUs the 'effective latency' as visible
5205 * to users decreases. But the relationship is not linear,
5206 * so pick a second-best guess by going with the log2 of the
5209 * This idea comes from the SD scheduler of Con Kolivas:
5211 static int get_update_sysctl_factor(void)
5213 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5214 unsigned int factor;
5216 switch (sysctl_sched_tunable_scaling) {
5217 case SCHED_TUNABLESCALING_NONE:
5220 case SCHED_TUNABLESCALING_LINEAR:
5223 case SCHED_TUNABLESCALING_LOG:
5225 factor = 1 + ilog2(cpus);
5232 static void update_sysctl(void)
5234 unsigned int factor = get_update_sysctl_factor();
5236 #define SET_SYSCTL(name) \
5237 (sysctl_##name = (factor) * normalized_sysctl_##name)
5238 SET_SYSCTL(sched_min_granularity);
5239 SET_SYSCTL(sched_latency);
5240 SET_SYSCTL(sched_wakeup_granularity);
5241 SET_SYSCTL(sched_shares_ratelimit);
5245 static inline void sched_init_granularity(void)
5252 * This is how migration works:
5254 * 1) we invoke migration_cpu_stop() on the target CPU using
5256 * 2) stopper starts to run (implicitly forcing the migrated thread
5258 * 3) it checks whether the migrated task is still in the wrong runqueue.
5259 * 4) if it's in the wrong runqueue then the migration thread removes
5260 * it and puts it into the right queue.
5261 * 5) stopper completes and stop_one_cpu() returns and the migration
5266 * Change a given task's CPU affinity. Migrate the thread to a
5267 * proper CPU and schedule it away if the CPU it's executing on
5268 * is removed from the allowed bitmask.
5270 * NOTE: the caller must have a valid reference to the task, the
5271 * task must not exit() & deallocate itself prematurely. The
5272 * call is not atomic; no spinlocks may be held.
5274 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5276 unsigned long flags;
5278 unsigned int dest_cpu;
5282 * Serialize against TASK_WAKING so that ttwu() and wunt() can
5283 * drop the rq->lock and still rely on ->cpus_allowed.
5286 while (task_is_waking(p))
5288 rq = task_rq_lock(p, &flags);
5289 if (task_is_waking(p)) {
5290 task_rq_unlock(rq, &flags);
5294 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5299 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5300 !cpumask_equal(&p->cpus_allowed, new_mask))) {
5305 if (p->sched_class->set_cpus_allowed)
5306 p->sched_class->set_cpus_allowed(p, new_mask);
5308 cpumask_copy(&p->cpus_allowed, new_mask);
5309 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5312 /* Can the task run on the task's current CPU? If so, we're done */
5313 if (cpumask_test_cpu(task_cpu(p), new_mask))
5316 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
5317 if (migrate_task(p, dest_cpu)) {
5318 struct migration_arg arg = { p, dest_cpu };
5319 /* Need help from migration thread: drop lock and wait. */
5320 task_rq_unlock(rq, &flags);
5321 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
5322 tlb_migrate_finish(p->mm);
5326 task_rq_unlock(rq, &flags);
5330 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5333 * Move (not current) task off this cpu, onto dest cpu. We're doing
5334 * this because either it can't run here any more (set_cpus_allowed()
5335 * away from this CPU, or CPU going down), or because we're
5336 * attempting to rebalance this task on exec (sched_exec).
5338 * So we race with normal scheduler movements, but that's OK, as long
5339 * as the task is no longer on this CPU.
5341 * Returns non-zero if task was successfully migrated.
5343 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5345 struct rq *rq_dest, *rq_src;
5348 if (unlikely(!cpu_active(dest_cpu)))
5351 rq_src = cpu_rq(src_cpu);
5352 rq_dest = cpu_rq(dest_cpu);
5354 double_rq_lock(rq_src, rq_dest);
5355 /* Already moved. */
5356 if (task_cpu(p) != src_cpu)
5358 /* Affinity changed (again). */
5359 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
5363 * If we're not on a rq, the next wake-up will ensure we're
5367 deactivate_task(rq_src, p, 0);
5368 set_task_cpu(p, dest_cpu);
5369 activate_task(rq_dest, p, 0);
5370 check_preempt_curr(rq_dest, p, 0);
5375 double_rq_unlock(rq_src, rq_dest);
5380 * migration_cpu_stop - this will be executed by a highprio stopper thread
5381 * and performs thread migration by bumping thread off CPU then
5382 * 'pushing' onto another runqueue.
5384 static int migration_cpu_stop(void *data)
5386 struct migration_arg *arg = data;
5389 * The original target cpu might have gone down and we might
5390 * be on another cpu but it doesn't matter.
5392 local_irq_disable();
5393 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
5398 #ifdef CONFIG_HOTPLUG_CPU
5400 * Figure out where task on dead CPU should go, use force if necessary.
5402 void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5404 struct rq *rq = cpu_rq(dead_cpu);
5405 int needs_cpu, uninitialized_var(dest_cpu);
5406 unsigned long flags;
5408 local_irq_save(flags);
5410 raw_spin_lock(&rq->lock);
5411 needs_cpu = (task_cpu(p) == dead_cpu) && (p->state != TASK_WAKING);
5413 dest_cpu = select_fallback_rq(dead_cpu, p);
5414 raw_spin_unlock(&rq->lock);
5416 * It can only fail if we race with set_cpus_allowed(),
5417 * in the racer should migrate the task anyway.
5420 __migrate_task(p, dead_cpu, dest_cpu);
5421 local_irq_restore(flags);
5425 * While a dead CPU has no uninterruptible tasks queued at this point,
5426 * it might still have a nonzero ->nr_uninterruptible counter, because
5427 * for performance reasons the counter is not stricly tracking tasks to
5428 * their home CPUs. So we just add the counter to another CPU's counter,
5429 * to keep the global sum constant after CPU-down:
5431 static void migrate_nr_uninterruptible(struct rq *rq_src)
5433 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5434 unsigned long flags;
5436 local_irq_save(flags);
5437 double_rq_lock(rq_src, rq_dest);
5438 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5439 rq_src->nr_uninterruptible = 0;
5440 double_rq_unlock(rq_src, rq_dest);
5441 local_irq_restore(flags);
5444 /* Run through task list and migrate tasks from the dead cpu. */
5445 static void migrate_live_tasks(int src_cpu)
5447 struct task_struct *p, *t;
5449 read_lock(&tasklist_lock);
5451 do_each_thread(t, p) {
5455 if (task_cpu(p) == src_cpu)
5456 move_task_off_dead_cpu(src_cpu, p);
5457 } while_each_thread(t, p);
5459 read_unlock(&tasklist_lock);
5463 * Schedules idle task to be the next runnable task on current CPU.
5464 * It does so by boosting its priority to highest possible.
5465 * Used by CPU offline code.
5467 void sched_idle_next(void)
5469 int this_cpu = smp_processor_id();
5470 struct rq *rq = cpu_rq(this_cpu);
5471 struct task_struct *p = rq->idle;
5472 unsigned long flags;
5474 /* cpu has to be offline */
5475 BUG_ON(cpu_online(this_cpu));
5478 * Strictly not necessary since rest of the CPUs are stopped by now
5479 * and interrupts disabled on the current cpu.
5481 raw_spin_lock_irqsave(&rq->lock, flags);
5483 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5485 activate_task(rq, p, 0);
5487 raw_spin_unlock_irqrestore(&rq->lock, flags);
5491 * Ensures that the idle task is using init_mm right before its cpu goes
5494 void idle_task_exit(void)
5496 struct mm_struct *mm = current->active_mm;
5498 BUG_ON(cpu_online(smp_processor_id()));
5501 switch_mm(mm, &init_mm, current);
5505 /* called under rq->lock with disabled interrupts */
5506 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5508 struct rq *rq = cpu_rq(dead_cpu);
5510 /* Must be exiting, otherwise would be on tasklist. */
5511 BUG_ON(!p->exit_state);
5513 /* Cannot have done final schedule yet: would have vanished. */
5514 BUG_ON(p->state == TASK_DEAD);
5519 * Drop lock around migration; if someone else moves it,
5520 * that's OK. No task can be added to this CPU, so iteration is
5523 raw_spin_unlock_irq(&rq->lock);
5524 move_task_off_dead_cpu(dead_cpu, p);
5525 raw_spin_lock_irq(&rq->lock);
5530 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5531 static void migrate_dead_tasks(unsigned int dead_cpu)
5533 struct rq *rq = cpu_rq(dead_cpu);
5534 struct task_struct *next;
5537 if (!rq->nr_running)
5539 next = pick_next_task(rq);
5542 next->sched_class->put_prev_task(rq, next);
5543 migrate_dead(dead_cpu, next);
5549 * remove the tasks which were accounted by rq from calc_load_tasks.
5551 static void calc_global_load_remove(struct rq *rq)
5553 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5554 rq->calc_load_active = 0;
5556 #endif /* CONFIG_HOTPLUG_CPU */
5558 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5560 static struct ctl_table sd_ctl_dir[] = {
5562 .procname = "sched_domain",
5568 static struct ctl_table sd_ctl_root[] = {
5570 .procname = "kernel",
5572 .child = sd_ctl_dir,
5577 static struct ctl_table *sd_alloc_ctl_entry(int n)
5579 struct ctl_table *entry =
5580 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5585 static void sd_free_ctl_entry(struct ctl_table **tablep)
5587 struct ctl_table *entry;
5590 * In the intermediate directories, both the child directory and
5591 * procname are dynamically allocated and could fail but the mode
5592 * will always be set. In the lowest directory the names are
5593 * static strings and all have proc handlers.
5595 for (entry = *tablep; entry->mode; entry++) {
5597 sd_free_ctl_entry(&entry->child);
5598 if (entry->proc_handler == NULL)
5599 kfree(entry->procname);
5607 set_table_entry(struct ctl_table *entry,
5608 const char *procname, void *data, int maxlen,
5609 mode_t mode, proc_handler *proc_handler)
5611 entry->procname = procname;
5613 entry->maxlen = maxlen;
5615 entry->proc_handler = proc_handler;
5618 static struct ctl_table *
5619 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5621 struct ctl_table *table = sd_alloc_ctl_entry(13);
5626 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5627 sizeof(long), 0644, proc_doulongvec_minmax);
5628 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5629 sizeof(long), 0644, proc_doulongvec_minmax);
5630 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5631 sizeof(int), 0644, proc_dointvec_minmax);
5632 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5633 sizeof(int), 0644, proc_dointvec_minmax);
5634 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5635 sizeof(int), 0644, proc_dointvec_minmax);
5636 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5637 sizeof(int), 0644, proc_dointvec_minmax);
5638 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5639 sizeof(int), 0644, proc_dointvec_minmax);
5640 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5641 sizeof(int), 0644, proc_dointvec_minmax);
5642 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5643 sizeof(int), 0644, proc_dointvec_minmax);
5644 set_table_entry(&table[9], "cache_nice_tries",
5645 &sd->cache_nice_tries,
5646 sizeof(int), 0644, proc_dointvec_minmax);
5647 set_table_entry(&table[10], "flags", &sd->flags,
5648 sizeof(int), 0644, proc_dointvec_minmax);
5649 set_table_entry(&table[11], "name", sd->name,
5650 CORENAME_MAX_SIZE, 0444, proc_dostring);
5651 /* &table[12] is terminator */
5656 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5658 struct ctl_table *entry, *table;
5659 struct sched_domain *sd;
5660 int domain_num = 0, i;
5663 for_each_domain(cpu, sd)
5665 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5670 for_each_domain(cpu, sd) {
5671 snprintf(buf, 32, "domain%d", i);
5672 entry->procname = kstrdup(buf, GFP_KERNEL);
5674 entry->child = sd_alloc_ctl_domain_table(sd);
5681 static struct ctl_table_header *sd_sysctl_header;
5682 static void register_sched_domain_sysctl(void)
5684 int i, cpu_num = num_possible_cpus();
5685 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5688 WARN_ON(sd_ctl_dir[0].child);
5689 sd_ctl_dir[0].child = entry;
5694 for_each_possible_cpu(i) {
5695 snprintf(buf, 32, "cpu%d", i);
5696 entry->procname = kstrdup(buf, GFP_KERNEL);
5698 entry->child = sd_alloc_ctl_cpu_table(i);
5702 WARN_ON(sd_sysctl_header);
5703 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5706 /* may be called multiple times per register */
5707 static void unregister_sched_domain_sysctl(void)
5709 if (sd_sysctl_header)
5710 unregister_sysctl_table(sd_sysctl_header);
5711 sd_sysctl_header = NULL;
5712 if (sd_ctl_dir[0].child)
5713 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5716 static void register_sched_domain_sysctl(void)
5719 static void unregister_sched_domain_sysctl(void)
5724 static void set_rq_online(struct rq *rq)
5727 const struct sched_class *class;
5729 cpumask_set_cpu(rq->cpu, rq->rd->online);
5732 for_each_class(class) {
5733 if (class->rq_online)
5734 class->rq_online(rq);
5739 static void set_rq_offline(struct rq *rq)
5742 const struct sched_class *class;
5744 for_each_class(class) {
5745 if (class->rq_offline)
5746 class->rq_offline(rq);
5749 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5755 * migration_call - callback that gets triggered when a CPU is added.
5756 * Here we can start up the necessary migration thread for the new CPU.
5758 static int __cpuinit
5759 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5761 int cpu = (long)hcpu;
5762 unsigned long flags;
5763 struct rq *rq = cpu_rq(cpu);
5767 case CPU_UP_PREPARE:
5768 case CPU_UP_PREPARE_FROZEN:
5769 rq->calc_load_update = calc_load_update;
5773 case CPU_ONLINE_FROZEN:
5774 /* Update our root-domain */
5775 raw_spin_lock_irqsave(&rq->lock, flags);
5777 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5781 raw_spin_unlock_irqrestore(&rq->lock, flags);
5784 #ifdef CONFIG_HOTPLUG_CPU
5786 case CPU_DEAD_FROZEN:
5787 migrate_live_tasks(cpu);
5788 /* Idle task back to normal (off runqueue, low prio) */
5789 raw_spin_lock_irq(&rq->lock);
5790 deactivate_task(rq, rq->idle, 0);
5791 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5792 rq->idle->sched_class = &idle_sched_class;
5793 migrate_dead_tasks(cpu);
5794 raw_spin_unlock_irq(&rq->lock);
5795 migrate_nr_uninterruptible(rq);
5796 BUG_ON(rq->nr_running != 0);
5797 calc_global_load_remove(rq);
5801 case CPU_DYING_FROZEN:
5802 /* Update our root-domain */
5803 raw_spin_lock_irqsave(&rq->lock, flags);
5805 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5808 raw_spin_unlock_irqrestore(&rq->lock, flags);
5816 * Register at high priority so that task migration (migrate_all_tasks)
5817 * happens before everything else. This has to be lower priority than
5818 * the notifier in the perf_event subsystem, though.
5820 static struct notifier_block __cpuinitdata migration_notifier = {
5821 .notifier_call = migration_call,
5825 static int __init migration_init(void)
5827 void *cpu = (void *)(long)smp_processor_id();
5830 /* Start one for the boot CPU: */
5831 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5832 BUG_ON(err == NOTIFY_BAD);
5833 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5834 register_cpu_notifier(&migration_notifier);
5838 early_initcall(migration_init);
5843 #ifdef CONFIG_SCHED_DEBUG
5845 static __read_mostly int sched_domain_debug_enabled;
5847 static int __init sched_domain_debug_setup(char *str)
5849 sched_domain_debug_enabled = 1;
5853 early_param("sched_debug", sched_domain_debug_setup);
5855 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5856 struct cpumask *groupmask)
5858 struct sched_group *group = sd->groups;
5861 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5862 cpumask_clear(groupmask);
5864 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5866 if (!(sd->flags & SD_LOAD_BALANCE)) {
5867 printk("does not load-balance\n");
5869 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5874 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5876 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5877 printk(KERN_ERR "ERROR: domain->span does not contain "
5880 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5881 printk(KERN_ERR "ERROR: domain->groups does not contain"
5885 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5889 printk(KERN_ERR "ERROR: group is NULL\n");
5893 if (!group->cpu_power) {
5894 printk(KERN_CONT "\n");
5895 printk(KERN_ERR "ERROR: domain->cpu_power not "
5900 if (!cpumask_weight(sched_group_cpus(group))) {
5901 printk(KERN_CONT "\n");
5902 printk(KERN_ERR "ERROR: empty group\n");
5906 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
5907 printk(KERN_CONT "\n");
5908 printk(KERN_ERR "ERROR: repeated CPUs\n");
5912 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5914 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5916 printk(KERN_CONT " %s", str);
5917 if (group->cpu_power != SCHED_LOAD_SCALE) {
5918 printk(KERN_CONT " (cpu_power = %d)",
5922 group = group->next;
5923 } while (group != sd->groups);
5924 printk(KERN_CONT "\n");
5926 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5927 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5930 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5931 printk(KERN_ERR "ERROR: parent span is not a superset "
5932 "of domain->span\n");
5936 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5938 cpumask_var_t groupmask;
5941 if (!sched_domain_debug_enabled)
5945 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5949 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5951 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
5952 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
5957 if (sched_domain_debug_one(sd, cpu, level, groupmask))
5964 free_cpumask_var(groupmask);
5966 #else /* !CONFIG_SCHED_DEBUG */
5967 # define sched_domain_debug(sd, cpu) do { } while (0)
5968 #endif /* CONFIG_SCHED_DEBUG */
5970 static int sd_degenerate(struct sched_domain *sd)
5972 if (cpumask_weight(sched_domain_span(sd)) == 1)
5975 /* Following flags need at least 2 groups */
5976 if (sd->flags & (SD_LOAD_BALANCE |
5977 SD_BALANCE_NEWIDLE |
5981 SD_SHARE_PKG_RESOURCES)) {
5982 if (sd->groups != sd->groups->next)
5986 /* Following flags don't use groups */
5987 if (sd->flags & (SD_WAKE_AFFINE))
5994 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5996 unsigned long cflags = sd->flags, pflags = parent->flags;
5998 if (sd_degenerate(parent))
6001 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6004 /* Flags needing groups don't count if only 1 group in parent */
6005 if (parent->groups == parent->groups->next) {
6006 pflags &= ~(SD_LOAD_BALANCE |
6007 SD_BALANCE_NEWIDLE |
6011 SD_SHARE_PKG_RESOURCES);
6012 if (nr_node_ids == 1)
6013 pflags &= ~SD_SERIALIZE;
6015 if (~cflags & pflags)
6021 static void free_rootdomain(struct root_domain *rd)
6023 synchronize_sched();
6025 cpupri_cleanup(&rd->cpupri);
6027 free_cpumask_var(rd->rto_mask);
6028 free_cpumask_var(rd->online);
6029 free_cpumask_var(rd->span);
6033 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6035 struct root_domain *old_rd = NULL;
6036 unsigned long flags;
6038 raw_spin_lock_irqsave(&rq->lock, flags);
6043 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6046 cpumask_clear_cpu(rq->cpu, old_rd->span);
6049 * If we dont want to free the old_rt yet then
6050 * set old_rd to NULL to skip the freeing later
6053 if (!atomic_dec_and_test(&old_rd->refcount))
6057 atomic_inc(&rd->refcount);
6060 cpumask_set_cpu(rq->cpu, rd->span);
6061 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6064 raw_spin_unlock_irqrestore(&rq->lock, flags);
6067 free_rootdomain(old_rd);
6070 static int init_rootdomain(struct root_domain *rd, bool bootmem)
6072 gfp_t gfp = GFP_KERNEL;
6074 memset(rd, 0, sizeof(*rd));
6079 if (!alloc_cpumask_var(&rd->span, gfp))
6081 if (!alloc_cpumask_var(&rd->online, gfp))
6083 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
6086 if (cpupri_init(&rd->cpupri, bootmem) != 0)
6091 free_cpumask_var(rd->rto_mask);
6093 free_cpumask_var(rd->online);
6095 free_cpumask_var(rd->span);
6100 static void init_defrootdomain(void)
6102 init_rootdomain(&def_root_domain, true);
6104 atomic_set(&def_root_domain.refcount, 1);
6107 static struct root_domain *alloc_rootdomain(void)
6109 struct root_domain *rd;
6111 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6115 if (init_rootdomain(rd, false) != 0) {
6124 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6125 * hold the hotplug lock.
6128 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6130 struct rq *rq = cpu_rq(cpu);
6131 struct sched_domain *tmp;
6133 for (tmp = sd; tmp; tmp = tmp->parent)
6134 tmp->span_weight = cpumask_weight(sched_domain_span(tmp));
6136 /* Remove the sched domains which do not contribute to scheduling. */
6137 for (tmp = sd; tmp; ) {
6138 struct sched_domain *parent = tmp->parent;
6142 if (sd_parent_degenerate(tmp, parent)) {
6143 tmp->parent = parent->parent;
6145 parent->parent->child = tmp;
6150 if (sd && sd_degenerate(sd)) {
6156 sched_domain_debug(sd, cpu);
6158 rq_attach_root(rq, rd);
6159 rcu_assign_pointer(rq->sd, sd);
6162 /* cpus with isolated domains */
6163 static cpumask_var_t cpu_isolated_map;
6165 /* Setup the mask of cpus configured for isolated domains */
6166 static int __init isolated_cpu_setup(char *str)
6168 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6169 cpulist_parse(str, cpu_isolated_map);
6173 __setup("isolcpus=", isolated_cpu_setup);
6176 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6177 * to a function which identifies what group(along with sched group) a CPU
6178 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6179 * (due to the fact that we keep track of groups covered with a struct cpumask).
6181 * init_sched_build_groups will build a circular linked list of the groups
6182 * covered by the given span, and will set each group's ->cpumask correctly,
6183 * and ->cpu_power to 0.
6186 init_sched_build_groups(const struct cpumask *span,
6187 const struct cpumask *cpu_map,
6188 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
6189 struct sched_group **sg,
6190 struct cpumask *tmpmask),
6191 struct cpumask *covered, struct cpumask *tmpmask)
6193 struct sched_group *first = NULL, *last = NULL;
6196 cpumask_clear(covered);
6198 for_each_cpu(i, span) {
6199 struct sched_group *sg;
6200 int group = group_fn(i, cpu_map, &sg, tmpmask);
6203 if (cpumask_test_cpu(i, covered))
6206 cpumask_clear(sched_group_cpus(sg));
6209 for_each_cpu(j, span) {
6210 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6213 cpumask_set_cpu(j, covered);
6214 cpumask_set_cpu(j, sched_group_cpus(sg));
6225 #define SD_NODES_PER_DOMAIN 16
6230 * find_next_best_node - find the next node to include in a sched_domain
6231 * @node: node whose sched_domain we're building
6232 * @used_nodes: nodes already in the sched_domain
6234 * Find the next node to include in a given scheduling domain. Simply
6235 * finds the closest node not already in the @used_nodes map.
6237 * Should use nodemask_t.
6239 static int find_next_best_node(int node, nodemask_t *used_nodes)
6241 int i, n, val, min_val, best_node = 0;
6245 for (i = 0; i < nr_node_ids; i++) {
6246 /* Start at @node */
6247 n = (node + i) % nr_node_ids;
6249 if (!nr_cpus_node(n))
6252 /* Skip already used nodes */
6253 if (node_isset(n, *used_nodes))
6256 /* Simple min distance search */
6257 val = node_distance(node, n);
6259 if (val < min_val) {
6265 node_set(best_node, *used_nodes);
6270 * sched_domain_node_span - get a cpumask for a node's sched_domain
6271 * @node: node whose cpumask we're constructing
6272 * @span: resulting cpumask
6274 * Given a node, construct a good cpumask for its sched_domain to span. It
6275 * should be one that prevents unnecessary balancing, but also spreads tasks
6278 static void sched_domain_node_span(int node, struct cpumask *span)
6280 nodemask_t used_nodes;
6283 cpumask_clear(span);
6284 nodes_clear(used_nodes);
6286 cpumask_or(span, span, cpumask_of_node(node));
6287 node_set(node, used_nodes);
6289 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6290 int next_node = find_next_best_node(node, &used_nodes);
6292 cpumask_or(span, span, cpumask_of_node(next_node));
6295 #endif /* CONFIG_NUMA */
6297 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6300 * The cpus mask in sched_group and sched_domain hangs off the end.
6302 * ( See the the comments in include/linux/sched.h:struct sched_group
6303 * and struct sched_domain. )
6305 struct static_sched_group {
6306 struct sched_group sg;
6307 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
6310 struct static_sched_domain {
6311 struct sched_domain sd;
6312 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
6318 cpumask_var_t domainspan;
6319 cpumask_var_t covered;
6320 cpumask_var_t notcovered;
6322 cpumask_var_t nodemask;
6323 cpumask_var_t this_sibling_map;
6324 cpumask_var_t this_core_map;
6325 cpumask_var_t send_covered;
6326 cpumask_var_t tmpmask;
6327 struct sched_group **sched_group_nodes;
6328 struct root_domain *rd;
6332 sa_sched_groups = 0,
6337 sa_this_sibling_map,
6339 sa_sched_group_nodes,
6349 * SMT sched-domains:
6351 #ifdef CONFIG_SCHED_SMT
6352 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
6353 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
6356 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
6357 struct sched_group **sg, struct cpumask *unused)
6360 *sg = &per_cpu(sched_groups, cpu).sg;
6363 #endif /* CONFIG_SCHED_SMT */
6366 * multi-core sched-domains:
6368 #ifdef CONFIG_SCHED_MC
6369 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
6370 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
6371 #endif /* CONFIG_SCHED_MC */
6373 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6375 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6376 struct sched_group **sg, struct cpumask *mask)
6380 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6381 group = cpumask_first(mask);
6383 *sg = &per_cpu(sched_group_core, group).sg;
6386 #elif defined(CONFIG_SCHED_MC)
6388 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6389 struct sched_group **sg, struct cpumask *unused)
6392 *sg = &per_cpu(sched_group_core, cpu).sg;
6397 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
6398 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
6401 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
6402 struct sched_group **sg, struct cpumask *mask)
6405 #ifdef CONFIG_SCHED_MC
6406 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6407 group = cpumask_first(mask);
6408 #elif defined(CONFIG_SCHED_SMT)
6409 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6410 group = cpumask_first(mask);
6415 *sg = &per_cpu(sched_group_phys, group).sg;
6421 * The init_sched_build_groups can't handle what we want to do with node
6422 * groups, so roll our own. Now each node has its own list of groups which
6423 * gets dynamically allocated.
6425 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
6426 static struct sched_group ***sched_group_nodes_bycpu;
6428 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
6429 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
6431 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
6432 struct sched_group **sg,
6433 struct cpumask *nodemask)
6437 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
6438 group = cpumask_first(nodemask);
6441 *sg = &per_cpu(sched_group_allnodes, group).sg;
6445 static void init_numa_sched_groups_power(struct sched_group *group_head)
6447 struct sched_group *sg = group_head;
6453 for_each_cpu(j, sched_group_cpus(sg)) {
6454 struct sched_domain *sd;
6456 sd = &per_cpu(phys_domains, j).sd;
6457 if (j != group_first_cpu(sd->groups)) {
6459 * Only add "power" once for each
6465 sg->cpu_power += sd->groups->cpu_power;
6468 } while (sg != group_head);
6471 static int build_numa_sched_groups(struct s_data *d,
6472 const struct cpumask *cpu_map, int num)
6474 struct sched_domain *sd;
6475 struct sched_group *sg, *prev;
6478 cpumask_clear(d->covered);
6479 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
6480 if (cpumask_empty(d->nodemask)) {
6481 d->sched_group_nodes[num] = NULL;
6485 sched_domain_node_span(num, d->domainspan);
6486 cpumask_and(d->domainspan, d->domainspan, cpu_map);
6488 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6491 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
6495 d->sched_group_nodes[num] = sg;
6497 for_each_cpu(j, d->nodemask) {
6498 sd = &per_cpu(node_domains, j).sd;
6503 cpumask_copy(sched_group_cpus(sg), d->nodemask);
6505 cpumask_or(d->covered, d->covered, d->nodemask);
6508 for (j = 0; j < nr_node_ids; j++) {
6509 n = (num + j) % nr_node_ids;
6510 cpumask_complement(d->notcovered, d->covered);
6511 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
6512 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
6513 if (cpumask_empty(d->tmpmask))
6515 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
6516 if (cpumask_empty(d->tmpmask))
6518 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6522 "Can not alloc domain group for node %d\n", j);
6526 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
6527 sg->next = prev->next;
6528 cpumask_or(d->covered, d->covered, d->tmpmask);
6535 #endif /* CONFIG_NUMA */
6538 /* Free memory allocated for various sched_group structures */
6539 static void free_sched_groups(const struct cpumask *cpu_map,
6540 struct cpumask *nodemask)
6544 for_each_cpu(cpu, cpu_map) {
6545 struct sched_group **sched_group_nodes
6546 = sched_group_nodes_bycpu[cpu];
6548 if (!sched_group_nodes)
6551 for (i = 0; i < nr_node_ids; i++) {
6552 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6554 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
6555 if (cpumask_empty(nodemask))
6565 if (oldsg != sched_group_nodes[i])
6568 kfree(sched_group_nodes);
6569 sched_group_nodes_bycpu[cpu] = NULL;
6572 #else /* !CONFIG_NUMA */
6573 static void free_sched_groups(const struct cpumask *cpu_map,
6574 struct cpumask *nodemask)
6577 #endif /* CONFIG_NUMA */
6580 * Initialize sched groups cpu_power.
6582 * cpu_power indicates the capacity of sched group, which is used while
6583 * distributing the load between different sched groups in a sched domain.
6584 * Typically cpu_power for all the groups in a sched domain will be same unless
6585 * there are asymmetries in the topology. If there are asymmetries, group
6586 * having more cpu_power will pickup more load compared to the group having
6589 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6591 struct sched_domain *child;
6592 struct sched_group *group;
6596 WARN_ON(!sd || !sd->groups);
6598 if (cpu != group_first_cpu(sd->groups))
6603 sd->groups->cpu_power = 0;
6606 power = SCHED_LOAD_SCALE;
6607 weight = cpumask_weight(sched_domain_span(sd));
6609 * SMT siblings share the power of a single core.
6610 * Usually multiple threads get a better yield out of
6611 * that one core than a single thread would have,
6612 * reflect that in sd->smt_gain.
6614 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
6615 power *= sd->smt_gain;
6617 power >>= SCHED_LOAD_SHIFT;
6619 sd->groups->cpu_power += power;
6624 * Add cpu_power of each child group to this groups cpu_power.
6626 group = child->groups;
6628 sd->groups->cpu_power += group->cpu_power;
6629 group = group->next;
6630 } while (group != child->groups);
6634 * Initializers for schedule domains
6635 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6638 #ifdef CONFIG_SCHED_DEBUG
6639 # define SD_INIT_NAME(sd, type) sd->name = #type
6641 # define SD_INIT_NAME(sd, type) do { } while (0)
6644 #define SD_INIT(sd, type) sd_init_##type(sd)
6646 #define SD_INIT_FUNC(type) \
6647 static noinline void sd_init_##type(struct sched_domain *sd) \
6649 memset(sd, 0, sizeof(*sd)); \
6650 *sd = SD_##type##_INIT; \
6651 sd->level = SD_LV_##type; \
6652 SD_INIT_NAME(sd, type); \
6657 SD_INIT_FUNC(ALLNODES)
6660 #ifdef CONFIG_SCHED_SMT
6661 SD_INIT_FUNC(SIBLING)
6663 #ifdef CONFIG_SCHED_MC
6667 static int default_relax_domain_level = -1;
6669 static int __init setup_relax_domain_level(char *str)
6673 val = simple_strtoul(str, NULL, 0);
6674 if (val < SD_LV_MAX)
6675 default_relax_domain_level = val;
6679 __setup("relax_domain_level=", setup_relax_domain_level);
6681 static void set_domain_attribute(struct sched_domain *sd,
6682 struct sched_domain_attr *attr)
6686 if (!attr || attr->relax_domain_level < 0) {
6687 if (default_relax_domain_level < 0)
6690 request = default_relax_domain_level;
6692 request = attr->relax_domain_level;
6693 if (request < sd->level) {
6694 /* turn off idle balance on this domain */
6695 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6697 /* turn on idle balance on this domain */
6698 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6702 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6703 const struct cpumask *cpu_map)
6706 case sa_sched_groups:
6707 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
6708 d->sched_group_nodes = NULL;
6710 free_rootdomain(d->rd); /* fall through */
6712 free_cpumask_var(d->tmpmask); /* fall through */
6713 case sa_send_covered:
6714 free_cpumask_var(d->send_covered); /* fall through */
6715 case sa_this_core_map:
6716 free_cpumask_var(d->this_core_map); /* fall through */
6717 case sa_this_sibling_map:
6718 free_cpumask_var(d->this_sibling_map); /* fall through */
6720 free_cpumask_var(d->nodemask); /* fall through */
6721 case sa_sched_group_nodes:
6723 kfree(d->sched_group_nodes); /* fall through */
6725 free_cpumask_var(d->notcovered); /* fall through */
6727 free_cpumask_var(d->covered); /* fall through */
6729 free_cpumask_var(d->domainspan); /* fall through */
6736 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6737 const struct cpumask *cpu_map)
6740 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
6742 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
6743 return sa_domainspan;
6744 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
6746 /* Allocate the per-node list of sched groups */
6747 d->sched_group_nodes = kcalloc(nr_node_ids,
6748 sizeof(struct sched_group *), GFP_KERNEL);
6749 if (!d->sched_group_nodes) {
6750 printk(KERN_WARNING "Can not alloc sched group node list\n");
6751 return sa_notcovered;
6753 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
6755 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
6756 return sa_sched_group_nodes;
6757 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
6759 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
6760 return sa_this_sibling_map;
6761 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
6762 return sa_this_core_map;
6763 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
6764 return sa_send_covered;
6765 d->rd = alloc_rootdomain();
6767 printk(KERN_WARNING "Cannot alloc root domain\n");
6770 return sa_rootdomain;
6773 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
6774 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
6776 struct sched_domain *sd = NULL;
6778 struct sched_domain *parent;
6781 if (cpumask_weight(cpu_map) >
6782 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
6783 sd = &per_cpu(allnodes_domains, i).sd;
6784 SD_INIT(sd, ALLNODES);
6785 set_domain_attribute(sd, attr);
6786 cpumask_copy(sched_domain_span(sd), cpu_map);
6787 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
6792 sd = &per_cpu(node_domains, i).sd;
6794 set_domain_attribute(sd, attr);
6795 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
6796 sd->parent = parent;
6799 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
6804 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
6805 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6806 struct sched_domain *parent, int i)
6808 struct sched_domain *sd;
6809 sd = &per_cpu(phys_domains, i).sd;
6811 set_domain_attribute(sd, attr);
6812 cpumask_copy(sched_domain_span(sd), d->nodemask);
6813 sd->parent = parent;
6816 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
6820 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
6821 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6822 struct sched_domain *parent, int i)
6824 struct sched_domain *sd = parent;
6825 #ifdef CONFIG_SCHED_MC
6826 sd = &per_cpu(core_domains, i).sd;
6828 set_domain_attribute(sd, attr);
6829 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
6830 sd->parent = parent;
6832 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
6837 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
6838 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6839 struct sched_domain *parent, int i)
6841 struct sched_domain *sd = parent;
6842 #ifdef CONFIG_SCHED_SMT
6843 sd = &per_cpu(cpu_domains, i).sd;
6844 SD_INIT(sd, SIBLING);
6845 set_domain_attribute(sd, attr);
6846 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
6847 sd->parent = parent;
6849 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
6854 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
6855 const struct cpumask *cpu_map, int cpu)
6858 #ifdef CONFIG_SCHED_SMT
6859 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
6860 cpumask_and(d->this_sibling_map, cpu_map,
6861 topology_thread_cpumask(cpu));
6862 if (cpu == cpumask_first(d->this_sibling_map))
6863 init_sched_build_groups(d->this_sibling_map, cpu_map,
6865 d->send_covered, d->tmpmask);
6868 #ifdef CONFIG_SCHED_MC
6869 case SD_LV_MC: /* set up multi-core groups */
6870 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
6871 if (cpu == cpumask_first(d->this_core_map))
6872 init_sched_build_groups(d->this_core_map, cpu_map,
6874 d->send_covered, d->tmpmask);
6877 case SD_LV_CPU: /* set up physical groups */
6878 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
6879 if (!cpumask_empty(d->nodemask))
6880 init_sched_build_groups(d->nodemask, cpu_map,
6882 d->send_covered, d->tmpmask);
6885 case SD_LV_ALLNODES:
6886 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
6887 d->send_covered, d->tmpmask);
6896 * Build sched domains for a given set of cpus and attach the sched domains
6897 * to the individual cpus
6899 static int __build_sched_domains(const struct cpumask *cpu_map,
6900 struct sched_domain_attr *attr)
6902 enum s_alloc alloc_state = sa_none;
6904 struct sched_domain *sd;
6910 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6911 if (alloc_state != sa_rootdomain)
6913 alloc_state = sa_sched_groups;
6916 * Set up domains for cpus specified by the cpu_map.
6918 for_each_cpu(i, cpu_map) {
6919 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
6922 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
6923 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
6924 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
6925 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
6928 for_each_cpu(i, cpu_map) {
6929 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
6930 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
6933 /* Set up physical groups */
6934 for (i = 0; i < nr_node_ids; i++)
6935 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
6938 /* Set up node groups */
6940 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
6942 for (i = 0; i < nr_node_ids; i++)
6943 if (build_numa_sched_groups(&d, cpu_map, i))
6947 /* Calculate CPU power for physical packages and nodes */
6948 #ifdef CONFIG_SCHED_SMT
6949 for_each_cpu(i, cpu_map) {
6950 sd = &per_cpu(cpu_domains, i).sd;
6951 init_sched_groups_power(i, sd);
6954 #ifdef CONFIG_SCHED_MC
6955 for_each_cpu(i, cpu_map) {
6956 sd = &per_cpu(core_domains, i).sd;
6957 init_sched_groups_power(i, sd);
6961 for_each_cpu(i, cpu_map) {
6962 sd = &per_cpu(phys_domains, i).sd;
6963 init_sched_groups_power(i, sd);
6967 for (i = 0; i < nr_node_ids; i++)
6968 init_numa_sched_groups_power(d.sched_group_nodes[i]);
6970 if (d.sd_allnodes) {
6971 struct sched_group *sg;
6973 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
6975 init_numa_sched_groups_power(sg);
6979 /* Attach the domains */
6980 for_each_cpu(i, cpu_map) {
6981 #ifdef CONFIG_SCHED_SMT
6982 sd = &per_cpu(cpu_domains, i).sd;
6983 #elif defined(CONFIG_SCHED_MC)
6984 sd = &per_cpu(core_domains, i).sd;
6986 sd = &per_cpu(phys_domains, i).sd;
6988 cpu_attach_domain(sd, d.rd, i);
6991 d.sched_group_nodes = NULL; /* don't free this we still need it */
6992 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
6996 __free_domain_allocs(&d, alloc_state, cpu_map);
7000 static int build_sched_domains(const struct cpumask *cpu_map)
7002 return __build_sched_domains(cpu_map, NULL);
7005 static cpumask_var_t *doms_cur; /* current sched domains */
7006 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7007 static struct sched_domain_attr *dattr_cur;
7008 /* attribues of custom domains in 'doms_cur' */
7011 * Special case: If a kmalloc of a doms_cur partition (array of
7012 * cpumask) fails, then fallback to a single sched domain,
7013 * as determined by the single cpumask fallback_doms.
7015 static cpumask_var_t fallback_doms;
7018 * arch_update_cpu_topology lets virtualized architectures update the
7019 * cpu core maps. It is supposed to return 1 if the topology changed
7020 * or 0 if it stayed the same.
7022 int __attribute__((weak)) arch_update_cpu_topology(void)
7027 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7030 cpumask_var_t *doms;
7032 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7035 for (i = 0; i < ndoms; i++) {
7036 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7037 free_sched_domains(doms, i);
7044 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7047 for (i = 0; i < ndoms; i++)
7048 free_cpumask_var(doms[i]);
7053 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7054 * For now this just excludes isolated cpus, but could be used to
7055 * exclude other special cases in the future.
7057 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7061 arch_update_cpu_topology();
7063 doms_cur = alloc_sched_domains(ndoms_cur);
7065 doms_cur = &fallback_doms;
7066 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7068 err = build_sched_domains(doms_cur[0]);
7069 register_sched_domain_sysctl();
7074 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7075 struct cpumask *tmpmask)
7077 free_sched_groups(cpu_map, tmpmask);
7081 * Detach sched domains from a group of cpus specified in cpu_map
7082 * These cpus will now be attached to the NULL domain
7084 static void detach_destroy_domains(const struct cpumask *cpu_map)
7086 /* Save because hotplug lock held. */
7087 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7090 for_each_cpu(i, cpu_map)
7091 cpu_attach_domain(NULL, &def_root_domain, i);
7092 synchronize_sched();
7093 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7096 /* handle null as "default" */
7097 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7098 struct sched_domain_attr *new, int idx_new)
7100 struct sched_domain_attr tmp;
7107 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7108 new ? (new + idx_new) : &tmp,
7109 sizeof(struct sched_domain_attr));
7113 * Partition sched domains as specified by the 'ndoms_new'
7114 * cpumasks in the array doms_new[] of cpumasks. This compares
7115 * doms_new[] to the current sched domain partitioning, doms_cur[].
7116 * It destroys each deleted domain and builds each new domain.
7118 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7119 * The masks don't intersect (don't overlap.) We should setup one
7120 * sched domain for each mask. CPUs not in any of the cpumasks will
7121 * not be load balanced. If the same cpumask appears both in the
7122 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7125 * The passed in 'doms_new' should be allocated using
7126 * alloc_sched_domains. This routine takes ownership of it and will
7127 * free_sched_domains it when done with it. If the caller failed the
7128 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7129 * and partition_sched_domains() will fallback to the single partition
7130 * 'fallback_doms', it also forces the domains to be rebuilt.
7132 * If doms_new == NULL it will be replaced with cpu_online_mask.
7133 * ndoms_new == 0 is a special case for destroying existing domains,
7134 * and it will not create the default domain.
7136 * Call with hotplug lock held
7138 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7139 struct sched_domain_attr *dattr_new)
7144 mutex_lock(&sched_domains_mutex);
7146 /* always unregister in case we don't destroy any domains */
7147 unregister_sched_domain_sysctl();
7149 /* Let architecture update cpu core mappings. */
7150 new_topology = arch_update_cpu_topology();
7152 n = doms_new ? ndoms_new : 0;
7154 /* Destroy deleted domains */
7155 for (i = 0; i < ndoms_cur; i++) {
7156 for (j = 0; j < n && !new_topology; j++) {
7157 if (cpumask_equal(doms_cur[i], doms_new[j])
7158 && dattrs_equal(dattr_cur, i, dattr_new, j))
7161 /* no match - a current sched domain not in new doms_new[] */
7162 detach_destroy_domains(doms_cur[i]);
7167 if (doms_new == NULL) {
7169 doms_new = &fallback_doms;
7170 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7171 WARN_ON_ONCE(dattr_new);
7174 /* Build new domains */
7175 for (i = 0; i < ndoms_new; i++) {
7176 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7177 if (cpumask_equal(doms_new[i], doms_cur[j])
7178 && dattrs_equal(dattr_new, i, dattr_cur, j))
7181 /* no match - add a new doms_new */
7182 __build_sched_domains(doms_new[i],
7183 dattr_new ? dattr_new + i : NULL);
7188 /* Remember the new sched domains */
7189 if (doms_cur != &fallback_doms)
7190 free_sched_domains(doms_cur, ndoms_cur);
7191 kfree(dattr_cur); /* kfree(NULL) is safe */
7192 doms_cur = doms_new;
7193 dattr_cur = dattr_new;
7194 ndoms_cur = ndoms_new;
7196 register_sched_domain_sysctl();
7198 mutex_unlock(&sched_domains_mutex);
7201 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7202 static void arch_reinit_sched_domains(void)
7206 /* Destroy domains first to force the rebuild */
7207 partition_sched_domains(0, NULL, NULL);
7209 rebuild_sched_domains();
7213 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7215 unsigned int level = 0;
7217 if (sscanf(buf, "%u", &level) != 1)
7221 * level is always be positive so don't check for
7222 * level < POWERSAVINGS_BALANCE_NONE which is 0
7223 * What happens on 0 or 1 byte write,
7224 * need to check for count as well?
7227 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7231 sched_smt_power_savings = level;
7233 sched_mc_power_savings = level;
7235 arch_reinit_sched_domains();
7240 #ifdef CONFIG_SCHED_MC
7241 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7242 struct sysdev_class_attribute *attr,
7245 return sprintf(page, "%u\n", sched_mc_power_savings);
7247 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7248 struct sysdev_class_attribute *attr,
7249 const char *buf, size_t count)
7251 return sched_power_savings_store(buf, count, 0);
7253 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7254 sched_mc_power_savings_show,
7255 sched_mc_power_savings_store);
7258 #ifdef CONFIG_SCHED_SMT
7259 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7260 struct sysdev_class_attribute *attr,
7263 return sprintf(page, "%u\n", sched_smt_power_savings);
7265 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7266 struct sysdev_class_attribute *attr,
7267 const char *buf, size_t count)
7269 return sched_power_savings_store(buf, count, 1);
7271 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7272 sched_smt_power_savings_show,
7273 sched_smt_power_savings_store);
7276 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7280 #ifdef CONFIG_SCHED_SMT
7282 err = sysfs_create_file(&cls->kset.kobj,
7283 &attr_sched_smt_power_savings.attr);
7285 #ifdef CONFIG_SCHED_MC
7286 if (!err && mc_capable())
7287 err = sysfs_create_file(&cls->kset.kobj,
7288 &attr_sched_mc_power_savings.attr);
7292 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7294 #ifndef CONFIG_CPUSETS
7296 * Add online and remove offline CPUs from the scheduler domains.
7297 * When cpusets are enabled they take over this function.
7299 static int update_sched_domains(struct notifier_block *nfb,
7300 unsigned long action, void *hcpu)
7304 case CPU_ONLINE_FROZEN:
7305 case CPU_DOWN_PREPARE:
7306 case CPU_DOWN_PREPARE_FROZEN:
7307 case CPU_DOWN_FAILED:
7308 case CPU_DOWN_FAILED_FROZEN:
7309 partition_sched_domains(1, NULL, NULL);
7318 static int update_runtime(struct notifier_block *nfb,
7319 unsigned long action, void *hcpu)
7321 int cpu = (int)(long)hcpu;
7324 case CPU_DOWN_PREPARE:
7325 case CPU_DOWN_PREPARE_FROZEN:
7326 disable_runtime(cpu_rq(cpu));
7329 case CPU_DOWN_FAILED:
7330 case CPU_DOWN_FAILED_FROZEN:
7332 case CPU_ONLINE_FROZEN:
7333 enable_runtime(cpu_rq(cpu));
7341 void __init sched_init_smp(void)
7343 cpumask_var_t non_isolated_cpus;
7345 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7346 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7348 #if defined(CONFIG_NUMA)
7349 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7351 BUG_ON(sched_group_nodes_bycpu == NULL);
7354 mutex_lock(&sched_domains_mutex);
7355 arch_init_sched_domains(cpu_active_mask);
7356 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7357 if (cpumask_empty(non_isolated_cpus))
7358 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7359 mutex_unlock(&sched_domains_mutex);
7362 #ifndef CONFIG_CPUSETS
7363 /* XXX: Theoretical race here - CPU may be hotplugged now */
7364 hotcpu_notifier(update_sched_domains, 0);
7367 /* RT runtime code needs to handle some hotplug events */
7368 hotcpu_notifier(update_runtime, 0);
7372 /* Move init over to a non-isolated CPU */
7373 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7375 sched_init_granularity();
7376 free_cpumask_var(non_isolated_cpus);
7378 init_sched_rt_class();
7381 void __init sched_init_smp(void)
7383 sched_init_granularity();
7385 #endif /* CONFIG_SMP */
7387 const_debug unsigned int sysctl_timer_migration = 1;
7389 int in_sched_functions(unsigned long addr)
7391 return in_lock_functions(addr) ||
7392 (addr >= (unsigned long)__sched_text_start
7393 && addr < (unsigned long)__sched_text_end);
7396 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7398 cfs_rq->tasks_timeline = RB_ROOT;
7399 INIT_LIST_HEAD(&cfs_rq->tasks);
7400 #ifdef CONFIG_FAIR_GROUP_SCHED
7403 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7406 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7408 struct rt_prio_array *array;
7411 array = &rt_rq->active;
7412 for (i = 0; i < MAX_RT_PRIO; i++) {
7413 INIT_LIST_HEAD(array->queue + i);
7414 __clear_bit(i, array->bitmap);
7416 /* delimiter for bitsearch: */
7417 __set_bit(MAX_RT_PRIO, array->bitmap);
7419 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7420 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7422 rt_rq->highest_prio.next = MAX_RT_PRIO;
7426 rt_rq->rt_nr_migratory = 0;
7427 rt_rq->overloaded = 0;
7428 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
7432 rt_rq->rt_throttled = 0;
7433 rt_rq->rt_runtime = 0;
7434 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7436 #ifdef CONFIG_RT_GROUP_SCHED
7437 rt_rq->rt_nr_boosted = 0;
7442 #ifdef CONFIG_FAIR_GROUP_SCHED
7443 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7444 struct sched_entity *se, int cpu, int add,
7445 struct sched_entity *parent)
7447 struct rq *rq = cpu_rq(cpu);
7448 tg->cfs_rq[cpu] = cfs_rq;
7449 init_cfs_rq(cfs_rq, rq);
7452 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7455 /* se could be NULL for init_task_group */
7460 se->cfs_rq = &rq->cfs;
7462 se->cfs_rq = parent->my_q;
7465 se->load.weight = tg->shares;
7466 se->load.inv_weight = 0;
7467 se->parent = parent;
7471 #ifdef CONFIG_RT_GROUP_SCHED
7472 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7473 struct sched_rt_entity *rt_se, int cpu, int add,
7474 struct sched_rt_entity *parent)
7476 struct rq *rq = cpu_rq(cpu);
7478 tg->rt_rq[cpu] = rt_rq;
7479 init_rt_rq(rt_rq, rq);
7481 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7483 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7485 tg->rt_se[cpu] = rt_se;
7490 rt_se->rt_rq = &rq->rt;
7492 rt_se->rt_rq = parent->my_q;
7494 rt_se->my_q = rt_rq;
7495 rt_se->parent = parent;
7496 INIT_LIST_HEAD(&rt_se->run_list);
7500 void __init sched_init(void)
7503 unsigned long alloc_size = 0, ptr;
7505 #ifdef CONFIG_FAIR_GROUP_SCHED
7506 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7508 #ifdef CONFIG_RT_GROUP_SCHED
7509 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7511 #ifdef CONFIG_CPUMASK_OFFSTACK
7512 alloc_size += num_possible_cpus() * cpumask_size();
7515 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7517 #ifdef CONFIG_FAIR_GROUP_SCHED
7518 init_task_group.se = (struct sched_entity **)ptr;
7519 ptr += nr_cpu_ids * sizeof(void **);
7521 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7522 ptr += nr_cpu_ids * sizeof(void **);
7524 #endif /* CONFIG_FAIR_GROUP_SCHED */
7525 #ifdef CONFIG_RT_GROUP_SCHED
7526 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7527 ptr += nr_cpu_ids * sizeof(void **);
7529 init_task_group.rt_rq = (struct rt_rq **)ptr;
7530 ptr += nr_cpu_ids * sizeof(void **);
7532 #endif /* CONFIG_RT_GROUP_SCHED */
7533 #ifdef CONFIG_CPUMASK_OFFSTACK
7534 for_each_possible_cpu(i) {
7535 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
7536 ptr += cpumask_size();
7538 #endif /* CONFIG_CPUMASK_OFFSTACK */
7542 init_defrootdomain();
7545 init_rt_bandwidth(&def_rt_bandwidth,
7546 global_rt_period(), global_rt_runtime());
7548 #ifdef CONFIG_RT_GROUP_SCHED
7549 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7550 global_rt_period(), global_rt_runtime());
7551 #endif /* CONFIG_RT_GROUP_SCHED */
7553 #ifdef CONFIG_CGROUP_SCHED
7554 list_add(&init_task_group.list, &task_groups);
7555 INIT_LIST_HEAD(&init_task_group.children);
7557 #endif /* CONFIG_CGROUP_SCHED */
7559 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
7560 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
7561 __alignof__(unsigned long));
7563 for_each_possible_cpu(i) {
7567 raw_spin_lock_init(&rq->lock);
7569 rq->calc_load_active = 0;
7570 rq->calc_load_update = jiffies + LOAD_FREQ;
7571 init_cfs_rq(&rq->cfs, rq);
7572 init_rt_rq(&rq->rt, rq);
7573 #ifdef CONFIG_FAIR_GROUP_SCHED
7574 init_task_group.shares = init_task_group_load;
7575 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7576 #ifdef CONFIG_CGROUP_SCHED
7578 * How much cpu bandwidth does init_task_group get?
7580 * In case of task-groups formed thr' the cgroup filesystem, it
7581 * gets 100% of the cpu resources in the system. This overall
7582 * system cpu resource is divided among the tasks of
7583 * init_task_group and its child task-groups in a fair manner,
7584 * based on each entity's (task or task-group's) weight
7585 * (se->load.weight).
7587 * In other words, if init_task_group has 10 tasks of weight
7588 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7589 * then A0's share of the cpu resource is:
7591 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7593 * We achieve this by letting init_task_group's tasks sit
7594 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7596 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
7598 #endif /* CONFIG_FAIR_GROUP_SCHED */
7600 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7601 #ifdef CONFIG_RT_GROUP_SCHED
7602 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7603 #ifdef CONFIG_CGROUP_SCHED
7604 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
7608 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7609 rq->cpu_load[j] = 0;
7613 rq->cpu_power = SCHED_LOAD_SCALE;
7614 rq->post_schedule = 0;
7615 rq->active_balance = 0;
7616 rq->next_balance = jiffies;
7621 rq->avg_idle = 2*sysctl_sched_migration_cost;
7622 rq_attach_root(rq, &def_root_domain);
7625 atomic_set(&rq->nr_iowait, 0);
7628 set_load_weight(&init_task);
7630 #ifdef CONFIG_PREEMPT_NOTIFIERS
7631 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7635 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7638 #ifdef CONFIG_RT_MUTEXES
7639 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
7643 * The boot idle thread does lazy MMU switching as well:
7645 atomic_inc(&init_mm.mm_count);
7646 enter_lazy_tlb(&init_mm, current);
7649 * Make us the idle thread. Technically, schedule() should not be
7650 * called from this thread, however somewhere below it might be,
7651 * but because we are the idle thread, we just pick up running again
7652 * when this runqueue becomes "idle".
7654 init_idle(current, smp_processor_id());
7656 calc_load_update = jiffies + LOAD_FREQ;
7659 * During early bootup we pretend to be a normal task:
7661 current->sched_class = &fair_sched_class;
7663 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
7664 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
7667 zalloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
7668 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
7670 /* May be allocated at isolcpus cmdline parse time */
7671 if (cpu_isolated_map == NULL)
7672 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7677 scheduler_running = 1;
7680 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7681 static inline int preempt_count_equals(int preempt_offset)
7683 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7685 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
7688 void __might_sleep(const char *file, int line, int preempt_offset)
7691 static unsigned long prev_jiffy; /* ratelimiting */
7693 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7694 system_state != SYSTEM_RUNNING || oops_in_progress)
7696 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7698 prev_jiffy = jiffies;
7701 "BUG: sleeping function called from invalid context at %s:%d\n",
7704 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7705 in_atomic(), irqs_disabled(),
7706 current->pid, current->comm);
7708 debug_show_held_locks(current);
7709 if (irqs_disabled())
7710 print_irqtrace_events(current);
7714 EXPORT_SYMBOL(__might_sleep);
7717 #ifdef CONFIG_MAGIC_SYSRQ
7718 static void normalize_task(struct rq *rq, struct task_struct *p)
7722 on_rq = p->se.on_rq;
7724 deactivate_task(rq, p, 0);
7725 __setscheduler(rq, p, SCHED_NORMAL, 0);
7727 activate_task(rq, p, 0);
7728 resched_task(rq->curr);
7732 void normalize_rt_tasks(void)
7734 struct task_struct *g, *p;
7735 unsigned long flags;
7738 read_lock_irqsave(&tasklist_lock, flags);
7739 do_each_thread(g, p) {
7741 * Only normalize user tasks:
7746 p->se.exec_start = 0;
7747 #ifdef CONFIG_SCHEDSTATS
7748 p->se.statistics.wait_start = 0;
7749 p->se.statistics.sleep_start = 0;
7750 p->se.statistics.block_start = 0;
7755 * Renice negative nice level userspace
7758 if (TASK_NICE(p) < 0 && p->mm)
7759 set_user_nice(p, 0);
7763 raw_spin_lock(&p->pi_lock);
7764 rq = __task_rq_lock(p);
7766 normalize_task(rq, p);
7768 __task_rq_unlock(rq);
7769 raw_spin_unlock(&p->pi_lock);
7770 } while_each_thread(g, p);
7772 read_unlock_irqrestore(&tasklist_lock, flags);
7775 #endif /* CONFIG_MAGIC_SYSRQ */
7777 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7779 * These functions are only useful for the IA64 MCA handling, or kdb.
7781 * They can only be called when the whole system has been
7782 * stopped - every CPU needs to be quiescent, and no scheduling
7783 * activity can take place. Using them for anything else would
7784 * be a serious bug, and as a result, they aren't even visible
7785 * under any other configuration.
7789 * curr_task - return the current task for a given cpu.
7790 * @cpu: the processor in question.
7792 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7794 struct task_struct *curr_task(int cpu)
7796 return cpu_curr(cpu);
7799 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7803 * set_curr_task - set the current task for a given cpu.
7804 * @cpu: the processor in question.
7805 * @p: the task pointer to set.
7807 * Description: This function must only be used when non-maskable interrupts
7808 * are serviced on a separate stack. It allows the architecture to switch the
7809 * notion of the current task on a cpu in a non-blocking manner. This function
7810 * must be called with all CPU's synchronized, and interrupts disabled, the
7811 * and caller must save the original value of the current task (see
7812 * curr_task() above) and restore that value before reenabling interrupts and
7813 * re-starting the system.
7815 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7817 void set_curr_task(int cpu, struct task_struct *p)
7824 #ifdef CONFIG_FAIR_GROUP_SCHED
7825 static void free_fair_sched_group(struct task_group *tg)
7829 for_each_possible_cpu(i) {
7831 kfree(tg->cfs_rq[i]);
7841 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7843 struct cfs_rq *cfs_rq;
7844 struct sched_entity *se;
7848 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7851 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7855 tg->shares = NICE_0_LOAD;
7857 for_each_possible_cpu(i) {
7860 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
7861 GFP_KERNEL, cpu_to_node(i));
7865 se = kzalloc_node(sizeof(struct sched_entity),
7866 GFP_KERNEL, cpu_to_node(i));
7870 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
7881 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
7883 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
7884 &cpu_rq(cpu)->leaf_cfs_rq_list);
7887 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
7889 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
7891 #else /* !CONFG_FAIR_GROUP_SCHED */
7892 static inline void free_fair_sched_group(struct task_group *tg)
7897 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7902 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
7906 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
7909 #endif /* CONFIG_FAIR_GROUP_SCHED */
7911 #ifdef CONFIG_RT_GROUP_SCHED
7912 static void free_rt_sched_group(struct task_group *tg)
7916 destroy_rt_bandwidth(&tg->rt_bandwidth);
7918 for_each_possible_cpu(i) {
7920 kfree(tg->rt_rq[i]);
7922 kfree(tg->rt_se[i]);
7930 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
7932 struct rt_rq *rt_rq;
7933 struct sched_rt_entity *rt_se;
7937 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
7940 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
7944 init_rt_bandwidth(&tg->rt_bandwidth,
7945 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
7947 for_each_possible_cpu(i) {
7950 rt_rq = kzalloc_node(sizeof(struct rt_rq),
7951 GFP_KERNEL, cpu_to_node(i));
7955 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
7956 GFP_KERNEL, cpu_to_node(i));
7960 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
7971 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
7973 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
7974 &cpu_rq(cpu)->leaf_rt_rq_list);
7977 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
7979 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
7981 #else /* !CONFIG_RT_GROUP_SCHED */
7982 static inline void free_rt_sched_group(struct task_group *tg)
7987 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
7992 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
7996 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
7999 #endif /* CONFIG_RT_GROUP_SCHED */
8001 #ifdef CONFIG_CGROUP_SCHED
8002 static void free_sched_group(struct task_group *tg)
8004 free_fair_sched_group(tg);
8005 free_rt_sched_group(tg);
8009 /* allocate runqueue etc for a new task group */
8010 struct task_group *sched_create_group(struct task_group *parent)
8012 struct task_group *tg;
8013 unsigned long flags;
8016 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8018 return ERR_PTR(-ENOMEM);
8020 if (!alloc_fair_sched_group(tg, parent))
8023 if (!alloc_rt_sched_group(tg, parent))
8026 spin_lock_irqsave(&task_group_lock, flags);
8027 for_each_possible_cpu(i) {
8028 register_fair_sched_group(tg, i);
8029 register_rt_sched_group(tg, i);
8031 list_add_rcu(&tg->list, &task_groups);
8033 WARN_ON(!parent); /* root should already exist */
8035 tg->parent = parent;
8036 INIT_LIST_HEAD(&tg->children);
8037 list_add_rcu(&tg->siblings, &parent->children);
8038 spin_unlock_irqrestore(&task_group_lock, flags);
8043 free_sched_group(tg);
8044 return ERR_PTR(-ENOMEM);
8047 /* rcu callback to free various structures associated with a task group */
8048 static void free_sched_group_rcu(struct rcu_head *rhp)
8050 /* now it should be safe to free those cfs_rqs */
8051 free_sched_group(container_of(rhp, struct task_group, rcu));
8054 /* Destroy runqueue etc associated with a task group */
8055 void sched_destroy_group(struct task_group *tg)
8057 unsigned long flags;
8060 spin_lock_irqsave(&task_group_lock, flags);
8061 for_each_possible_cpu(i) {
8062 unregister_fair_sched_group(tg, i);
8063 unregister_rt_sched_group(tg, i);
8065 list_del_rcu(&tg->list);
8066 list_del_rcu(&tg->siblings);
8067 spin_unlock_irqrestore(&task_group_lock, flags);
8069 /* wait for possible concurrent references to cfs_rqs complete */
8070 call_rcu(&tg->rcu, free_sched_group_rcu);
8073 /* change task's runqueue when it moves between groups.
8074 * The caller of this function should have put the task in its new group
8075 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8076 * reflect its new group.
8078 void sched_move_task(struct task_struct *tsk)
8081 unsigned long flags;
8084 rq = task_rq_lock(tsk, &flags);
8086 running = task_current(rq, tsk);
8087 on_rq = tsk->se.on_rq;
8090 dequeue_task(rq, tsk, 0);
8091 if (unlikely(running))
8092 tsk->sched_class->put_prev_task(rq, tsk);
8094 set_task_rq(tsk, task_cpu(tsk));
8096 #ifdef CONFIG_FAIR_GROUP_SCHED
8097 if (tsk->sched_class->moved_group)
8098 tsk->sched_class->moved_group(tsk, on_rq);
8101 if (unlikely(running))
8102 tsk->sched_class->set_curr_task(rq);
8104 enqueue_task(rq, tsk, 0);
8106 task_rq_unlock(rq, &flags);
8108 #endif /* CONFIG_CGROUP_SCHED */
8110 #ifdef CONFIG_FAIR_GROUP_SCHED
8111 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8113 struct cfs_rq *cfs_rq = se->cfs_rq;
8118 dequeue_entity(cfs_rq, se, 0);
8120 se->load.weight = shares;
8121 se->load.inv_weight = 0;
8124 enqueue_entity(cfs_rq, se, 0);
8127 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8129 struct cfs_rq *cfs_rq = se->cfs_rq;
8130 struct rq *rq = cfs_rq->rq;
8131 unsigned long flags;
8133 raw_spin_lock_irqsave(&rq->lock, flags);
8134 __set_se_shares(se, shares);
8135 raw_spin_unlock_irqrestore(&rq->lock, flags);
8138 static DEFINE_MUTEX(shares_mutex);
8140 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8143 unsigned long flags;
8146 * We can't change the weight of the root cgroup.
8151 if (shares < MIN_SHARES)
8152 shares = MIN_SHARES;
8153 else if (shares > MAX_SHARES)
8154 shares = MAX_SHARES;
8156 mutex_lock(&shares_mutex);
8157 if (tg->shares == shares)
8160 spin_lock_irqsave(&task_group_lock, flags);
8161 for_each_possible_cpu(i)
8162 unregister_fair_sched_group(tg, i);
8163 list_del_rcu(&tg->siblings);
8164 spin_unlock_irqrestore(&task_group_lock, flags);
8166 /* wait for any ongoing reference to this group to finish */
8167 synchronize_sched();
8170 * Now we are free to modify the group's share on each cpu
8171 * w/o tripping rebalance_share or load_balance_fair.
8173 tg->shares = shares;
8174 for_each_possible_cpu(i) {
8178 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8179 set_se_shares(tg->se[i], shares);
8183 * Enable load balance activity on this group, by inserting it back on
8184 * each cpu's rq->leaf_cfs_rq_list.
8186 spin_lock_irqsave(&task_group_lock, flags);
8187 for_each_possible_cpu(i)
8188 register_fair_sched_group(tg, i);
8189 list_add_rcu(&tg->siblings, &tg->parent->children);
8190 spin_unlock_irqrestore(&task_group_lock, flags);
8192 mutex_unlock(&shares_mutex);
8196 unsigned long sched_group_shares(struct task_group *tg)
8202 #ifdef CONFIG_RT_GROUP_SCHED
8204 * Ensure that the real time constraints are schedulable.
8206 static DEFINE_MUTEX(rt_constraints_mutex);
8208 static unsigned long to_ratio(u64 period, u64 runtime)
8210 if (runtime == RUNTIME_INF)
8213 return div64_u64(runtime << 20, period);
8216 /* Must be called with tasklist_lock held */
8217 static inline int tg_has_rt_tasks(struct task_group *tg)
8219 struct task_struct *g, *p;
8221 do_each_thread(g, p) {
8222 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8224 } while_each_thread(g, p);
8229 struct rt_schedulable_data {
8230 struct task_group *tg;
8235 static int tg_schedulable(struct task_group *tg, void *data)
8237 struct rt_schedulable_data *d = data;
8238 struct task_group *child;
8239 unsigned long total, sum = 0;
8240 u64 period, runtime;
8242 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8243 runtime = tg->rt_bandwidth.rt_runtime;
8246 period = d->rt_period;
8247 runtime = d->rt_runtime;
8251 * Cannot have more runtime than the period.
8253 if (runtime > period && runtime != RUNTIME_INF)
8257 * Ensure we don't starve existing RT tasks.
8259 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8262 total = to_ratio(period, runtime);
8265 * Nobody can have more than the global setting allows.
8267 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8271 * The sum of our children's runtime should not exceed our own.
8273 list_for_each_entry_rcu(child, &tg->children, siblings) {
8274 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8275 runtime = child->rt_bandwidth.rt_runtime;
8277 if (child == d->tg) {
8278 period = d->rt_period;
8279 runtime = d->rt_runtime;
8282 sum += to_ratio(period, runtime);
8291 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8293 struct rt_schedulable_data data = {
8295 .rt_period = period,
8296 .rt_runtime = runtime,
8299 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8302 static int tg_set_bandwidth(struct task_group *tg,
8303 u64 rt_period, u64 rt_runtime)
8307 mutex_lock(&rt_constraints_mutex);
8308 read_lock(&tasklist_lock);
8309 err = __rt_schedulable(tg, rt_period, rt_runtime);
8313 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8314 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8315 tg->rt_bandwidth.rt_runtime = rt_runtime;
8317 for_each_possible_cpu(i) {
8318 struct rt_rq *rt_rq = tg->rt_rq[i];
8320 raw_spin_lock(&rt_rq->rt_runtime_lock);
8321 rt_rq->rt_runtime = rt_runtime;
8322 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8324 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8326 read_unlock(&tasklist_lock);
8327 mutex_unlock(&rt_constraints_mutex);
8332 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8334 u64 rt_runtime, rt_period;
8336 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8337 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8338 if (rt_runtime_us < 0)
8339 rt_runtime = RUNTIME_INF;
8341 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8344 long sched_group_rt_runtime(struct task_group *tg)
8348 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8351 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8352 do_div(rt_runtime_us, NSEC_PER_USEC);
8353 return rt_runtime_us;
8356 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8358 u64 rt_runtime, rt_period;
8360 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8361 rt_runtime = tg->rt_bandwidth.rt_runtime;
8366 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8369 long sched_group_rt_period(struct task_group *tg)
8373 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8374 do_div(rt_period_us, NSEC_PER_USEC);
8375 return rt_period_us;
8378 static int sched_rt_global_constraints(void)
8380 u64 runtime, period;
8383 if (sysctl_sched_rt_period <= 0)
8386 runtime = global_rt_runtime();
8387 period = global_rt_period();
8390 * Sanity check on the sysctl variables.
8392 if (runtime > period && runtime != RUNTIME_INF)
8395 mutex_lock(&rt_constraints_mutex);
8396 read_lock(&tasklist_lock);
8397 ret = __rt_schedulable(NULL, 0, 0);
8398 read_unlock(&tasklist_lock);
8399 mutex_unlock(&rt_constraints_mutex);
8404 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8406 /* Don't accept realtime tasks when there is no way for them to run */
8407 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8413 #else /* !CONFIG_RT_GROUP_SCHED */
8414 static int sched_rt_global_constraints(void)
8416 unsigned long flags;
8419 if (sysctl_sched_rt_period <= 0)
8423 * There's always some RT tasks in the root group
8424 * -- migration, kstopmachine etc..
8426 if (sysctl_sched_rt_runtime == 0)
8429 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8430 for_each_possible_cpu(i) {
8431 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8433 raw_spin_lock(&rt_rq->rt_runtime_lock);
8434 rt_rq->rt_runtime = global_rt_runtime();
8435 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8437 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8441 #endif /* CONFIG_RT_GROUP_SCHED */
8443 int sched_rt_handler(struct ctl_table *table, int write,
8444 void __user *buffer, size_t *lenp,
8448 int old_period, old_runtime;
8449 static DEFINE_MUTEX(mutex);
8452 old_period = sysctl_sched_rt_period;
8453 old_runtime = sysctl_sched_rt_runtime;
8455 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8457 if (!ret && write) {
8458 ret = sched_rt_global_constraints();
8460 sysctl_sched_rt_period = old_period;
8461 sysctl_sched_rt_runtime = old_runtime;
8463 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8464 def_rt_bandwidth.rt_period =
8465 ns_to_ktime(global_rt_period());
8468 mutex_unlock(&mutex);
8473 #ifdef CONFIG_CGROUP_SCHED
8475 /* return corresponding task_group object of a cgroup */
8476 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8478 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8479 struct task_group, css);
8482 static struct cgroup_subsys_state *
8483 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8485 struct task_group *tg, *parent;
8487 if (!cgrp->parent) {
8488 /* This is early initialization for the top cgroup */
8489 return &init_task_group.css;
8492 parent = cgroup_tg(cgrp->parent);
8493 tg = sched_create_group(parent);
8495 return ERR_PTR(-ENOMEM);
8501 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8503 struct task_group *tg = cgroup_tg(cgrp);
8505 sched_destroy_group(tg);
8509 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8511 #ifdef CONFIG_RT_GROUP_SCHED
8512 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
8515 /* We don't support RT-tasks being in separate groups */
8516 if (tsk->sched_class != &fair_sched_class)
8523 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8524 struct task_struct *tsk, bool threadgroup)
8526 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
8530 struct task_struct *c;
8532 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8533 retval = cpu_cgroup_can_attach_task(cgrp, c);
8545 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8546 struct cgroup *old_cont, struct task_struct *tsk,
8549 sched_move_task(tsk);
8551 struct task_struct *c;
8553 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8560 #ifdef CONFIG_FAIR_GROUP_SCHED
8561 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8564 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8567 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8569 struct task_group *tg = cgroup_tg(cgrp);
8571 return (u64) tg->shares;
8573 #endif /* CONFIG_FAIR_GROUP_SCHED */
8575 #ifdef CONFIG_RT_GROUP_SCHED
8576 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8579 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8582 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8584 return sched_group_rt_runtime(cgroup_tg(cgrp));
8587 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8590 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8593 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8595 return sched_group_rt_period(cgroup_tg(cgrp));
8597 #endif /* CONFIG_RT_GROUP_SCHED */
8599 static struct cftype cpu_files[] = {
8600 #ifdef CONFIG_FAIR_GROUP_SCHED
8603 .read_u64 = cpu_shares_read_u64,
8604 .write_u64 = cpu_shares_write_u64,
8607 #ifdef CONFIG_RT_GROUP_SCHED
8609 .name = "rt_runtime_us",
8610 .read_s64 = cpu_rt_runtime_read,
8611 .write_s64 = cpu_rt_runtime_write,
8614 .name = "rt_period_us",
8615 .read_u64 = cpu_rt_period_read_uint,
8616 .write_u64 = cpu_rt_period_write_uint,
8621 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8623 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8626 struct cgroup_subsys cpu_cgroup_subsys = {
8628 .create = cpu_cgroup_create,
8629 .destroy = cpu_cgroup_destroy,
8630 .can_attach = cpu_cgroup_can_attach,
8631 .attach = cpu_cgroup_attach,
8632 .populate = cpu_cgroup_populate,
8633 .subsys_id = cpu_cgroup_subsys_id,
8637 #endif /* CONFIG_CGROUP_SCHED */
8639 #ifdef CONFIG_CGROUP_CPUACCT
8642 * CPU accounting code for task groups.
8644 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8645 * (balbir@in.ibm.com).
8648 /* track cpu usage of a group of tasks and its child groups */
8650 struct cgroup_subsys_state css;
8651 /* cpuusage holds pointer to a u64-type object on every cpu */
8652 u64 __percpu *cpuusage;
8653 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
8654 struct cpuacct *parent;
8657 struct cgroup_subsys cpuacct_subsys;
8659 /* return cpu accounting group corresponding to this container */
8660 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8662 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8663 struct cpuacct, css);
8666 /* return cpu accounting group to which this task belongs */
8667 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8669 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8670 struct cpuacct, css);
8673 /* create a new cpu accounting group */
8674 static struct cgroup_subsys_state *cpuacct_create(
8675 struct cgroup_subsys *ss, struct cgroup *cgrp)
8677 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8683 ca->cpuusage = alloc_percpu(u64);
8687 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8688 if (percpu_counter_init(&ca->cpustat[i], 0))
8689 goto out_free_counters;
8692 ca->parent = cgroup_ca(cgrp->parent);
8698 percpu_counter_destroy(&ca->cpustat[i]);
8699 free_percpu(ca->cpuusage);
8703 return ERR_PTR(-ENOMEM);
8706 /* destroy an existing cpu accounting group */
8708 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8710 struct cpuacct *ca = cgroup_ca(cgrp);
8713 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8714 percpu_counter_destroy(&ca->cpustat[i]);
8715 free_percpu(ca->cpuusage);
8719 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
8721 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8724 #ifndef CONFIG_64BIT
8726 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8728 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8730 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8738 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
8740 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8742 #ifndef CONFIG_64BIT
8744 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8746 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8748 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8754 /* return total cpu usage (in nanoseconds) of a group */
8755 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8757 struct cpuacct *ca = cgroup_ca(cgrp);
8758 u64 totalcpuusage = 0;
8761 for_each_present_cpu(i)
8762 totalcpuusage += cpuacct_cpuusage_read(ca, i);
8764 return totalcpuusage;
8767 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8770 struct cpuacct *ca = cgroup_ca(cgrp);
8779 for_each_present_cpu(i)
8780 cpuacct_cpuusage_write(ca, i, 0);
8786 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
8789 struct cpuacct *ca = cgroup_ca(cgroup);
8793 for_each_present_cpu(i) {
8794 percpu = cpuacct_cpuusage_read(ca, i);
8795 seq_printf(m, "%llu ", (unsigned long long) percpu);
8797 seq_printf(m, "\n");
8801 static const char *cpuacct_stat_desc[] = {
8802 [CPUACCT_STAT_USER] = "user",
8803 [CPUACCT_STAT_SYSTEM] = "system",
8806 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
8807 struct cgroup_map_cb *cb)
8809 struct cpuacct *ca = cgroup_ca(cgrp);
8812 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
8813 s64 val = percpu_counter_read(&ca->cpustat[i]);
8814 val = cputime64_to_clock_t(val);
8815 cb->fill(cb, cpuacct_stat_desc[i], val);
8820 static struct cftype files[] = {
8823 .read_u64 = cpuusage_read,
8824 .write_u64 = cpuusage_write,
8827 .name = "usage_percpu",
8828 .read_seq_string = cpuacct_percpu_seq_read,
8832 .read_map = cpuacct_stats_show,
8836 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
8838 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
8842 * charge this task's execution time to its accounting group.
8844 * called with rq->lock held.
8846 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8851 if (unlikely(!cpuacct_subsys.active))
8854 cpu = task_cpu(tsk);
8860 for (; ca; ca = ca->parent) {
8861 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8862 *cpuusage += cputime;
8869 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
8870 * in cputime_t units. As a result, cpuacct_update_stats calls
8871 * percpu_counter_add with values large enough to always overflow the
8872 * per cpu batch limit causing bad SMP scalability.
8874 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
8875 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
8876 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
8879 #define CPUACCT_BATCH \
8880 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
8882 #define CPUACCT_BATCH 0
8886 * Charge the system/user time to the task's accounting group.
8888 static void cpuacct_update_stats(struct task_struct *tsk,
8889 enum cpuacct_stat_index idx, cputime_t val)
8892 int batch = CPUACCT_BATCH;
8894 if (unlikely(!cpuacct_subsys.active))
8901 __percpu_counter_add(&ca->cpustat[idx], val, batch);
8907 struct cgroup_subsys cpuacct_subsys = {
8909 .create = cpuacct_create,
8910 .destroy = cpuacct_destroy,
8911 .populate = cpuacct_populate,
8912 .subsys_id = cpuacct_subsys_id,
8914 #endif /* CONFIG_CGROUP_CPUACCT */
8918 void synchronize_sched_expedited(void)
8922 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
8924 #else /* #ifndef CONFIG_SMP */
8926 static atomic_t synchronize_sched_expedited_count = ATOMIC_INIT(0);
8928 static int synchronize_sched_expedited_cpu_stop(void *data)
8931 * There must be a full memory barrier on each affected CPU
8932 * between the time that try_stop_cpus() is called and the
8933 * time that it returns.
8935 * In the current initial implementation of cpu_stop, the
8936 * above condition is already met when the control reaches
8937 * this point and the following smp_mb() is not strictly
8938 * necessary. Do smp_mb() anyway for documentation and
8939 * robustness against future implementation changes.
8941 smp_mb(); /* See above comment block. */
8946 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
8947 * approach to force grace period to end quickly. This consumes
8948 * significant time on all CPUs, and is thus not recommended for
8949 * any sort of common-case code.
8951 * Note that it is illegal to call this function while holding any
8952 * lock that is acquired by a CPU-hotplug notifier. Failing to
8953 * observe this restriction will result in deadlock.
8955 void synchronize_sched_expedited(void)
8957 int snap, trycount = 0;
8959 smp_mb(); /* ensure prior mod happens before capturing snap. */
8960 snap = atomic_read(&synchronize_sched_expedited_count) + 1;
8962 while (try_stop_cpus(cpu_online_mask,
8963 synchronize_sched_expedited_cpu_stop,
8966 if (trycount++ < 10)
8967 udelay(trycount * num_online_cpus());
8969 synchronize_sched();
8972 if (atomic_read(&synchronize_sched_expedited_count) - snap > 0) {
8973 smp_mb(); /* ensure test happens before caller kfree */
8978 atomic_inc(&synchronize_sched_expedited_count);
8979 smp_mb__after_atomic_inc(); /* ensure post-GP actions seen after GP. */
8982 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
8984 #endif /* #else #ifndef CONFIG_SMP */