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))
736 if (strncmp(buf, "NO_", 3) == 0) {
741 for (i = 0; sched_feat_names[i]; i++) {
742 int len = strlen(sched_feat_names[i]);
744 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
746 sysctl_sched_features &= ~(1UL << i);
748 sysctl_sched_features |= (1UL << i);
753 if (!sched_feat_names[i])
761 static int sched_feat_open(struct inode *inode, struct file *filp)
763 return single_open(filp, sched_feat_show, NULL);
766 static const struct file_operations sched_feat_fops = {
767 .open = sched_feat_open,
768 .write = sched_feat_write,
771 .release = single_release,
774 static __init int sched_init_debug(void)
776 debugfs_create_file("sched_features", 0644, NULL, NULL,
781 late_initcall(sched_init_debug);
785 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
788 * Number of tasks to iterate in a single balance run.
789 * Limited because this is done with IRQs disabled.
791 const_debug unsigned int sysctl_sched_nr_migrate = 32;
794 * ratelimit for updating the group shares.
797 unsigned int sysctl_sched_shares_ratelimit = 250000;
798 unsigned int normalized_sysctl_sched_shares_ratelimit = 250000;
801 * Inject some fuzzyness into changing the per-cpu group shares
802 * this avoids remote rq-locks at the expense of fairness.
805 unsigned int sysctl_sched_shares_thresh = 4;
808 * period over which we average the RT time consumption, measured
813 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
816 * period over which we measure -rt task cpu usage in us.
819 unsigned int sysctl_sched_rt_period = 1000000;
821 static __read_mostly int scheduler_running;
824 * part of the period that we allow rt tasks to run in us.
827 int sysctl_sched_rt_runtime = 950000;
829 static inline u64 global_rt_period(void)
831 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
834 static inline u64 global_rt_runtime(void)
836 if (sysctl_sched_rt_runtime < 0)
839 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
842 #ifndef prepare_arch_switch
843 # define prepare_arch_switch(next) do { } while (0)
845 #ifndef finish_arch_switch
846 # define finish_arch_switch(prev) do { } while (0)
849 static inline int task_current(struct rq *rq, struct task_struct *p)
851 return rq->curr == p;
854 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
855 static inline int task_running(struct rq *rq, struct task_struct *p)
857 return task_current(rq, p);
860 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
864 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
866 #ifdef CONFIG_DEBUG_SPINLOCK
867 /* this is a valid case when another task releases the spinlock */
868 rq->lock.owner = current;
871 * If we are tracking spinlock dependencies then we have to
872 * fix up the runqueue lock - which gets 'carried over' from
875 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
877 raw_spin_unlock_irq(&rq->lock);
880 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
881 static inline int task_running(struct rq *rq, struct task_struct *p)
886 return task_current(rq, p);
890 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
894 * We can optimise this out completely for !SMP, because the
895 * SMP rebalancing from interrupt is the only thing that cares
900 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
901 raw_spin_unlock_irq(&rq->lock);
903 raw_spin_unlock(&rq->lock);
907 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
911 * After ->oncpu is cleared, the task can be moved to a different CPU.
912 * We must ensure this doesn't happen until the switch is completely
918 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
922 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
925 * Check whether the task is waking, we use this to synchronize ->cpus_allowed
928 static inline int task_is_waking(struct task_struct *p)
930 return unlikely(p->state == TASK_WAKING);
934 * __task_rq_lock - lock the runqueue a given task resides on.
935 * Must be called interrupts disabled.
937 static inline struct rq *__task_rq_lock(struct task_struct *p)
944 raw_spin_lock(&rq->lock);
945 if (likely(rq == task_rq(p)))
947 raw_spin_unlock(&rq->lock);
952 * task_rq_lock - lock the runqueue a given task resides on and disable
953 * interrupts. Note the ordering: we can safely lookup the task_rq without
954 * explicitly disabling preemption.
956 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
962 local_irq_save(*flags);
964 raw_spin_lock(&rq->lock);
965 if (likely(rq == task_rq(p)))
967 raw_spin_unlock_irqrestore(&rq->lock, *flags);
971 static void __task_rq_unlock(struct rq *rq)
974 raw_spin_unlock(&rq->lock);
977 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
980 raw_spin_unlock_irqrestore(&rq->lock, *flags);
984 * this_rq_lock - lock this runqueue and disable interrupts.
986 static struct rq *this_rq_lock(void)
993 raw_spin_lock(&rq->lock);
998 #ifdef CONFIG_SCHED_HRTICK
1000 * Use HR-timers to deliver accurate preemption points.
1002 * Its all a bit involved since we cannot program an hrt while holding the
1003 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1006 * When we get rescheduled we reprogram the hrtick_timer outside of the
1012 * - enabled by features
1013 * - hrtimer is actually high res
1015 static inline int hrtick_enabled(struct rq *rq)
1017 if (!sched_feat(HRTICK))
1019 if (!cpu_active(cpu_of(rq)))
1021 return hrtimer_is_hres_active(&rq->hrtick_timer);
1024 static void hrtick_clear(struct rq *rq)
1026 if (hrtimer_active(&rq->hrtick_timer))
1027 hrtimer_cancel(&rq->hrtick_timer);
1031 * High-resolution timer tick.
1032 * Runs from hardirq context with interrupts disabled.
1034 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1036 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1038 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1040 raw_spin_lock(&rq->lock);
1041 update_rq_clock(rq);
1042 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1043 raw_spin_unlock(&rq->lock);
1045 return HRTIMER_NORESTART;
1050 * called from hardirq (IPI) context
1052 static void __hrtick_start(void *arg)
1054 struct rq *rq = arg;
1056 raw_spin_lock(&rq->lock);
1057 hrtimer_restart(&rq->hrtick_timer);
1058 rq->hrtick_csd_pending = 0;
1059 raw_spin_unlock(&rq->lock);
1063 * Called to set the hrtick timer state.
1065 * called with rq->lock held and irqs disabled
1067 static void hrtick_start(struct rq *rq, u64 delay)
1069 struct hrtimer *timer = &rq->hrtick_timer;
1070 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1072 hrtimer_set_expires(timer, time);
1074 if (rq == this_rq()) {
1075 hrtimer_restart(timer);
1076 } else if (!rq->hrtick_csd_pending) {
1077 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1078 rq->hrtick_csd_pending = 1;
1083 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1085 int cpu = (int)(long)hcpu;
1088 case CPU_UP_CANCELED:
1089 case CPU_UP_CANCELED_FROZEN:
1090 case CPU_DOWN_PREPARE:
1091 case CPU_DOWN_PREPARE_FROZEN:
1093 case CPU_DEAD_FROZEN:
1094 hrtick_clear(cpu_rq(cpu));
1101 static __init void init_hrtick(void)
1103 hotcpu_notifier(hotplug_hrtick, 0);
1107 * Called to set the hrtick timer state.
1109 * called with rq->lock held and irqs disabled
1111 static void hrtick_start(struct rq *rq, u64 delay)
1113 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1114 HRTIMER_MODE_REL_PINNED, 0);
1117 static inline void init_hrtick(void)
1120 #endif /* CONFIG_SMP */
1122 static void init_rq_hrtick(struct rq *rq)
1125 rq->hrtick_csd_pending = 0;
1127 rq->hrtick_csd.flags = 0;
1128 rq->hrtick_csd.func = __hrtick_start;
1129 rq->hrtick_csd.info = rq;
1132 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1133 rq->hrtick_timer.function = hrtick;
1135 #else /* CONFIG_SCHED_HRTICK */
1136 static inline void hrtick_clear(struct rq *rq)
1140 static inline void init_rq_hrtick(struct rq *rq)
1144 static inline void init_hrtick(void)
1147 #endif /* CONFIG_SCHED_HRTICK */
1150 * resched_task - mark a task 'to be rescheduled now'.
1152 * On UP this means the setting of the need_resched flag, on SMP it
1153 * might also involve a cross-CPU call to trigger the scheduler on
1158 #ifndef tsk_is_polling
1159 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1162 static void resched_task(struct task_struct *p)
1166 assert_raw_spin_locked(&task_rq(p)->lock);
1168 if (test_tsk_need_resched(p))
1171 set_tsk_need_resched(p);
1174 if (cpu == smp_processor_id())
1177 /* NEED_RESCHED must be visible before we test polling */
1179 if (!tsk_is_polling(p))
1180 smp_send_reschedule(cpu);
1183 static void resched_cpu(int cpu)
1185 struct rq *rq = cpu_rq(cpu);
1186 unsigned long flags;
1188 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1190 resched_task(cpu_curr(cpu));
1191 raw_spin_unlock_irqrestore(&rq->lock, flags);
1196 * When add_timer_on() enqueues a timer into the timer wheel of an
1197 * idle CPU then this timer might expire before the next timer event
1198 * which is scheduled to wake up that CPU. In case of a completely
1199 * idle system the next event might even be infinite time into the
1200 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1201 * leaves the inner idle loop so the newly added timer is taken into
1202 * account when the CPU goes back to idle and evaluates the timer
1203 * wheel for the next timer event.
1205 void wake_up_idle_cpu(int cpu)
1207 struct rq *rq = cpu_rq(cpu);
1209 if (cpu == smp_processor_id())
1213 * This is safe, as this function is called with the timer
1214 * wheel base lock of (cpu) held. When the CPU is on the way
1215 * to idle and has not yet set rq->curr to idle then it will
1216 * be serialized on the timer wheel base lock and take the new
1217 * timer into account automatically.
1219 if (rq->curr != rq->idle)
1223 * We can set TIF_RESCHED on the idle task of the other CPU
1224 * lockless. The worst case is that the other CPU runs the
1225 * idle task through an additional NOOP schedule()
1227 set_tsk_need_resched(rq->idle);
1229 /* NEED_RESCHED must be visible before we test polling */
1231 if (!tsk_is_polling(rq->idle))
1232 smp_send_reschedule(cpu);
1235 int nohz_ratelimit(int cpu)
1237 struct rq *rq = cpu_rq(cpu);
1238 u64 diff = rq->clock - rq->nohz_stamp;
1240 rq->nohz_stamp = rq->clock;
1242 return diff < (NSEC_PER_SEC / HZ) >> 1;
1245 #endif /* CONFIG_NO_HZ */
1247 static u64 sched_avg_period(void)
1249 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1252 static void sched_avg_update(struct rq *rq)
1254 s64 period = sched_avg_period();
1256 while ((s64)(rq->clock - rq->age_stamp) > period) {
1257 rq->age_stamp += period;
1262 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1264 rq->rt_avg += rt_delta;
1265 sched_avg_update(rq);
1268 #else /* !CONFIG_SMP */
1269 static void resched_task(struct task_struct *p)
1271 assert_raw_spin_locked(&task_rq(p)->lock);
1272 set_tsk_need_resched(p);
1275 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1278 #endif /* CONFIG_SMP */
1280 #if BITS_PER_LONG == 32
1281 # define WMULT_CONST (~0UL)
1283 # define WMULT_CONST (1UL << 32)
1286 #define WMULT_SHIFT 32
1289 * Shift right and round:
1291 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1294 * delta *= weight / lw
1296 static unsigned long
1297 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1298 struct load_weight *lw)
1302 if (!lw->inv_weight) {
1303 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1306 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1310 tmp = (u64)delta_exec * weight;
1312 * Check whether we'd overflow the 64-bit multiplication:
1314 if (unlikely(tmp > WMULT_CONST))
1315 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1318 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1320 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1323 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1329 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1336 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1337 * of tasks with abnormal "nice" values across CPUs the contribution that
1338 * each task makes to its run queue's load is weighted according to its
1339 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1340 * scaled version of the new time slice allocation that they receive on time
1344 #define WEIGHT_IDLEPRIO 3
1345 #define WMULT_IDLEPRIO 1431655765
1348 * Nice levels are multiplicative, with a gentle 10% change for every
1349 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1350 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1351 * that remained on nice 0.
1353 * The "10% effect" is relative and cumulative: from _any_ nice level,
1354 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1355 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1356 * If a task goes up by ~10% and another task goes down by ~10% then
1357 * the relative distance between them is ~25%.)
1359 static const int prio_to_weight[40] = {
1360 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1361 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1362 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1363 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1364 /* 0 */ 1024, 820, 655, 526, 423,
1365 /* 5 */ 335, 272, 215, 172, 137,
1366 /* 10 */ 110, 87, 70, 56, 45,
1367 /* 15 */ 36, 29, 23, 18, 15,
1371 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1373 * In cases where the weight does not change often, we can use the
1374 * precalculated inverse to speed up arithmetics by turning divisions
1375 * into multiplications:
1377 static const u32 prio_to_wmult[40] = {
1378 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1379 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1380 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1381 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1382 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1383 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1384 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1385 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1388 /* Time spent by the tasks of the cpu accounting group executing in ... */
1389 enum cpuacct_stat_index {
1390 CPUACCT_STAT_USER, /* ... user mode */
1391 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1393 CPUACCT_STAT_NSTATS,
1396 #ifdef CONFIG_CGROUP_CPUACCT
1397 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1398 static void cpuacct_update_stats(struct task_struct *tsk,
1399 enum cpuacct_stat_index idx, cputime_t val);
1401 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1402 static inline void cpuacct_update_stats(struct task_struct *tsk,
1403 enum cpuacct_stat_index idx, cputime_t val) {}
1406 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1408 update_load_add(&rq->load, load);
1411 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1413 update_load_sub(&rq->load, load);
1416 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1417 typedef int (*tg_visitor)(struct task_group *, void *);
1420 * Iterate the full tree, calling @down when first entering a node and @up when
1421 * leaving it for the final time.
1423 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1425 struct task_group *parent, *child;
1429 parent = &root_task_group;
1431 ret = (*down)(parent, data);
1434 list_for_each_entry_rcu(child, &parent->children, siblings) {
1441 ret = (*up)(parent, data);
1446 parent = parent->parent;
1455 static int tg_nop(struct task_group *tg, void *data)
1462 /* Used instead of source_load when we know the type == 0 */
1463 static unsigned long weighted_cpuload(const int cpu)
1465 return cpu_rq(cpu)->load.weight;
1469 * Return a low guess at the load of a migration-source cpu weighted
1470 * according to the scheduling class and "nice" value.
1472 * We want to under-estimate the load of migration sources, to
1473 * balance conservatively.
1475 static unsigned long source_load(int cpu, int type)
1477 struct rq *rq = cpu_rq(cpu);
1478 unsigned long total = weighted_cpuload(cpu);
1480 if (type == 0 || !sched_feat(LB_BIAS))
1483 return min(rq->cpu_load[type-1], total);
1487 * Return a high guess at the load of a migration-target cpu weighted
1488 * according to the scheduling class and "nice" value.
1490 static unsigned long target_load(int cpu, int type)
1492 struct rq *rq = cpu_rq(cpu);
1493 unsigned long total = weighted_cpuload(cpu);
1495 if (type == 0 || !sched_feat(LB_BIAS))
1498 return max(rq->cpu_load[type-1], total);
1501 static unsigned long power_of(int cpu)
1503 return cpu_rq(cpu)->cpu_power;
1506 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1508 static unsigned long cpu_avg_load_per_task(int cpu)
1510 struct rq *rq = cpu_rq(cpu);
1511 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1514 rq->avg_load_per_task = rq->load.weight / nr_running;
1516 rq->avg_load_per_task = 0;
1518 return rq->avg_load_per_task;
1521 #ifdef CONFIG_FAIR_GROUP_SCHED
1523 static __read_mostly unsigned long __percpu *update_shares_data;
1525 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1528 * Calculate and set the cpu's group shares.
1530 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1531 unsigned long sd_shares,
1532 unsigned long sd_rq_weight,
1533 unsigned long *usd_rq_weight)
1535 unsigned long shares, rq_weight;
1538 rq_weight = usd_rq_weight[cpu];
1541 rq_weight = NICE_0_LOAD;
1545 * \Sum_j shares_j * rq_weight_i
1546 * shares_i = -----------------------------
1547 * \Sum_j rq_weight_j
1549 shares = (sd_shares * rq_weight) / sd_rq_weight;
1550 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1552 if (abs(shares - tg->se[cpu]->load.weight) >
1553 sysctl_sched_shares_thresh) {
1554 struct rq *rq = cpu_rq(cpu);
1555 unsigned long flags;
1557 raw_spin_lock_irqsave(&rq->lock, flags);
1558 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1559 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1560 __set_se_shares(tg->se[cpu], shares);
1561 raw_spin_unlock_irqrestore(&rq->lock, flags);
1566 * Re-compute the task group their per cpu shares over the given domain.
1567 * This needs to be done in a bottom-up fashion because the rq weight of a
1568 * parent group depends on the shares of its child groups.
1570 static int tg_shares_up(struct task_group *tg, void *data)
1572 unsigned long weight, rq_weight = 0, sum_weight = 0, shares = 0;
1573 unsigned long *usd_rq_weight;
1574 struct sched_domain *sd = data;
1575 unsigned long flags;
1581 local_irq_save(flags);
1582 usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id());
1584 for_each_cpu(i, sched_domain_span(sd)) {
1585 weight = tg->cfs_rq[i]->load.weight;
1586 usd_rq_weight[i] = weight;
1588 rq_weight += weight;
1590 * If there are currently no tasks on the cpu pretend there
1591 * is one of average load so that when a new task gets to
1592 * run here it will not get delayed by group starvation.
1595 weight = NICE_0_LOAD;
1597 sum_weight += weight;
1598 shares += tg->cfs_rq[i]->shares;
1602 rq_weight = sum_weight;
1604 if ((!shares && rq_weight) || shares > tg->shares)
1605 shares = tg->shares;
1607 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1608 shares = tg->shares;
1610 for_each_cpu(i, sched_domain_span(sd))
1611 update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight);
1613 local_irq_restore(flags);
1619 * Compute the cpu's hierarchical load factor for each task group.
1620 * This needs to be done in a top-down fashion because the load of a child
1621 * group is a fraction of its parents load.
1623 static int tg_load_down(struct task_group *tg, void *data)
1626 long cpu = (long)data;
1629 load = cpu_rq(cpu)->load.weight;
1631 load = tg->parent->cfs_rq[cpu]->h_load;
1632 load *= tg->cfs_rq[cpu]->shares;
1633 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1636 tg->cfs_rq[cpu]->h_load = load;
1641 static void update_shares(struct sched_domain *sd)
1646 if (root_task_group_empty())
1649 now = cpu_clock(raw_smp_processor_id());
1650 elapsed = now - sd->last_update;
1652 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1653 sd->last_update = now;
1654 walk_tg_tree(tg_nop, tg_shares_up, sd);
1658 static void update_h_load(long cpu)
1660 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1665 static inline void update_shares(struct sched_domain *sd)
1671 #ifdef CONFIG_PREEMPT
1673 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1676 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1677 * way at the expense of forcing extra atomic operations in all
1678 * invocations. This assures that the double_lock is acquired using the
1679 * same underlying policy as the spinlock_t on this architecture, which
1680 * reduces latency compared to the unfair variant below. However, it
1681 * also adds more overhead and therefore may reduce throughput.
1683 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1684 __releases(this_rq->lock)
1685 __acquires(busiest->lock)
1686 __acquires(this_rq->lock)
1688 raw_spin_unlock(&this_rq->lock);
1689 double_rq_lock(this_rq, busiest);
1696 * Unfair double_lock_balance: Optimizes throughput at the expense of
1697 * latency by eliminating extra atomic operations when the locks are
1698 * already in proper order on entry. This favors lower cpu-ids and will
1699 * grant the double lock to lower cpus over higher ids under contention,
1700 * regardless of entry order into the function.
1702 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1703 __releases(this_rq->lock)
1704 __acquires(busiest->lock)
1705 __acquires(this_rq->lock)
1709 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1710 if (busiest < this_rq) {
1711 raw_spin_unlock(&this_rq->lock);
1712 raw_spin_lock(&busiest->lock);
1713 raw_spin_lock_nested(&this_rq->lock,
1714 SINGLE_DEPTH_NESTING);
1717 raw_spin_lock_nested(&busiest->lock,
1718 SINGLE_DEPTH_NESTING);
1723 #endif /* CONFIG_PREEMPT */
1726 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1728 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1730 if (unlikely(!irqs_disabled())) {
1731 /* printk() doesn't work good under rq->lock */
1732 raw_spin_unlock(&this_rq->lock);
1736 return _double_lock_balance(this_rq, busiest);
1739 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1740 __releases(busiest->lock)
1742 raw_spin_unlock(&busiest->lock);
1743 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1747 * double_rq_lock - safely lock two runqueues
1749 * Note this does not disable interrupts like task_rq_lock,
1750 * you need to do so manually before calling.
1752 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1753 __acquires(rq1->lock)
1754 __acquires(rq2->lock)
1756 BUG_ON(!irqs_disabled());
1758 raw_spin_lock(&rq1->lock);
1759 __acquire(rq2->lock); /* Fake it out ;) */
1762 raw_spin_lock(&rq1->lock);
1763 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1765 raw_spin_lock(&rq2->lock);
1766 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1772 * double_rq_unlock - safely unlock two runqueues
1774 * Note this does not restore interrupts like task_rq_unlock,
1775 * you need to do so manually after calling.
1777 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1778 __releases(rq1->lock)
1779 __releases(rq2->lock)
1781 raw_spin_unlock(&rq1->lock);
1783 raw_spin_unlock(&rq2->lock);
1785 __release(rq2->lock);
1790 #ifdef CONFIG_FAIR_GROUP_SCHED
1791 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1794 cfs_rq->shares = shares;
1799 static void calc_load_account_idle(struct rq *this_rq);
1800 static void update_sysctl(void);
1801 static int get_update_sysctl_factor(void);
1803 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1805 set_task_rq(p, cpu);
1808 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1809 * successfuly executed on another CPU. We must ensure that updates of
1810 * per-task data have been completed by this moment.
1813 task_thread_info(p)->cpu = cpu;
1817 static const struct sched_class rt_sched_class;
1819 #define sched_class_highest (&rt_sched_class)
1820 #define for_each_class(class) \
1821 for (class = sched_class_highest; class; class = class->next)
1823 #include "sched_stats.h"
1825 static void inc_nr_running(struct rq *rq)
1830 static void dec_nr_running(struct rq *rq)
1835 static void set_load_weight(struct task_struct *p)
1837 if (task_has_rt_policy(p)) {
1838 p->se.load.weight = 0;
1839 p->se.load.inv_weight = WMULT_CONST;
1844 * SCHED_IDLE tasks get minimal weight:
1846 if (p->policy == SCHED_IDLE) {
1847 p->se.load.weight = WEIGHT_IDLEPRIO;
1848 p->se.load.inv_weight = WMULT_IDLEPRIO;
1852 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1853 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1856 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1858 update_rq_clock(rq);
1859 sched_info_queued(p);
1860 p->sched_class->enqueue_task(rq, p, flags);
1864 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1866 update_rq_clock(rq);
1867 sched_info_dequeued(p);
1868 p->sched_class->dequeue_task(rq, p, flags);
1873 * activate_task - move a task to the runqueue.
1875 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1877 if (task_contributes_to_load(p))
1878 rq->nr_uninterruptible--;
1880 enqueue_task(rq, p, flags);
1885 * deactivate_task - remove a task from the runqueue.
1887 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1889 if (task_contributes_to_load(p))
1890 rq->nr_uninterruptible++;
1892 dequeue_task(rq, p, flags);
1896 #include "sched_idletask.c"
1897 #include "sched_fair.c"
1898 #include "sched_rt.c"
1899 #ifdef CONFIG_SCHED_DEBUG
1900 # include "sched_debug.c"
1904 * __normal_prio - return the priority that is based on the static prio
1906 static inline int __normal_prio(struct task_struct *p)
1908 return p->static_prio;
1912 * Calculate the expected normal priority: i.e. priority
1913 * without taking RT-inheritance into account. Might be
1914 * boosted by interactivity modifiers. Changes upon fork,
1915 * setprio syscalls, and whenever the interactivity
1916 * estimator recalculates.
1918 static inline int normal_prio(struct task_struct *p)
1922 if (task_has_rt_policy(p))
1923 prio = MAX_RT_PRIO-1 - p->rt_priority;
1925 prio = __normal_prio(p);
1930 * Calculate the current priority, i.e. the priority
1931 * taken into account by the scheduler. This value might
1932 * be boosted by RT tasks, or might be boosted by
1933 * interactivity modifiers. Will be RT if the task got
1934 * RT-boosted. If not then it returns p->normal_prio.
1936 static int effective_prio(struct task_struct *p)
1938 p->normal_prio = normal_prio(p);
1940 * If we are RT tasks or we were boosted to RT priority,
1941 * keep the priority unchanged. Otherwise, update priority
1942 * to the normal priority:
1944 if (!rt_prio(p->prio))
1945 return p->normal_prio;
1950 * task_curr - is this task currently executing on a CPU?
1951 * @p: the task in question.
1953 inline int task_curr(const struct task_struct *p)
1955 return cpu_curr(task_cpu(p)) == p;
1958 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1959 const struct sched_class *prev_class,
1960 int oldprio, int running)
1962 if (prev_class != p->sched_class) {
1963 if (prev_class->switched_from)
1964 prev_class->switched_from(rq, p, running);
1965 p->sched_class->switched_to(rq, p, running);
1967 p->sched_class->prio_changed(rq, p, oldprio, running);
1972 * Is this task likely cache-hot:
1975 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1979 if (p->sched_class != &fair_sched_class)
1983 * Buddy candidates are cache hot:
1985 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
1986 (&p->se == cfs_rq_of(&p->se)->next ||
1987 &p->se == cfs_rq_of(&p->se)->last))
1990 if (sysctl_sched_migration_cost == -1)
1992 if (sysctl_sched_migration_cost == 0)
1995 delta = now - p->se.exec_start;
1997 return delta < (s64)sysctl_sched_migration_cost;
2000 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2002 #ifdef CONFIG_SCHED_DEBUG
2004 * We should never call set_task_cpu() on a blocked task,
2005 * ttwu() will sort out the placement.
2007 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2008 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2011 trace_sched_migrate_task(p, new_cpu);
2013 if (task_cpu(p) != new_cpu) {
2014 p->se.nr_migrations++;
2015 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2018 __set_task_cpu(p, new_cpu);
2021 struct migration_arg {
2022 struct task_struct *task;
2026 static int migration_cpu_stop(void *data);
2029 * The task's runqueue lock must be held.
2030 * Returns true if you have to wait for migration thread.
2032 static bool migrate_task(struct task_struct *p, int dest_cpu)
2034 struct rq *rq = task_rq(p);
2037 * If the task is not on a runqueue (and not running), then
2038 * the next wake-up will properly place the task.
2040 return p->se.on_rq || task_running(rq, p);
2044 * wait_task_inactive - wait for a thread to unschedule.
2046 * If @match_state is nonzero, it's the @p->state value just checked and
2047 * not expected to change. If it changes, i.e. @p might have woken up,
2048 * then return zero. When we succeed in waiting for @p to be off its CPU,
2049 * we return a positive number (its total switch count). If a second call
2050 * a short while later returns the same number, the caller can be sure that
2051 * @p has remained unscheduled the whole time.
2053 * The caller must ensure that the task *will* unschedule sometime soon,
2054 * else this function might spin for a *long* time. This function can't
2055 * be called with interrupts off, or it may introduce deadlock with
2056 * smp_call_function() if an IPI is sent by the same process we are
2057 * waiting to become inactive.
2059 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2061 unsigned long flags;
2068 * We do the initial early heuristics without holding
2069 * any task-queue locks at all. We'll only try to get
2070 * the runqueue lock when things look like they will
2076 * If the task is actively running on another CPU
2077 * still, just relax and busy-wait without holding
2080 * NOTE! Since we don't hold any locks, it's not
2081 * even sure that "rq" stays as the right runqueue!
2082 * But we don't care, since "task_running()" will
2083 * return false if the runqueue has changed and p
2084 * is actually now running somewhere else!
2086 while (task_running(rq, p)) {
2087 if (match_state && unlikely(p->state != match_state))
2093 * Ok, time to look more closely! We need the rq
2094 * lock now, to be *sure*. If we're wrong, we'll
2095 * just go back and repeat.
2097 rq = task_rq_lock(p, &flags);
2098 trace_sched_wait_task(p);
2099 running = task_running(rq, p);
2100 on_rq = p->se.on_rq;
2102 if (!match_state || p->state == match_state)
2103 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2104 task_rq_unlock(rq, &flags);
2107 * If it changed from the expected state, bail out now.
2109 if (unlikely(!ncsw))
2113 * Was it really running after all now that we
2114 * checked with the proper locks actually held?
2116 * Oops. Go back and try again..
2118 if (unlikely(running)) {
2124 * It's not enough that it's not actively running,
2125 * it must be off the runqueue _entirely_, and not
2128 * So if it was still runnable (but just not actively
2129 * running right now), it's preempted, and we should
2130 * yield - it could be a while.
2132 if (unlikely(on_rq)) {
2133 schedule_timeout_uninterruptible(1);
2138 * Ahh, all good. It wasn't running, and it wasn't
2139 * runnable, which means that it will never become
2140 * running in the future either. We're all done!
2149 * kick_process - kick a running thread to enter/exit the kernel
2150 * @p: the to-be-kicked thread
2152 * Cause a process which is running on another CPU to enter
2153 * kernel-mode, without any delay. (to get signals handled.)
2155 * NOTE: this function doesnt have to take the runqueue lock,
2156 * because all it wants to ensure is that the remote task enters
2157 * the kernel. If the IPI races and the task has been migrated
2158 * to another CPU then no harm is done and the purpose has been
2161 void kick_process(struct task_struct *p)
2167 if ((cpu != smp_processor_id()) && task_curr(p))
2168 smp_send_reschedule(cpu);
2171 EXPORT_SYMBOL_GPL(kick_process);
2172 #endif /* CONFIG_SMP */
2175 * task_oncpu_function_call - call a function on the cpu on which a task runs
2176 * @p: the task to evaluate
2177 * @func: the function to be called
2178 * @info: the function call argument
2180 * Calls the function @func when the task is currently running. This might
2181 * be on the current CPU, which just calls the function directly
2183 void task_oncpu_function_call(struct task_struct *p,
2184 void (*func) (void *info), void *info)
2191 smp_call_function_single(cpu, func, info, 1);
2197 * ->cpus_allowed is protected by either TASK_WAKING or rq->lock held.
2199 static int select_fallback_rq(int cpu, struct task_struct *p)
2202 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2204 /* Look for allowed, online CPU in same node. */
2205 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2206 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2209 /* Any allowed, online CPU? */
2210 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2211 if (dest_cpu < nr_cpu_ids)
2214 /* No more Mr. Nice Guy. */
2215 if (unlikely(dest_cpu >= nr_cpu_ids)) {
2216 dest_cpu = cpuset_cpus_allowed_fallback(p);
2218 * Don't tell them about moving exiting tasks or
2219 * kernel threads (both mm NULL), since they never
2222 if (p->mm && printk_ratelimit()) {
2223 printk(KERN_INFO "process %d (%s) no "
2224 "longer affine to cpu%d\n",
2225 task_pid_nr(p), p->comm, cpu);
2233 * The caller (fork, wakeup) owns TASK_WAKING, ->cpus_allowed is stable.
2236 int select_task_rq(struct rq *rq, struct task_struct *p, int sd_flags, int wake_flags)
2238 int cpu = p->sched_class->select_task_rq(rq, p, sd_flags, wake_flags);
2241 * In order not to call set_task_cpu() on a blocking task we need
2242 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2245 * Since this is common to all placement strategies, this lives here.
2247 * [ this allows ->select_task() to simply return task_cpu(p) and
2248 * not worry about this generic constraint ]
2250 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2252 cpu = select_fallback_rq(task_cpu(p), p);
2257 static void update_avg(u64 *avg, u64 sample)
2259 s64 diff = sample - *avg;
2265 * try_to_wake_up - wake up a thread
2266 * @p: the to-be-woken-up thread
2267 * @state: the mask of task states that can be woken
2268 * @sync: do a synchronous wakeup?
2270 * Put it on the run-queue if it's not already there. The "current"
2271 * thread is always on the run-queue (except when the actual
2272 * re-schedule is in progress), and as such you're allowed to do
2273 * the simpler "current->state = TASK_RUNNING" to mark yourself
2274 * runnable without the overhead of this.
2276 * returns failure only if the task is already active.
2278 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2281 int cpu, orig_cpu, this_cpu, success = 0;
2282 unsigned long flags;
2283 unsigned long en_flags = ENQUEUE_WAKEUP;
2286 this_cpu = get_cpu();
2289 rq = task_rq_lock(p, &flags);
2290 if (!(p->state & state))
2300 if (unlikely(task_running(rq, p)))
2304 * In order to handle concurrent wakeups and release the rq->lock
2305 * we put the task in TASK_WAKING state.
2307 * First fix up the nr_uninterruptible count:
2309 if (task_contributes_to_load(p)) {
2310 if (likely(cpu_online(orig_cpu)))
2311 rq->nr_uninterruptible--;
2313 this_rq()->nr_uninterruptible--;
2315 p->state = TASK_WAKING;
2317 if (p->sched_class->task_waking) {
2318 p->sched_class->task_waking(rq, p);
2319 en_flags |= ENQUEUE_WAKING;
2322 cpu = select_task_rq(rq, p, SD_BALANCE_WAKE, wake_flags);
2323 if (cpu != orig_cpu)
2324 set_task_cpu(p, cpu);
2325 __task_rq_unlock(rq);
2328 raw_spin_lock(&rq->lock);
2331 * We migrated the task without holding either rq->lock, however
2332 * since the task is not on the task list itself, nobody else
2333 * will try and migrate the task, hence the rq should match the
2334 * cpu we just moved it to.
2336 WARN_ON(task_cpu(p) != cpu);
2337 WARN_ON(p->state != TASK_WAKING);
2339 #ifdef CONFIG_SCHEDSTATS
2340 schedstat_inc(rq, ttwu_count);
2341 if (cpu == this_cpu)
2342 schedstat_inc(rq, ttwu_local);
2344 struct sched_domain *sd;
2345 for_each_domain(this_cpu, sd) {
2346 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2347 schedstat_inc(sd, ttwu_wake_remote);
2352 #endif /* CONFIG_SCHEDSTATS */
2355 #endif /* CONFIG_SMP */
2356 schedstat_inc(p, se.statistics.nr_wakeups);
2357 if (wake_flags & WF_SYNC)
2358 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2359 if (orig_cpu != cpu)
2360 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2361 if (cpu == this_cpu)
2362 schedstat_inc(p, se.statistics.nr_wakeups_local);
2364 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2365 activate_task(rq, p, en_flags);
2369 trace_sched_wakeup(p, success);
2370 check_preempt_curr(rq, p, wake_flags);
2372 p->state = TASK_RUNNING;
2374 if (p->sched_class->task_woken)
2375 p->sched_class->task_woken(rq, p);
2377 if (unlikely(rq->idle_stamp)) {
2378 u64 delta = rq->clock - rq->idle_stamp;
2379 u64 max = 2*sysctl_sched_migration_cost;
2384 update_avg(&rq->avg_idle, delta);
2389 task_rq_unlock(rq, &flags);
2396 * wake_up_process - Wake up a specific process
2397 * @p: The process to be woken up.
2399 * Attempt to wake up the nominated process and move it to the set of runnable
2400 * processes. Returns 1 if the process was woken up, 0 if it was already
2403 * It may be assumed that this function implies a write memory barrier before
2404 * changing the task state if and only if any tasks are woken up.
2406 int wake_up_process(struct task_struct *p)
2408 return try_to_wake_up(p, TASK_ALL, 0);
2410 EXPORT_SYMBOL(wake_up_process);
2412 int wake_up_state(struct task_struct *p, unsigned int state)
2414 return try_to_wake_up(p, state, 0);
2418 * Perform scheduler related setup for a newly forked process p.
2419 * p is forked by current.
2421 * __sched_fork() is basic setup used by init_idle() too:
2423 static void __sched_fork(struct task_struct *p)
2425 p->se.exec_start = 0;
2426 p->se.sum_exec_runtime = 0;
2427 p->se.prev_sum_exec_runtime = 0;
2428 p->se.nr_migrations = 0;
2430 #ifdef CONFIG_SCHEDSTATS
2431 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2434 INIT_LIST_HEAD(&p->rt.run_list);
2436 INIT_LIST_HEAD(&p->se.group_node);
2438 #ifdef CONFIG_PREEMPT_NOTIFIERS
2439 INIT_HLIST_HEAD(&p->preempt_notifiers);
2444 * fork()/clone()-time setup:
2446 void sched_fork(struct task_struct *p, int clone_flags)
2448 int cpu = get_cpu();
2452 * We mark the process as running here. This guarantees that
2453 * nobody will actually run it, and a signal or other external
2454 * event cannot wake it up and insert it on the runqueue either.
2456 p->state = TASK_RUNNING;
2459 * Revert to default priority/policy on fork if requested.
2461 if (unlikely(p->sched_reset_on_fork)) {
2462 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2463 p->policy = SCHED_NORMAL;
2464 p->normal_prio = p->static_prio;
2467 if (PRIO_TO_NICE(p->static_prio) < 0) {
2468 p->static_prio = NICE_TO_PRIO(0);
2469 p->normal_prio = p->static_prio;
2474 * We don't need the reset flag anymore after the fork. It has
2475 * fulfilled its duty:
2477 p->sched_reset_on_fork = 0;
2481 * Make sure we do not leak PI boosting priority to the child.
2483 p->prio = current->normal_prio;
2485 if (!rt_prio(p->prio))
2486 p->sched_class = &fair_sched_class;
2488 if (p->sched_class->task_fork)
2489 p->sched_class->task_fork(p);
2491 set_task_cpu(p, cpu);
2493 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2494 if (likely(sched_info_on()))
2495 memset(&p->sched_info, 0, sizeof(p->sched_info));
2497 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2500 #ifdef CONFIG_PREEMPT
2501 /* Want to start with kernel preemption disabled. */
2502 task_thread_info(p)->preempt_count = 1;
2504 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2510 * wake_up_new_task - wake up a newly created task for the first time.
2512 * This function will do some initial scheduler statistics housekeeping
2513 * that must be done for every newly created context, then puts the task
2514 * on the runqueue and wakes it.
2516 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2518 unsigned long flags;
2520 int cpu __maybe_unused = get_cpu();
2523 rq = task_rq_lock(p, &flags);
2524 p->state = TASK_WAKING;
2527 * Fork balancing, do it here and not earlier because:
2528 * - cpus_allowed can change in the fork path
2529 * - any previously selected cpu might disappear through hotplug
2531 * We set TASK_WAKING so that select_task_rq() can drop rq->lock
2532 * without people poking at ->cpus_allowed.
2534 cpu = select_task_rq(rq, p, SD_BALANCE_FORK, 0);
2535 set_task_cpu(p, cpu);
2537 p->state = TASK_RUNNING;
2538 task_rq_unlock(rq, &flags);
2541 rq = task_rq_lock(p, &flags);
2542 activate_task(rq, p, 0);
2543 trace_sched_wakeup_new(p, 1);
2544 check_preempt_curr(rq, p, WF_FORK);
2546 if (p->sched_class->task_woken)
2547 p->sched_class->task_woken(rq, p);
2549 task_rq_unlock(rq, &flags);
2553 #ifdef CONFIG_PREEMPT_NOTIFIERS
2556 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2557 * @notifier: notifier struct to register
2559 void preempt_notifier_register(struct preempt_notifier *notifier)
2561 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2563 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2566 * preempt_notifier_unregister - no longer interested in preemption notifications
2567 * @notifier: notifier struct to unregister
2569 * This is safe to call from within a preemption notifier.
2571 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2573 hlist_del(¬ifier->link);
2575 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2577 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2579 struct preempt_notifier *notifier;
2580 struct hlist_node *node;
2582 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2583 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2587 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2588 struct task_struct *next)
2590 struct preempt_notifier *notifier;
2591 struct hlist_node *node;
2593 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2594 notifier->ops->sched_out(notifier, next);
2597 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2599 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2604 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2605 struct task_struct *next)
2609 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2612 * prepare_task_switch - prepare to switch tasks
2613 * @rq: the runqueue preparing to switch
2614 * @prev: the current task that is being switched out
2615 * @next: the task we are going to switch to.
2617 * This is called with the rq lock held and interrupts off. It must
2618 * be paired with a subsequent finish_task_switch after the context
2621 * prepare_task_switch sets up locking and calls architecture specific
2625 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2626 struct task_struct *next)
2628 fire_sched_out_preempt_notifiers(prev, next);
2629 prepare_lock_switch(rq, next);
2630 prepare_arch_switch(next);
2634 * finish_task_switch - clean up after a task-switch
2635 * @rq: runqueue associated with task-switch
2636 * @prev: the thread we just switched away from.
2638 * finish_task_switch must be called after the context switch, paired
2639 * with a prepare_task_switch call before the context switch.
2640 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2641 * and do any other architecture-specific cleanup actions.
2643 * Note that we may have delayed dropping an mm in context_switch(). If
2644 * so, we finish that here outside of the runqueue lock. (Doing it
2645 * with the lock held can cause deadlocks; see schedule() for
2648 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2649 __releases(rq->lock)
2651 struct mm_struct *mm = rq->prev_mm;
2657 * A task struct has one reference for the use as "current".
2658 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2659 * schedule one last time. The schedule call will never return, and
2660 * the scheduled task must drop that reference.
2661 * The test for TASK_DEAD must occur while the runqueue locks are
2662 * still held, otherwise prev could be scheduled on another cpu, die
2663 * there before we look at prev->state, and then the reference would
2665 * Manfred Spraul <manfred@colorfullife.com>
2667 prev_state = prev->state;
2668 finish_arch_switch(prev);
2669 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2670 local_irq_disable();
2671 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2672 perf_event_task_sched_in(current);
2673 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2675 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2676 finish_lock_switch(rq, prev);
2678 fire_sched_in_preempt_notifiers(current);
2681 if (unlikely(prev_state == TASK_DEAD)) {
2683 * Remove function-return probe instances associated with this
2684 * task and put them back on the free list.
2686 kprobe_flush_task(prev);
2687 put_task_struct(prev);
2693 /* assumes rq->lock is held */
2694 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2696 if (prev->sched_class->pre_schedule)
2697 prev->sched_class->pre_schedule(rq, prev);
2700 /* rq->lock is NOT held, but preemption is disabled */
2701 static inline void post_schedule(struct rq *rq)
2703 if (rq->post_schedule) {
2704 unsigned long flags;
2706 raw_spin_lock_irqsave(&rq->lock, flags);
2707 if (rq->curr->sched_class->post_schedule)
2708 rq->curr->sched_class->post_schedule(rq);
2709 raw_spin_unlock_irqrestore(&rq->lock, flags);
2711 rq->post_schedule = 0;
2717 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2721 static inline void post_schedule(struct rq *rq)
2728 * schedule_tail - first thing a freshly forked thread must call.
2729 * @prev: the thread we just switched away from.
2731 asmlinkage void schedule_tail(struct task_struct *prev)
2732 __releases(rq->lock)
2734 struct rq *rq = this_rq();
2736 finish_task_switch(rq, prev);
2739 * FIXME: do we need to worry about rq being invalidated by the
2744 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2745 /* In this case, finish_task_switch does not reenable preemption */
2748 if (current->set_child_tid)
2749 put_user(task_pid_vnr(current), current->set_child_tid);
2753 * context_switch - switch to the new MM and the new
2754 * thread's register state.
2757 context_switch(struct rq *rq, struct task_struct *prev,
2758 struct task_struct *next)
2760 struct mm_struct *mm, *oldmm;
2762 prepare_task_switch(rq, prev, next);
2763 trace_sched_switch(prev, next);
2765 oldmm = prev->active_mm;
2767 * For paravirt, this is coupled with an exit in switch_to to
2768 * combine the page table reload and the switch backend into
2771 arch_start_context_switch(prev);
2774 next->active_mm = oldmm;
2775 atomic_inc(&oldmm->mm_count);
2776 enter_lazy_tlb(oldmm, next);
2778 switch_mm(oldmm, mm, next);
2780 if (likely(!prev->mm)) {
2781 prev->active_mm = NULL;
2782 rq->prev_mm = oldmm;
2785 * Since the runqueue lock will be released by the next
2786 * task (which is an invalid locking op but in the case
2787 * of the scheduler it's an obvious special-case), so we
2788 * do an early lockdep release here:
2790 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2791 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2794 /* Here we just switch the register state and the stack. */
2795 switch_to(prev, next, prev);
2799 * this_rq must be evaluated again because prev may have moved
2800 * CPUs since it called schedule(), thus the 'rq' on its stack
2801 * frame will be invalid.
2803 finish_task_switch(this_rq(), prev);
2807 * nr_running, nr_uninterruptible and nr_context_switches:
2809 * externally visible scheduler statistics: current number of runnable
2810 * threads, current number of uninterruptible-sleeping threads, total
2811 * number of context switches performed since bootup.
2813 unsigned long nr_running(void)
2815 unsigned long i, sum = 0;
2817 for_each_online_cpu(i)
2818 sum += cpu_rq(i)->nr_running;
2823 unsigned long nr_uninterruptible(void)
2825 unsigned long i, sum = 0;
2827 for_each_possible_cpu(i)
2828 sum += cpu_rq(i)->nr_uninterruptible;
2831 * Since we read the counters lockless, it might be slightly
2832 * inaccurate. Do not allow it to go below zero though:
2834 if (unlikely((long)sum < 0))
2840 unsigned long long nr_context_switches(void)
2843 unsigned long long sum = 0;
2845 for_each_possible_cpu(i)
2846 sum += cpu_rq(i)->nr_switches;
2851 unsigned long nr_iowait(void)
2853 unsigned long i, sum = 0;
2855 for_each_possible_cpu(i)
2856 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2861 unsigned long nr_iowait_cpu(void)
2863 struct rq *this = this_rq();
2864 return atomic_read(&this->nr_iowait);
2867 unsigned long this_cpu_load(void)
2869 struct rq *this = this_rq();
2870 return this->cpu_load[0];
2874 /* Variables and functions for calc_load */
2875 static atomic_long_t calc_load_tasks;
2876 static unsigned long calc_load_update;
2877 unsigned long avenrun[3];
2878 EXPORT_SYMBOL(avenrun);
2880 static long calc_load_fold_active(struct rq *this_rq)
2882 long nr_active, delta = 0;
2884 nr_active = this_rq->nr_running;
2885 nr_active += (long) this_rq->nr_uninterruptible;
2887 if (nr_active != this_rq->calc_load_active) {
2888 delta = nr_active - this_rq->calc_load_active;
2889 this_rq->calc_load_active = nr_active;
2897 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
2899 * When making the ILB scale, we should try to pull this in as well.
2901 static atomic_long_t calc_load_tasks_idle;
2903 static void calc_load_account_idle(struct rq *this_rq)
2907 delta = calc_load_fold_active(this_rq);
2909 atomic_long_add(delta, &calc_load_tasks_idle);
2912 static long calc_load_fold_idle(void)
2917 * Its got a race, we don't care...
2919 if (atomic_long_read(&calc_load_tasks_idle))
2920 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
2925 static void calc_load_account_idle(struct rq *this_rq)
2929 static inline long calc_load_fold_idle(void)
2936 * get_avenrun - get the load average array
2937 * @loads: pointer to dest load array
2938 * @offset: offset to add
2939 * @shift: shift count to shift the result left
2941 * These values are estimates at best, so no need for locking.
2943 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2945 loads[0] = (avenrun[0] + offset) << shift;
2946 loads[1] = (avenrun[1] + offset) << shift;
2947 loads[2] = (avenrun[2] + offset) << shift;
2950 static unsigned long
2951 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2954 load += active * (FIXED_1 - exp);
2955 return load >> FSHIFT;
2959 * calc_load - update the avenrun load estimates 10 ticks after the
2960 * CPUs have updated calc_load_tasks.
2962 void calc_global_load(void)
2964 unsigned long upd = calc_load_update + 10;
2967 if (time_before(jiffies, upd))
2970 active = atomic_long_read(&calc_load_tasks);
2971 active = active > 0 ? active * FIXED_1 : 0;
2973 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2974 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2975 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2977 calc_load_update += LOAD_FREQ;
2981 * Called from update_cpu_load() to periodically update this CPU's
2984 static void calc_load_account_active(struct rq *this_rq)
2988 if (time_before(jiffies, this_rq->calc_load_update))
2991 delta = calc_load_fold_active(this_rq);
2992 delta += calc_load_fold_idle();
2994 atomic_long_add(delta, &calc_load_tasks);
2996 this_rq->calc_load_update += LOAD_FREQ;
3000 * Update rq->cpu_load[] statistics. This function is usually called every
3001 * scheduler tick (TICK_NSEC).
3003 static void update_cpu_load(struct rq *this_rq)
3005 unsigned long this_load = this_rq->load.weight;
3008 this_rq->nr_load_updates++;
3010 /* Update our load: */
3011 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3012 unsigned long old_load, new_load;
3014 /* scale is effectively 1 << i now, and >> i divides by scale */
3016 old_load = this_rq->cpu_load[i];
3017 new_load = this_load;
3019 * Round up the averaging division if load is increasing. This
3020 * prevents us from getting stuck on 9 if the load is 10, for
3023 if (new_load > old_load)
3024 new_load += scale-1;
3025 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3028 calc_load_account_active(this_rq);
3034 * sched_exec - execve() is a valuable balancing opportunity, because at
3035 * this point the task has the smallest effective memory and cache footprint.
3037 void sched_exec(void)
3039 struct task_struct *p = current;
3040 unsigned long flags;
3044 rq = task_rq_lock(p, &flags);
3045 dest_cpu = p->sched_class->select_task_rq(rq, p, SD_BALANCE_EXEC, 0);
3046 if (dest_cpu == smp_processor_id())
3050 * select_task_rq() can race against ->cpus_allowed
3052 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed) &&
3053 likely(cpu_active(dest_cpu)) && migrate_task(p, dest_cpu)) {
3054 struct migration_arg arg = { p, dest_cpu };
3056 task_rq_unlock(rq, &flags);
3057 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
3061 task_rq_unlock(rq, &flags);
3066 DEFINE_PER_CPU(struct kernel_stat, kstat);
3068 EXPORT_PER_CPU_SYMBOL(kstat);
3071 * Return any ns on the sched_clock that have not yet been accounted in
3072 * @p in case that task is currently running.
3074 * Called with task_rq_lock() held on @rq.
3076 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3080 if (task_current(rq, p)) {
3081 update_rq_clock(rq);
3082 ns = rq->clock - p->se.exec_start;
3090 unsigned long long task_delta_exec(struct task_struct *p)
3092 unsigned long flags;
3096 rq = task_rq_lock(p, &flags);
3097 ns = do_task_delta_exec(p, rq);
3098 task_rq_unlock(rq, &flags);
3104 * Return accounted runtime for the task.
3105 * In case the task is currently running, return the runtime plus current's
3106 * pending runtime that have not been accounted yet.
3108 unsigned long long task_sched_runtime(struct task_struct *p)
3110 unsigned long flags;
3114 rq = task_rq_lock(p, &flags);
3115 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3116 task_rq_unlock(rq, &flags);
3122 * Return sum_exec_runtime for the thread group.
3123 * In case the task is currently running, return the sum plus current's
3124 * pending runtime that have not been accounted yet.
3126 * Note that the thread group might have other running tasks as well,
3127 * so the return value not includes other pending runtime that other
3128 * running tasks might have.
3130 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3132 struct task_cputime totals;
3133 unsigned long flags;
3137 rq = task_rq_lock(p, &flags);
3138 thread_group_cputime(p, &totals);
3139 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3140 task_rq_unlock(rq, &flags);
3146 * Account user cpu time to a process.
3147 * @p: the process that the cpu time gets accounted to
3148 * @cputime: the cpu time spent in user space since the last update
3149 * @cputime_scaled: cputime scaled by cpu frequency
3151 void account_user_time(struct task_struct *p, cputime_t cputime,
3152 cputime_t cputime_scaled)
3154 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3157 /* Add user time to process. */
3158 p->utime = cputime_add(p->utime, cputime);
3159 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3160 account_group_user_time(p, cputime);
3162 /* Add user time to cpustat. */
3163 tmp = cputime_to_cputime64(cputime);
3164 if (TASK_NICE(p) > 0)
3165 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3167 cpustat->user = cputime64_add(cpustat->user, tmp);
3169 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3170 /* Account for user time used */
3171 acct_update_integrals(p);
3175 * Account guest cpu time to a process.
3176 * @p: the process that the cpu time gets accounted to
3177 * @cputime: the cpu time spent in virtual machine since the last update
3178 * @cputime_scaled: cputime scaled by cpu frequency
3180 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3181 cputime_t cputime_scaled)
3184 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3186 tmp = cputime_to_cputime64(cputime);
3188 /* Add guest time to process. */
3189 p->utime = cputime_add(p->utime, cputime);
3190 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3191 account_group_user_time(p, cputime);
3192 p->gtime = cputime_add(p->gtime, cputime);
3194 /* Add guest time to cpustat. */
3195 if (TASK_NICE(p) > 0) {
3196 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3197 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3199 cpustat->user = cputime64_add(cpustat->user, tmp);
3200 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3205 * Account system cpu time to a process.
3206 * @p: the process that the cpu time gets accounted to
3207 * @hardirq_offset: the offset to subtract from hardirq_count()
3208 * @cputime: the cpu time spent in kernel space since the last update
3209 * @cputime_scaled: cputime scaled by cpu frequency
3211 void account_system_time(struct task_struct *p, int hardirq_offset,
3212 cputime_t cputime, cputime_t cputime_scaled)
3214 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3217 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3218 account_guest_time(p, cputime, cputime_scaled);
3222 /* Add system time to process. */
3223 p->stime = cputime_add(p->stime, cputime);
3224 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3225 account_group_system_time(p, cputime);
3227 /* Add system time to cpustat. */
3228 tmp = cputime_to_cputime64(cputime);
3229 if (hardirq_count() - hardirq_offset)
3230 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3231 else if (softirq_count())
3232 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3234 cpustat->system = cputime64_add(cpustat->system, tmp);
3236 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3238 /* Account for system time used */
3239 acct_update_integrals(p);
3243 * Account for involuntary wait time.
3244 * @steal: the cpu time spent in involuntary wait
3246 void account_steal_time(cputime_t cputime)
3248 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3249 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3251 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3255 * Account for idle time.
3256 * @cputime: the cpu time spent in idle wait
3258 void account_idle_time(cputime_t cputime)
3260 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3261 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3262 struct rq *rq = this_rq();
3264 if (atomic_read(&rq->nr_iowait) > 0)
3265 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3267 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3270 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3273 * Account a single tick of cpu time.
3274 * @p: the process that the cpu time gets accounted to
3275 * @user_tick: indicates if the tick is a user or a system tick
3277 void account_process_tick(struct task_struct *p, int user_tick)
3279 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3280 struct rq *rq = this_rq();
3283 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3284 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3285 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3288 account_idle_time(cputime_one_jiffy);
3292 * Account multiple ticks of steal time.
3293 * @p: the process from which the cpu time has been stolen
3294 * @ticks: number of stolen ticks
3296 void account_steal_ticks(unsigned long ticks)
3298 account_steal_time(jiffies_to_cputime(ticks));
3302 * Account multiple ticks of idle time.
3303 * @ticks: number of stolen ticks
3305 void account_idle_ticks(unsigned long ticks)
3307 account_idle_time(jiffies_to_cputime(ticks));
3313 * Use precise platform statistics if available:
3315 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3316 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3322 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3324 struct task_cputime cputime;
3326 thread_group_cputime(p, &cputime);
3328 *ut = cputime.utime;
3329 *st = cputime.stime;
3333 #ifndef nsecs_to_cputime
3334 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3337 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3339 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
3342 * Use CFS's precise accounting:
3344 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3349 temp = (u64)(rtime * utime);
3350 do_div(temp, total);
3351 utime = (cputime_t)temp;
3356 * Compare with previous values, to keep monotonicity:
3358 p->prev_utime = max(p->prev_utime, utime);
3359 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
3361 *ut = p->prev_utime;
3362 *st = p->prev_stime;
3366 * Must be called with siglock held.
3368 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3370 struct signal_struct *sig = p->signal;
3371 struct task_cputime cputime;
3372 cputime_t rtime, utime, total;
3374 thread_group_cputime(p, &cputime);
3376 total = cputime_add(cputime.utime, cputime.stime);
3377 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3382 temp = (u64)(rtime * cputime.utime);
3383 do_div(temp, total);
3384 utime = (cputime_t)temp;
3388 sig->prev_utime = max(sig->prev_utime, utime);
3389 sig->prev_stime = max(sig->prev_stime,
3390 cputime_sub(rtime, sig->prev_utime));
3392 *ut = sig->prev_utime;
3393 *st = sig->prev_stime;
3398 * This function gets called by the timer code, with HZ frequency.
3399 * We call it with interrupts disabled.
3401 * It also gets called by the fork code, when changing the parent's
3404 void scheduler_tick(void)
3406 int cpu = smp_processor_id();
3407 struct rq *rq = cpu_rq(cpu);
3408 struct task_struct *curr = rq->curr;
3412 raw_spin_lock(&rq->lock);
3413 update_rq_clock(rq);
3414 update_cpu_load(rq);
3415 curr->sched_class->task_tick(rq, curr, 0);
3416 raw_spin_unlock(&rq->lock);
3418 perf_event_task_tick(curr);
3421 rq->idle_at_tick = idle_cpu(cpu);
3422 trigger_load_balance(rq, cpu);
3426 notrace unsigned long get_parent_ip(unsigned long addr)
3428 if (in_lock_functions(addr)) {
3429 addr = CALLER_ADDR2;
3430 if (in_lock_functions(addr))
3431 addr = CALLER_ADDR3;
3436 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3437 defined(CONFIG_PREEMPT_TRACER))
3439 void __kprobes add_preempt_count(int val)
3441 #ifdef CONFIG_DEBUG_PREEMPT
3445 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3448 preempt_count() += val;
3449 #ifdef CONFIG_DEBUG_PREEMPT
3451 * Spinlock count overflowing soon?
3453 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3456 if (preempt_count() == val)
3457 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3459 EXPORT_SYMBOL(add_preempt_count);
3461 void __kprobes sub_preempt_count(int val)
3463 #ifdef CONFIG_DEBUG_PREEMPT
3467 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3470 * Is the spinlock portion underflowing?
3472 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3473 !(preempt_count() & PREEMPT_MASK)))
3477 if (preempt_count() == val)
3478 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3479 preempt_count() -= val;
3481 EXPORT_SYMBOL(sub_preempt_count);
3486 * Print scheduling while atomic bug:
3488 static noinline void __schedule_bug(struct task_struct *prev)
3490 struct pt_regs *regs = get_irq_regs();
3492 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3493 prev->comm, prev->pid, preempt_count());
3495 debug_show_held_locks(prev);
3497 if (irqs_disabled())
3498 print_irqtrace_events(prev);
3507 * Various schedule()-time debugging checks and statistics:
3509 static inline void schedule_debug(struct task_struct *prev)
3512 * Test if we are atomic. Since do_exit() needs to call into
3513 * schedule() atomically, we ignore that path for now.
3514 * Otherwise, whine if we are scheduling when we should not be.
3516 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3517 __schedule_bug(prev);
3519 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3521 schedstat_inc(this_rq(), sched_count);
3522 #ifdef CONFIG_SCHEDSTATS
3523 if (unlikely(prev->lock_depth >= 0)) {
3524 schedstat_inc(this_rq(), bkl_count);
3525 schedstat_inc(prev, sched_info.bkl_count);
3530 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3533 update_rq_clock(rq);
3534 rq->skip_clock_update = 0;
3535 prev->sched_class->put_prev_task(rq, prev);
3539 * Pick up the highest-prio task:
3541 static inline struct task_struct *
3542 pick_next_task(struct rq *rq)
3544 const struct sched_class *class;
3545 struct task_struct *p;
3548 * Optimization: we know that if all tasks are in
3549 * the fair class we can call that function directly:
3551 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3552 p = fair_sched_class.pick_next_task(rq);
3557 class = sched_class_highest;
3559 p = class->pick_next_task(rq);
3563 * Will never be NULL as the idle class always
3564 * returns a non-NULL p:
3566 class = class->next;
3571 * schedule() is the main scheduler function.
3573 asmlinkage void __sched schedule(void)
3575 struct task_struct *prev, *next;
3576 unsigned long *switch_count;
3582 cpu = smp_processor_id();
3584 rcu_note_context_switch(cpu);
3586 switch_count = &prev->nivcsw;
3588 release_kernel_lock(prev);
3589 need_resched_nonpreemptible:
3591 schedule_debug(prev);
3593 if (sched_feat(HRTICK))
3596 raw_spin_lock_irq(&rq->lock);
3597 clear_tsk_need_resched(prev);
3599 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3600 if (unlikely(signal_pending_state(prev->state, prev)))
3601 prev->state = TASK_RUNNING;
3603 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3604 switch_count = &prev->nvcsw;
3607 pre_schedule(rq, prev);
3609 if (unlikely(!rq->nr_running))
3610 idle_balance(cpu, rq);
3612 put_prev_task(rq, prev);
3613 next = pick_next_task(rq);
3615 if (likely(prev != next)) {
3616 sched_info_switch(prev, next);
3617 perf_event_task_sched_out(prev, next);
3623 context_switch(rq, prev, next); /* unlocks the rq */
3625 * the context switch might have flipped the stack from under
3626 * us, hence refresh the local variables.
3628 cpu = smp_processor_id();
3631 raw_spin_unlock_irq(&rq->lock);
3635 if (unlikely(reacquire_kernel_lock(current) < 0)) {
3637 switch_count = &prev->nivcsw;
3638 goto need_resched_nonpreemptible;
3641 preempt_enable_no_resched();
3645 EXPORT_SYMBOL(schedule);
3647 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3649 * Look out! "owner" is an entirely speculative pointer
3650 * access and not reliable.
3652 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
3657 if (!sched_feat(OWNER_SPIN))
3660 #ifdef CONFIG_DEBUG_PAGEALLOC
3662 * Need to access the cpu field knowing that
3663 * DEBUG_PAGEALLOC could have unmapped it if
3664 * the mutex owner just released it and exited.
3666 if (probe_kernel_address(&owner->cpu, cpu))
3673 * Even if the access succeeded (likely case),
3674 * the cpu field may no longer be valid.
3676 if (cpu >= nr_cpumask_bits)
3680 * We need to validate that we can do a
3681 * get_cpu() and that we have the percpu area.
3683 if (!cpu_online(cpu))
3690 * Owner changed, break to re-assess state.
3692 if (lock->owner != owner)
3696 * Is that owner really running on that cpu?
3698 if (task_thread_info(rq->curr) != owner || need_resched())
3708 #ifdef CONFIG_PREEMPT
3710 * this is the entry point to schedule() from in-kernel preemption
3711 * off of preempt_enable. Kernel preemptions off return from interrupt
3712 * occur there and call schedule directly.
3714 asmlinkage void __sched preempt_schedule(void)
3716 struct thread_info *ti = current_thread_info();
3719 * If there is a non-zero preempt_count or interrupts are disabled,
3720 * we do not want to preempt the current task. Just return..
3722 if (likely(ti->preempt_count || irqs_disabled()))
3726 add_preempt_count(PREEMPT_ACTIVE);
3728 sub_preempt_count(PREEMPT_ACTIVE);
3731 * Check again in case we missed a preemption opportunity
3732 * between schedule and now.
3735 } while (need_resched());
3737 EXPORT_SYMBOL(preempt_schedule);
3740 * this is the entry point to schedule() from kernel preemption
3741 * off of irq context.
3742 * Note, that this is called and return with irqs disabled. This will
3743 * protect us against recursive calling from irq.
3745 asmlinkage void __sched preempt_schedule_irq(void)
3747 struct thread_info *ti = current_thread_info();
3749 /* Catch callers which need to be fixed */
3750 BUG_ON(ti->preempt_count || !irqs_disabled());
3753 add_preempt_count(PREEMPT_ACTIVE);
3756 local_irq_disable();
3757 sub_preempt_count(PREEMPT_ACTIVE);
3760 * Check again in case we missed a preemption opportunity
3761 * between schedule and now.
3764 } while (need_resched());
3767 #endif /* CONFIG_PREEMPT */
3769 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3772 return try_to_wake_up(curr->private, mode, wake_flags);
3774 EXPORT_SYMBOL(default_wake_function);
3777 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3778 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3779 * number) then we wake all the non-exclusive tasks and one exclusive task.
3781 * There are circumstances in which we can try to wake a task which has already
3782 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3783 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3785 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3786 int nr_exclusive, int wake_flags, void *key)
3788 wait_queue_t *curr, *next;
3790 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3791 unsigned flags = curr->flags;
3793 if (curr->func(curr, mode, wake_flags, key) &&
3794 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3800 * __wake_up - wake up threads blocked on a waitqueue.
3802 * @mode: which threads
3803 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3804 * @key: is directly passed to the wakeup function
3806 * It may be assumed that this function implies a write memory barrier before
3807 * changing the task state if and only if any tasks are woken up.
3809 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3810 int nr_exclusive, void *key)
3812 unsigned long flags;
3814 spin_lock_irqsave(&q->lock, flags);
3815 __wake_up_common(q, mode, nr_exclusive, 0, key);
3816 spin_unlock_irqrestore(&q->lock, flags);
3818 EXPORT_SYMBOL(__wake_up);
3821 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3823 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3825 __wake_up_common(q, mode, 1, 0, NULL);
3827 EXPORT_SYMBOL_GPL(__wake_up_locked);
3829 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3831 __wake_up_common(q, mode, 1, 0, key);
3835 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3837 * @mode: which threads
3838 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3839 * @key: opaque value to be passed to wakeup targets
3841 * The sync wakeup differs that the waker knows that it will schedule
3842 * away soon, so while the target thread will be woken up, it will not
3843 * be migrated to another CPU - ie. the two threads are 'synchronized'
3844 * with each other. This can prevent needless bouncing between CPUs.
3846 * On UP it can prevent extra preemption.
3848 * It may be assumed that this function implies a write memory barrier before
3849 * changing the task state if and only if any tasks are woken up.
3851 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3852 int nr_exclusive, void *key)
3854 unsigned long flags;
3855 int wake_flags = WF_SYNC;
3860 if (unlikely(!nr_exclusive))
3863 spin_lock_irqsave(&q->lock, flags);
3864 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3865 spin_unlock_irqrestore(&q->lock, flags);
3867 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3870 * __wake_up_sync - see __wake_up_sync_key()
3872 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3874 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
3876 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3879 * complete: - signals a single thread waiting on this completion
3880 * @x: holds the state of this particular completion
3882 * This will wake up a single thread waiting on this completion. Threads will be
3883 * awakened in the same order in which they were queued.
3885 * See also complete_all(), wait_for_completion() and related routines.
3887 * It may be assumed that this function implies a write memory barrier before
3888 * changing the task state if and only if any tasks are woken up.
3890 void complete(struct completion *x)
3892 unsigned long flags;
3894 spin_lock_irqsave(&x->wait.lock, flags);
3896 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3897 spin_unlock_irqrestore(&x->wait.lock, flags);
3899 EXPORT_SYMBOL(complete);
3902 * complete_all: - signals all threads waiting on this completion
3903 * @x: holds the state of this particular completion
3905 * This will wake up all threads waiting on this particular completion event.
3907 * It may be assumed that this function implies a write memory barrier before
3908 * changing the task state if and only if any tasks are woken up.
3910 void complete_all(struct completion *x)
3912 unsigned long flags;
3914 spin_lock_irqsave(&x->wait.lock, flags);
3915 x->done += UINT_MAX/2;
3916 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3917 spin_unlock_irqrestore(&x->wait.lock, flags);
3919 EXPORT_SYMBOL(complete_all);
3921 static inline long __sched
3922 do_wait_for_common(struct completion *x, long timeout, int state)
3925 DECLARE_WAITQUEUE(wait, current);
3927 __add_wait_queue_tail_exclusive(&x->wait, &wait);
3929 if (signal_pending_state(state, current)) {
3930 timeout = -ERESTARTSYS;
3933 __set_current_state(state);
3934 spin_unlock_irq(&x->wait.lock);
3935 timeout = schedule_timeout(timeout);
3936 spin_lock_irq(&x->wait.lock);
3937 } while (!x->done && timeout);
3938 __remove_wait_queue(&x->wait, &wait);
3943 return timeout ?: 1;
3947 wait_for_common(struct completion *x, long timeout, int state)
3951 spin_lock_irq(&x->wait.lock);
3952 timeout = do_wait_for_common(x, timeout, state);
3953 spin_unlock_irq(&x->wait.lock);
3958 * wait_for_completion: - waits for completion of a task
3959 * @x: holds the state of this particular completion
3961 * This waits to be signaled for completion of a specific task. It is NOT
3962 * interruptible and there is no timeout.
3964 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3965 * and interrupt capability. Also see complete().
3967 void __sched wait_for_completion(struct completion *x)
3969 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3971 EXPORT_SYMBOL(wait_for_completion);
3974 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3975 * @x: holds the state of this particular completion
3976 * @timeout: timeout value in jiffies
3978 * This waits for either a completion of a specific task to be signaled or for a
3979 * specified timeout to expire. The timeout is in jiffies. It is not
3982 unsigned long __sched
3983 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3985 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3987 EXPORT_SYMBOL(wait_for_completion_timeout);
3990 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3991 * @x: holds the state of this particular completion
3993 * This waits for completion of a specific task to be signaled. It is
3996 int __sched wait_for_completion_interruptible(struct completion *x)
3998 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3999 if (t == -ERESTARTSYS)
4003 EXPORT_SYMBOL(wait_for_completion_interruptible);
4006 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4007 * @x: holds the state of this particular completion
4008 * @timeout: timeout value in jiffies
4010 * This waits for either a completion of a specific task to be signaled or for a
4011 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4013 unsigned long __sched
4014 wait_for_completion_interruptible_timeout(struct completion *x,
4015 unsigned long timeout)
4017 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4019 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4022 * wait_for_completion_killable: - waits for completion of a task (killable)
4023 * @x: holds the state of this particular completion
4025 * This waits to be signaled for completion of a specific task. It can be
4026 * interrupted by a kill signal.
4028 int __sched wait_for_completion_killable(struct completion *x)
4030 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4031 if (t == -ERESTARTSYS)
4035 EXPORT_SYMBOL(wait_for_completion_killable);
4038 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4039 * @x: holds the state of this particular completion
4040 * @timeout: timeout value in jiffies
4042 * This waits for either a completion of a specific task to be
4043 * signaled or for a specified timeout to expire. It can be
4044 * interrupted by a kill signal. The timeout is in jiffies.
4046 unsigned long __sched
4047 wait_for_completion_killable_timeout(struct completion *x,
4048 unsigned long timeout)
4050 return wait_for_common(x, timeout, TASK_KILLABLE);
4052 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
4055 * try_wait_for_completion - try to decrement a completion without blocking
4056 * @x: completion structure
4058 * Returns: 0 if a decrement cannot be done without blocking
4059 * 1 if a decrement succeeded.
4061 * If a completion is being used as a counting completion,
4062 * attempt to decrement the counter without blocking. This
4063 * enables us to avoid waiting if the resource the completion
4064 * is protecting is not available.
4066 bool try_wait_for_completion(struct completion *x)
4068 unsigned long flags;
4071 spin_lock_irqsave(&x->wait.lock, flags);
4076 spin_unlock_irqrestore(&x->wait.lock, flags);
4079 EXPORT_SYMBOL(try_wait_for_completion);
4082 * completion_done - Test to see if a completion has any waiters
4083 * @x: completion structure
4085 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4086 * 1 if there are no waiters.
4089 bool completion_done(struct completion *x)
4091 unsigned long flags;
4094 spin_lock_irqsave(&x->wait.lock, flags);
4097 spin_unlock_irqrestore(&x->wait.lock, flags);
4100 EXPORT_SYMBOL(completion_done);
4103 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4105 unsigned long flags;
4108 init_waitqueue_entry(&wait, current);
4110 __set_current_state(state);
4112 spin_lock_irqsave(&q->lock, flags);
4113 __add_wait_queue(q, &wait);
4114 spin_unlock(&q->lock);
4115 timeout = schedule_timeout(timeout);
4116 spin_lock_irq(&q->lock);
4117 __remove_wait_queue(q, &wait);
4118 spin_unlock_irqrestore(&q->lock, flags);
4123 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4125 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4127 EXPORT_SYMBOL(interruptible_sleep_on);
4130 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4132 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4134 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4136 void __sched sleep_on(wait_queue_head_t *q)
4138 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4140 EXPORT_SYMBOL(sleep_on);
4142 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4144 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4146 EXPORT_SYMBOL(sleep_on_timeout);
4148 #ifdef CONFIG_RT_MUTEXES
4151 * rt_mutex_setprio - set the current priority of a task
4153 * @prio: prio value (kernel-internal form)
4155 * This function changes the 'effective' priority of a task. It does
4156 * not touch ->normal_prio like __setscheduler().
4158 * Used by the rt_mutex code to implement priority inheritance logic.
4160 void rt_mutex_setprio(struct task_struct *p, int prio)
4162 unsigned long flags;
4163 int oldprio, on_rq, running;
4165 const struct sched_class *prev_class;
4167 BUG_ON(prio < 0 || prio > MAX_PRIO);
4169 rq = task_rq_lock(p, &flags);
4172 prev_class = p->sched_class;
4173 on_rq = p->se.on_rq;
4174 running = task_current(rq, p);
4176 dequeue_task(rq, p, 0);
4178 p->sched_class->put_prev_task(rq, p);
4181 p->sched_class = &rt_sched_class;
4183 p->sched_class = &fair_sched_class;
4188 p->sched_class->set_curr_task(rq);
4190 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4192 check_class_changed(rq, p, prev_class, oldprio, running);
4194 task_rq_unlock(rq, &flags);
4199 void set_user_nice(struct task_struct *p, long nice)
4201 int old_prio, delta, on_rq;
4202 unsigned long flags;
4205 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4208 * We have to be careful, if called from sys_setpriority(),
4209 * the task might be in the middle of scheduling on another CPU.
4211 rq = task_rq_lock(p, &flags);
4213 * The RT priorities are set via sched_setscheduler(), but we still
4214 * allow the 'normal' nice value to be set - but as expected
4215 * it wont have any effect on scheduling until the task is
4216 * SCHED_FIFO/SCHED_RR:
4218 if (task_has_rt_policy(p)) {
4219 p->static_prio = NICE_TO_PRIO(nice);
4222 on_rq = p->se.on_rq;
4224 dequeue_task(rq, p, 0);
4226 p->static_prio = NICE_TO_PRIO(nice);
4229 p->prio = effective_prio(p);
4230 delta = p->prio - old_prio;
4233 enqueue_task(rq, p, 0);
4235 * If the task increased its priority or is running and
4236 * lowered its priority, then reschedule its CPU:
4238 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4239 resched_task(rq->curr);
4242 task_rq_unlock(rq, &flags);
4244 EXPORT_SYMBOL(set_user_nice);
4247 * can_nice - check if a task can reduce its nice value
4251 int can_nice(const struct task_struct *p, const int nice)
4253 /* convert nice value [19,-20] to rlimit style value [1,40] */
4254 int nice_rlim = 20 - nice;
4256 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4257 capable(CAP_SYS_NICE));
4260 #ifdef __ARCH_WANT_SYS_NICE
4263 * sys_nice - change the priority of the current process.
4264 * @increment: priority increment
4266 * sys_setpriority is a more generic, but much slower function that
4267 * does similar things.
4269 SYSCALL_DEFINE1(nice, int, increment)
4274 * Setpriority might change our priority at the same moment.
4275 * We don't have to worry. Conceptually one call occurs first
4276 * and we have a single winner.
4278 if (increment < -40)
4283 nice = TASK_NICE(current) + increment;
4289 if (increment < 0 && !can_nice(current, nice))
4292 retval = security_task_setnice(current, nice);
4296 set_user_nice(current, nice);
4303 * task_prio - return the priority value of a given task.
4304 * @p: the task in question.
4306 * This is the priority value as seen by users in /proc.
4307 * RT tasks are offset by -200. Normal tasks are centered
4308 * around 0, value goes from -16 to +15.
4310 int task_prio(const struct task_struct *p)
4312 return p->prio - MAX_RT_PRIO;
4316 * task_nice - return the nice value of a given task.
4317 * @p: the task in question.
4319 int task_nice(const struct task_struct *p)
4321 return TASK_NICE(p);
4323 EXPORT_SYMBOL(task_nice);
4326 * idle_cpu - is a given cpu idle currently?
4327 * @cpu: the processor in question.
4329 int idle_cpu(int cpu)
4331 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4335 * idle_task - return the idle task for a given cpu.
4336 * @cpu: the processor in question.
4338 struct task_struct *idle_task(int cpu)
4340 return cpu_rq(cpu)->idle;
4344 * find_process_by_pid - find a process with a matching PID value.
4345 * @pid: the pid in question.
4347 static struct task_struct *find_process_by_pid(pid_t pid)
4349 return pid ? find_task_by_vpid(pid) : current;
4352 /* Actually do priority change: must hold rq lock. */
4354 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4356 BUG_ON(p->se.on_rq);
4359 p->rt_priority = prio;
4360 p->normal_prio = normal_prio(p);
4361 /* we are holding p->pi_lock already */
4362 p->prio = rt_mutex_getprio(p);
4363 if (rt_prio(p->prio))
4364 p->sched_class = &rt_sched_class;
4366 p->sched_class = &fair_sched_class;
4371 * check the target process has a UID that matches the current process's
4373 static bool check_same_owner(struct task_struct *p)
4375 const struct cred *cred = current_cred(), *pcred;
4379 pcred = __task_cred(p);
4380 match = (cred->euid == pcred->euid ||
4381 cred->euid == pcred->uid);
4386 static int __sched_setscheduler(struct task_struct *p, int policy,
4387 struct sched_param *param, bool user)
4389 int retval, oldprio, oldpolicy = -1, on_rq, running;
4390 unsigned long flags;
4391 const struct sched_class *prev_class;
4395 /* may grab non-irq protected spin_locks */
4396 BUG_ON(in_interrupt());
4398 /* double check policy once rq lock held */
4400 reset_on_fork = p->sched_reset_on_fork;
4401 policy = oldpolicy = p->policy;
4403 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4404 policy &= ~SCHED_RESET_ON_FORK;
4406 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4407 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4408 policy != SCHED_IDLE)
4413 * Valid priorities for SCHED_FIFO and SCHED_RR are
4414 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4415 * SCHED_BATCH and SCHED_IDLE is 0.
4417 if (param->sched_priority < 0 ||
4418 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4419 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4421 if (rt_policy(policy) != (param->sched_priority != 0))
4425 * Allow unprivileged RT tasks to decrease priority:
4427 if (user && !capable(CAP_SYS_NICE)) {
4428 if (rt_policy(policy)) {
4429 unsigned long rlim_rtprio;
4431 if (!lock_task_sighand(p, &flags))
4433 rlim_rtprio = task_rlimit(p, RLIMIT_RTPRIO);
4434 unlock_task_sighand(p, &flags);
4436 /* can't set/change the rt policy */
4437 if (policy != p->policy && !rlim_rtprio)
4440 /* can't increase priority */
4441 if (param->sched_priority > p->rt_priority &&
4442 param->sched_priority > rlim_rtprio)
4446 * Like positive nice levels, dont allow tasks to
4447 * move out of SCHED_IDLE either:
4449 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4452 /* can't change other user's priorities */
4453 if (!check_same_owner(p))
4456 /* Normal users shall not reset the sched_reset_on_fork flag */
4457 if (p->sched_reset_on_fork && !reset_on_fork)
4462 retval = security_task_setscheduler(p, policy, param);
4468 * make sure no PI-waiters arrive (or leave) while we are
4469 * changing the priority of the task:
4471 raw_spin_lock_irqsave(&p->pi_lock, flags);
4473 * To be able to change p->policy safely, the apropriate
4474 * runqueue lock must be held.
4476 rq = __task_rq_lock(p);
4478 #ifdef CONFIG_RT_GROUP_SCHED
4481 * Do not allow realtime tasks into groups that have no runtime
4484 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4485 task_group(p)->rt_bandwidth.rt_runtime == 0) {
4486 __task_rq_unlock(rq);
4487 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4493 /* recheck policy now with rq lock held */
4494 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4495 policy = oldpolicy = -1;
4496 __task_rq_unlock(rq);
4497 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4500 on_rq = p->se.on_rq;
4501 running = task_current(rq, p);
4503 deactivate_task(rq, p, 0);
4505 p->sched_class->put_prev_task(rq, p);
4507 p->sched_reset_on_fork = reset_on_fork;
4510 prev_class = p->sched_class;
4511 __setscheduler(rq, p, policy, param->sched_priority);
4514 p->sched_class->set_curr_task(rq);
4516 activate_task(rq, p, 0);
4518 check_class_changed(rq, p, prev_class, oldprio, running);
4520 __task_rq_unlock(rq);
4521 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4523 rt_mutex_adjust_pi(p);
4529 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4530 * @p: the task in question.
4531 * @policy: new policy.
4532 * @param: structure containing the new RT priority.
4534 * NOTE that the task may be already dead.
4536 int sched_setscheduler(struct task_struct *p, int policy,
4537 struct sched_param *param)
4539 return __sched_setscheduler(p, policy, param, true);
4541 EXPORT_SYMBOL_GPL(sched_setscheduler);
4544 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4545 * @p: the task in question.
4546 * @policy: new policy.
4547 * @param: structure containing the new RT priority.
4549 * Just like sched_setscheduler, only don't bother checking if the
4550 * current context has permission. For example, this is needed in
4551 * stop_machine(): we create temporary high priority worker threads,
4552 * but our caller might not have that capability.
4554 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4555 struct sched_param *param)
4557 return __sched_setscheduler(p, policy, param, false);
4561 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4563 struct sched_param lparam;
4564 struct task_struct *p;
4567 if (!param || pid < 0)
4569 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4574 p = find_process_by_pid(pid);
4576 retval = sched_setscheduler(p, policy, &lparam);
4583 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4584 * @pid: the pid in question.
4585 * @policy: new policy.
4586 * @param: structure containing the new RT priority.
4588 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4589 struct sched_param __user *, param)
4591 /* negative values for policy are not valid */
4595 return do_sched_setscheduler(pid, policy, param);
4599 * sys_sched_setparam - set/change the RT priority of a thread
4600 * @pid: the pid in question.
4601 * @param: structure containing the new RT priority.
4603 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4605 return do_sched_setscheduler(pid, -1, param);
4609 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4610 * @pid: the pid in question.
4612 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4614 struct task_struct *p;
4622 p = find_process_by_pid(pid);
4624 retval = security_task_getscheduler(p);
4627 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4634 * sys_sched_getparam - get the RT priority of a thread
4635 * @pid: the pid in question.
4636 * @param: structure containing the RT priority.
4638 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4640 struct sched_param lp;
4641 struct task_struct *p;
4644 if (!param || pid < 0)
4648 p = find_process_by_pid(pid);
4653 retval = security_task_getscheduler(p);
4657 lp.sched_priority = p->rt_priority;
4661 * This one might sleep, we cannot do it with a spinlock held ...
4663 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4672 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4674 cpumask_var_t cpus_allowed, new_mask;
4675 struct task_struct *p;
4681 p = find_process_by_pid(pid);
4688 /* Prevent p going away */
4692 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4696 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4698 goto out_free_cpus_allowed;
4701 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
4704 retval = security_task_setscheduler(p, 0, NULL);
4708 cpuset_cpus_allowed(p, cpus_allowed);
4709 cpumask_and(new_mask, in_mask, cpus_allowed);
4711 retval = set_cpus_allowed_ptr(p, new_mask);
4714 cpuset_cpus_allowed(p, cpus_allowed);
4715 if (!cpumask_subset(new_mask, cpus_allowed)) {
4717 * We must have raced with a concurrent cpuset
4718 * update. Just reset the cpus_allowed to the
4719 * cpuset's cpus_allowed
4721 cpumask_copy(new_mask, cpus_allowed);
4726 free_cpumask_var(new_mask);
4727 out_free_cpus_allowed:
4728 free_cpumask_var(cpus_allowed);
4735 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4736 struct cpumask *new_mask)
4738 if (len < cpumask_size())
4739 cpumask_clear(new_mask);
4740 else if (len > cpumask_size())
4741 len = cpumask_size();
4743 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4747 * sys_sched_setaffinity - set the cpu affinity of a process
4748 * @pid: pid of the process
4749 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4750 * @user_mask_ptr: user-space pointer to the new cpu mask
4752 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4753 unsigned long __user *, user_mask_ptr)
4755 cpumask_var_t new_mask;
4758 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4761 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4763 retval = sched_setaffinity(pid, new_mask);
4764 free_cpumask_var(new_mask);
4768 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4770 struct task_struct *p;
4771 unsigned long flags;
4779 p = find_process_by_pid(pid);
4783 retval = security_task_getscheduler(p);
4787 rq = task_rq_lock(p, &flags);
4788 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4789 task_rq_unlock(rq, &flags);
4799 * sys_sched_getaffinity - get the cpu affinity of a process
4800 * @pid: pid of the process
4801 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4802 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4804 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4805 unsigned long __user *, user_mask_ptr)
4810 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4812 if (len & (sizeof(unsigned long)-1))
4815 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4818 ret = sched_getaffinity(pid, mask);
4820 size_t retlen = min_t(size_t, len, cpumask_size());
4822 if (copy_to_user(user_mask_ptr, mask, retlen))
4827 free_cpumask_var(mask);
4833 * sys_sched_yield - yield the current processor to other threads.
4835 * This function yields the current CPU to other tasks. If there are no
4836 * other threads running on this CPU then this function will return.
4838 SYSCALL_DEFINE0(sched_yield)
4840 struct rq *rq = this_rq_lock();
4842 schedstat_inc(rq, yld_count);
4843 current->sched_class->yield_task(rq);
4846 * Since we are going to call schedule() anyway, there's
4847 * no need to preempt or enable interrupts:
4849 __release(rq->lock);
4850 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4851 do_raw_spin_unlock(&rq->lock);
4852 preempt_enable_no_resched();
4859 static inline int should_resched(void)
4861 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4864 static void __cond_resched(void)
4866 add_preempt_count(PREEMPT_ACTIVE);
4868 sub_preempt_count(PREEMPT_ACTIVE);
4871 int __sched _cond_resched(void)
4873 if (should_resched()) {
4879 EXPORT_SYMBOL(_cond_resched);
4882 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4883 * call schedule, and on return reacquire the lock.
4885 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4886 * operations here to prevent schedule() from being called twice (once via
4887 * spin_unlock(), once by hand).
4889 int __cond_resched_lock(spinlock_t *lock)
4891 int resched = should_resched();
4894 lockdep_assert_held(lock);
4896 if (spin_needbreak(lock) || resched) {
4907 EXPORT_SYMBOL(__cond_resched_lock);
4909 int __sched __cond_resched_softirq(void)
4911 BUG_ON(!in_softirq());
4913 if (should_resched()) {
4921 EXPORT_SYMBOL(__cond_resched_softirq);
4924 * yield - yield the current processor to other threads.
4926 * This is a shortcut for kernel-space yielding - it marks the
4927 * thread runnable and calls sys_sched_yield().
4929 void __sched yield(void)
4931 set_current_state(TASK_RUNNING);
4934 EXPORT_SYMBOL(yield);
4937 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4938 * that process accounting knows that this is a task in IO wait state.
4940 void __sched io_schedule(void)
4942 struct rq *rq = raw_rq();
4944 delayacct_blkio_start();
4945 atomic_inc(&rq->nr_iowait);
4946 current->in_iowait = 1;
4948 current->in_iowait = 0;
4949 atomic_dec(&rq->nr_iowait);
4950 delayacct_blkio_end();
4952 EXPORT_SYMBOL(io_schedule);
4954 long __sched io_schedule_timeout(long timeout)
4956 struct rq *rq = raw_rq();
4959 delayacct_blkio_start();
4960 atomic_inc(&rq->nr_iowait);
4961 current->in_iowait = 1;
4962 ret = schedule_timeout(timeout);
4963 current->in_iowait = 0;
4964 atomic_dec(&rq->nr_iowait);
4965 delayacct_blkio_end();
4970 * sys_sched_get_priority_max - return maximum RT priority.
4971 * @policy: scheduling class.
4973 * this syscall returns the maximum rt_priority that can be used
4974 * by a given scheduling class.
4976 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4983 ret = MAX_USER_RT_PRIO-1;
4995 * sys_sched_get_priority_min - return minimum RT priority.
4996 * @policy: scheduling class.
4998 * this syscall returns the minimum rt_priority that can be used
4999 * by a given scheduling class.
5001 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5019 * sys_sched_rr_get_interval - return the default timeslice of a process.
5020 * @pid: pid of the process.
5021 * @interval: userspace pointer to the timeslice value.
5023 * this syscall writes the default timeslice value of a given process
5024 * into the user-space timespec buffer. A value of '0' means infinity.
5026 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5027 struct timespec __user *, interval)
5029 struct task_struct *p;
5030 unsigned int time_slice;
5031 unsigned long flags;
5041 p = find_process_by_pid(pid);
5045 retval = security_task_getscheduler(p);
5049 rq = task_rq_lock(p, &flags);
5050 time_slice = p->sched_class->get_rr_interval(rq, p);
5051 task_rq_unlock(rq, &flags);
5054 jiffies_to_timespec(time_slice, &t);
5055 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5063 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5065 void sched_show_task(struct task_struct *p)
5067 unsigned long free = 0;
5070 state = p->state ? __ffs(p->state) + 1 : 0;
5071 printk(KERN_INFO "%-13.13s %c", p->comm,
5072 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5073 #if BITS_PER_LONG == 32
5074 if (state == TASK_RUNNING)
5075 printk(KERN_CONT " running ");
5077 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5079 if (state == TASK_RUNNING)
5080 printk(KERN_CONT " running task ");
5082 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5084 #ifdef CONFIG_DEBUG_STACK_USAGE
5085 free = stack_not_used(p);
5087 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5088 task_pid_nr(p), task_pid_nr(p->real_parent),
5089 (unsigned long)task_thread_info(p)->flags);
5091 show_stack(p, NULL);
5094 void show_state_filter(unsigned long state_filter)
5096 struct task_struct *g, *p;
5098 #if BITS_PER_LONG == 32
5100 " task PC stack pid father\n");
5103 " task PC stack pid father\n");
5105 read_lock(&tasklist_lock);
5106 do_each_thread(g, p) {
5108 * reset the NMI-timeout, listing all files on a slow
5109 * console might take alot of time:
5111 touch_nmi_watchdog();
5112 if (!state_filter || (p->state & state_filter))
5114 } while_each_thread(g, p);
5116 touch_all_softlockup_watchdogs();
5118 #ifdef CONFIG_SCHED_DEBUG
5119 sysrq_sched_debug_show();
5121 read_unlock(&tasklist_lock);
5123 * Only show locks if all tasks are dumped:
5126 debug_show_all_locks();
5129 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5131 idle->sched_class = &idle_sched_class;
5135 * init_idle - set up an idle thread for a given CPU
5136 * @idle: task in question
5137 * @cpu: cpu the idle task belongs to
5139 * NOTE: this function does not set the idle thread's NEED_RESCHED
5140 * flag, to make booting more robust.
5142 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5144 struct rq *rq = cpu_rq(cpu);
5145 unsigned long flags;
5147 raw_spin_lock_irqsave(&rq->lock, flags);
5150 idle->state = TASK_RUNNING;
5151 idle->se.exec_start = sched_clock();
5153 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
5154 __set_task_cpu(idle, cpu);
5156 rq->curr = rq->idle = idle;
5157 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5160 raw_spin_unlock_irqrestore(&rq->lock, flags);
5162 /* Set the preempt count _outside_ the spinlocks! */
5163 #if defined(CONFIG_PREEMPT)
5164 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5166 task_thread_info(idle)->preempt_count = 0;
5169 * The idle tasks have their own, simple scheduling class:
5171 idle->sched_class = &idle_sched_class;
5172 ftrace_graph_init_task(idle);
5176 * In a system that switches off the HZ timer nohz_cpu_mask
5177 * indicates which cpus entered this state. This is used
5178 * in the rcu update to wait only for active cpus. For system
5179 * which do not switch off the HZ timer nohz_cpu_mask should
5180 * always be CPU_BITS_NONE.
5182 cpumask_var_t nohz_cpu_mask;
5185 * Increase the granularity value when there are more CPUs,
5186 * because with more CPUs the 'effective latency' as visible
5187 * to users decreases. But the relationship is not linear,
5188 * so pick a second-best guess by going with the log2 of the
5191 * This idea comes from the SD scheduler of Con Kolivas:
5193 static int get_update_sysctl_factor(void)
5195 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5196 unsigned int factor;
5198 switch (sysctl_sched_tunable_scaling) {
5199 case SCHED_TUNABLESCALING_NONE:
5202 case SCHED_TUNABLESCALING_LINEAR:
5205 case SCHED_TUNABLESCALING_LOG:
5207 factor = 1 + ilog2(cpus);
5214 static void update_sysctl(void)
5216 unsigned int factor = get_update_sysctl_factor();
5218 #define SET_SYSCTL(name) \
5219 (sysctl_##name = (factor) * normalized_sysctl_##name)
5220 SET_SYSCTL(sched_min_granularity);
5221 SET_SYSCTL(sched_latency);
5222 SET_SYSCTL(sched_wakeup_granularity);
5223 SET_SYSCTL(sched_shares_ratelimit);
5227 static inline void sched_init_granularity(void)
5234 * This is how migration works:
5236 * 1) we invoke migration_cpu_stop() on the target CPU using
5238 * 2) stopper starts to run (implicitly forcing the migrated thread
5240 * 3) it checks whether the migrated task is still in the wrong runqueue.
5241 * 4) if it's in the wrong runqueue then the migration thread removes
5242 * it and puts it into the right queue.
5243 * 5) stopper completes and stop_one_cpu() returns and the migration
5248 * Change a given task's CPU affinity. Migrate the thread to a
5249 * proper CPU and schedule it away if the CPU it's executing on
5250 * is removed from the allowed bitmask.
5252 * NOTE: the caller must have a valid reference to the task, the
5253 * task must not exit() & deallocate itself prematurely. The
5254 * call is not atomic; no spinlocks may be held.
5256 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5258 unsigned long flags;
5260 unsigned int dest_cpu;
5264 * Serialize against TASK_WAKING so that ttwu() and wunt() can
5265 * drop the rq->lock and still rely on ->cpus_allowed.
5268 while (task_is_waking(p))
5270 rq = task_rq_lock(p, &flags);
5271 if (task_is_waking(p)) {
5272 task_rq_unlock(rq, &flags);
5276 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5281 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5282 !cpumask_equal(&p->cpus_allowed, new_mask))) {
5287 if (p->sched_class->set_cpus_allowed)
5288 p->sched_class->set_cpus_allowed(p, new_mask);
5290 cpumask_copy(&p->cpus_allowed, new_mask);
5291 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5294 /* Can the task run on the task's current CPU? If so, we're done */
5295 if (cpumask_test_cpu(task_cpu(p), new_mask))
5298 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
5299 if (migrate_task(p, dest_cpu)) {
5300 struct migration_arg arg = { p, dest_cpu };
5301 /* Need help from migration thread: drop lock and wait. */
5302 task_rq_unlock(rq, &flags);
5303 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
5304 tlb_migrate_finish(p->mm);
5308 task_rq_unlock(rq, &flags);
5312 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5315 * Move (not current) task off this cpu, onto dest cpu. We're doing
5316 * this because either it can't run here any more (set_cpus_allowed()
5317 * away from this CPU, or CPU going down), or because we're
5318 * attempting to rebalance this task on exec (sched_exec).
5320 * So we race with normal scheduler movements, but that's OK, as long
5321 * as the task is no longer on this CPU.
5323 * Returns non-zero if task was successfully migrated.
5325 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5327 struct rq *rq_dest, *rq_src;
5330 if (unlikely(!cpu_active(dest_cpu)))
5333 rq_src = cpu_rq(src_cpu);
5334 rq_dest = cpu_rq(dest_cpu);
5336 double_rq_lock(rq_src, rq_dest);
5337 /* Already moved. */
5338 if (task_cpu(p) != src_cpu)
5340 /* Affinity changed (again). */
5341 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
5345 * If we're not on a rq, the next wake-up will ensure we're
5349 deactivate_task(rq_src, p, 0);
5350 set_task_cpu(p, dest_cpu);
5351 activate_task(rq_dest, p, 0);
5352 check_preempt_curr(rq_dest, p, 0);
5357 double_rq_unlock(rq_src, rq_dest);
5362 * migration_cpu_stop - this will be executed by a highprio stopper thread
5363 * and performs thread migration by bumping thread off CPU then
5364 * 'pushing' onto another runqueue.
5366 static int migration_cpu_stop(void *data)
5368 struct migration_arg *arg = data;
5371 * The original target cpu might have gone down and we might
5372 * be on another cpu but it doesn't matter.
5374 local_irq_disable();
5375 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
5380 #ifdef CONFIG_HOTPLUG_CPU
5382 * Figure out where task on dead CPU should go, use force if necessary.
5384 void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5386 struct rq *rq = cpu_rq(dead_cpu);
5387 int needs_cpu, uninitialized_var(dest_cpu);
5388 unsigned long flags;
5390 local_irq_save(flags);
5392 raw_spin_lock(&rq->lock);
5393 needs_cpu = (task_cpu(p) == dead_cpu) && (p->state != TASK_WAKING);
5395 dest_cpu = select_fallback_rq(dead_cpu, p);
5396 raw_spin_unlock(&rq->lock);
5398 * It can only fail if we race with set_cpus_allowed(),
5399 * in the racer should migrate the task anyway.
5402 __migrate_task(p, dead_cpu, dest_cpu);
5403 local_irq_restore(flags);
5407 * While a dead CPU has no uninterruptible tasks queued at this point,
5408 * it might still have a nonzero ->nr_uninterruptible counter, because
5409 * for performance reasons the counter is not stricly tracking tasks to
5410 * their home CPUs. So we just add the counter to another CPU's counter,
5411 * to keep the global sum constant after CPU-down:
5413 static void migrate_nr_uninterruptible(struct rq *rq_src)
5415 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5416 unsigned long flags;
5418 local_irq_save(flags);
5419 double_rq_lock(rq_src, rq_dest);
5420 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5421 rq_src->nr_uninterruptible = 0;
5422 double_rq_unlock(rq_src, rq_dest);
5423 local_irq_restore(flags);
5426 /* Run through task list and migrate tasks from the dead cpu. */
5427 static void migrate_live_tasks(int src_cpu)
5429 struct task_struct *p, *t;
5431 read_lock(&tasklist_lock);
5433 do_each_thread(t, p) {
5437 if (task_cpu(p) == src_cpu)
5438 move_task_off_dead_cpu(src_cpu, p);
5439 } while_each_thread(t, p);
5441 read_unlock(&tasklist_lock);
5445 * Schedules idle task to be the next runnable task on current CPU.
5446 * It does so by boosting its priority to highest possible.
5447 * Used by CPU offline code.
5449 void sched_idle_next(void)
5451 int this_cpu = smp_processor_id();
5452 struct rq *rq = cpu_rq(this_cpu);
5453 struct task_struct *p = rq->idle;
5454 unsigned long flags;
5456 /* cpu has to be offline */
5457 BUG_ON(cpu_online(this_cpu));
5460 * Strictly not necessary since rest of the CPUs are stopped by now
5461 * and interrupts disabled on the current cpu.
5463 raw_spin_lock_irqsave(&rq->lock, flags);
5465 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5467 activate_task(rq, p, 0);
5469 raw_spin_unlock_irqrestore(&rq->lock, flags);
5473 * Ensures that the idle task is using init_mm right before its cpu goes
5476 void idle_task_exit(void)
5478 struct mm_struct *mm = current->active_mm;
5480 BUG_ON(cpu_online(smp_processor_id()));
5483 switch_mm(mm, &init_mm, current);
5487 /* called under rq->lock with disabled interrupts */
5488 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5490 struct rq *rq = cpu_rq(dead_cpu);
5492 /* Must be exiting, otherwise would be on tasklist. */
5493 BUG_ON(!p->exit_state);
5495 /* Cannot have done final schedule yet: would have vanished. */
5496 BUG_ON(p->state == TASK_DEAD);
5501 * Drop lock around migration; if someone else moves it,
5502 * that's OK. No task can be added to this CPU, so iteration is
5505 raw_spin_unlock_irq(&rq->lock);
5506 move_task_off_dead_cpu(dead_cpu, p);
5507 raw_spin_lock_irq(&rq->lock);
5512 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5513 static void migrate_dead_tasks(unsigned int dead_cpu)
5515 struct rq *rq = cpu_rq(dead_cpu);
5516 struct task_struct *next;
5519 if (!rq->nr_running)
5521 next = pick_next_task(rq);
5524 next->sched_class->put_prev_task(rq, next);
5525 migrate_dead(dead_cpu, next);
5531 * remove the tasks which were accounted by rq from calc_load_tasks.
5533 static void calc_global_load_remove(struct rq *rq)
5535 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5536 rq->calc_load_active = 0;
5538 #endif /* CONFIG_HOTPLUG_CPU */
5540 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5542 static struct ctl_table sd_ctl_dir[] = {
5544 .procname = "sched_domain",
5550 static struct ctl_table sd_ctl_root[] = {
5552 .procname = "kernel",
5554 .child = sd_ctl_dir,
5559 static struct ctl_table *sd_alloc_ctl_entry(int n)
5561 struct ctl_table *entry =
5562 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5567 static void sd_free_ctl_entry(struct ctl_table **tablep)
5569 struct ctl_table *entry;
5572 * In the intermediate directories, both the child directory and
5573 * procname are dynamically allocated and could fail but the mode
5574 * will always be set. In the lowest directory the names are
5575 * static strings and all have proc handlers.
5577 for (entry = *tablep; entry->mode; entry++) {
5579 sd_free_ctl_entry(&entry->child);
5580 if (entry->proc_handler == NULL)
5581 kfree(entry->procname);
5589 set_table_entry(struct ctl_table *entry,
5590 const char *procname, void *data, int maxlen,
5591 mode_t mode, proc_handler *proc_handler)
5593 entry->procname = procname;
5595 entry->maxlen = maxlen;
5597 entry->proc_handler = proc_handler;
5600 static struct ctl_table *
5601 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5603 struct ctl_table *table = sd_alloc_ctl_entry(13);
5608 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5609 sizeof(long), 0644, proc_doulongvec_minmax);
5610 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5611 sizeof(long), 0644, proc_doulongvec_minmax);
5612 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5613 sizeof(int), 0644, proc_dointvec_minmax);
5614 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5615 sizeof(int), 0644, proc_dointvec_minmax);
5616 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5617 sizeof(int), 0644, proc_dointvec_minmax);
5618 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5619 sizeof(int), 0644, proc_dointvec_minmax);
5620 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5621 sizeof(int), 0644, proc_dointvec_minmax);
5622 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5623 sizeof(int), 0644, proc_dointvec_minmax);
5624 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5625 sizeof(int), 0644, proc_dointvec_minmax);
5626 set_table_entry(&table[9], "cache_nice_tries",
5627 &sd->cache_nice_tries,
5628 sizeof(int), 0644, proc_dointvec_minmax);
5629 set_table_entry(&table[10], "flags", &sd->flags,
5630 sizeof(int), 0644, proc_dointvec_minmax);
5631 set_table_entry(&table[11], "name", sd->name,
5632 CORENAME_MAX_SIZE, 0444, proc_dostring);
5633 /* &table[12] is terminator */
5638 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5640 struct ctl_table *entry, *table;
5641 struct sched_domain *sd;
5642 int domain_num = 0, i;
5645 for_each_domain(cpu, sd)
5647 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5652 for_each_domain(cpu, sd) {
5653 snprintf(buf, 32, "domain%d", i);
5654 entry->procname = kstrdup(buf, GFP_KERNEL);
5656 entry->child = sd_alloc_ctl_domain_table(sd);
5663 static struct ctl_table_header *sd_sysctl_header;
5664 static void register_sched_domain_sysctl(void)
5666 int i, cpu_num = num_possible_cpus();
5667 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5670 WARN_ON(sd_ctl_dir[0].child);
5671 sd_ctl_dir[0].child = entry;
5676 for_each_possible_cpu(i) {
5677 snprintf(buf, 32, "cpu%d", i);
5678 entry->procname = kstrdup(buf, GFP_KERNEL);
5680 entry->child = sd_alloc_ctl_cpu_table(i);
5684 WARN_ON(sd_sysctl_header);
5685 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5688 /* may be called multiple times per register */
5689 static void unregister_sched_domain_sysctl(void)
5691 if (sd_sysctl_header)
5692 unregister_sysctl_table(sd_sysctl_header);
5693 sd_sysctl_header = NULL;
5694 if (sd_ctl_dir[0].child)
5695 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5698 static void register_sched_domain_sysctl(void)
5701 static void unregister_sched_domain_sysctl(void)
5706 static void set_rq_online(struct rq *rq)
5709 const struct sched_class *class;
5711 cpumask_set_cpu(rq->cpu, rq->rd->online);
5714 for_each_class(class) {
5715 if (class->rq_online)
5716 class->rq_online(rq);
5721 static void set_rq_offline(struct rq *rq)
5724 const struct sched_class *class;
5726 for_each_class(class) {
5727 if (class->rq_offline)
5728 class->rq_offline(rq);
5731 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5737 * migration_call - callback that gets triggered when a CPU is added.
5738 * Here we can start up the necessary migration thread for the new CPU.
5740 static int __cpuinit
5741 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5743 int cpu = (long)hcpu;
5744 unsigned long flags;
5745 struct rq *rq = cpu_rq(cpu);
5749 case CPU_UP_PREPARE:
5750 case CPU_UP_PREPARE_FROZEN:
5751 rq->calc_load_update = calc_load_update;
5755 case CPU_ONLINE_FROZEN:
5756 /* Update our root-domain */
5757 raw_spin_lock_irqsave(&rq->lock, flags);
5759 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5763 raw_spin_unlock_irqrestore(&rq->lock, flags);
5766 #ifdef CONFIG_HOTPLUG_CPU
5768 case CPU_DEAD_FROZEN:
5769 migrate_live_tasks(cpu);
5770 /* Idle task back to normal (off runqueue, low prio) */
5771 raw_spin_lock_irq(&rq->lock);
5772 deactivate_task(rq, rq->idle, 0);
5773 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5774 rq->idle->sched_class = &idle_sched_class;
5775 migrate_dead_tasks(cpu);
5776 raw_spin_unlock_irq(&rq->lock);
5777 migrate_nr_uninterruptible(rq);
5778 BUG_ON(rq->nr_running != 0);
5779 calc_global_load_remove(rq);
5783 case CPU_DYING_FROZEN:
5784 /* Update our root-domain */
5785 raw_spin_lock_irqsave(&rq->lock, flags);
5787 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5790 raw_spin_unlock_irqrestore(&rq->lock, flags);
5798 * Register at high priority so that task migration (migrate_all_tasks)
5799 * happens before everything else. This has to be lower priority than
5800 * the notifier in the perf_event subsystem, though.
5802 static struct notifier_block __cpuinitdata migration_notifier = {
5803 .notifier_call = migration_call,
5807 static int __init migration_init(void)
5809 void *cpu = (void *)(long)smp_processor_id();
5812 /* Start one for the boot CPU: */
5813 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5814 BUG_ON(err == NOTIFY_BAD);
5815 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5816 register_cpu_notifier(&migration_notifier);
5820 early_initcall(migration_init);
5825 #ifdef CONFIG_SCHED_DEBUG
5827 static __read_mostly int sched_domain_debug_enabled;
5829 static int __init sched_domain_debug_setup(char *str)
5831 sched_domain_debug_enabled = 1;
5835 early_param("sched_debug", sched_domain_debug_setup);
5837 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5838 struct cpumask *groupmask)
5840 struct sched_group *group = sd->groups;
5843 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5844 cpumask_clear(groupmask);
5846 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5848 if (!(sd->flags & SD_LOAD_BALANCE)) {
5849 printk("does not load-balance\n");
5851 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5856 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5858 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5859 printk(KERN_ERR "ERROR: domain->span does not contain "
5862 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5863 printk(KERN_ERR "ERROR: domain->groups does not contain"
5867 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5871 printk(KERN_ERR "ERROR: group is NULL\n");
5875 if (!group->cpu_power) {
5876 printk(KERN_CONT "\n");
5877 printk(KERN_ERR "ERROR: domain->cpu_power not "
5882 if (!cpumask_weight(sched_group_cpus(group))) {
5883 printk(KERN_CONT "\n");
5884 printk(KERN_ERR "ERROR: empty group\n");
5888 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
5889 printk(KERN_CONT "\n");
5890 printk(KERN_ERR "ERROR: repeated CPUs\n");
5894 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5896 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5898 printk(KERN_CONT " %s", str);
5899 if (group->cpu_power != SCHED_LOAD_SCALE) {
5900 printk(KERN_CONT " (cpu_power = %d)",
5904 group = group->next;
5905 } while (group != sd->groups);
5906 printk(KERN_CONT "\n");
5908 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5909 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5912 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5913 printk(KERN_ERR "ERROR: parent span is not a superset "
5914 "of domain->span\n");
5918 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5920 cpumask_var_t groupmask;
5923 if (!sched_domain_debug_enabled)
5927 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5931 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5933 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
5934 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
5939 if (sched_domain_debug_one(sd, cpu, level, groupmask))
5946 free_cpumask_var(groupmask);
5948 #else /* !CONFIG_SCHED_DEBUG */
5949 # define sched_domain_debug(sd, cpu) do { } while (0)
5950 #endif /* CONFIG_SCHED_DEBUG */
5952 static int sd_degenerate(struct sched_domain *sd)
5954 if (cpumask_weight(sched_domain_span(sd)) == 1)
5957 /* Following flags need at least 2 groups */
5958 if (sd->flags & (SD_LOAD_BALANCE |
5959 SD_BALANCE_NEWIDLE |
5963 SD_SHARE_PKG_RESOURCES)) {
5964 if (sd->groups != sd->groups->next)
5968 /* Following flags don't use groups */
5969 if (sd->flags & (SD_WAKE_AFFINE))
5976 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5978 unsigned long cflags = sd->flags, pflags = parent->flags;
5980 if (sd_degenerate(parent))
5983 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5986 /* Flags needing groups don't count if only 1 group in parent */
5987 if (parent->groups == parent->groups->next) {
5988 pflags &= ~(SD_LOAD_BALANCE |
5989 SD_BALANCE_NEWIDLE |
5993 SD_SHARE_PKG_RESOURCES);
5994 if (nr_node_ids == 1)
5995 pflags &= ~SD_SERIALIZE;
5997 if (~cflags & pflags)
6003 static void free_rootdomain(struct root_domain *rd)
6005 synchronize_sched();
6007 cpupri_cleanup(&rd->cpupri);
6009 free_cpumask_var(rd->rto_mask);
6010 free_cpumask_var(rd->online);
6011 free_cpumask_var(rd->span);
6015 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6017 struct root_domain *old_rd = NULL;
6018 unsigned long flags;
6020 raw_spin_lock_irqsave(&rq->lock, flags);
6025 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6028 cpumask_clear_cpu(rq->cpu, old_rd->span);
6031 * If we dont want to free the old_rt yet then
6032 * set old_rd to NULL to skip the freeing later
6035 if (!atomic_dec_and_test(&old_rd->refcount))
6039 atomic_inc(&rd->refcount);
6042 cpumask_set_cpu(rq->cpu, rd->span);
6043 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6046 raw_spin_unlock_irqrestore(&rq->lock, flags);
6049 free_rootdomain(old_rd);
6052 static int init_rootdomain(struct root_domain *rd, bool bootmem)
6054 gfp_t gfp = GFP_KERNEL;
6056 memset(rd, 0, sizeof(*rd));
6061 if (!alloc_cpumask_var(&rd->span, gfp))
6063 if (!alloc_cpumask_var(&rd->online, gfp))
6065 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
6068 if (cpupri_init(&rd->cpupri, bootmem) != 0)
6073 free_cpumask_var(rd->rto_mask);
6075 free_cpumask_var(rd->online);
6077 free_cpumask_var(rd->span);
6082 static void init_defrootdomain(void)
6084 init_rootdomain(&def_root_domain, true);
6086 atomic_set(&def_root_domain.refcount, 1);
6089 static struct root_domain *alloc_rootdomain(void)
6091 struct root_domain *rd;
6093 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6097 if (init_rootdomain(rd, false) != 0) {
6106 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6107 * hold the hotplug lock.
6110 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6112 struct rq *rq = cpu_rq(cpu);
6113 struct sched_domain *tmp;
6115 for (tmp = sd; tmp; tmp = tmp->parent)
6116 tmp->span_weight = cpumask_weight(sched_domain_span(tmp));
6118 /* Remove the sched domains which do not contribute to scheduling. */
6119 for (tmp = sd; tmp; ) {
6120 struct sched_domain *parent = tmp->parent;
6124 if (sd_parent_degenerate(tmp, parent)) {
6125 tmp->parent = parent->parent;
6127 parent->parent->child = tmp;
6132 if (sd && sd_degenerate(sd)) {
6138 sched_domain_debug(sd, cpu);
6140 rq_attach_root(rq, rd);
6141 rcu_assign_pointer(rq->sd, sd);
6144 /* cpus with isolated domains */
6145 static cpumask_var_t cpu_isolated_map;
6147 /* Setup the mask of cpus configured for isolated domains */
6148 static int __init isolated_cpu_setup(char *str)
6150 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6151 cpulist_parse(str, cpu_isolated_map);
6155 __setup("isolcpus=", isolated_cpu_setup);
6158 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6159 * to a function which identifies what group(along with sched group) a CPU
6160 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6161 * (due to the fact that we keep track of groups covered with a struct cpumask).
6163 * init_sched_build_groups will build a circular linked list of the groups
6164 * covered by the given span, and will set each group's ->cpumask correctly,
6165 * and ->cpu_power to 0.
6168 init_sched_build_groups(const struct cpumask *span,
6169 const struct cpumask *cpu_map,
6170 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
6171 struct sched_group **sg,
6172 struct cpumask *tmpmask),
6173 struct cpumask *covered, struct cpumask *tmpmask)
6175 struct sched_group *first = NULL, *last = NULL;
6178 cpumask_clear(covered);
6180 for_each_cpu(i, span) {
6181 struct sched_group *sg;
6182 int group = group_fn(i, cpu_map, &sg, tmpmask);
6185 if (cpumask_test_cpu(i, covered))
6188 cpumask_clear(sched_group_cpus(sg));
6191 for_each_cpu(j, span) {
6192 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6195 cpumask_set_cpu(j, covered);
6196 cpumask_set_cpu(j, sched_group_cpus(sg));
6207 #define SD_NODES_PER_DOMAIN 16
6212 * find_next_best_node - find the next node to include in a sched_domain
6213 * @node: node whose sched_domain we're building
6214 * @used_nodes: nodes already in the sched_domain
6216 * Find the next node to include in a given scheduling domain. Simply
6217 * finds the closest node not already in the @used_nodes map.
6219 * Should use nodemask_t.
6221 static int find_next_best_node(int node, nodemask_t *used_nodes)
6223 int i, n, val, min_val, best_node = 0;
6227 for (i = 0; i < nr_node_ids; i++) {
6228 /* Start at @node */
6229 n = (node + i) % nr_node_ids;
6231 if (!nr_cpus_node(n))
6234 /* Skip already used nodes */
6235 if (node_isset(n, *used_nodes))
6238 /* Simple min distance search */
6239 val = node_distance(node, n);
6241 if (val < min_val) {
6247 node_set(best_node, *used_nodes);
6252 * sched_domain_node_span - get a cpumask for a node's sched_domain
6253 * @node: node whose cpumask we're constructing
6254 * @span: resulting cpumask
6256 * Given a node, construct a good cpumask for its sched_domain to span. It
6257 * should be one that prevents unnecessary balancing, but also spreads tasks
6260 static void sched_domain_node_span(int node, struct cpumask *span)
6262 nodemask_t used_nodes;
6265 cpumask_clear(span);
6266 nodes_clear(used_nodes);
6268 cpumask_or(span, span, cpumask_of_node(node));
6269 node_set(node, used_nodes);
6271 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6272 int next_node = find_next_best_node(node, &used_nodes);
6274 cpumask_or(span, span, cpumask_of_node(next_node));
6277 #endif /* CONFIG_NUMA */
6279 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6282 * The cpus mask in sched_group and sched_domain hangs off the end.
6284 * ( See the the comments in include/linux/sched.h:struct sched_group
6285 * and struct sched_domain. )
6287 struct static_sched_group {
6288 struct sched_group sg;
6289 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
6292 struct static_sched_domain {
6293 struct sched_domain sd;
6294 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
6300 cpumask_var_t domainspan;
6301 cpumask_var_t covered;
6302 cpumask_var_t notcovered;
6304 cpumask_var_t nodemask;
6305 cpumask_var_t this_sibling_map;
6306 cpumask_var_t this_core_map;
6307 cpumask_var_t send_covered;
6308 cpumask_var_t tmpmask;
6309 struct sched_group **sched_group_nodes;
6310 struct root_domain *rd;
6314 sa_sched_groups = 0,
6319 sa_this_sibling_map,
6321 sa_sched_group_nodes,
6331 * SMT sched-domains:
6333 #ifdef CONFIG_SCHED_SMT
6334 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
6335 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
6338 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
6339 struct sched_group **sg, struct cpumask *unused)
6342 *sg = &per_cpu(sched_groups, cpu).sg;
6345 #endif /* CONFIG_SCHED_SMT */
6348 * multi-core sched-domains:
6350 #ifdef CONFIG_SCHED_MC
6351 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
6352 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
6353 #endif /* CONFIG_SCHED_MC */
6355 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6357 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6358 struct sched_group **sg, struct cpumask *mask)
6362 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6363 group = cpumask_first(mask);
6365 *sg = &per_cpu(sched_group_core, group).sg;
6368 #elif defined(CONFIG_SCHED_MC)
6370 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6371 struct sched_group **sg, struct cpumask *unused)
6374 *sg = &per_cpu(sched_group_core, cpu).sg;
6379 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
6380 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
6383 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
6384 struct sched_group **sg, struct cpumask *mask)
6387 #ifdef CONFIG_SCHED_MC
6388 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6389 group = cpumask_first(mask);
6390 #elif defined(CONFIG_SCHED_SMT)
6391 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6392 group = cpumask_first(mask);
6397 *sg = &per_cpu(sched_group_phys, group).sg;
6403 * The init_sched_build_groups can't handle what we want to do with node
6404 * groups, so roll our own. Now each node has its own list of groups which
6405 * gets dynamically allocated.
6407 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
6408 static struct sched_group ***sched_group_nodes_bycpu;
6410 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
6411 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
6413 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
6414 struct sched_group **sg,
6415 struct cpumask *nodemask)
6419 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
6420 group = cpumask_first(nodemask);
6423 *sg = &per_cpu(sched_group_allnodes, group).sg;
6427 static void init_numa_sched_groups_power(struct sched_group *group_head)
6429 struct sched_group *sg = group_head;
6435 for_each_cpu(j, sched_group_cpus(sg)) {
6436 struct sched_domain *sd;
6438 sd = &per_cpu(phys_domains, j).sd;
6439 if (j != group_first_cpu(sd->groups)) {
6441 * Only add "power" once for each
6447 sg->cpu_power += sd->groups->cpu_power;
6450 } while (sg != group_head);
6453 static int build_numa_sched_groups(struct s_data *d,
6454 const struct cpumask *cpu_map, int num)
6456 struct sched_domain *sd;
6457 struct sched_group *sg, *prev;
6460 cpumask_clear(d->covered);
6461 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
6462 if (cpumask_empty(d->nodemask)) {
6463 d->sched_group_nodes[num] = NULL;
6467 sched_domain_node_span(num, d->domainspan);
6468 cpumask_and(d->domainspan, d->domainspan, cpu_map);
6470 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6473 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
6477 d->sched_group_nodes[num] = sg;
6479 for_each_cpu(j, d->nodemask) {
6480 sd = &per_cpu(node_domains, j).sd;
6485 cpumask_copy(sched_group_cpus(sg), d->nodemask);
6487 cpumask_or(d->covered, d->covered, d->nodemask);
6490 for (j = 0; j < nr_node_ids; j++) {
6491 n = (num + j) % nr_node_ids;
6492 cpumask_complement(d->notcovered, d->covered);
6493 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
6494 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
6495 if (cpumask_empty(d->tmpmask))
6497 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
6498 if (cpumask_empty(d->tmpmask))
6500 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6504 "Can not alloc domain group for node %d\n", j);
6508 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
6509 sg->next = prev->next;
6510 cpumask_or(d->covered, d->covered, d->tmpmask);
6517 #endif /* CONFIG_NUMA */
6520 /* Free memory allocated for various sched_group structures */
6521 static void free_sched_groups(const struct cpumask *cpu_map,
6522 struct cpumask *nodemask)
6526 for_each_cpu(cpu, cpu_map) {
6527 struct sched_group **sched_group_nodes
6528 = sched_group_nodes_bycpu[cpu];
6530 if (!sched_group_nodes)
6533 for (i = 0; i < nr_node_ids; i++) {
6534 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6536 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
6537 if (cpumask_empty(nodemask))
6547 if (oldsg != sched_group_nodes[i])
6550 kfree(sched_group_nodes);
6551 sched_group_nodes_bycpu[cpu] = NULL;
6554 #else /* !CONFIG_NUMA */
6555 static void free_sched_groups(const struct cpumask *cpu_map,
6556 struct cpumask *nodemask)
6559 #endif /* CONFIG_NUMA */
6562 * Initialize sched groups cpu_power.
6564 * cpu_power indicates the capacity of sched group, which is used while
6565 * distributing the load between different sched groups in a sched domain.
6566 * Typically cpu_power for all the groups in a sched domain will be same unless
6567 * there are asymmetries in the topology. If there are asymmetries, group
6568 * having more cpu_power will pickup more load compared to the group having
6571 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6573 struct sched_domain *child;
6574 struct sched_group *group;
6578 WARN_ON(!sd || !sd->groups);
6580 if (cpu != group_first_cpu(sd->groups))
6585 sd->groups->cpu_power = 0;
6588 power = SCHED_LOAD_SCALE;
6589 weight = cpumask_weight(sched_domain_span(sd));
6591 * SMT siblings share the power of a single core.
6592 * Usually multiple threads get a better yield out of
6593 * that one core than a single thread would have,
6594 * reflect that in sd->smt_gain.
6596 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
6597 power *= sd->smt_gain;
6599 power >>= SCHED_LOAD_SHIFT;
6601 sd->groups->cpu_power += power;
6606 * Add cpu_power of each child group to this groups cpu_power.
6608 group = child->groups;
6610 sd->groups->cpu_power += group->cpu_power;
6611 group = group->next;
6612 } while (group != child->groups);
6616 * Initializers for schedule domains
6617 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6620 #ifdef CONFIG_SCHED_DEBUG
6621 # define SD_INIT_NAME(sd, type) sd->name = #type
6623 # define SD_INIT_NAME(sd, type) do { } while (0)
6626 #define SD_INIT(sd, type) sd_init_##type(sd)
6628 #define SD_INIT_FUNC(type) \
6629 static noinline void sd_init_##type(struct sched_domain *sd) \
6631 memset(sd, 0, sizeof(*sd)); \
6632 *sd = SD_##type##_INIT; \
6633 sd->level = SD_LV_##type; \
6634 SD_INIT_NAME(sd, type); \
6639 SD_INIT_FUNC(ALLNODES)
6642 #ifdef CONFIG_SCHED_SMT
6643 SD_INIT_FUNC(SIBLING)
6645 #ifdef CONFIG_SCHED_MC
6649 static int default_relax_domain_level = -1;
6651 static int __init setup_relax_domain_level(char *str)
6655 val = simple_strtoul(str, NULL, 0);
6656 if (val < SD_LV_MAX)
6657 default_relax_domain_level = val;
6661 __setup("relax_domain_level=", setup_relax_domain_level);
6663 static void set_domain_attribute(struct sched_domain *sd,
6664 struct sched_domain_attr *attr)
6668 if (!attr || attr->relax_domain_level < 0) {
6669 if (default_relax_domain_level < 0)
6672 request = default_relax_domain_level;
6674 request = attr->relax_domain_level;
6675 if (request < sd->level) {
6676 /* turn off idle balance on this domain */
6677 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6679 /* turn on idle balance on this domain */
6680 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6684 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6685 const struct cpumask *cpu_map)
6688 case sa_sched_groups:
6689 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
6690 d->sched_group_nodes = NULL;
6692 free_rootdomain(d->rd); /* fall through */
6694 free_cpumask_var(d->tmpmask); /* fall through */
6695 case sa_send_covered:
6696 free_cpumask_var(d->send_covered); /* fall through */
6697 case sa_this_core_map:
6698 free_cpumask_var(d->this_core_map); /* fall through */
6699 case sa_this_sibling_map:
6700 free_cpumask_var(d->this_sibling_map); /* fall through */
6702 free_cpumask_var(d->nodemask); /* fall through */
6703 case sa_sched_group_nodes:
6705 kfree(d->sched_group_nodes); /* fall through */
6707 free_cpumask_var(d->notcovered); /* fall through */
6709 free_cpumask_var(d->covered); /* fall through */
6711 free_cpumask_var(d->domainspan); /* fall through */
6718 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6719 const struct cpumask *cpu_map)
6722 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
6724 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
6725 return sa_domainspan;
6726 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
6728 /* Allocate the per-node list of sched groups */
6729 d->sched_group_nodes = kcalloc(nr_node_ids,
6730 sizeof(struct sched_group *), GFP_KERNEL);
6731 if (!d->sched_group_nodes) {
6732 printk(KERN_WARNING "Can not alloc sched group node list\n");
6733 return sa_notcovered;
6735 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
6737 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
6738 return sa_sched_group_nodes;
6739 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
6741 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
6742 return sa_this_sibling_map;
6743 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
6744 return sa_this_core_map;
6745 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
6746 return sa_send_covered;
6747 d->rd = alloc_rootdomain();
6749 printk(KERN_WARNING "Cannot alloc root domain\n");
6752 return sa_rootdomain;
6755 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
6756 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
6758 struct sched_domain *sd = NULL;
6760 struct sched_domain *parent;
6763 if (cpumask_weight(cpu_map) >
6764 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
6765 sd = &per_cpu(allnodes_domains, i).sd;
6766 SD_INIT(sd, ALLNODES);
6767 set_domain_attribute(sd, attr);
6768 cpumask_copy(sched_domain_span(sd), cpu_map);
6769 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
6774 sd = &per_cpu(node_domains, i).sd;
6776 set_domain_attribute(sd, attr);
6777 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
6778 sd->parent = parent;
6781 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
6786 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
6787 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6788 struct sched_domain *parent, int i)
6790 struct sched_domain *sd;
6791 sd = &per_cpu(phys_domains, i).sd;
6793 set_domain_attribute(sd, attr);
6794 cpumask_copy(sched_domain_span(sd), d->nodemask);
6795 sd->parent = parent;
6798 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
6802 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
6803 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6804 struct sched_domain *parent, int i)
6806 struct sched_domain *sd = parent;
6807 #ifdef CONFIG_SCHED_MC
6808 sd = &per_cpu(core_domains, i).sd;
6810 set_domain_attribute(sd, attr);
6811 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
6812 sd->parent = parent;
6814 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
6819 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
6820 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6821 struct sched_domain *parent, int i)
6823 struct sched_domain *sd = parent;
6824 #ifdef CONFIG_SCHED_SMT
6825 sd = &per_cpu(cpu_domains, i).sd;
6826 SD_INIT(sd, SIBLING);
6827 set_domain_attribute(sd, attr);
6828 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
6829 sd->parent = parent;
6831 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
6836 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
6837 const struct cpumask *cpu_map, int cpu)
6840 #ifdef CONFIG_SCHED_SMT
6841 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
6842 cpumask_and(d->this_sibling_map, cpu_map,
6843 topology_thread_cpumask(cpu));
6844 if (cpu == cpumask_first(d->this_sibling_map))
6845 init_sched_build_groups(d->this_sibling_map, cpu_map,
6847 d->send_covered, d->tmpmask);
6850 #ifdef CONFIG_SCHED_MC
6851 case SD_LV_MC: /* set up multi-core groups */
6852 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
6853 if (cpu == cpumask_first(d->this_core_map))
6854 init_sched_build_groups(d->this_core_map, cpu_map,
6856 d->send_covered, d->tmpmask);
6859 case SD_LV_CPU: /* set up physical groups */
6860 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
6861 if (!cpumask_empty(d->nodemask))
6862 init_sched_build_groups(d->nodemask, cpu_map,
6864 d->send_covered, d->tmpmask);
6867 case SD_LV_ALLNODES:
6868 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
6869 d->send_covered, d->tmpmask);
6878 * Build sched domains for a given set of cpus and attach the sched domains
6879 * to the individual cpus
6881 static int __build_sched_domains(const struct cpumask *cpu_map,
6882 struct sched_domain_attr *attr)
6884 enum s_alloc alloc_state = sa_none;
6886 struct sched_domain *sd;
6892 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6893 if (alloc_state != sa_rootdomain)
6895 alloc_state = sa_sched_groups;
6898 * Set up domains for cpus specified by the cpu_map.
6900 for_each_cpu(i, cpu_map) {
6901 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
6904 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
6905 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
6906 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
6907 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
6910 for_each_cpu(i, cpu_map) {
6911 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
6912 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
6915 /* Set up physical groups */
6916 for (i = 0; i < nr_node_ids; i++)
6917 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
6920 /* Set up node groups */
6922 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
6924 for (i = 0; i < nr_node_ids; i++)
6925 if (build_numa_sched_groups(&d, cpu_map, i))
6929 /* Calculate CPU power for physical packages and nodes */
6930 #ifdef CONFIG_SCHED_SMT
6931 for_each_cpu(i, cpu_map) {
6932 sd = &per_cpu(cpu_domains, i).sd;
6933 init_sched_groups_power(i, sd);
6936 #ifdef CONFIG_SCHED_MC
6937 for_each_cpu(i, cpu_map) {
6938 sd = &per_cpu(core_domains, i).sd;
6939 init_sched_groups_power(i, sd);
6943 for_each_cpu(i, cpu_map) {
6944 sd = &per_cpu(phys_domains, i).sd;
6945 init_sched_groups_power(i, sd);
6949 for (i = 0; i < nr_node_ids; i++)
6950 init_numa_sched_groups_power(d.sched_group_nodes[i]);
6952 if (d.sd_allnodes) {
6953 struct sched_group *sg;
6955 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
6957 init_numa_sched_groups_power(sg);
6961 /* Attach the domains */
6962 for_each_cpu(i, cpu_map) {
6963 #ifdef CONFIG_SCHED_SMT
6964 sd = &per_cpu(cpu_domains, i).sd;
6965 #elif defined(CONFIG_SCHED_MC)
6966 sd = &per_cpu(core_domains, i).sd;
6968 sd = &per_cpu(phys_domains, i).sd;
6970 cpu_attach_domain(sd, d.rd, i);
6973 d.sched_group_nodes = NULL; /* don't free this we still need it */
6974 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
6978 __free_domain_allocs(&d, alloc_state, cpu_map);
6982 static int build_sched_domains(const struct cpumask *cpu_map)
6984 return __build_sched_domains(cpu_map, NULL);
6987 static cpumask_var_t *doms_cur; /* current sched domains */
6988 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6989 static struct sched_domain_attr *dattr_cur;
6990 /* attribues of custom domains in 'doms_cur' */
6993 * Special case: If a kmalloc of a doms_cur partition (array of
6994 * cpumask) fails, then fallback to a single sched domain,
6995 * as determined by the single cpumask fallback_doms.
6997 static cpumask_var_t fallback_doms;
7000 * arch_update_cpu_topology lets virtualized architectures update the
7001 * cpu core maps. It is supposed to return 1 if the topology changed
7002 * or 0 if it stayed the same.
7004 int __attribute__((weak)) arch_update_cpu_topology(void)
7009 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7012 cpumask_var_t *doms;
7014 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7017 for (i = 0; i < ndoms; i++) {
7018 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7019 free_sched_domains(doms, i);
7026 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7029 for (i = 0; i < ndoms; i++)
7030 free_cpumask_var(doms[i]);
7035 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7036 * For now this just excludes isolated cpus, but could be used to
7037 * exclude other special cases in the future.
7039 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7043 arch_update_cpu_topology();
7045 doms_cur = alloc_sched_domains(ndoms_cur);
7047 doms_cur = &fallback_doms;
7048 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7050 err = build_sched_domains(doms_cur[0]);
7051 register_sched_domain_sysctl();
7056 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7057 struct cpumask *tmpmask)
7059 free_sched_groups(cpu_map, tmpmask);
7063 * Detach sched domains from a group of cpus specified in cpu_map
7064 * These cpus will now be attached to the NULL domain
7066 static void detach_destroy_domains(const struct cpumask *cpu_map)
7068 /* Save because hotplug lock held. */
7069 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7072 for_each_cpu(i, cpu_map)
7073 cpu_attach_domain(NULL, &def_root_domain, i);
7074 synchronize_sched();
7075 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7078 /* handle null as "default" */
7079 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7080 struct sched_domain_attr *new, int idx_new)
7082 struct sched_domain_attr tmp;
7089 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7090 new ? (new + idx_new) : &tmp,
7091 sizeof(struct sched_domain_attr));
7095 * Partition sched domains as specified by the 'ndoms_new'
7096 * cpumasks in the array doms_new[] of cpumasks. This compares
7097 * doms_new[] to the current sched domain partitioning, doms_cur[].
7098 * It destroys each deleted domain and builds each new domain.
7100 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7101 * The masks don't intersect (don't overlap.) We should setup one
7102 * sched domain for each mask. CPUs not in any of the cpumasks will
7103 * not be load balanced. If the same cpumask appears both in the
7104 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7107 * The passed in 'doms_new' should be allocated using
7108 * alloc_sched_domains. This routine takes ownership of it and will
7109 * free_sched_domains it when done with it. If the caller failed the
7110 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7111 * and partition_sched_domains() will fallback to the single partition
7112 * 'fallback_doms', it also forces the domains to be rebuilt.
7114 * If doms_new == NULL it will be replaced with cpu_online_mask.
7115 * ndoms_new == 0 is a special case for destroying existing domains,
7116 * and it will not create the default domain.
7118 * Call with hotplug lock held
7120 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7121 struct sched_domain_attr *dattr_new)
7126 mutex_lock(&sched_domains_mutex);
7128 /* always unregister in case we don't destroy any domains */
7129 unregister_sched_domain_sysctl();
7131 /* Let architecture update cpu core mappings. */
7132 new_topology = arch_update_cpu_topology();
7134 n = doms_new ? ndoms_new : 0;
7136 /* Destroy deleted domains */
7137 for (i = 0; i < ndoms_cur; i++) {
7138 for (j = 0; j < n && !new_topology; j++) {
7139 if (cpumask_equal(doms_cur[i], doms_new[j])
7140 && dattrs_equal(dattr_cur, i, dattr_new, j))
7143 /* no match - a current sched domain not in new doms_new[] */
7144 detach_destroy_domains(doms_cur[i]);
7149 if (doms_new == NULL) {
7151 doms_new = &fallback_doms;
7152 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7153 WARN_ON_ONCE(dattr_new);
7156 /* Build new domains */
7157 for (i = 0; i < ndoms_new; i++) {
7158 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7159 if (cpumask_equal(doms_new[i], doms_cur[j])
7160 && dattrs_equal(dattr_new, i, dattr_cur, j))
7163 /* no match - add a new doms_new */
7164 __build_sched_domains(doms_new[i],
7165 dattr_new ? dattr_new + i : NULL);
7170 /* Remember the new sched domains */
7171 if (doms_cur != &fallback_doms)
7172 free_sched_domains(doms_cur, ndoms_cur);
7173 kfree(dattr_cur); /* kfree(NULL) is safe */
7174 doms_cur = doms_new;
7175 dattr_cur = dattr_new;
7176 ndoms_cur = ndoms_new;
7178 register_sched_domain_sysctl();
7180 mutex_unlock(&sched_domains_mutex);
7183 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7184 static void arch_reinit_sched_domains(void)
7188 /* Destroy domains first to force the rebuild */
7189 partition_sched_domains(0, NULL, NULL);
7191 rebuild_sched_domains();
7195 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7197 unsigned int level = 0;
7199 if (sscanf(buf, "%u", &level) != 1)
7203 * level is always be positive so don't check for
7204 * level < POWERSAVINGS_BALANCE_NONE which is 0
7205 * What happens on 0 or 1 byte write,
7206 * need to check for count as well?
7209 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7213 sched_smt_power_savings = level;
7215 sched_mc_power_savings = level;
7217 arch_reinit_sched_domains();
7222 #ifdef CONFIG_SCHED_MC
7223 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7224 struct sysdev_class_attribute *attr,
7227 return sprintf(page, "%u\n", sched_mc_power_savings);
7229 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7230 struct sysdev_class_attribute *attr,
7231 const char *buf, size_t count)
7233 return sched_power_savings_store(buf, count, 0);
7235 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7236 sched_mc_power_savings_show,
7237 sched_mc_power_savings_store);
7240 #ifdef CONFIG_SCHED_SMT
7241 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7242 struct sysdev_class_attribute *attr,
7245 return sprintf(page, "%u\n", sched_smt_power_savings);
7247 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7248 struct sysdev_class_attribute *attr,
7249 const char *buf, size_t count)
7251 return sched_power_savings_store(buf, count, 1);
7253 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7254 sched_smt_power_savings_show,
7255 sched_smt_power_savings_store);
7258 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7262 #ifdef CONFIG_SCHED_SMT
7264 err = sysfs_create_file(&cls->kset.kobj,
7265 &attr_sched_smt_power_savings.attr);
7267 #ifdef CONFIG_SCHED_MC
7268 if (!err && mc_capable())
7269 err = sysfs_create_file(&cls->kset.kobj,
7270 &attr_sched_mc_power_savings.attr);
7274 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7276 #ifndef CONFIG_CPUSETS
7278 * Add online and remove offline CPUs from the scheduler domains.
7279 * When cpusets are enabled they take over this function.
7281 static int update_sched_domains(struct notifier_block *nfb,
7282 unsigned long action, void *hcpu)
7286 case CPU_ONLINE_FROZEN:
7287 case CPU_DOWN_PREPARE:
7288 case CPU_DOWN_PREPARE_FROZEN:
7289 case CPU_DOWN_FAILED:
7290 case CPU_DOWN_FAILED_FROZEN:
7291 partition_sched_domains(1, NULL, NULL);
7300 static int update_runtime(struct notifier_block *nfb,
7301 unsigned long action, void *hcpu)
7303 int cpu = (int)(long)hcpu;
7306 case CPU_DOWN_PREPARE:
7307 case CPU_DOWN_PREPARE_FROZEN:
7308 disable_runtime(cpu_rq(cpu));
7311 case CPU_DOWN_FAILED:
7312 case CPU_DOWN_FAILED_FROZEN:
7314 case CPU_ONLINE_FROZEN:
7315 enable_runtime(cpu_rq(cpu));
7323 void __init sched_init_smp(void)
7325 cpumask_var_t non_isolated_cpus;
7327 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7328 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7330 #if defined(CONFIG_NUMA)
7331 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7333 BUG_ON(sched_group_nodes_bycpu == NULL);
7336 mutex_lock(&sched_domains_mutex);
7337 arch_init_sched_domains(cpu_active_mask);
7338 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7339 if (cpumask_empty(non_isolated_cpus))
7340 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7341 mutex_unlock(&sched_domains_mutex);
7344 #ifndef CONFIG_CPUSETS
7345 /* XXX: Theoretical race here - CPU may be hotplugged now */
7346 hotcpu_notifier(update_sched_domains, 0);
7349 /* RT runtime code needs to handle some hotplug events */
7350 hotcpu_notifier(update_runtime, 0);
7354 /* Move init over to a non-isolated CPU */
7355 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7357 sched_init_granularity();
7358 free_cpumask_var(non_isolated_cpus);
7360 init_sched_rt_class();
7363 void __init sched_init_smp(void)
7365 sched_init_granularity();
7367 #endif /* CONFIG_SMP */
7369 const_debug unsigned int sysctl_timer_migration = 1;
7371 int in_sched_functions(unsigned long addr)
7373 return in_lock_functions(addr) ||
7374 (addr >= (unsigned long)__sched_text_start
7375 && addr < (unsigned long)__sched_text_end);
7378 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7380 cfs_rq->tasks_timeline = RB_ROOT;
7381 INIT_LIST_HEAD(&cfs_rq->tasks);
7382 #ifdef CONFIG_FAIR_GROUP_SCHED
7385 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7388 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7390 struct rt_prio_array *array;
7393 array = &rt_rq->active;
7394 for (i = 0; i < MAX_RT_PRIO; i++) {
7395 INIT_LIST_HEAD(array->queue + i);
7396 __clear_bit(i, array->bitmap);
7398 /* delimiter for bitsearch: */
7399 __set_bit(MAX_RT_PRIO, array->bitmap);
7401 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7402 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7404 rt_rq->highest_prio.next = MAX_RT_PRIO;
7408 rt_rq->rt_nr_migratory = 0;
7409 rt_rq->overloaded = 0;
7410 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
7414 rt_rq->rt_throttled = 0;
7415 rt_rq->rt_runtime = 0;
7416 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7418 #ifdef CONFIG_RT_GROUP_SCHED
7419 rt_rq->rt_nr_boosted = 0;
7424 #ifdef CONFIG_FAIR_GROUP_SCHED
7425 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7426 struct sched_entity *se, int cpu, int add,
7427 struct sched_entity *parent)
7429 struct rq *rq = cpu_rq(cpu);
7430 tg->cfs_rq[cpu] = cfs_rq;
7431 init_cfs_rq(cfs_rq, rq);
7434 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7437 /* se could be NULL for init_task_group */
7442 se->cfs_rq = &rq->cfs;
7444 se->cfs_rq = parent->my_q;
7447 se->load.weight = tg->shares;
7448 se->load.inv_weight = 0;
7449 se->parent = parent;
7453 #ifdef CONFIG_RT_GROUP_SCHED
7454 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7455 struct sched_rt_entity *rt_se, int cpu, int add,
7456 struct sched_rt_entity *parent)
7458 struct rq *rq = cpu_rq(cpu);
7460 tg->rt_rq[cpu] = rt_rq;
7461 init_rt_rq(rt_rq, rq);
7463 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7465 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7467 tg->rt_se[cpu] = rt_se;
7472 rt_se->rt_rq = &rq->rt;
7474 rt_se->rt_rq = parent->my_q;
7476 rt_se->my_q = rt_rq;
7477 rt_se->parent = parent;
7478 INIT_LIST_HEAD(&rt_se->run_list);
7482 void __init sched_init(void)
7485 unsigned long alloc_size = 0, ptr;
7487 #ifdef CONFIG_FAIR_GROUP_SCHED
7488 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7490 #ifdef CONFIG_RT_GROUP_SCHED
7491 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7493 #ifdef CONFIG_CPUMASK_OFFSTACK
7494 alloc_size += num_possible_cpus() * cpumask_size();
7497 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7499 #ifdef CONFIG_FAIR_GROUP_SCHED
7500 init_task_group.se = (struct sched_entity **)ptr;
7501 ptr += nr_cpu_ids * sizeof(void **);
7503 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7504 ptr += nr_cpu_ids * sizeof(void **);
7506 #endif /* CONFIG_FAIR_GROUP_SCHED */
7507 #ifdef CONFIG_RT_GROUP_SCHED
7508 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7509 ptr += nr_cpu_ids * sizeof(void **);
7511 init_task_group.rt_rq = (struct rt_rq **)ptr;
7512 ptr += nr_cpu_ids * sizeof(void **);
7514 #endif /* CONFIG_RT_GROUP_SCHED */
7515 #ifdef CONFIG_CPUMASK_OFFSTACK
7516 for_each_possible_cpu(i) {
7517 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
7518 ptr += cpumask_size();
7520 #endif /* CONFIG_CPUMASK_OFFSTACK */
7524 init_defrootdomain();
7527 init_rt_bandwidth(&def_rt_bandwidth,
7528 global_rt_period(), global_rt_runtime());
7530 #ifdef CONFIG_RT_GROUP_SCHED
7531 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7532 global_rt_period(), global_rt_runtime());
7533 #endif /* CONFIG_RT_GROUP_SCHED */
7535 #ifdef CONFIG_CGROUP_SCHED
7536 list_add(&init_task_group.list, &task_groups);
7537 INIT_LIST_HEAD(&init_task_group.children);
7539 #endif /* CONFIG_CGROUP_SCHED */
7541 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
7542 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
7543 __alignof__(unsigned long));
7545 for_each_possible_cpu(i) {
7549 raw_spin_lock_init(&rq->lock);
7551 rq->calc_load_active = 0;
7552 rq->calc_load_update = jiffies + LOAD_FREQ;
7553 init_cfs_rq(&rq->cfs, rq);
7554 init_rt_rq(&rq->rt, rq);
7555 #ifdef CONFIG_FAIR_GROUP_SCHED
7556 init_task_group.shares = init_task_group_load;
7557 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7558 #ifdef CONFIG_CGROUP_SCHED
7560 * How much cpu bandwidth does init_task_group get?
7562 * In case of task-groups formed thr' the cgroup filesystem, it
7563 * gets 100% of the cpu resources in the system. This overall
7564 * system cpu resource is divided among the tasks of
7565 * init_task_group and its child task-groups in a fair manner,
7566 * based on each entity's (task or task-group's) weight
7567 * (se->load.weight).
7569 * In other words, if init_task_group has 10 tasks of weight
7570 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7571 * then A0's share of the cpu resource is:
7573 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7575 * We achieve this by letting init_task_group's tasks sit
7576 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7578 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
7580 #endif /* CONFIG_FAIR_GROUP_SCHED */
7582 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7583 #ifdef CONFIG_RT_GROUP_SCHED
7584 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7585 #ifdef CONFIG_CGROUP_SCHED
7586 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
7590 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7591 rq->cpu_load[j] = 0;
7595 rq->cpu_power = SCHED_LOAD_SCALE;
7596 rq->post_schedule = 0;
7597 rq->active_balance = 0;
7598 rq->next_balance = jiffies;
7603 rq->avg_idle = 2*sysctl_sched_migration_cost;
7604 rq_attach_root(rq, &def_root_domain);
7607 atomic_set(&rq->nr_iowait, 0);
7610 set_load_weight(&init_task);
7612 #ifdef CONFIG_PREEMPT_NOTIFIERS
7613 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7617 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7620 #ifdef CONFIG_RT_MUTEXES
7621 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
7625 * The boot idle thread does lazy MMU switching as well:
7627 atomic_inc(&init_mm.mm_count);
7628 enter_lazy_tlb(&init_mm, current);
7631 * Make us the idle thread. Technically, schedule() should not be
7632 * called from this thread, however somewhere below it might be,
7633 * but because we are the idle thread, we just pick up running again
7634 * when this runqueue becomes "idle".
7636 init_idle(current, smp_processor_id());
7638 calc_load_update = jiffies + LOAD_FREQ;
7641 * During early bootup we pretend to be a normal task:
7643 current->sched_class = &fair_sched_class;
7645 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
7646 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
7649 zalloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
7650 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
7652 /* May be allocated at isolcpus cmdline parse time */
7653 if (cpu_isolated_map == NULL)
7654 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7659 scheduler_running = 1;
7662 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7663 static inline int preempt_count_equals(int preempt_offset)
7665 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7667 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
7670 void __might_sleep(const char *file, int line, int preempt_offset)
7673 static unsigned long prev_jiffy; /* ratelimiting */
7675 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7676 system_state != SYSTEM_RUNNING || oops_in_progress)
7678 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7680 prev_jiffy = jiffies;
7683 "BUG: sleeping function called from invalid context at %s:%d\n",
7686 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7687 in_atomic(), irqs_disabled(),
7688 current->pid, current->comm);
7690 debug_show_held_locks(current);
7691 if (irqs_disabled())
7692 print_irqtrace_events(current);
7696 EXPORT_SYMBOL(__might_sleep);
7699 #ifdef CONFIG_MAGIC_SYSRQ
7700 static void normalize_task(struct rq *rq, struct task_struct *p)
7704 on_rq = p->se.on_rq;
7706 deactivate_task(rq, p, 0);
7707 __setscheduler(rq, p, SCHED_NORMAL, 0);
7709 activate_task(rq, p, 0);
7710 resched_task(rq->curr);
7714 void normalize_rt_tasks(void)
7716 struct task_struct *g, *p;
7717 unsigned long flags;
7720 read_lock_irqsave(&tasklist_lock, flags);
7721 do_each_thread(g, p) {
7723 * Only normalize user tasks:
7728 p->se.exec_start = 0;
7729 #ifdef CONFIG_SCHEDSTATS
7730 p->se.statistics.wait_start = 0;
7731 p->se.statistics.sleep_start = 0;
7732 p->se.statistics.block_start = 0;
7737 * Renice negative nice level userspace
7740 if (TASK_NICE(p) < 0 && p->mm)
7741 set_user_nice(p, 0);
7745 raw_spin_lock(&p->pi_lock);
7746 rq = __task_rq_lock(p);
7748 normalize_task(rq, p);
7750 __task_rq_unlock(rq);
7751 raw_spin_unlock(&p->pi_lock);
7752 } while_each_thread(g, p);
7754 read_unlock_irqrestore(&tasklist_lock, flags);
7757 #endif /* CONFIG_MAGIC_SYSRQ */
7759 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7761 * These functions are only useful for the IA64 MCA handling, or kdb.
7763 * They can only be called when the whole system has been
7764 * stopped - every CPU needs to be quiescent, and no scheduling
7765 * activity can take place. Using them for anything else would
7766 * be a serious bug, and as a result, they aren't even visible
7767 * under any other configuration.
7771 * curr_task - return the current task for a given cpu.
7772 * @cpu: the processor in question.
7774 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7776 struct task_struct *curr_task(int cpu)
7778 return cpu_curr(cpu);
7781 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7785 * set_curr_task - set the current task for a given cpu.
7786 * @cpu: the processor in question.
7787 * @p: the task pointer to set.
7789 * Description: This function must only be used when non-maskable interrupts
7790 * are serviced on a separate stack. It allows the architecture to switch the
7791 * notion of the current task on a cpu in a non-blocking manner. This function
7792 * must be called with all CPU's synchronized, and interrupts disabled, the
7793 * and caller must save the original value of the current task (see
7794 * curr_task() above) and restore that value before reenabling interrupts and
7795 * re-starting the system.
7797 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7799 void set_curr_task(int cpu, struct task_struct *p)
7806 #ifdef CONFIG_FAIR_GROUP_SCHED
7807 static void free_fair_sched_group(struct task_group *tg)
7811 for_each_possible_cpu(i) {
7813 kfree(tg->cfs_rq[i]);
7823 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7825 struct cfs_rq *cfs_rq;
7826 struct sched_entity *se;
7830 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7833 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7837 tg->shares = NICE_0_LOAD;
7839 for_each_possible_cpu(i) {
7842 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
7843 GFP_KERNEL, cpu_to_node(i));
7847 se = kzalloc_node(sizeof(struct sched_entity),
7848 GFP_KERNEL, cpu_to_node(i));
7852 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
7863 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
7865 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
7866 &cpu_rq(cpu)->leaf_cfs_rq_list);
7869 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
7871 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
7873 #else /* !CONFG_FAIR_GROUP_SCHED */
7874 static inline void free_fair_sched_group(struct task_group *tg)
7879 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7884 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
7888 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
7891 #endif /* CONFIG_FAIR_GROUP_SCHED */
7893 #ifdef CONFIG_RT_GROUP_SCHED
7894 static void free_rt_sched_group(struct task_group *tg)
7898 destroy_rt_bandwidth(&tg->rt_bandwidth);
7900 for_each_possible_cpu(i) {
7902 kfree(tg->rt_rq[i]);
7904 kfree(tg->rt_se[i]);
7912 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
7914 struct rt_rq *rt_rq;
7915 struct sched_rt_entity *rt_se;
7919 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
7922 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
7926 init_rt_bandwidth(&tg->rt_bandwidth,
7927 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
7929 for_each_possible_cpu(i) {
7932 rt_rq = kzalloc_node(sizeof(struct rt_rq),
7933 GFP_KERNEL, cpu_to_node(i));
7937 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
7938 GFP_KERNEL, cpu_to_node(i));
7942 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
7953 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
7955 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
7956 &cpu_rq(cpu)->leaf_rt_rq_list);
7959 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
7961 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
7963 #else /* !CONFIG_RT_GROUP_SCHED */
7964 static inline void free_rt_sched_group(struct task_group *tg)
7969 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
7974 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
7978 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
7981 #endif /* CONFIG_RT_GROUP_SCHED */
7983 #ifdef CONFIG_CGROUP_SCHED
7984 static void free_sched_group(struct task_group *tg)
7986 free_fair_sched_group(tg);
7987 free_rt_sched_group(tg);
7991 /* allocate runqueue etc for a new task group */
7992 struct task_group *sched_create_group(struct task_group *parent)
7994 struct task_group *tg;
7995 unsigned long flags;
7998 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8000 return ERR_PTR(-ENOMEM);
8002 if (!alloc_fair_sched_group(tg, parent))
8005 if (!alloc_rt_sched_group(tg, parent))
8008 spin_lock_irqsave(&task_group_lock, flags);
8009 for_each_possible_cpu(i) {
8010 register_fair_sched_group(tg, i);
8011 register_rt_sched_group(tg, i);
8013 list_add_rcu(&tg->list, &task_groups);
8015 WARN_ON(!parent); /* root should already exist */
8017 tg->parent = parent;
8018 INIT_LIST_HEAD(&tg->children);
8019 list_add_rcu(&tg->siblings, &parent->children);
8020 spin_unlock_irqrestore(&task_group_lock, flags);
8025 free_sched_group(tg);
8026 return ERR_PTR(-ENOMEM);
8029 /* rcu callback to free various structures associated with a task group */
8030 static void free_sched_group_rcu(struct rcu_head *rhp)
8032 /* now it should be safe to free those cfs_rqs */
8033 free_sched_group(container_of(rhp, struct task_group, rcu));
8036 /* Destroy runqueue etc associated with a task group */
8037 void sched_destroy_group(struct task_group *tg)
8039 unsigned long flags;
8042 spin_lock_irqsave(&task_group_lock, flags);
8043 for_each_possible_cpu(i) {
8044 unregister_fair_sched_group(tg, i);
8045 unregister_rt_sched_group(tg, i);
8047 list_del_rcu(&tg->list);
8048 list_del_rcu(&tg->siblings);
8049 spin_unlock_irqrestore(&task_group_lock, flags);
8051 /* wait for possible concurrent references to cfs_rqs complete */
8052 call_rcu(&tg->rcu, free_sched_group_rcu);
8055 /* change task's runqueue when it moves between groups.
8056 * The caller of this function should have put the task in its new group
8057 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8058 * reflect its new group.
8060 void sched_move_task(struct task_struct *tsk)
8063 unsigned long flags;
8066 rq = task_rq_lock(tsk, &flags);
8068 running = task_current(rq, tsk);
8069 on_rq = tsk->se.on_rq;
8072 dequeue_task(rq, tsk, 0);
8073 if (unlikely(running))
8074 tsk->sched_class->put_prev_task(rq, tsk);
8076 set_task_rq(tsk, task_cpu(tsk));
8078 #ifdef CONFIG_FAIR_GROUP_SCHED
8079 if (tsk->sched_class->moved_group)
8080 tsk->sched_class->moved_group(tsk, on_rq);
8083 if (unlikely(running))
8084 tsk->sched_class->set_curr_task(rq);
8086 enqueue_task(rq, tsk, 0);
8088 task_rq_unlock(rq, &flags);
8090 #endif /* CONFIG_CGROUP_SCHED */
8092 #ifdef CONFIG_FAIR_GROUP_SCHED
8093 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8095 struct cfs_rq *cfs_rq = se->cfs_rq;
8100 dequeue_entity(cfs_rq, se, 0);
8102 se->load.weight = shares;
8103 se->load.inv_weight = 0;
8106 enqueue_entity(cfs_rq, se, 0);
8109 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8111 struct cfs_rq *cfs_rq = se->cfs_rq;
8112 struct rq *rq = cfs_rq->rq;
8113 unsigned long flags;
8115 raw_spin_lock_irqsave(&rq->lock, flags);
8116 __set_se_shares(se, shares);
8117 raw_spin_unlock_irqrestore(&rq->lock, flags);
8120 static DEFINE_MUTEX(shares_mutex);
8122 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8125 unsigned long flags;
8128 * We can't change the weight of the root cgroup.
8133 if (shares < MIN_SHARES)
8134 shares = MIN_SHARES;
8135 else if (shares > MAX_SHARES)
8136 shares = MAX_SHARES;
8138 mutex_lock(&shares_mutex);
8139 if (tg->shares == shares)
8142 spin_lock_irqsave(&task_group_lock, flags);
8143 for_each_possible_cpu(i)
8144 unregister_fair_sched_group(tg, i);
8145 list_del_rcu(&tg->siblings);
8146 spin_unlock_irqrestore(&task_group_lock, flags);
8148 /* wait for any ongoing reference to this group to finish */
8149 synchronize_sched();
8152 * Now we are free to modify the group's share on each cpu
8153 * w/o tripping rebalance_share or load_balance_fair.
8155 tg->shares = shares;
8156 for_each_possible_cpu(i) {
8160 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8161 set_se_shares(tg->se[i], shares);
8165 * Enable load balance activity on this group, by inserting it back on
8166 * each cpu's rq->leaf_cfs_rq_list.
8168 spin_lock_irqsave(&task_group_lock, flags);
8169 for_each_possible_cpu(i)
8170 register_fair_sched_group(tg, i);
8171 list_add_rcu(&tg->siblings, &tg->parent->children);
8172 spin_unlock_irqrestore(&task_group_lock, flags);
8174 mutex_unlock(&shares_mutex);
8178 unsigned long sched_group_shares(struct task_group *tg)
8184 #ifdef CONFIG_RT_GROUP_SCHED
8186 * Ensure that the real time constraints are schedulable.
8188 static DEFINE_MUTEX(rt_constraints_mutex);
8190 static unsigned long to_ratio(u64 period, u64 runtime)
8192 if (runtime == RUNTIME_INF)
8195 return div64_u64(runtime << 20, period);
8198 /* Must be called with tasklist_lock held */
8199 static inline int tg_has_rt_tasks(struct task_group *tg)
8201 struct task_struct *g, *p;
8203 do_each_thread(g, p) {
8204 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8206 } while_each_thread(g, p);
8211 struct rt_schedulable_data {
8212 struct task_group *tg;
8217 static int tg_schedulable(struct task_group *tg, void *data)
8219 struct rt_schedulable_data *d = data;
8220 struct task_group *child;
8221 unsigned long total, sum = 0;
8222 u64 period, runtime;
8224 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8225 runtime = tg->rt_bandwidth.rt_runtime;
8228 period = d->rt_period;
8229 runtime = d->rt_runtime;
8233 * Cannot have more runtime than the period.
8235 if (runtime > period && runtime != RUNTIME_INF)
8239 * Ensure we don't starve existing RT tasks.
8241 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8244 total = to_ratio(period, runtime);
8247 * Nobody can have more than the global setting allows.
8249 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8253 * The sum of our children's runtime should not exceed our own.
8255 list_for_each_entry_rcu(child, &tg->children, siblings) {
8256 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8257 runtime = child->rt_bandwidth.rt_runtime;
8259 if (child == d->tg) {
8260 period = d->rt_period;
8261 runtime = d->rt_runtime;
8264 sum += to_ratio(period, runtime);
8273 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8275 struct rt_schedulable_data data = {
8277 .rt_period = period,
8278 .rt_runtime = runtime,
8281 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8284 static int tg_set_bandwidth(struct task_group *tg,
8285 u64 rt_period, u64 rt_runtime)
8289 mutex_lock(&rt_constraints_mutex);
8290 read_lock(&tasklist_lock);
8291 err = __rt_schedulable(tg, rt_period, rt_runtime);
8295 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8296 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8297 tg->rt_bandwidth.rt_runtime = rt_runtime;
8299 for_each_possible_cpu(i) {
8300 struct rt_rq *rt_rq = tg->rt_rq[i];
8302 raw_spin_lock(&rt_rq->rt_runtime_lock);
8303 rt_rq->rt_runtime = rt_runtime;
8304 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8306 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8308 read_unlock(&tasklist_lock);
8309 mutex_unlock(&rt_constraints_mutex);
8314 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8316 u64 rt_runtime, rt_period;
8318 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8319 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8320 if (rt_runtime_us < 0)
8321 rt_runtime = RUNTIME_INF;
8323 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8326 long sched_group_rt_runtime(struct task_group *tg)
8330 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8333 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8334 do_div(rt_runtime_us, NSEC_PER_USEC);
8335 return rt_runtime_us;
8338 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8340 u64 rt_runtime, rt_period;
8342 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8343 rt_runtime = tg->rt_bandwidth.rt_runtime;
8348 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8351 long sched_group_rt_period(struct task_group *tg)
8355 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8356 do_div(rt_period_us, NSEC_PER_USEC);
8357 return rt_period_us;
8360 static int sched_rt_global_constraints(void)
8362 u64 runtime, period;
8365 if (sysctl_sched_rt_period <= 0)
8368 runtime = global_rt_runtime();
8369 period = global_rt_period();
8372 * Sanity check on the sysctl variables.
8374 if (runtime > period && runtime != RUNTIME_INF)
8377 mutex_lock(&rt_constraints_mutex);
8378 read_lock(&tasklist_lock);
8379 ret = __rt_schedulable(NULL, 0, 0);
8380 read_unlock(&tasklist_lock);
8381 mutex_unlock(&rt_constraints_mutex);
8386 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8388 /* Don't accept realtime tasks when there is no way for them to run */
8389 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8395 #else /* !CONFIG_RT_GROUP_SCHED */
8396 static int sched_rt_global_constraints(void)
8398 unsigned long flags;
8401 if (sysctl_sched_rt_period <= 0)
8405 * There's always some RT tasks in the root group
8406 * -- migration, kstopmachine etc..
8408 if (sysctl_sched_rt_runtime == 0)
8411 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8412 for_each_possible_cpu(i) {
8413 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8415 raw_spin_lock(&rt_rq->rt_runtime_lock);
8416 rt_rq->rt_runtime = global_rt_runtime();
8417 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8419 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8423 #endif /* CONFIG_RT_GROUP_SCHED */
8425 int sched_rt_handler(struct ctl_table *table, int write,
8426 void __user *buffer, size_t *lenp,
8430 int old_period, old_runtime;
8431 static DEFINE_MUTEX(mutex);
8434 old_period = sysctl_sched_rt_period;
8435 old_runtime = sysctl_sched_rt_runtime;
8437 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8439 if (!ret && write) {
8440 ret = sched_rt_global_constraints();
8442 sysctl_sched_rt_period = old_period;
8443 sysctl_sched_rt_runtime = old_runtime;
8445 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8446 def_rt_bandwidth.rt_period =
8447 ns_to_ktime(global_rt_period());
8450 mutex_unlock(&mutex);
8455 #ifdef CONFIG_CGROUP_SCHED
8457 /* return corresponding task_group object of a cgroup */
8458 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8460 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8461 struct task_group, css);
8464 static struct cgroup_subsys_state *
8465 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8467 struct task_group *tg, *parent;
8469 if (!cgrp->parent) {
8470 /* This is early initialization for the top cgroup */
8471 return &init_task_group.css;
8474 parent = cgroup_tg(cgrp->parent);
8475 tg = sched_create_group(parent);
8477 return ERR_PTR(-ENOMEM);
8483 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8485 struct task_group *tg = cgroup_tg(cgrp);
8487 sched_destroy_group(tg);
8491 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8493 #ifdef CONFIG_RT_GROUP_SCHED
8494 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
8497 /* We don't support RT-tasks being in separate groups */
8498 if (tsk->sched_class != &fair_sched_class)
8505 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8506 struct task_struct *tsk, bool threadgroup)
8508 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
8512 struct task_struct *c;
8514 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8515 retval = cpu_cgroup_can_attach_task(cgrp, c);
8527 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8528 struct cgroup *old_cont, struct task_struct *tsk,
8531 sched_move_task(tsk);
8533 struct task_struct *c;
8535 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8542 #ifdef CONFIG_FAIR_GROUP_SCHED
8543 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8546 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8549 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8551 struct task_group *tg = cgroup_tg(cgrp);
8553 return (u64) tg->shares;
8555 #endif /* CONFIG_FAIR_GROUP_SCHED */
8557 #ifdef CONFIG_RT_GROUP_SCHED
8558 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8561 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8564 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8566 return sched_group_rt_runtime(cgroup_tg(cgrp));
8569 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8572 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8575 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8577 return sched_group_rt_period(cgroup_tg(cgrp));
8579 #endif /* CONFIG_RT_GROUP_SCHED */
8581 static struct cftype cpu_files[] = {
8582 #ifdef CONFIG_FAIR_GROUP_SCHED
8585 .read_u64 = cpu_shares_read_u64,
8586 .write_u64 = cpu_shares_write_u64,
8589 #ifdef CONFIG_RT_GROUP_SCHED
8591 .name = "rt_runtime_us",
8592 .read_s64 = cpu_rt_runtime_read,
8593 .write_s64 = cpu_rt_runtime_write,
8596 .name = "rt_period_us",
8597 .read_u64 = cpu_rt_period_read_uint,
8598 .write_u64 = cpu_rt_period_write_uint,
8603 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8605 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8608 struct cgroup_subsys cpu_cgroup_subsys = {
8610 .create = cpu_cgroup_create,
8611 .destroy = cpu_cgroup_destroy,
8612 .can_attach = cpu_cgroup_can_attach,
8613 .attach = cpu_cgroup_attach,
8614 .populate = cpu_cgroup_populate,
8615 .subsys_id = cpu_cgroup_subsys_id,
8619 #endif /* CONFIG_CGROUP_SCHED */
8621 #ifdef CONFIG_CGROUP_CPUACCT
8624 * CPU accounting code for task groups.
8626 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8627 * (balbir@in.ibm.com).
8630 /* track cpu usage of a group of tasks and its child groups */
8632 struct cgroup_subsys_state css;
8633 /* cpuusage holds pointer to a u64-type object on every cpu */
8634 u64 __percpu *cpuusage;
8635 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
8636 struct cpuacct *parent;
8639 struct cgroup_subsys cpuacct_subsys;
8641 /* return cpu accounting group corresponding to this container */
8642 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8644 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8645 struct cpuacct, css);
8648 /* return cpu accounting group to which this task belongs */
8649 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8651 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8652 struct cpuacct, css);
8655 /* create a new cpu accounting group */
8656 static struct cgroup_subsys_state *cpuacct_create(
8657 struct cgroup_subsys *ss, struct cgroup *cgrp)
8659 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8665 ca->cpuusage = alloc_percpu(u64);
8669 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8670 if (percpu_counter_init(&ca->cpustat[i], 0))
8671 goto out_free_counters;
8674 ca->parent = cgroup_ca(cgrp->parent);
8680 percpu_counter_destroy(&ca->cpustat[i]);
8681 free_percpu(ca->cpuusage);
8685 return ERR_PTR(-ENOMEM);
8688 /* destroy an existing cpu accounting group */
8690 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8692 struct cpuacct *ca = cgroup_ca(cgrp);
8695 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8696 percpu_counter_destroy(&ca->cpustat[i]);
8697 free_percpu(ca->cpuusage);
8701 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
8703 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8706 #ifndef CONFIG_64BIT
8708 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8710 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8712 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8720 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
8722 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8724 #ifndef CONFIG_64BIT
8726 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8728 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8730 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8736 /* return total cpu usage (in nanoseconds) of a group */
8737 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8739 struct cpuacct *ca = cgroup_ca(cgrp);
8740 u64 totalcpuusage = 0;
8743 for_each_present_cpu(i)
8744 totalcpuusage += cpuacct_cpuusage_read(ca, i);
8746 return totalcpuusage;
8749 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8752 struct cpuacct *ca = cgroup_ca(cgrp);
8761 for_each_present_cpu(i)
8762 cpuacct_cpuusage_write(ca, i, 0);
8768 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
8771 struct cpuacct *ca = cgroup_ca(cgroup);
8775 for_each_present_cpu(i) {
8776 percpu = cpuacct_cpuusage_read(ca, i);
8777 seq_printf(m, "%llu ", (unsigned long long) percpu);
8779 seq_printf(m, "\n");
8783 static const char *cpuacct_stat_desc[] = {
8784 [CPUACCT_STAT_USER] = "user",
8785 [CPUACCT_STAT_SYSTEM] = "system",
8788 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
8789 struct cgroup_map_cb *cb)
8791 struct cpuacct *ca = cgroup_ca(cgrp);
8794 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
8795 s64 val = percpu_counter_read(&ca->cpustat[i]);
8796 val = cputime64_to_clock_t(val);
8797 cb->fill(cb, cpuacct_stat_desc[i], val);
8802 static struct cftype files[] = {
8805 .read_u64 = cpuusage_read,
8806 .write_u64 = cpuusage_write,
8809 .name = "usage_percpu",
8810 .read_seq_string = cpuacct_percpu_seq_read,
8814 .read_map = cpuacct_stats_show,
8818 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
8820 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
8824 * charge this task's execution time to its accounting group.
8826 * called with rq->lock held.
8828 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8833 if (unlikely(!cpuacct_subsys.active))
8836 cpu = task_cpu(tsk);
8842 for (; ca; ca = ca->parent) {
8843 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8844 *cpuusage += cputime;
8851 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
8852 * in cputime_t units. As a result, cpuacct_update_stats calls
8853 * percpu_counter_add with values large enough to always overflow the
8854 * per cpu batch limit causing bad SMP scalability.
8856 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
8857 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
8858 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
8861 #define CPUACCT_BATCH \
8862 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
8864 #define CPUACCT_BATCH 0
8868 * Charge the system/user time to the task's accounting group.
8870 static void cpuacct_update_stats(struct task_struct *tsk,
8871 enum cpuacct_stat_index idx, cputime_t val)
8874 int batch = CPUACCT_BATCH;
8876 if (unlikely(!cpuacct_subsys.active))
8883 __percpu_counter_add(&ca->cpustat[idx], val, batch);
8889 struct cgroup_subsys cpuacct_subsys = {
8891 .create = cpuacct_create,
8892 .destroy = cpuacct_destroy,
8893 .populate = cpuacct_populate,
8894 .subsys_id = cpuacct_subsys_id,
8896 #endif /* CONFIG_CGROUP_CPUACCT */
8900 void synchronize_sched_expedited(void)
8904 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
8906 #else /* #ifndef CONFIG_SMP */
8908 static atomic_t synchronize_sched_expedited_count = ATOMIC_INIT(0);
8910 static int synchronize_sched_expedited_cpu_stop(void *data)
8913 * There must be a full memory barrier on each affected CPU
8914 * between the time that try_stop_cpus() is called and the
8915 * time that it returns.
8917 * In the current initial implementation of cpu_stop, the
8918 * above condition is already met when the control reaches
8919 * this point and the following smp_mb() is not strictly
8920 * necessary. Do smp_mb() anyway for documentation and
8921 * robustness against future implementation changes.
8923 smp_mb(); /* See above comment block. */
8928 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
8929 * approach to force grace period to end quickly. This consumes
8930 * significant time on all CPUs, and is thus not recommended for
8931 * any sort of common-case code.
8933 * Note that it is illegal to call this function while holding any
8934 * lock that is acquired by a CPU-hotplug notifier. Failing to
8935 * observe this restriction will result in deadlock.
8937 void synchronize_sched_expedited(void)
8939 int snap, trycount = 0;
8941 smp_mb(); /* ensure prior mod happens before capturing snap. */
8942 snap = atomic_read(&synchronize_sched_expedited_count) + 1;
8944 while (try_stop_cpus(cpu_online_mask,
8945 synchronize_sched_expedited_cpu_stop,
8948 if (trycount++ < 10)
8949 udelay(trycount * num_online_cpus());
8951 synchronize_sched();
8954 if (atomic_read(&synchronize_sched_expedited_count) - snap > 0) {
8955 smp_mb(); /* ensure test happens before caller kfree */
8960 atomic_inc(&synchronize_sched_expedited_count);
8961 smp_mb__after_atomic_inc(); /* ensure post-GP actions seen after GP. */
8964 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
8966 #endif /* #else #ifndef CONFIG_SMP */