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 if (root_task_group_empty())
1663 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1668 static inline void update_shares(struct sched_domain *sd)
1674 #ifdef CONFIG_PREEMPT
1676 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1679 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1680 * way at the expense of forcing extra atomic operations in all
1681 * invocations. This assures that the double_lock is acquired using the
1682 * same underlying policy as the spinlock_t on this architecture, which
1683 * reduces latency compared to the unfair variant below. However, it
1684 * also adds more overhead and therefore may reduce throughput.
1686 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1687 __releases(this_rq->lock)
1688 __acquires(busiest->lock)
1689 __acquires(this_rq->lock)
1691 raw_spin_unlock(&this_rq->lock);
1692 double_rq_lock(this_rq, busiest);
1699 * Unfair double_lock_balance: Optimizes throughput at the expense of
1700 * latency by eliminating extra atomic operations when the locks are
1701 * already in proper order on entry. This favors lower cpu-ids and will
1702 * grant the double lock to lower cpus over higher ids under contention,
1703 * regardless of entry order into the function.
1705 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1706 __releases(this_rq->lock)
1707 __acquires(busiest->lock)
1708 __acquires(this_rq->lock)
1712 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1713 if (busiest < this_rq) {
1714 raw_spin_unlock(&this_rq->lock);
1715 raw_spin_lock(&busiest->lock);
1716 raw_spin_lock_nested(&this_rq->lock,
1717 SINGLE_DEPTH_NESTING);
1720 raw_spin_lock_nested(&busiest->lock,
1721 SINGLE_DEPTH_NESTING);
1726 #endif /* CONFIG_PREEMPT */
1729 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1731 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1733 if (unlikely(!irqs_disabled())) {
1734 /* printk() doesn't work good under rq->lock */
1735 raw_spin_unlock(&this_rq->lock);
1739 return _double_lock_balance(this_rq, busiest);
1742 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1743 __releases(busiest->lock)
1745 raw_spin_unlock(&busiest->lock);
1746 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1750 * double_rq_lock - safely lock two runqueues
1752 * Note this does not disable interrupts like task_rq_lock,
1753 * you need to do so manually before calling.
1755 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1756 __acquires(rq1->lock)
1757 __acquires(rq2->lock)
1759 BUG_ON(!irqs_disabled());
1761 raw_spin_lock(&rq1->lock);
1762 __acquire(rq2->lock); /* Fake it out ;) */
1765 raw_spin_lock(&rq1->lock);
1766 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1768 raw_spin_lock(&rq2->lock);
1769 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1775 * double_rq_unlock - safely unlock two runqueues
1777 * Note this does not restore interrupts like task_rq_unlock,
1778 * you need to do so manually after calling.
1780 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1781 __releases(rq1->lock)
1782 __releases(rq2->lock)
1784 raw_spin_unlock(&rq1->lock);
1786 raw_spin_unlock(&rq2->lock);
1788 __release(rq2->lock);
1793 #ifdef CONFIG_FAIR_GROUP_SCHED
1794 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1797 cfs_rq->shares = shares;
1802 static void calc_load_account_idle(struct rq *this_rq);
1803 static void update_sysctl(void);
1804 static int get_update_sysctl_factor(void);
1806 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1808 set_task_rq(p, cpu);
1811 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1812 * successfuly executed on another CPU. We must ensure that updates of
1813 * per-task data have been completed by this moment.
1816 task_thread_info(p)->cpu = cpu;
1820 static const struct sched_class rt_sched_class;
1822 #define sched_class_highest (&rt_sched_class)
1823 #define for_each_class(class) \
1824 for (class = sched_class_highest; class; class = class->next)
1826 #include "sched_stats.h"
1828 static void inc_nr_running(struct rq *rq)
1833 static void dec_nr_running(struct rq *rq)
1838 static void set_load_weight(struct task_struct *p)
1840 if (task_has_rt_policy(p)) {
1841 p->se.load.weight = 0;
1842 p->se.load.inv_weight = WMULT_CONST;
1847 * SCHED_IDLE tasks get minimal weight:
1849 if (p->policy == SCHED_IDLE) {
1850 p->se.load.weight = WEIGHT_IDLEPRIO;
1851 p->se.load.inv_weight = WMULT_IDLEPRIO;
1855 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1856 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1859 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1861 update_rq_clock(rq);
1862 sched_info_queued(p);
1863 p->sched_class->enqueue_task(rq, p, flags);
1867 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1869 update_rq_clock(rq);
1870 sched_info_dequeued(p);
1871 p->sched_class->dequeue_task(rq, p, flags);
1876 * activate_task - move a task to the runqueue.
1878 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1880 if (task_contributes_to_load(p))
1881 rq->nr_uninterruptible--;
1883 enqueue_task(rq, p, flags);
1888 * deactivate_task - remove a task from the runqueue.
1890 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1892 if (task_contributes_to_load(p))
1893 rq->nr_uninterruptible++;
1895 dequeue_task(rq, p, flags);
1899 #include "sched_idletask.c"
1900 #include "sched_fair.c"
1901 #include "sched_rt.c"
1902 #ifdef CONFIG_SCHED_DEBUG
1903 # include "sched_debug.c"
1907 * __normal_prio - return the priority that is based on the static prio
1909 static inline int __normal_prio(struct task_struct *p)
1911 return p->static_prio;
1915 * Calculate the expected normal priority: i.e. priority
1916 * without taking RT-inheritance into account. Might be
1917 * boosted by interactivity modifiers. Changes upon fork,
1918 * setprio syscalls, and whenever the interactivity
1919 * estimator recalculates.
1921 static inline int normal_prio(struct task_struct *p)
1925 if (task_has_rt_policy(p))
1926 prio = MAX_RT_PRIO-1 - p->rt_priority;
1928 prio = __normal_prio(p);
1933 * Calculate the current priority, i.e. the priority
1934 * taken into account by the scheduler. This value might
1935 * be boosted by RT tasks, or might be boosted by
1936 * interactivity modifiers. Will be RT if the task got
1937 * RT-boosted. If not then it returns p->normal_prio.
1939 static int effective_prio(struct task_struct *p)
1941 p->normal_prio = normal_prio(p);
1943 * If we are RT tasks or we were boosted to RT priority,
1944 * keep the priority unchanged. Otherwise, update priority
1945 * to the normal priority:
1947 if (!rt_prio(p->prio))
1948 return p->normal_prio;
1953 * task_curr - is this task currently executing on a CPU?
1954 * @p: the task in question.
1956 inline int task_curr(const struct task_struct *p)
1958 return cpu_curr(task_cpu(p)) == p;
1961 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1962 const struct sched_class *prev_class,
1963 int oldprio, int running)
1965 if (prev_class != p->sched_class) {
1966 if (prev_class->switched_from)
1967 prev_class->switched_from(rq, p, running);
1968 p->sched_class->switched_to(rq, p, running);
1970 p->sched_class->prio_changed(rq, p, oldprio, running);
1975 * Is this task likely cache-hot:
1978 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1982 if (p->sched_class != &fair_sched_class)
1986 * Buddy candidates are cache hot:
1988 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
1989 (&p->se == cfs_rq_of(&p->se)->next ||
1990 &p->se == cfs_rq_of(&p->se)->last))
1993 if (sysctl_sched_migration_cost == -1)
1995 if (sysctl_sched_migration_cost == 0)
1998 delta = now - p->se.exec_start;
2000 return delta < (s64)sysctl_sched_migration_cost;
2003 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2005 #ifdef CONFIG_SCHED_DEBUG
2007 * We should never call set_task_cpu() on a blocked task,
2008 * ttwu() will sort out the placement.
2010 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2011 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2014 trace_sched_migrate_task(p, new_cpu);
2016 if (task_cpu(p) != new_cpu) {
2017 p->se.nr_migrations++;
2018 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2021 __set_task_cpu(p, new_cpu);
2024 struct migration_arg {
2025 struct task_struct *task;
2029 static int migration_cpu_stop(void *data);
2032 * The task's runqueue lock must be held.
2033 * Returns true if you have to wait for migration thread.
2035 static bool migrate_task(struct task_struct *p, int dest_cpu)
2037 struct rq *rq = task_rq(p);
2040 * If the task is not on a runqueue (and not running), then
2041 * the next wake-up will properly place the task.
2043 return p->se.on_rq || task_running(rq, p);
2047 * wait_task_inactive - wait for a thread to unschedule.
2049 * If @match_state is nonzero, it's the @p->state value just checked and
2050 * not expected to change. If it changes, i.e. @p might have woken up,
2051 * then return zero. When we succeed in waiting for @p to be off its CPU,
2052 * we return a positive number (its total switch count). If a second call
2053 * a short while later returns the same number, the caller can be sure that
2054 * @p has remained unscheduled the whole time.
2056 * The caller must ensure that the task *will* unschedule sometime soon,
2057 * else this function might spin for a *long* time. This function can't
2058 * be called with interrupts off, or it may introduce deadlock with
2059 * smp_call_function() if an IPI is sent by the same process we are
2060 * waiting to become inactive.
2062 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2064 unsigned long flags;
2071 * We do the initial early heuristics without holding
2072 * any task-queue locks at all. We'll only try to get
2073 * the runqueue lock when things look like they will
2079 * If the task is actively running on another CPU
2080 * still, just relax and busy-wait without holding
2083 * NOTE! Since we don't hold any locks, it's not
2084 * even sure that "rq" stays as the right runqueue!
2085 * But we don't care, since "task_running()" will
2086 * return false if the runqueue has changed and p
2087 * is actually now running somewhere else!
2089 while (task_running(rq, p)) {
2090 if (match_state && unlikely(p->state != match_state))
2096 * Ok, time to look more closely! We need the rq
2097 * lock now, to be *sure*. If we're wrong, we'll
2098 * just go back and repeat.
2100 rq = task_rq_lock(p, &flags);
2101 trace_sched_wait_task(p);
2102 running = task_running(rq, p);
2103 on_rq = p->se.on_rq;
2105 if (!match_state || p->state == match_state)
2106 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2107 task_rq_unlock(rq, &flags);
2110 * If it changed from the expected state, bail out now.
2112 if (unlikely(!ncsw))
2116 * Was it really running after all now that we
2117 * checked with the proper locks actually held?
2119 * Oops. Go back and try again..
2121 if (unlikely(running)) {
2127 * It's not enough that it's not actively running,
2128 * it must be off the runqueue _entirely_, and not
2131 * So if it was still runnable (but just not actively
2132 * running right now), it's preempted, and we should
2133 * yield - it could be a while.
2135 if (unlikely(on_rq)) {
2136 schedule_timeout_uninterruptible(1);
2141 * Ahh, all good. It wasn't running, and it wasn't
2142 * runnable, which means that it will never become
2143 * running in the future either. We're all done!
2152 * kick_process - kick a running thread to enter/exit the kernel
2153 * @p: the to-be-kicked thread
2155 * Cause a process which is running on another CPU to enter
2156 * kernel-mode, without any delay. (to get signals handled.)
2158 * NOTE: this function doesnt have to take the runqueue lock,
2159 * because all it wants to ensure is that the remote task enters
2160 * the kernel. If the IPI races and the task has been migrated
2161 * to another CPU then no harm is done and the purpose has been
2164 void kick_process(struct task_struct *p)
2170 if ((cpu != smp_processor_id()) && task_curr(p))
2171 smp_send_reschedule(cpu);
2174 EXPORT_SYMBOL_GPL(kick_process);
2175 #endif /* CONFIG_SMP */
2178 * task_oncpu_function_call - call a function on the cpu on which a task runs
2179 * @p: the task to evaluate
2180 * @func: the function to be called
2181 * @info: the function call argument
2183 * Calls the function @func when the task is currently running. This might
2184 * be on the current CPU, which just calls the function directly
2186 void task_oncpu_function_call(struct task_struct *p,
2187 void (*func) (void *info), void *info)
2194 smp_call_function_single(cpu, func, info, 1);
2200 * ->cpus_allowed is protected by either TASK_WAKING or rq->lock held.
2202 static int select_fallback_rq(int cpu, struct task_struct *p)
2205 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2207 /* Look for allowed, online CPU in same node. */
2208 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2209 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2212 /* Any allowed, online CPU? */
2213 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2214 if (dest_cpu < nr_cpu_ids)
2217 /* No more Mr. Nice Guy. */
2218 if (unlikely(dest_cpu >= nr_cpu_ids)) {
2219 dest_cpu = cpuset_cpus_allowed_fallback(p);
2221 * Don't tell them about moving exiting tasks or
2222 * kernel threads (both mm NULL), since they never
2225 if (p->mm && printk_ratelimit()) {
2226 printk(KERN_INFO "process %d (%s) no "
2227 "longer affine to cpu%d\n",
2228 task_pid_nr(p), p->comm, cpu);
2236 * The caller (fork, wakeup) owns TASK_WAKING, ->cpus_allowed is stable.
2239 int select_task_rq(struct rq *rq, struct task_struct *p, int sd_flags, int wake_flags)
2241 int cpu = p->sched_class->select_task_rq(rq, p, sd_flags, wake_flags);
2244 * In order not to call set_task_cpu() on a blocking task we need
2245 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2248 * Since this is common to all placement strategies, this lives here.
2250 * [ this allows ->select_task() to simply return task_cpu(p) and
2251 * not worry about this generic constraint ]
2253 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2255 cpu = select_fallback_rq(task_cpu(p), p);
2260 static void update_avg(u64 *avg, u64 sample)
2262 s64 diff = sample - *avg;
2268 * try_to_wake_up - wake up a thread
2269 * @p: the to-be-woken-up thread
2270 * @state: the mask of task states that can be woken
2271 * @sync: do a synchronous wakeup?
2273 * Put it on the run-queue if it's not already there. The "current"
2274 * thread is always on the run-queue (except when the actual
2275 * re-schedule is in progress), and as such you're allowed to do
2276 * the simpler "current->state = TASK_RUNNING" to mark yourself
2277 * runnable without the overhead of this.
2279 * returns failure only if the task is already active.
2281 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2284 int cpu, orig_cpu, this_cpu, success = 0;
2285 unsigned long flags;
2286 unsigned long en_flags = ENQUEUE_WAKEUP;
2289 this_cpu = get_cpu();
2292 rq = task_rq_lock(p, &flags);
2293 if (!(p->state & state))
2303 if (unlikely(task_running(rq, p)))
2307 * In order to handle concurrent wakeups and release the rq->lock
2308 * we put the task in TASK_WAKING state.
2310 * First fix up the nr_uninterruptible count:
2312 if (task_contributes_to_load(p)) {
2313 if (likely(cpu_online(orig_cpu)))
2314 rq->nr_uninterruptible--;
2316 this_rq()->nr_uninterruptible--;
2318 p->state = TASK_WAKING;
2320 if (p->sched_class->task_waking) {
2321 p->sched_class->task_waking(rq, p);
2322 en_flags |= ENQUEUE_WAKING;
2325 cpu = select_task_rq(rq, p, SD_BALANCE_WAKE, wake_flags);
2326 if (cpu != orig_cpu)
2327 set_task_cpu(p, cpu);
2328 __task_rq_unlock(rq);
2331 raw_spin_lock(&rq->lock);
2334 * We migrated the task without holding either rq->lock, however
2335 * since the task is not on the task list itself, nobody else
2336 * will try and migrate the task, hence the rq should match the
2337 * cpu we just moved it to.
2339 WARN_ON(task_cpu(p) != cpu);
2340 WARN_ON(p->state != TASK_WAKING);
2342 #ifdef CONFIG_SCHEDSTATS
2343 schedstat_inc(rq, ttwu_count);
2344 if (cpu == this_cpu)
2345 schedstat_inc(rq, ttwu_local);
2347 struct sched_domain *sd;
2348 for_each_domain(this_cpu, sd) {
2349 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2350 schedstat_inc(sd, ttwu_wake_remote);
2355 #endif /* CONFIG_SCHEDSTATS */
2358 #endif /* CONFIG_SMP */
2359 schedstat_inc(p, se.statistics.nr_wakeups);
2360 if (wake_flags & WF_SYNC)
2361 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2362 if (orig_cpu != cpu)
2363 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2364 if (cpu == this_cpu)
2365 schedstat_inc(p, se.statistics.nr_wakeups_local);
2367 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2368 activate_task(rq, p, en_flags);
2372 trace_sched_wakeup(p, success);
2373 check_preempt_curr(rq, p, wake_flags);
2375 p->state = TASK_RUNNING;
2377 if (p->sched_class->task_woken)
2378 p->sched_class->task_woken(rq, p);
2380 if (unlikely(rq->idle_stamp)) {
2381 u64 delta = rq->clock - rq->idle_stamp;
2382 u64 max = 2*sysctl_sched_migration_cost;
2387 update_avg(&rq->avg_idle, delta);
2392 task_rq_unlock(rq, &flags);
2399 * wake_up_process - Wake up a specific process
2400 * @p: The process to be woken up.
2402 * Attempt to wake up the nominated process and move it to the set of runnable
2403 * processes. Returns 1 if the process was woken up, 0 if it was already
2406 * It may be assumed that this function implies a write memory barrier before
2407 * changing the task state if and only if any tasks are woken up.
2409 int wake_up_process(struct task_struct *p)
2411 return try_to_wake_up(p, TASK_ALL, 0);
2413 EXPORT_SYMBOL(wake_up_process);
2415 int wake_up_state(struct task_struct *p, unsigned int state)
2417 return try_to_wake_up(p, state, 0);
2421 * Perform scheduler related setup for a newly forked process p.
2422 * p is forked by current.
2424 * __sched_fork() is basic setup used by init_idle() too:
2426 static void __sched_fork(struct task_struct *p)
2428 p->se.exec_start = 0;
2429 p->se.sum_exec_runtime = 0;
2430 p->se.prev_sum_exec_runtime = 0;
2431 p->se.nr_migrations = 0;
2433 #ifdef CONFIG_SCHEDSTATS
2434 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2437 INIT_LIST_HEAD(&p->rt.run_list);
2439 INIT_LIST_HEAD(&p->se.group_node);
2441 #ifdef CONFIG_PREEMPT_NOTIFIERS
2442 INIT_HLIST_HEAD(&p->preempt_notifiers);
2447 * fork()/clone()-time setup:
2449 void sched_fork(struct task_struct *p, int clone_flags)
2451 int cpu = get_cpu();
2455 * We mark the process as running here. This guarantees that
2456 * nobody will actually run it, and a signal or other external
2457 * event cannot wake it up and insert it on the runqueue either.
2459 p->state = TASK_RUNNING;
2462 * Revert to default priority/policy on fork if requested.
2464 if (unlikely(p->sched_reset_on_fork)) {
2465 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2466 p->policy = SCHED_NORMAL;
2467 p->normal_prio = p->static_prio;
2470 if (PRIO_TO_NICE(p->static_prio) < 0) {
2471 p->static_prio = NICE_TO_PRIO(0);
2472 p->normal_prio = p->static_prio;
2477 * We don't need the reset flag anymore after the fork. It has
2478 * fulfilled its duty:
2480 p->sched_reset_on_fork = 0;
2484 * Make sure we do not leak PI boosting priority to the child.
2486 p->prio = current->normal_prio;
2488 if (!rt_prio(p->prio))
2489 p->sched_class = &fair_sched_class;
2491 if (p->sched_class->task_fork)
2492 p->sched_class->task_fork(p);
2494 set_task_cpu(p, cpu);
2496 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2497 if (likely(sched_info_on()))
2498 memset(&p->sched_info, 0, sizeof(p->sched_info));
2500 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2503 #ifdef CONFIG_PREEMPT
2504 /* Want to start with kernel preemption disabled. */
2505 task_thread_info(p)->preempt_count = 1;
2507 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2513 * wake_up_new_task - wake up a newly created task for the first time.
2515 * This function will do some initial scheduler statistics housekeeping
2516 * that must be done for every newly created context, then puts the task
2517 * on the runqueue and wakes it.
2519 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2521 unsigned long flags;
2523 int cpu __maybe_unused = get_cpu();
2526 rq = task_rq_lock(p, &flags);
2527 p->state = TASK_WAKING;
2530 * Fork balancing, do it here and not earlier because:
2531 * - cpus_allowed can change in the fork path
2532 * - any previously selected cpu might disappear through hotplug
2534 * We set TASK_WAKING so that select_task_rq() can drop rq->lock
2535 * without people poking at ->cpus_allowed.
2537 cpu = select_task_rq(rq, p, SD_BALANCE_FORK, 0);
2538 set_task_cpu(p, cpu);
2540 p->state = TASK_RUNNING;
2541 task_rq_unlock(rq, &flags);
2544 rq = task_rq_lock(p, &flags);
2545 activate_task(rq, p, 0);
2546 trace_sched_wakeup_new(p, 1);
2547 check_preempt_curr(rq, p, WF_FORK);
2549 if (p->sched_class->task_woken)
2550 p->sched_class->task_woken(rq, p);
2552 task_rq_unlock(rq, &flags);
2556 #ifdef CONFIG_PREEMPT_NOTIFIERS
2559 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2560 * @notifier: notifier struct to register
2562 void preempt_notifier_register(struct preempt_notifier *notifier)
2564 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2566 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2569 * preempt_notifier_unregister - no longer interested in preemption notifications
2570 * @notifier: notifier struct to unregister
2572 * This is safe to call from within a preemption notifier.
2574 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2576 hlist_del(¬ifier->link);
2578 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2580 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2582 struct preempt_notifier *notifier;
2583 struct hlist_node *node;
2585 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2586 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2590 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2591 struct task_struct *next)
2593 struct preempt_notifier *notifier;
2594 struct hlist_node *node;
2596 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2597 notifier->ops->sched_out(notifier, next);
2600 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2602 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2607 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2608 struct task_struct *next)
2612 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2615 * prepare_task_switch - prepare to switch tasks
2616 * @rq: the runqueue preparing to switch
2617 * @prev: the current task that is being switched out
2618 * @next: the task we are going to switch to.
2620 * This is called with the rq lock held and interrupts off. It must
2621 * be paired with a subsequent finish_task_switch after the context
2624 * prepare_task_switch sets up locking and calls architecture specific
2628 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2629 struct task_struct *next)
2631 fire_sched_out_preempt_notifiers(prev, next);
2632 prepare_lock_switch(rq, next);
2633 prepare_arch_switch(next);
2637 * finish_task_switch - clean up after a task-switch
2638 * @rq: runqueue associated with task-switch
2639 * @prev: the thread we just switched away from.
2641 * finish_task_switch must be called after the context switch, paired
2642 * with a prepare_task_switch call before the context switch.
2643 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2644 * and do any other architecture-specific cleanup actions.
2646 * Note that we may have delayed dropping an mm in context_switch(). If
2647 * so, we finish that here outside of the runqueue lock. (Doing it
2648 * with the lock held can cause deadlocks; see schedule() for
2651 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2652 __releases(rq->lock)
2654 struct mm_struct *mm = rq->prev_mm;
2660 * A task struct has one reference for the use as "current".
2661 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2662 * schedule one last time. The schedule call will never return, and
2663 * the scheduled task must drop that reference.
2664 * The test for TASK_DEAD must occur while the runqueue locks are
2665 * still held, otherwise prev could be scheduled on another cpu, die
2666 * there before we look at prev->state, and then the reference would
2668 * Manfred Spraul <manfred@colorfullife.com>
2670 prev_state = prev->state;
2671 finish_arch_switch(prev);
2672 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2673 local_irq_disable();
2674 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2675 perf_event_task_sched_in(current);
2676 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2678 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2679 finish_lock_switch(rq, prev);
2681 fire_sched_in_preempt_notifiers(current);
2684 if (unlikely(prev_state == TASK_DEAD)) {
2686 * Remove function-return probe instances associated with this
2687 * task and put them back on the free list.
2689 kprobe_flush_task(prev);
2690 put_task_struct(prev);
2696 /* assumes rq->lock is held */
2697 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2699 if (prev->sched_class->pre_schedule)
2700 prev->sched_class->pre_schedule(rq, prev);
2703 /* rq->lock is NOT held, but preemption is disabled */
2704 static inline void post_schedule(struct rq *rq)
2706 if (rq->post_schedule) {
2707 unsigned long flags;
2709 raw_spin_lock_irqsave(&rq->lock, flags);
2710 if (rq->curr->sched_class->post_schedule)
2711 rq->curr->sched_class->post_schedule(rq);
2712 raw_spin_unlock_irqrestore(&rq->lock, flags);
2714 rq->post_schedule = 0;
2720 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2724 static inline void post_schedule(struct rq *rq)
2731 * schedule_tail - first thing a freshly forked thread must call.
2732 * @prev: the thread we just switched away from.
2734 asmlinkage void schedule_tail(struct task_struct *prev)
2735 __releases(rq->lock)
2737 struct rq *rq = this_rq();
2739 finish_task_switch(rq, prev);
2742 * FIXME: do we need to worry about rq being invalidated by the
2747 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2748 /* In this case, finish_task_switch does not reenable preemption */
2751 if (current->set_child_tid)
2752 put_user(task_pid_vnr(current), current->set_child_tid);
2756 * context_switch - switch to the new MM and the new
2757 * thread's register state.
2760 context_switch(struct rq *rq, struct task_struct *prev,
2761 struct task_struct *next)
2763 struct mm_struct *mm, *oldmm;
2765 prepare_task_switch(rq, prev, next);
2766 trace_sched_switch(prev, next);
2768 oldmm = prev->active_mm;
2770 * For paravirt, this is coupled with an exit in switch_to to
2771 * combine the page table reload and the switch backend into
2774 arch_start_context_switch(prev);
2777 next->active_mm = oldmm;
2778 atomic_inc(&oldmm->mm_count);
2779 enter_lazy_tlb(oldmm, next);
2781 switch_mm(oldmm, mm, next);
2783 if (likely(!prev->mm)) {
2784 prev->active_mm = NULL;
2785 rq->prev_mm = oldmm;
2788 * Since the runqueue lock will be released by the next
2789 * task (which is an invalid locking op but in the case
2790 * of the scheduler it's an obvious special-case), so we
2791 * do an early lockdep release here:
2793 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2794 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2797 /* Here we just switch the register state and the stack. */
2798 switch_to(prev, next, prev);
2802 * this_rq must be evaluated again because prev may have moved
2803 * CPUs since it called schedule(), thus the 'rq' on its stack
2804 * frame will be invalid.
2806 finish_task_switch(this_rq(), prev);
2810 * nr_running, nr_uninterruptible and nr_context_switches:
2812 * externally visible scheduler statistics: current number of runnable
2813 * threads, current number of uninterruptible-sleeping threads, total
2814 * number of context switches performed since bootup.
2816 unsigned long nr_running(void)
2818 unsigned long i, sum = 0;
2820 for_each_online_cpu(i)
2821 sum += cpu_rq(i)->nr_running;
2826 unsigned long nr_uninterruptible(void)
2828 unsigned long i, sum = 0;
2830 for_each_possible_cpu(i)
2831 sum += cpu_rq(i)->nr_uninterruptible;
2834 * Since we read the counters lockless, it might be slightly
2835 * inaccurate. Do not allow it to go below zero though:
2837 if (unlikely((long)sum < 0))
2843 unsigned long long nr_context_switches(void)
2846 unsigned long long sum = 0;
2848 for_each_possible_cpu(i)
2849 sum += cpu_rq(i)->nr_switches;
2854 unsigned long nr_iowait(void)
2856 unsigned long i, sum = 0;
2858 for_each_possible_cpu(i)
2859 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2864 unsigned long nr_iowait_cpu(void)
2866 struct rq *this = this_rq();
2867 return atomic_read(&this->nr_iowait);
2870 unsigned long this_cpu_load(void)
2872 struct rq *this = this_rq();
2873 return this->cpu_load[0];
2877 /* Variables and functions for calc_load */
2878 static atomic_long_t calc_load_tasks;
2879 static unsigned long calc_load_update;
2880 unsigned long avenrun[3];
2881 EXPORT_SYMBOL(avenrun);
2883 static long calc_load_fold_active(struct rq *this_rq)
2885 long nr_active, delta = 0;
2887 nr_active = this_rq->nr_running;
2888 nr_active += (long) this_rq->nr_uninterruptible;
2890 if (nr_active != this_rq->calc_load_active) {
2891 delta = nr_active - this_rq->calc_load_active;
2892 this_rq->calc_load_active = nr_active;
2900 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
2902 * When making the ILB scale, we should try to pull this in as well.
2904 static atomic_long_t calc_load_tasks_idle;
2906 static void calc_load_account_idle(struct rq *this_rq)
2910 delta = calc_load_fold_active(this_rq);
2912 atomic_long_add(delta, &calc_load_tasks_idle);
2915 static long calc_load_fold_idle(void)
2920 * Its got a race, we don't care...
2922 if (atomic_long_read(&calc_load_tasks_idle))
2923 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
2928 static void calc_load_account_idle(struct rq *this_rq)
2932 static inline long calc_load_fold_idle(void)
2939 * get_avenrun - get the load average array
2940 * @loads: pointer to dest load array
2941 * @offset: offset to add
2942 * @shift: shift count to shift the result left
2944 * These values are estimates at best, so no need for locking.
2946 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2948 loads[0] = (avenrun[0] + offset) << shift;
2949 loads[1] = (avenrun[1] + offset) << shift;
2950 loads[2] = (avenrun[2] + offset) << shift;
2953 static unsigned long
2954 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2957 load += active * (FIXED_1 - exp);
2958 return load >> FSHIFT;
2962 * calc_load - update the avenrun load estimates 10 ticks after the
2963 * CPUs have updated calc_load_tasks.
2965 void calc_global_load(void)
2967 unsigned long upd = calc_load_update + 10;
2970 if (time_before(jiffies, upd))
2973 active = atomic_long_read(&calc_load_tasks);
2974 active = active > 0 ? active * FIXED_1 : 0;
2976 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2977 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2978 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2980 calc_load_update += LOAD_FREQ;
2984 * Called from update_cpu_load() to periodically update this CPU's
2987 static void calc_load_account_active(struct rq *this_rq)
2991 if (time_before(jiffies, this_rq->calc_load_update))
2994 delta = calc_load_fold_active(this_rq);
2995 delta += calc_load_fold_idle();
2997 atomic_long_add(delta, &calc_load_tasks);
2999 this_rq->calc_load_update += LOAD_FREQ;
3003 * Update rq->cpu_load[] statistics. This function is usually called every
3004 * scheduler tick (TICK_NSEC).
3006 static void update_cpu_load(struct rq *this_rq)
3008 unsigned long this_load = this_rq->load.weight;
3011 this_rq->nr_load_updates++;
3013 /* Update our load: */
3014 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3015 unsigned long old_load, new_load;
3017 /* scale is effectively 1 << i now, and >> i divides by scale */
3019 old_load = this_rq->cpu_load[i];
3020 new_load = this_load;
3022 * Round up the averaging division if load is increasing. This
3023 * prevents us from getting stuck on 9 if the load is 10, for
3026 if (new_load > old_load)
3027 new_load += scale-1;
3028 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3031 calc_load_account_active(this_rq);
3037 * sched_exec - execve() is a valuable balancing opportunity, because at
3038 * this point the task has the smallest effective memory and cache footprint.
3040 void sched_exec(void)
3042 struct task_struct *p = current;
3043 unsigned long flags;
3047 rq = task_rq_lock(p, &flags);
3048 dest_cpu = p->sched_class->select_task_rq(rq, p, SD_BALANCE_EXEC, 0);
3049 if (dest_cpu == smp_processor_id())
3053 * select_task_rq() can race against ->cpus_allowed
3055 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed) &&
3056 likely(cpu_active(dest_cpu)) && migrate_task(p, dest_cpu)) {
3057 struct migration_arg arg = { p, dest_cpu };
3059 task_rq_unlock(rq, &flags);
3060 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
3064 task_rq_unlock(rq, &flags);
3069 DEFINE_PER_CPU(struct kernel_stat, kstat);
3071 EXPORT_PER_CPU_SYMBOL(kstat);
3074 * Return any ns on the sched_clock that have not yet been accounted in
3075 * @p in case that task is currently running.
3077 * Called with task_rq_lock() held on @rq.
3079 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3083 if (task_current(rq, p)) {
3084 update_rq_clock(rq);
3085 ns = rq->clock - p->se.exec_start;
3093 unsigned long long task_delta_exec(struct task_struct *p)
3095 unsigned long flags;
3099 rq = task_rq_lock(p, &flags);
3100 ns = do_task_delta_exec(p, rq);
3101 task_rq_unlock(rq, &flags);
3107 * Return accounted runtime for the task.
3108 * In case the task is currently running, return the runtime plus current's
3109 * pending runtime that have not been accounted yet.
3111 unsigned long long task_sched_runtime(struct task_struct *p)
3113 unsigned long flags;
3117 rq = task_rq_lock(p, &flags);
3118 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3119 task_rq_unlock(rq, &flags);
3125 * Return sum_exec_runtime for the thread group.
3126 * In case the task is currently running, return the sum plus current's
3127 * pending runtime that have not been accounted yet.
3129 * Note that the thread group might have other running tasks as well,
3130 * so the return value not includes other pending runtime that other
3131 * running tasks might have.
3133 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3135 struct task_cputime totals;
3136 unsigned long flags;
3140 rq = task_rq_lock(p, &flags);
3141 thread_group_cputime(p, &totals);
3142 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3143 task_rq_unlock(rq, &flags);
3149 * Account user cpu time to a process.
3150 * @p: the process that the cpu time gets accounted to
3151 * @cputime: the cpu time spent in user space since the last update
3152 * @cputime_scaled: cputime scaled by cpu frequency
3154 void account_user_time(struct task_struct *p, cputime_t cputime,
3155 cputime_t cputime_scaled)
3157 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3160 /* Add user time to process. */
3161 p->utime = cputime_add(p->utime, cputime);
3162 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3163 account_group_user_time(p, cputime);
3165 /* Add user time to cpustat. */
3166 tmp = cputime_to_cputime64(cputime);
3167 if (TASK_NICE(p) > 0)
3168 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3170 cpustat->user = cputime64_add(cpustat->user, tmp);
3172 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3173 /* Account for user time used */
3174 acct_update_integrals(p);
3178 * Account guest cpu time to a process.
3179 * @p: the process that the cpu time gets accounted to
3180 * @cputime: the cpu time spent in virtual machine since the last update
3181 * @cputime_scaled: cputime scaled by cpu frequency
3183 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3184 cputime_t cputime_scaled)
3187 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3189 tmp = cputime_to_cputime64(cputime);
3191 /* Add guest time to process. */
3192 p->utime = cputime_add(p->utime, cputime);
3193 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3194 account_group_user_time(p, cputime);
3195 p->gtime = cputime_add(p->gtime, cputime);
3197 /* Add guest time to cpustat. */
3198 if (TASK_NICE(p) > 0) {
3199 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3200 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3202 cpustat->user = cputime64_add(cpustat->user, tmp);
3203 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3208 * Account system cpu time to a process.
3209 * @p: the process that the cpu time gets accounted to
3210 * @hardirq_offset: the offset to subtract from hardirq_count()
3211 * @cputime: the cpu time spent in kernel space since the last update
3212 * @cputime_scaled: cputime scaled by cpu frequency
3214 void account_system_time(struct task_struct *p, int hardirq_offset,
3215 cputime_t cputime, cputime_t cputime_scaled)
3217 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3220 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3221 account_guest_time(p, cputime, cputime_scaled);
3225 /* Add system time to process. */
3226 p->stime = cputime_add(p->stime, cputime);
3227 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3228 account_group_system_time(p, cputime);
3230 /* Add system time to cpustat. */
3231 tmp = cputime_to_cputime64(cputime);
3232 if (hardirq_count() - hardirq_offset)
3233 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3234 else if (softirq_count())
3235 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3237 cpustat->system = cputime64_add(cpustat->system, tmp);
3239 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3241 /* Account for system time used */
3242 acct_update_integrals(p);
3246 * Account for involuntary wait time.
3247 * @steal: the cpu time spent in involuntary wait
3249 void account_steal_time(cputime_t cputime)
3251 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3252 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3254 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3258 * Account for idle time.
3259 * @cputime: the cpu time spent in idle wait
3261 void account_idle_time(cputime_t cputime)
3263 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3264 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3265 struct rq *rq = this_rq();
3267 if (atomic_read(&rq->nr_iowait) > 0)
3268 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3270 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3273 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3276 * Account a single tick of cpu time.
3277 * @p: the process that the cpu time gets accounted to
3278 * @user_tick: indicates if the tick is a user or a system tick
3280 void account_process_tick(struct task_struct *p, int user_tick)
3282 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3283 struct rq *rq = this_rq();
3286 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3287 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3288 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3291 account_idle_time(cputime_one_jiffy);
3295 * Account multiple ticks of steal time.
3296 * @p: the process from which the cpu time has been stolen
3297 * @ticks: number of stolen ticks
3299 void account_steal_ticks(unsigned long ticks)
3301 account_steal_time(jiffies_to_cputime(ticks));
3305 * Account multiple ticks of idle time.
3306 * @ticks: number of stolen ticks
3308 void account_idle_ticks(unsigned long ticks)
3310 account_idle_time(jiffies_to_cputime(ticks));
3316 * Use precise platform statistics if available:
3318 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3319 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3325 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3327 struct task_cputime cputime;
3329 thread_group_cputime(p, &cputime);
3331 *ut = cputime.utime;
3332 *st = cputime.stime;
3336 #ifndef nsecs_to_cputime
3337 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3340 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3342 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
3345 * Use CFS's precise accounting:
3347 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3352 temp = (u64)(rtime * utime);
3353 do_div(temp, total);
3354 utime = (cputime_t)temp;
3359 * Compare with previous values, to keep monotonicity:
3361 p->prev_utime = max(p->prev_utime, utime);
3362 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
3364 *ut = p->prev_utime;
3365 *st = p->prev_stime;
3369 * Must be called with siglock held.
3371 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3373 struct signal_struct *sig = p->signal;
3374 struct task_cputime cputime;
3375 cputime_t rtime, utime, total;
3377 thread_group_cputime(p, &cputime);
3379 total = cputime_add(cputime.utime, cputime.stime);
3380 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3385 temp = (u64)(rtime * cputime.utime);
3386 do_div(temp, total);
3387 utime = (cputime_t)temp;
3391 sig->prev_utime = max(sig->prev_utime, utime);
3392 sig->prev_stime = max(sig->prev_stime,
3393 cputime_sub(rtime, sig->prev_utime));
3395 *ut = sig->prev_utime;
3396 *st = sig->prev_stime;
3401 * This function gets called by the timer code, with HZ frequency.
3402 * We call it with interrupts disabled.
3404 * It also gets called by the fork code, when changing the parent's
3407 void scheduler_tick(void)
3409 int cpu = smp_processor_id();
3410 struct rq *rq = cpu_rq(cpu);
3411 struct task_struct *curr = rq->curr;
3415 raw_spin_lock(&rq->lock);
3416 update_rq_clock(rq);
3417 update_cpu_load(rq);
3418 curr->sched_class->task_tick(rq, curr, 0);
3419 raw_spin_unlock(&rq->lock);
3421 perf_event_task_tick(curr);
3424 rq->idle_at_tick = idle_cpu(cpu);
3425 trigger_load_balance(rq, cpu);
3429 notrace unsigned long get_parent_ip(unsigned long addr)
3431 if (in_lock_functions(addr)) {
3432 addr = CALLER_ADDR2;
3433 if (in_lock_functions(addr))
3434 addr = CALLER_ADDR3;
3439 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3440 defined(CONFIG_PREEMPT_TRACER))
3442 void __kprobes add_preempt_count(int val)
3444 #ifdef CONFIG_DEBUG_PREEMPT
3448 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3451 preempt_count() += val;
3452 #ifdef CONFIG_DEBUG_PREEMPT
3454 * Spinlock count overflowing soon?
3456 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3459 if (preempt_count() == val)
3460 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3462 EXPORT_SYMBOL(add_preempt_count);
3464 void __kprobes sub_preempt_count(int val)
3466 #ifdef CONFIG_DEBUG_PREEMPT
3470 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3473 * Is the spinlock portion underflowing?
3475 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3476 !(preempt_count() & PREEMPT_MASK)))
3480 if (preempt_count() == val)
3481 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3482 preempt_count() -= val;
3484 EXPORT_SYMBOL(sub_preempt_count);
3489 * Print scheduling while atomic bug:
3491 static noinline void __schedule_bug(struct task_struct *prev)
3493 struct pt_regs *regs = get_irq_regs();
3495 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3496 prev->comm, prev->pid, preempt_count());
3498 debug_show_held_locks(prev);
3500 if (irqs_disabled())
3501 print_irqtrace_events(prev);
3510 * Various schedule()-time debugging checks and statistics:
3512 static inline void schedule_debug(struct task_struct *prev)
3515 * Test if we are atomic. Since do_exit() needs to call into
3516 * schedule() atomically, we ignore that path for now.
3517 * Otherwise, whine if we are scheduling when we should not be.
3519 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3520 __schedule_bug(prev);
3522 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3524 schedstat_inc(this_rq(), sched_count);
3525 #ifdef CONFIG_SCHEDSTATS
3526 if (unlikely(prev->lock_depth >= 0)) {
3527 schedstat_inc(this_rq(), bkl_count);
3528 schedstat_inc(prev, sched_info.bkl_count);
3533 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3536 update_rq_clock(rq);
3537 rq->skip_clock_update = 0;
3538 prev->sched_class->put_prev_task(rq, prev);
3542 * Pick up the highest-prio task:
3544 static inline struct task_struct *
3545 pick_next_task(struct rq *rq)
3547 const struct sched_class *class;
3548 struct task_struct *p;
3551 * Optimization: we know that if all tasks are in
3552 * the fair class we can call that function directly:
3554 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3555 p = fair_sched_class.pick_next_task(rq);
3560 class = sched_class_highest;
3562 p = class->pick_next_task(rq);
3566 * Will never be NULL as the idle class always
3567 * returns a non-NULL p:
3569 class = class->next;
3574 * schedule() is the main scheduler function.
3576 asmlinkage void __sched schedule(void)
3578 struct task_struct *prev, *next;
3579 unsigned long *switch_count;
3585 cpu = smp_processor_id();
3587 rcu_note_context_switch(cpu);
3589 switch_count = &prev->nivcsw;
3591 release_kernel_lock(prev);
3592 need_resched_nonpreemptible:
3594 schedule_debug(prev);
3596 if (sched_feat(HRTICK))
3599 raw_spin_lock_irq(&rq->lock);
3600 clear_tsk_need_resched(prev);
3602 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3603 if (unlikely(signal_pending_state(prev->state, prev)))
3604 prev->state = TASK_RUNNING;
3606 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3607 switch_count = &prev->nvcsw;
3610 pre_schedule(rq, prev);
3612 if (unlikely(!rq->nr_running))
3613 idle_balance(cpu, rq);
3615 put_prev_task(rq, prev);
3616 next = pick_next_task(rq);
3618 if (likely(prev != next)) {
3619 sched_info_switch(prev, next);
3620 perf_event_task_sched_out(prev, next);
3626 context_switch(rq, prev, next); /* unlocks the rq */
3628 * the context switch might have flipped the stack from under
3629 * us, hence refresh the local variables.
3631 cpu = smp_processor_id();
3634 raw_spin_unlock_irq(&rq->lock);
3638 if (unlikely(reacquire_kernel_lock(current) < 0)) {
3640 switch_count = &prev->nivcsw;
3641 goto need_resched_nonpreemptible;
3644 preempt_enable_no_resched();
3648 EXPORT_SYMBOL(schedule);
3650 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3652 * Look out! "owner" is an entirely speculative pointer
3653 * access and not reliable.
3655 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
3660 if (!sched_feat(OWNER_SPIN))
3663 #ifdef CONFIG_DEBUG_PAGEALLOC
3665 * Need to access the cpu field knowing that
3666 * DEBUG_PAGEALLOC could have unmapped it if
3667 * the mutex owner just released it and exited.
3669 if (probe_kernel_address(&owner->cpu, cpu))
3676 * Even if the access succeeded (likely case),
3677 * the cpu field may no longer be valid.
3679 if (cpu >= nr_cpumask_bits)
3683 * We need to validate that we can do a
3684 * get_cpu() and that we have the percpu area.
3686 if (!cpu_online(cpu))
3693 * Owner changed, break to re-assess state.
3695 if (lock->owner != owner)
3699 * Is that owner really running on that cpu?
3701 if (task_thread_info(rq->curr) != owner || need_resched())
3711 #ifdef CONFIG_PREEMPT
3713 * this is the entry point to schedule() from in-kernel preemption
3714 * off of preempt_enable. Kernel preemptions off return from interrupt
3715 * occur there and call schedule directly.
3717 asmlinkage void __sched preempt_schedule(void)
3719 struct thread_info *ti = current_thread_info();
3722 * If there is a non-zero preempt_count or interrupts are disabled,
3723 * we do not want to preempt the current task. Just return..
3725 if (likely(ti->preempt_count || irqs_disabled()))
3729 add_preempt_count(PREEMPT_ACTIVE);
3731 sub_preempt_count(PREEMPT_ACTIVE);
3734 * Check again in case we missed a preemption opportunity
3735 * between schedule and now.
3738 } while (need_resched());
3740 EXPORT_SYMBOL(preempt_schedule);
3743 * this is the entry point to schedule() from kernel preemption
3744 * off of irq context.
3745 * Note, that this is called and return with irqs disabled. This will
3746 * protect us against recursive calling from irq.
3748 asmlinkage void __sched preempt_schedule_irq(void)
3750 struct thread_info *ti = current_thread_info();
3752 /* Catch callers which need to be fixed */
3753 BUG_ON(ti->preempt_count || !irqs_disabled());
3756 add_preempt_count(PREEMPT_ACTIVE);
3759 local_irq_disable();
3760 sub_preempt_count(PREEMPT_ACTIVE);
3763 * Check again in case we missed a preemption opportunity
3764 * between schedule and now.
3767 } while (need_resched());
3770 #endif /* CONFIG_PREEMPT */
3772 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3775 return try_to_wake_up(curr->private, mode, wake_flags);
3777 EXPORT_SYMBOL(default_wake_function);
3780 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3781 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3782 * number) then we wake all the non-exclusive tasks and one exclusive task.
3784 * There are circumstances in which we can try to wake a task which has already
3785 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3786 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3788 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3789 int nr_exclusive, int wake_flags, void *key)
3791 wait_queue_t *curr, *next;
3793 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3794 unsigned flags = curr->flags;
3796 if (curr->func(curr, mode, wake_flags, key) &&
3797 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3803 * __wake_up - wake up threads blocked on a waitqueue.
3805 * @mode: which threads
3806 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3807 * @key: is directly passed to the wakeup function
3809 * It may be assumed that this function implies a write memory barrier before
3810 * changing the task state if and only if any tasks are woken up.
3812 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3813 int nr_exclusive, void *key)
3815 unsigned long flags;
3817 spin_lock_irqsave(&q->lock, flags);
3818 __wake_up_common(q, mode, nr_exclusive, 0, key);
3819 spin_unlock_irqrestore(&q->lock, flags);
3821 EXPORT_SYMBOL(__wake_up);
3824 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3826 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3828 __wake_up_common(q, mode, 1, 0, NULL);
3830 EXPORT_SYMBOL_GPL(__wake_up_locked);
3832 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3834 __wake_up_common(q, mode, 1, 0, key);
3838 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3840 * @mode: which threads
3841 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3842 * @key: opaque value to be passed to wakeup targets
3844 * The sync wakeup differs that the waker knows that it will schedule
3845 * away soon, so while the target thread will be woken up, it will not
3846 * be migrated to another CPU - ie. the two threads are 'synchronized'
3847 * with each other. This can prevent needless bouncing between CPUs.
3849 * On UP it can prevent extra preemption.
3851 * It may be assumed that this function implies a write memory barrier before
3852 * changing the task state if and only if any tasks are woken up.
3854 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3855 int nr_exclusive, void *key)
3857 unsigned long flags;
3858 int wake_flags = WF_SYNC;
3863 if (unlikely(!nr_exclusive))
3866 spin_lock_irqsave(&q->lock, flags);
3867 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3868 spin_unlock_irqrestore(&q->lock, flags);
3870 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3873 * __wake_up_sync - see __wake_up_sync_key()
3875 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3877 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
3879 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3882 * complete: - signals a single thread waiting on this completion
3883 * @x: holds the state of this particular completion
3885 * This will wake up a single thread waiting on this completion. Threads will be
3886 * awakened in the same order in which they were queued.
3888 * See also complete_all(), wait_for_completion() and related routines.
3890 * It may be assumed that this function implies a write memory barrier before
3891 * changing the task state if and only if any tasks are woken up.
3893 void complete(struct completion *x)
3895 unsigned long flags;
3897 spin_lock_irqsave(&x->wait.lock, flags);
3899 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3900 spin_unlock_irqrestore(&x->wait.lock, flags);
3902 EXPORT_SYMBOL(complete);
3905 * complete_all: - signals all threads waiting on this completion
3906 * @x: holds the state of this particular completion
3908 * This will wake up all threads waiting on this particular completion event.
3910 * It may be assumed that this function implies a write memory barrier before
3911 * changing the task state if and only if any tasks are woken up.
3913 void complete_all(struct completion *x)
3915 unsigned long flags;
3917 spin_lock_irqsave(&x->wait.lock, flags);
3918 x->done += UINT_MAX/2;
3919 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3920 spin_unlock_irqrestore(&x->wait.lock, flags);
3922 EXPORT_SYMBOL(complete_all);
3924 static inline long __sched
3925 do_wait_for_common(struct completion *x, long timeout, int state)
3928 DECLARE_WAITQUEUE(wait, current);
3930 __add_wait_queue_tail_exclusive(&x->wait, &wait);
3932 if (signal_pending_state(state, current)) {
3933 timeout = -ERESTARTSYS;
3936 __set_current_state(state);
3937 spin_unlock_irq(&x->wait.lock);
3938 timeout = schedule_timeout(timeout);
3939 spin_lock_irq(&x->wait.lock);
3940 } while (!x->done && timeout);
3941 __remove_wait_queue(&x->wait, &wait);
3946 return timeout ?: 1;
3950 wait_for_common(struct completion *x, long timeout, int state)
3954 spin_lock_irq(&x->wait.lock);
3955 timeout = do_wait_for_common(x, timeout, state);
3956 spin_unlock_irq(&x->wait.lock);
3961 * wait_for_completion: - waits for completion of a task
3962 * @x: holds the state of this particular completion
3964 * This waits to be signaled for completion of a specific task. It is NOT
3965 * interruptible and there is no timeout.
3967 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3968 * and interrupt capability. Also see complete().
3970 void __sched wait_for_completion(struct completion *x)
3972 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3974 EXPORT_SYMBOL(wait_for_completion);
3977 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3978 * @x: holds the state of this particular completion
3979 * @timeout: timeout value in jiffies
3981 * This waits for either a completion of a specific task to be signaled or for a
3982 * specified timeout to expire. The timeout is in jiffies. It is not
3985 unsigned long __sched
3986 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3988 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3990 EXPORT_SYMBOL(wait_for_completion_timeout);
3993 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3994 * @x: holds the state of this particular completion
3996 * This waits for completion of a specific task to be signaled. It is
3999 int __sched wait_for_completion_interruptible(struct completion *x)
4001 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4002 if (t == -ERESTARTSYS)
4006 EXPORT_SYMBOL(wait_for_completion_interruptible);
4009 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4010 * @x: holds the state of this particular completion
4011 * @timeout: timeout value in jiffies
4013 * This waits for either a completion of a specific task to be signaled or for a
4014 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4016 unsigned long __sched
4017 wait_for_completion_interruptible_timeout(struct completion *x,
4018 unsigned long timeout)
4020 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4022 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4025 * wait_for_completion_killable: - waits for completion of a task (killable)
4026 * @x: holds the state of this particular completion
4028 * This waits to be signaled for completion of a specific task. It can be
4029 * interrupted by a kill signal.
4031 int __sched wait_for_completion_killable(struct completion *x)
4033 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4034 if (t == -ERESTARTSYS)
4038 EXPORT_SYMBOL(wait_for_completion_killable);
4041 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4042 * @x: holds the state of this particular completion
4043 * @timeout: timeout value in jiffies
4045 * This waits for either a completion of a specific task to be
4046 * signaled or for a specified timeout to expire. It can be
4047 * interrupted by a kill signal. The timeout is in jiffies.
4049 unsigned long __sched
4050 wait_for_completion_killable_timeout(struct completion *x,
4051 unsigned long timeout)
4053 return wait_for_common(x, timeout, TASK_KILLABLE);
4055 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
4058 * try_wait_for_completion - try to decrement a completion without blocking
4059 * @x: completion structure
4061 * Returns: 0 if a decrement cannot be done without blocking
4062 * 1 if a decrement succeeded.
4064 * If a completion is being used as a counting completion,
4065 * attempt to decrement the counter without blocking. This
4066 * enables us to avoid waiting if the resource the completion
4067 * is protecting is not available.
4069 bool try_wait_for_completion(struct completion *x)
4071 unsigned long flags;
4074 spin_lock_irqsave(&x->wait.lock, flags);
4079 spin_unlock_irqrestore(&x->wait.lock, flags);
4082 EXPORT_SYMBOL(try_wait_for_completion);
4085 * completion_done - Test to see if a completion has any waiters
4086 * @x: completion structure
4088 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4089 * 1 if there are no waiters.
4092 bool completion_done(struct completion *x)
4094 unsigned long flags;
4097 spin_lock_irqsave(&x->wait.lock, flags);
4100 spin_unlock_irqrestore(&x->wait.lock, flags);
4103 EXPORT_SYMBOL(completion_done);
4106 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4108 unsigned long flags;
4111 init_waitqueue_entry(&wait, current);
4113 __set_current_state(state);
4115 spin_lock_irqsave(&q->lock, flags);
4116 __add_wait_queue(q, &wait);
4117 spin_unlock(&q->lock);
4118 timeout = schedule_timeout(timeout);
4119 spin_lock_irq(&q->lock);
4120 __remove_wait_queue(q, &wait);
4121 spin_unlock_irqrestore(&q->lock, flags);
4126 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4128 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4130 EXPORT_SYMBOL(interruptible_sleep_on);
4133 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4135 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4137 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4139 void __sched sleep_on(wait_queue_head_t *q)
4141 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4143 EXPORT_SYMBOL(sleep_on);
4145 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4147 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4149 EXPORT_SYMBOL(sleep_on_timeout);
4151 #ifdef CONFIG_RT_MUTEXES
4154 * rt_mutex_setprio - set the current priority of a task
4156 * @prio: prio value (kernel-internal form)
4158 * This function changes the 'effective' priority of a task. It does
4159 * not touch ->normal_prio like __setscheduler().
4161 * Used by the rt_mutex code to implement priority inheritance logic.
4163 void rt_mutex_setprio(struct task_struct *p, int prio)
4165 unsigned long flags;
4166 int oldprio, on_rq, running;
4168 const struct sched_class *prev_class;
4170 BUG_ON(prio < 0 || prio > MAX_PRIO);
4172 rq = task_rq_lock(p, &flags);
4175 prev_class = p->sched_class;
4176 on_rq = p->se.on_rq;
4177 running = task_current(rq, p);
4179 dequeue_task(rq, p, 0);
4181 p->sched_class->put_prev_task(rq, p);
4184 p->sched_class = &rt_sched_class;
4186 p->sched_class = &fair_sched_class;
4191 p->sched_class->set_curr_task(rq);
4193 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4195 check_class_changed(rq, p, prev_class, oldprio, running);
4197 task_rq_unlock(rq, &flags);
4202 void set_user_nice(struct task_struct *p, long nice)
4204 int old_prio, delta, on_rq;
4205 unsigned long flags;
4208 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4211 * We have to be careful, if called from sys_setpriority(),
4212 * the task might be in the middle of scheduling on another CPU.
4214 rq = task_rq_lock(p, &flags);
4216 * The RT priorities are set via sched_setscheduler(), but we still
4217 * allow the 'normal' nice value to be set - but as expected
4218 * it wont have any effect on scheduling until the task is
4219 * SCHED_FIFO/SCHED_RR:
4221 if (task_has_rt_policy(p)) {
4222 p->static_prio = NICE_TO_PRIO(nice);
4225 on_rq = p->se.on_rq;
4227 dequeue_task(rq, p, 0);
4229 p->static_prio = NICE_TO_PRIO(nice);
4232 p->prio = effective_prio(p);
4233 delta = p->prio - old_prio;
4236 enqueue_task(rq, p, 0);
4238 * If the task increased its priority or is running and
4239 * lowered its priority, then reschedule its CPU:
4241 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4242 resched_task(rq->curr);
4245 task_rq_unlock(rq, &flags);
4247 EXPORT_SYMBOL(set_user_nice);
4250 * can_nice - check if a task can reduce its nice value
4254 int can_nice(const struct task_struct *p, const int nice)
4256 /* convert nice value [19,-20] to rlimit style value [1,40] */
4257 int nice_rlim = 20 - nice;
4259 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4260 capable(CAP_SYS_NICE));
4263 #ifdef __ARCH_WANT_SYS_NICE
4266 * sys_nice - change the priority of the current process.
4267 * @increment: priority increment
4269 * sys_setpriority is a more generic, but much slower function that
4270 * does similar things.
4272 SYSCALL_DEFINE1(nice, int, increment)
4277 * Setpriority might change our priority at the same moment.
4278 * We don't have to worry. Conceptually one call occurs first
4279 * and we have a single winner.
4281 if (increment < -40)
4286 nice = TASK_NICE(current) + increment;
4292 if (increment < 0 && !can_nice(current, nice))
4295 retval = security_task_setnice(current, nice);
4299 set_user_nice(current, nice);
4306 * task_prio - return the priority value of a given task.
4307 * @p: the task in question.
4309 * This is the priority value as seen by users in /proc.
4310 * RT tasks are offset by -200. Normal tasks are centered
4311 * around 0, value goes from -16 to +15.
4313 int task_prio(const struct task_struct *p)
4315 return p->prio - MAX_RT_PRIO;
4319 * task_nice - return the nice value of a given task.
4320 * @p: the task in question.
4322 int task_nice(const struct task_struct *p)
4324 return TASK_NICE(p);
4326 EXPORT_SYMBOL(task_nice);
4329 * idle_cpu - is a given cpu idle currently?
4330 * @cpu: the processor in question.
4332 int idle_cpu(int cpu)
4334 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4338 * idle_task - return the idle task for a given cpu.
4339 * @cpu: the processor in question.
4341 struct task_struct *idle_task(int cpu)
4343 return cpu_rq(cpu)->idle;
4347 * find_process_by_pid - find a process with a matching PID value.
4348 * @pid: the pid in question.
4350 static struct task_struct *find_process_by_pid(pid_t pid)
4352 return pid ? find_task_by_vpid(pid) : current;
4355 /* Actually do priority change: must hold rq lock. */
4357 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4359 BUG_ON(p->se.on_rq);
4362 p->rt_priority = prio;
4363 p->normal_prio = normal_prio(p);
4364 /* we are holding p->pi_lock already */
4365 p->prio = rt_mutex_getprio(p);
4366 if (rt_prio(p->prio))
4367 p->sched_class = &rt_sched_class;
4369 p->sched_class = &fair_sched_class;
4374 * check the target process has a UID that matches the current process's
4376 static bool check_same_owner(struct task_struct *p)
4378 const struct cred *cred = current_cred(), *pcred;
4382 pcred = __task_cred(p);
4383 match = (cred->euid == pcred->euid ||
4384 cred->euid == pcred->uid);
4389 static int __sched_setscheduler(struct task_struct *p, int policy,
4390 struct sched_param *param, bool user)
4392 int retval, oldprio, oldpolicy = -1, on_rq, running;
4393 unsigned long flags;
4394 const struct sched_class *prev_class;
4398 /* may grab non-irq protected spin_locks */
4399 BUG_ON(in_interrupt());
4401 /* double check policy once rq lock held */
4403 reset_on_fork = p->sched_reset_on_fork;
4404 policy = oldpolicy = p->policy;
4406 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4407 policy &= ~SCHED_RESET_ON_FORK;
4409 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4410 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4411 policy != SCHED_IDLE)
4416 * Valid priorities for SCHED_FIFO and SCHED_RR are
4417 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4418 * SCHED_BATCH and SCHED_IDLE is 0.
4420 if (param->sched_priority < 0 ||
4421 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4422 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4424 if (rt_policy(policy) != (param->sched_priority != 0))
4428 * Allow unprivileged RT tasks to decrease priority:
4430 if (user && !capable(CAP_SYS_NICE)) {
4431 if (rt_policy(policy)) {
4432 unsigned long rlim_rtprio;
4434 if (!lock_task_sighand(p, &flags))
4436 rlim_rtprio = task_rlimit(p, RLIMIT_RTPRIO);
4437 unlock_task_sighand(p, &flags);
4439 /* can't set/change the rt policy */
4440 if (policy != p->policy && !rlim_rtprio)
4443 /* can't increase priority */
4444 if (param->sched_priority > p->rt_priority &&
4445 param->sched_priority > rlim_rtprio)
4449 * Like positive nice levels, dont allow tasks to
4450 * move out of SCHED_IDLE either:
4452 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4455 /* can't change other user's priorities */
4456 if (!check_same_owner(p))
4459 /* Normal users shall not reset the sched_reset_on_fork flag */
4460 if (p->sched_reset_on_fork && !reset_on_fork)
4465 retval = security_task_setscheduler(p, policy, param);
4471 * make sure no PI-waiters arrive (or leave) while we are
4472 * changing the priority of the task:
4474 raw_spin_lock_irqsave(&p->pi_lock, flags);
4476 * To be able to change p->policy safely, the apropriate
4477 * runqueue lock must be held.
4479 rq = __task_rq_lock(p);
4481 #ifdef CONFIG_RT_GROUP_SCHED
4484 * Do not allow realtime tasks into groups that have no runtime
4487 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4488 task_group(p)->rt_bandwidth.rt_runtime == 0) {
4489 __task_rq_unlock(rq);
4490 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4496 /* recheck policy now with rq lock held */
4497 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4498 policy = oldpolicy = -1;
4499 __task_rq_unlock(rq);
4500 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4503 on_rq = p->se.on_rq;
4504 running = task_current(rq, p);
4506 deactivate_task(rq, p, 0);
4508 p->sched_class->put_prev_task(rq, p);
4510 p->sched_reset_on_fork = reset_on_fork;
4513 prev_class = p->sched_class;
4514 __setscheduler(rq, p, policy, param->sched_priority);
4517 p->sched_class->set_curr_task(rq);
4519 activate_task(rq, p, 0);
4521 check_class_changed(rq, p, prev_class, oldprio, running);
4523 __task_rq_unlock(rq);
4524 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4526 rt_mutex_adjust_pi(p);
4532 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4533 * @p: the task in question.
4534 * @policy: new policy.
4535 * @param: structure containing the new RT priority.
4537 * NOTE that the task may be already dead.
4539 int sched_setscheduler(struct task_struct *p, int policy,
4540 struct sched_param *param)
4542 return __sched_setscheduler(p, policy, param, true);
4544 EXPORT_SYMBOL_GPL(sched_setscheduler);
4547 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4548 * @p: the task in question.
4549 * @policy: new policy.
4550 * @param: structure containing the new RT priority.
4552 * Just like sched_setscheduler, only don't bother checking if the
4553 * current context has permission. For example, this is needed in
4554 * stop_machine(): we create temporary high priority worker threads,
4555 * but our caller might not have that capability.
4557 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4558 struct sched_param *param)
4560 return __sched_setscheduler(p, policy, param, false);
4564 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4566 struct sched_param lparam;
4567 struct task_struct *p;
4570 if (!param || pid < 0)
4572 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4577 p = find_process_by_pid(pid);
4579 retval = sched_setscheduler(p, policy, &lparam);
4586 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4587 * @pid: the pid in question.
4588 * @policy: new policy.
4589 * @param: structure containing the new RT priority.
4591 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4592 struct sched_param __user *, param)
4594 /* negative values for policy are not valid */
4598 return do_sched_setscheduler(pid, policy, param);
4602 * sys_sched_setparam - set/change the RT priority of a thread
4603 * @pid: the pid in question.
4604 * @param: structure containing the new RT priority.
4606 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4608 return do_sched_setscheduler(pid, -1, param);
4612 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4613 * @pid: the pid in question.
4615 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4617 struct task_struct *p;
4625 p = find_process_by_pid(pid);
4627 retval = security_task_getscheduler(p);
4630 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4637 * sys_sched_getparam - get the RT priority of a thread
4638 * @pid: the pid in question.
4639 * @param: structure containing the RT priority.
4641 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4643 struct sched_param lp;
4644 struct task_struct *p;
4647 if (!param || pid < 0)
4651 p = find_process_by_pid(pid);
4656 retval = security_task_getscheduler(p);
4660 lp.sched_priority = p->rt_priority;
4664 * This one might sleep, we cannot do it with a spinlock held ...
4666 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4675 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4677 cpumask_var_t cpus_allowed, new_mask;
4678 struct task_struct *p;
4684 p = find_process_by_pid(pid);
4691 /* Prevent p going away */
4695 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4699 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4701 goto out_free_cpus_allowed;
4704 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
4707 retval = security_task_setscheduler(p, 0, NULL);
4711 cpuset_cpus_allowed(p, cpus_allowed);
4712 cpumask_and(new_mask, in_mask, cpus_allowed);
4714 retval = set_cpus_allowed_ptr(p, new_mask);
4717 cpuset_cpus_allowed(p, cpus_allowed);
4718 if (!cpumask_subset(new_mask, cpus_allowed)) {
4720 * We must have raced with a concurrent cpuset
4721 * update. Just reset the cpus_allowed to the
4722 * cpuset's cpus_allowed
4724 cpumask_copy(new_mask, cpus_allowed);
4729 free_cpumask_var(new_mask);
4730 out_free_cpus_allowed:
4731 free_cpumask_var(cpus_allowed);
4738 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4739 struct cpumask *new_mask)
4741 if (len < cpumask_size())
4742 cpumask_clear(new_mask);
4743 else if (len > cpumask_size())
4744 len = cpumask_size();
4746 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4750 * sys_sched_setaffinity - set the cpu affinity of a process
4751 * @pid: pid of the process
4752 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4753 * @user_mask_ptr: user-space pointer to the new cpu mask
4755 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4756 unsigned long __user *, user_mask_ptr)
4758 cpumask_var_t new_mask;
4761 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4764 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4766 retval = sched_setaffinity(pid, new_mask);
4767 free_cpumask_var(new_mask);
4771 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4773 struct task_struct *p;
4774 unsigned long flags;
4782 p = find_process_by_pid(pid);
4786 retval = security_task_getscheduler(p);
4790 rq = task_rq_lock(p, &flags);
4791 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4792 task_rq_unlock(rq, &flags);
4802 * sys_sched_getaffinity - get the cpu affinity of a process
4803 * @pid: pid of the process
4804 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4805 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4807 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4808 unsigned long __user *, user_mask_ptr)
4813 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4815 if (len & (sizeof(unsigned long)-1))
4818 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4821 ret = sched_getaffinity(pid, mask);
4823 size_t retlen = min_t(size_t, len, cpumask_size());
4825 if (copy_to_user(user_mask_ptr, mask, retlen))
4830 free_cpumask_var(mask);
4836 * sys_sched_yield - yield the current processor to other threads.
4838 * This function yields the current CPU to other tasks. If there are no
4839 * other threads running on this CPU then this function will return.
4841 SYSCALL_DEFINE0(sched_yield)
4843 struct rq *rq = this_rq_lock();
4845 schedstat_inc(rq, yld_count);
4846 current->sched_class->yield_task(rq);
4849 * Since we are going to call schedule() anyway, there's
4850 * no need to preempt or enable interrupts:
4852 __release(rq->lock);
4853 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4854 do_raw_spin_unlock(&rq->lock);
4855 preempt_enable_no_resched();
4862 static inline int should_resched(void)
4864 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4867 static void __cond_resched(void)
4869 add_preempt_count(PREEMPT_ACTIVE);
4871 sub_preempt_count(PREEMPT_ACTIVE);
4874 int __sched _cond_resched(void)
4876 if (should_resched()) {
4882 EXPORT_SYMBOL(_cond_resched);
4885 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4886 * call schedule, and on return reacquire the lock.
4888 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4889 * operations here to prevent schedule() from being called twice (once via
4890 * spin_unlock(), once by hand).
4892 int __cond_resched_lock(spinlock_t *lock)
4894 int resched = should_resched();
4897 lockdep_assert_held(lock);
4899 if (spin_needbreak(lock) || resched) {
4910 EXPORT_SYMBOL(__cond_resched_lock);
4912 int __sched __cond_resched_softirq(void)
4914 BUG_ON(!in_softirq());
4916 if (should_resched()) {
4924 EXPORT_SYMBOL(__cond_resched_softirq);
4927 * yield - yield the current processor to other threads.
4929 * This is a shortcut for kernel-space yielding - it marks the
4930 * thread runnable and calls sys_sched_yield().
4932 void __sched yield(void)
4934 set_current_state(TASK_RUNNING);
4937 EXPORT_SYMBOL(yield);
4940 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4941 * that process accounting knows that this is a task in IO wait state.
4943 void __sched io_schedule(void)
4945 struct rq *rq = raw_rq();
4947 delayacct_blkio_start();
4948 atomic_inc(&rq->nr_iowait);
4949 current->in_iowait = 1;
4951 current->in_iowait = 0;
4952 atomic_dec(&rq->nr_iowait);
4953 delayacct_blkio_end();
4955 EXPORT_SYMBOL(io_schedule);
4957 long __sched io_schedule_timeout(long timeout)
4959 struct rq *rq = raw_rq();
4962 delayacct_blkio_start();
4963 atomic_inc(&rq->nr_iowait);
4964 current->in_iowait = 1;
4965 ret = schedule_timeout(timeout);
4966 current->in_iowait = 0;
4967 atomic_dec(&rq->nr_iowait);
4968 delayacct_blkio_end();
4973 * sys_sched_get_priority_max - return maximum RT priority.
4974 * @policy: scheduling class.
4976 * this syscall returns the maximum rt_priority that can be used
4977 * by a given scheduling class.
4979 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4986 ret = MAX_USER_RT_PRIO-1;
4998 * sys_sched_get_priority_min - return minimum RT priority.
4999 * @policy: scheduling class.
5001 * this syscall returns the minimum rt_priority that can be used
5002 * by a given scheduling class.
5004 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5022 * sys_sched_rr_get_interval - return the default timeslice of a process.
5023 * @pid: pid of the process.
5024 * @interval: userspace pointer to the timeslice value.
5026 * this syscall writes the default timeslice value of a given process
5027 * into the user-space timespec buffer. A value of '0' means infinity.
5029 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5030 struct timespec __user *, interval)
5032 struct task_struct *p;
5033 unsigned int time_slice;
5034 unsigned long flags;
5044 p = find_process_by_pid(pid);
5048 retval = security_task_getscheduler(p);
5052 rq = task_rq_lock(p, &flags);
5053 time_slice = p->sched_class->get_rr_interval(rq, p);
5054 task_rq_unlock(rq, &flags);
5057 jiffies_to_timespec(time_slice, &t);
5058 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5066 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5068 void sched_show_task(struct task_struct *p)
5070 unsigned long free = 0;
5073 state = p->state ? __ffs(p->state) + 1 : 0;
5074 printk(KERN_INFO "%-13.13s %c", p->comm,
5075 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5076 #if BITS_PER_LONG == 32
5077 if (state == TASK_RUNNING)
5078 printk(KERN_CONT " running ");
5080 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5082 if (state == TASK_RUNNING)
5083 printk(KERN_CONT " running task ");
5085 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5087 #ifdef CONFIG_DEBUG_STACK_USAGE
5088 free = stack_not_used(p);
5090 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5091 task_pid_nr(p), task_pid_nr(p->real_parent),
5092 (unsigned long)task_thread_info(p)->flags);
5094 show_stack(p, NULL);
5097 void show_state_filter(unsigned long state_filter)
5099 struct task_struct *g, *p;
5101 #if BITS_PER_LONG == 32
5103 " task PC stack pid father\n");
5106 " task PC stack pid father\n");
5108 read_lock(&tasklist_lock);
5109 do_each_thread(g, p) {
5111 * reset the NMI-timeout, listing all files on a slow
5112 * console might take alot of time:
5114 touch_nmi_watchdog();
5115 if (!state_filter || (p->state & state_filter))
5117 } while_each_thread(g, p);
5119 touch_all_softlockup_watchdogs();
5121 #ifdef CONFIG_SCHED_DEBUG
5122 sysrq_sched_debug_show();
5124 read_unlock(&tasklist_lock);
5126 * Only show locks if all tasks are dumped:
5129 debug_show_all_locks();
5132 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5134 idle->sched_class = &idle_sched_class;
5138 * init_idle - set up an idle thread for a given CPU
5139 * @idle: task in question
5140 * @cpu: cpu the idle task belongs to
5142 * NOTE: this function does not set the idle thread's NEED_RESCHED
5143 * flag, to make booting more robust.
5145 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5147 struct rq *rq = cpu_rq(cpu);
5148 unsigned long flags;
5150 raw_spin_lock_irqsave(&rq->lock, flags);
5153 idle->state = TASK_RUNNING;
5154 idle->se.exec_start = sched_clock();
5156 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
5157 __set_task_cpu(idle, cpu);
5159 rq->curr = rq->idle = idle;
5160 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5163 raw_spin_unlock_irqrestore(&rq->lock, flags);
5165 /* Set the preempt count _outside_ the spinlocks! */
5166 #if defined(CONFIG_PREEMPT)
5167 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5169 task_thread_info(idle)->preempt_count = 0;
5172 * The idle tasks have their own, simple scheduling class:
5174 idle->sched_class = &idle_sched_class;
5175 ftrace_graph_init_task(idle);
5179 * In a system that switches off the HZ timer nohz_cpu_mask
5180 * indicates which cpus entered this state. This is used
5181 * in the rcu update to wait only for active cpus. For system
5182 * which do not switch off the HZ timer nohz_cpu_mask should
5183 * always be CPU_BITS_NONE.
5185 cpumask_var_t nohz_cpu_mask;
5188 * Increase the granularity value when there are more CPUs,
5189 * because with more CPUs the 'effective latency' as visible
5190 * to users decreases. But the relationship is not linear,
5191 * so pick a second-best guess by going with the log2 of the
5194 * This idea comes from the SD scheduler of Con Kolivas:
5196 static int get_update_sysctl_factor(void)
5198 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5199 unsigned int factor;
5201 switch (sysctl_sched_tunable_scaling) {
5202 case SCHED_TUNABLESCALING_NONE:
5205 case SCHED_TUNABLESCALING_LINEAR:
5208 case SCHED_TUNABLESCALING_LOG:
5210 factor = 1 + ilog2(cpus);
5217 static void update_sysctl(void)
5219 unsigned int factor = get_update_sysctl_factor();
5221 #define SET_SYSCTL(name) \
5222 (sysctl_##name = (factor) * normalized_sysctl_##name)
5223 SET_SYSCTL(sched_min_granularity);
5224 SET_SYSCTL(sched_latency);
5225 SET_SYSCTL(sched_wakeup_granularity);
5226 SET_SYSCTL(sched_shares_ratelimit);
5230 static inline void sched_init_granularity(void)
5237 * This is how migration works:
5239 * 1) we invoke migration_cpu_stop() on the target CPU using
5241 * 2) stopper starts to run (implicitly forcing the migrated thread
5243 * 3) it checks whether the migrated task is still in the wrong runqueue.
5244 * 4) if it's in the wrong runqueue then the migration thread removes
5245 * it and puts it into the right queue.
5246 * 5) stopper completes and stop_one_cpu() returns and the migration
5251 * Change a given task's CPU affinity. Migrate the thread to a
5252 * proper CPU and schedule it away if the CPU it's executing on
5253 * is removed from the allowed bitmask.
5255 * NOTE: the caller must have a valid reference to the task, the
5256 * task must not exit() & deallocate itself prematurely. The
5257 * call is not atomic; no spinlocks may be held.
5259 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5261 unsigned long flags;
5263 unsigned int dest_cpu;
5267 * Serialize against TASK_WAKING so that ttwu() and wunt() can
5268 * drop the rq->lock and still rely on ->cpus_allowed.
5271 while (task_is_waking(p))
5273 rq = task_rq_lock(p, &flags);
5274 if (task_is_waking(p)) {
5275 task_rq_unlock(rq, &flags);
5279 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5284 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5285 !cpumask_equal(&p->cpus_allowed, new_mask))) {
5290 if (p->sched_class->set_cpus_allowed)
5291 p->sched_class->set_cpus_allowed(p, new_mask);
5293 cpumask_copy(&p->cpus_allowed, new_mask);
5294 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5297 /* Can the task run on the task's current CPU? If so, we're done */
5298 if (cpumask_test_cpu(task_cpu(p), new_mask))
5301 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
5302 if (migrate_task(p, dest_cpu)) {
5303 struct migration_arg arg = { p, dest_cpu };
5304 /* Need help from migration thread: drop lock and wait. */
5305 task_rq_unlock(rq, &flags);
5306 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
5307 tlb_migrate_finish(p->mm);
5311 task_rq_unlock(rq, &flags);
5315 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5318 * Move (not current) task off this cpu, onto dest cpu. We're doing
5319 * this because either it can't run here any more (set_cpus_allowed()
5320 * away from this CPU, or CPU going down), or because we're
5321 * attempting to rebalance this task on exec (sched_exec).
5323 * So we race with normal scheduler movements, but that's OK, as long
5324 * as the task is no longer on this CPU.
5326 * Returns non-zero if task was successfully migrated.
5328 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5330 struct rq *rq_dest, *rq_src;
5333 if (unlikely(!cpu_active(dest_cpu)))
5336 rq_src = cpu_rq(src_cpu);
5337 rq_dest = cpu_rq(dest_cpu);
5339 double_rq_lock(rq_src, rq_dest);
5340 /* Already moved. */
5341 if (task_cpu(p) != src_cpu)
5343 /* Affinity changed (again). */
5344 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
5348 * If we're not on a rq, the next wake-up will ensure we're
5352 deactivate_task(rq_src, p, 0);
5353 set_task_cpu(p, dest_cpu);
5354 activate_task(rq_dest, p, 0);
5355 check_preempt_curr(rq_dest, p, 0);
5360 double_rq_unlock(rq_src, rq_dest);
5365 * migration_cpu_stop - this will be executed by a highprio stopper thread
5366 * and performs thread migration by bumping thread off CPU then
5367 * 'pushing' onto another runqueue.
5369 static int migration_cpu_stop(void *data)
5371 struct migration_arg *arg = data;
5374 * The original target cpu might have gone down and we might
5375 * be on another cpu but it doesn't matter.
5377 local_irq_disable();
5378 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
5383 #ifdef CONFIG_HOTPLUG_CPU
5385 * Figure out where task on dead CPU should go, use force if necessary.
5387 void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5389 struct rq *rq = cpu_rq(dead_cpu);
5390 int needs_cpu, uninitialized_var(dest_cpu);
5391 unsigned long flags;
5393 local_irq_save(flags);
5395 raw_spin_lock(&rq->lock);
5396 needs_cpu = (task_cpu(p) == dead_cpu) && (p->state != TASK_WAKING);
5398 dest_cpu = select_fallback_rq(dead_cpu, p);
5399 raw_spin_unlock(&rq->lock);
5401 * It can only fail if we race with set_cpus_allowed(),
5402 * in the racer should migrate the task anyway.
5405 __migrate_task(p, dead_cpu, dest_cpu);
5406 local_irq_restore(flags);
5410 * While a dead CPU has no uninterruptible tasks queued at this point,
5411 * it might still have a nonzero ->nr_uninterruptible counter, because
5412 * for performance reasons the counter is not stricly tracking tasks to
5413 * their home CPUs. So we just add the counter to another CPU's counter,
5414 * to keep the global sum constant after CPU-down:
5416 static void migrate_nr_uninterruptible(struct rq *rq_src)
5418 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5419 unsigned long flags;
5421 local_irq_save(flags);
5422 double_rq_lock(rq_src, rq_dest);
5423 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5424 rq_src->nr_uninterruptible = 0;
5425 double_rq_unlock(rq_src, rq_dest);
5426 local_irq_restore(flags);
5429 /* Run through task list and migrate tasks from the dead cpu. */
5430 static void migrate_live_tasks(int src_cpu)
5432 struct task_struct *p, *t;
5434 read_lock(&tasklist_lock);
5436 do_each_thread(t, p) {
5440 if (task_cpu(p) == src_cpu)
5441 move_task_off_dead_cpu(src_cpu, p);
5442 } while_each_thread(t, p);
5444 read_unlock(&tasklist_lock);
5448 * Schedules idle task to be the next runnable task on current CPU.
5449 * It does so by boosting its priority to highest possible.
5450 * Used by CPU offline code.
5452 void sched_idle_next(void)
5454 int this_cpu = smp_processor_id();
5455 struct rq *rq = cpu_rq(this_cpu);
5456 struct task_struct *p = rq->idle;
5457 unsigned long flags;
5459 /* cpu has to be offline */
5460 BUG_ON(cpu_online(this_cpu));
5463 * Strictly not necessary since rest of the CPUs are stopped by now
5464 * and interrupts disabled on the current cpu.
5466 raw_spin_lock_irqsave(&rq->lock, flags);
5468 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5470 activate_task(rq, p, 0);
5472 raw_spin_unlock_irqrestore(&rq->lock, flags);
5476 * Ensures that the idle task is using init_mm right before its cpu goes
5479 void idle_task_exit(void)
5481 struct mm_struct *mm = current->active_mm;
5483 BUG_ON(cpu_online(smp_processor_id()));
5486 switch_mm(mm, &init_mm, current);
5490 /* called under rq->lock with disabled interrupts */
5491 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5493 struct rq *rq = cpu_rq(dead_cpu);
5495 /* Must be exiting, otherwise would be on tasklist. */
5496 BUG_ON(!p->exit_state);
5498 /* Cannot have done final schedule yet: would have vanished. */
5499 BUG_ON(p->state == TASK_DEAD);
5504 * Drop lock around migration; if someone else moves it,
5505 * that's OK. No task can be added to this CPU, so iteration is
5508 raw_spin_unlock_irq(&rq->lock);
5509 move_task_off_dead_cpu(dead_cpu, p);
5510 raw_spin_lock_irq(&rq->lock);
5515 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5516 static void migrate_dead_tasks(unsigned int dead_cpu)
5518 struct rq *rq = cpu_rq(dead_cpu);
5519 struct task_struct *next;
5522 if (!rq->nr_running)
5524 next = pick_next_task(rq);
5527 next->sched_class->put_prev_task(rq, next);
5528 migrate_dead(dead_cpu, next);
5534 * remove the tasks which were accounted by rq from calc_load_tasks.
5536 static void calc_global_load_remove(struct rq *rq)
5538 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5539 rq->calc_load_active = 0;
5541 #endif /* CONFIG_HOTPLUG_CPU */
5543 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5545 static struct ctl_table sd_ctl_dir[] = {
5547 .procname = "sched_domain",
5553 static struct ctl_table sd_ctl_root[] = {
5555 .procname = "kernel",
5557 .child = sd_ctl_dir,
5562 static struct ctl_table *sd_alloc_ctl_entry(int n)
5564 struct ctl_table *entry =
5565 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5570 static void sd_free_ctl_entry(struct ctl_table **tablep)
5572 struct ctl_table *entry;
5575 * In the intermediate directories, both the child directory and
5576 * procname are dynamically allocated and could fail but the mode
5577 * will always be set. In the lowest directory the names are
5578 * static strings and all have proc handlers.
5580 for (entry = *tablep; entry->mode; entry++) {
5582 sd_free_ctl_entry(&entry->child);
5583 if (entry->proc_handler == NULL)
5584 kfree(entry->procname);
5592 set_table_entry(struct ctl_table *entry,
5593 const char *procname, void *data, int maxlen,
5594 mode_t mode, proc_handler *proc_handler)
5596 entry->procname = procname;
5598 entry->maxlen = maxlen;
5600 entry->proc_handler = proc_handler;
5603 static struct ctl_table *
5604 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5606 struct ctl_table *table = sd_alloc_ctl_entry(13);
5611 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5612 sizeof(long), 0644, proc_doulongvec_minmax);
5613 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5614 sizeof(long), 0644, proc_doulongvec_minmax);
5615 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5616 sizeof(int), 0644, proc_dointvec_minmax);
5617 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5618 sizeof(int), 0644, proc_dointvec_minmax);
5619 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5620 sizeof(int), 0644, proc_dointvec_minmax);
5621 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5622 sizeof(int), 0644, proc_dointvec_minmax);
5623 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5624 sizeof(int), 0644, proc_dointvec_minmax);
5625 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5626 sizeof(int), 0644, proc_dointvec_minmax);
5627 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5628 sizeof(int), 0644, proc_dointvec_minmax);
5629 set_table_entry(&table[9], "cache_nice_tries",
5630 &sd->cache_nice_tries,
5631 sizeof(int), 0644, proc_dointvec_minmax);
5632 set_table_entry(&table[10], "flags", &sd->flags,
5633 sizeof(int), 0644, proc_dointvec_minmax);
5634 set_table_entry(&table[11], "name", sd->name,
5635 CORENAME_MAX_SIZE, 0444, proc_dostring);
5636 /* &table[12] is terminator */
5641 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5643 struct ctl_table *entry, *table;
5644 struct sched_domain *sd;
5645 int domain_num = 0, i;
5648 for_each_domain(cpu, sd)
5650 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5655 for_each_domain(cpu, sd) {
5656 snprintf(buf, 32, "domain%d", i);
5657 entry->procname = kstrdup(buf, GFP_KERNEL);
5659 entry->child = sd_alloc_ctl_domain_table(sd);
5666 static struct ctl_table_header *sd_sysctl_header;
5667 static void register_sched_domain_sysctl(void)
5669 int i, cpu_num = num_possible_cpus();
5670 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5673 WARN_ON(sd_ctl_dir[0].child);
5674 sd_ctl_dir[0].child = entry;
5679 for_each_possible_cpu(i) {
5680 snprintf(buf, 32, "cpu%d", i);
5681 entry->procname = kstrdup(buf, GFP_KERNEL);
5683 entry->child = sd_alloc_ctl_cpu_table(i);
5687 WARN_ON(sd_sysctl_header);
5688 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5691 /* may be called multiple times per register */
5692 static void unregister_sched_domain_sysctl(void)
5694 if (sd_sysctl_header)
5695 unregister_sysctl_table(sd_sysctl_header);
5696 sd_sysctl_header = NULL;
5697 if (sd_ctl_dir[0].child)
5698 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5701 static void register_sched_domain_sysctl(void)
5704 static void unregister_sched_domain_sysctl(void)
5709 static void set_rq_online(struct rq *rq)
5712 const struct sched_class *class;
5714 cpumask_set_cpu(rq->cpu, rq->rd->online);
5717 for_each_class(class) {
5718 if (class->rq_online)
5719 class->rq_online(rq);
5724 static void set_rq_offline(struct rq *rq)
5727 const struct sched_class *class;
5729 for_each_class(class) {
5730 if (class->rq_offline)
5731 class->rq_offline(rq);
5734 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5740 * migration_call - callback that gets triggered when a CPU is added.
5741 * Here we can start up the necessary migration thread for the new CPU.
5743 static int __cpuinit
5744 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5746 int cpu = (long)hcpu;
5747 unsigned long flags;
5748 struct rq *rq = cpu_rq(cpu);
5752 case CPU_UP_PREPARE:
5753 case CPU_UP_PREPARE_FROZEN:
5754 rq->calc_load_update = calc_load_update;
5758 case CPU_ONLINE_FROZEN:
5759 /* Update our root-domain */
5760 raw_spin_lock_irqsave(&rq->lock, flags);
5762 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5766 raw_spin_unlock_irqrestore(&rq->lock, flags);
5769 #ifdef CONFIG_HOTPLUG_CPU
5771 case CPU_DEAD_FROZEN:
5772 migrate_live_tasks(cpu);
5773 /* Idle task back to normal (off runqueue, low prio) */
5774 raw_spin_lock_irq(&rq->lock);
5775 deactivate_task(rq, rq->idle, 0);
5776 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5777 rq->idle->sched_class = &idle_sched_class;
5778 migrate_dead_tasks(cpu);
5779 raw_spin_unlock_irq(&rq->lock);
5780 migrate_nr_uninterruptible(rq);
5781 BUG_ON(rq->nr_running != 0);
5782 calc_global_load_remove(rq);
5786 case CPU_DYING_FROZEN:
5787 /* Update our root-domain */
5788 raw_spin_lock_irqsave(&rq->lock, flags);
5790 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5793 raw_spin_unlock_irqrestore(&rq->lock, flags);
5801 * Register at high priority so that task migration (migrate_all_tasks)
5802 * happens before everything else. This has to be lower priority than
5803 * the notifier in the perf_event subsystem, though.
5805 static struct notifier_block __cpuinitdata migration_notifier = {
5806 .notifier_call = migration_call,
5810 static int __init migration_init(void)
5812 void *cpu = (void *)(long)smp_processor_id();
5815 /* Start one for the boot CPU: */
5816 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5817 BUG_ON(err == NOTIFY_BAD);
5818 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5819 register_cpu_notifier(&migration_notifier);
5823 early_initcall(migration_init);
5828 #ifdef CONFIG_SCHED_DEBUG
5830 static __read_mostly int sched_domain_debug_enabled;
5832 static int __init sched_domain_debug_setup(char *str)
5834 sched_domain_debug_enabled = 1;
5838 early_param("sched_debug", sched_domain_debug_setup);
5840 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5841 struct cpumask *groupmask)
5843 struct sched_group *group = sd->groups;
5846 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5847 cpumask_clear(groupmask);
5849 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5851 if (!(sd->flags & SD_LOAD_BALANCE)) {
5852 printk("does not load-balance\n");
5854 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5859 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5861 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5862 printk(KERN_ERR "ERROR: domain->span does not contain "
5865 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5866 printk(KERN_ERR "ERROR: domain->groups does not contain"
5870 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5874 printk(KERN_ERR "ERROR: group is NULL\n");
5878 if (!group->cpu_power) {
5879 printk(KERN_CONT "\n");
5880 printk(KERN_ERR "ERROR: domain->cpu_power not "
5885 if (!cpumask_weight(sched_group_cpus(group))) {
5886 printk(KERN_CONT "\n");
5887 printk(KERN_ERR "ERROR: empty group\n");
5891 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
5892 printk(KERN_CONT "\n");
5893 printk(KERN_ERR "ERROR: repeated CPUs\n");
5897 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5899 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5901 printk(KERN_CONT " %s", str);
5902 if (group->cpu_power != SCHED_LOAD_SCALE) {
5903 printk(KERN_CONT " (cpu_power = %d)",
5907 group = group->next;
5908 } while (group != sd->groups);
5909 printk(KERN_CONT "\n");
5911 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5912 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5915 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5916 printk(KERN_ERR "ERROR: parent span is not a superset "
5917 "of domain->span\n");
5921 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5923 cpumask_var_t groupmask;
5926 if (!sched_domain_debug_enabled)
5930 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5934 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5936 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
5937 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
5942 if (sched_domain_debug_one(sd, cpu, level, groupmask))
5949 free_cpumask_var(groupmask);
5951 #else /* !CONFIG_SCHED_DEBUG */
5952 # define sched_domain_debug(sd, cpu) do { } while (0)
5953 #endif /* CONFIG_SCHED_DEBUG */
5955 static int sd_degenerate(struct sched_domain *sd)
5957 if (cpumask_weight(sched_domain_span(sd)) == 1)
5960 /* Following flags need at least 2 groups */
5961 if (sd->flags & (SD_LOAD_BALANCE |
5962 SD_BALANCE_NEWIDLE |
5966 SD_SHARE_PKG_RESOURCES)) {
5967 if (sd->groups != sd->groups->next)
5971 /* Following flags don't use groups */
5972 if (sd->flags & (SD_WAKE_AFFINE))
5979 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5981 unsigned long cflags = sd->flags, pflags = parent->flags;
5983 if (sd_degenerate(parent))
5986 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5989 /* Flags needing groups don't count if only 1 group in parent */
5990 if (parent->groups == parent->groups->next) {
5991 pflags &= ~(SD_LOAD_BALANCE |
5992 SD_BALANCE_NEWIDLE |
5996 SD_SHARE_PKG_RESOURCES);
5997 if (nr_node_ids == 1)
5998 pflags &= ~SD_SERIALIZE;
6000 if (~cflags & pflags)
6006 static void free_rootdomain(struct root_domain *rd)
6008 synchronize_sched();
6010 cpupri_cleanup(&rd->cpupri);
6012 free_cpumask_var(rd->rto_mask);
6013 free_cpumask_var(rd->online);
6014 free_cpumask_var(rd->span);
6018 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6020 struct root_domain *old_rd = NULL;
6021 unsigned long flags;
6023 raw_spin_lock_irqsave(&rq->lock, flags);
6028 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6031 cpumask_clear_cpu(rq->cpu, old_rd->span);
6034 * If we dont want to free the old_rt yet then
6035 * set old_rd to NULL to skip the freeing later
6038 if (!atomic_dec_and_test(&old_rd->refcount))
6042 atomic_inc(&rd->refcount);
6045 cpumask_set_cpu(rq->cpu, rd->span);
6046 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6049 raw_spin_unlock_irqrestore(&rq->lock, flags);
6052 free_rootdomain(old_rd);
6055 static int init_rootdomain(struct root_domain *rd, bool bootmem)
6057 gfp_t gfp = GFP_KERNEL;
6059 memset(rd, 0, sizeof(*rd));
6064 if (!alloc_cpumask_var(&rd->span, gfp))
6066 if (!alloc_cpumask_var(&rd->online, gfp))
6068 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
6071 if (cpupri_init(&rd->cpupri, bootmem) != 0)
6076 free_cpumask_var(rd->rto_mask);
6078 free_cpumask_var(rd->online);
6080 free_cpumask_var(rd->span);
6085 static void init_defrootdomain(void)
6087 init_rootdomain(&def_root_domain, true);
6089 atomic_set(&def_root_domain.refcount, 1);
6092 static struct root_domain *alloc_rootdomain(void)
6094 struct root_domain *rd;
6096 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6100 if (init_rootdomain(rd, false) != 0) {
6109 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6110 * hold the hotplug lock.
6113 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6115 struct rq *rq = cpu_rq(cpu);
6116 struct sched_domain *tmp;
6118 for (tmp = sd; tmp; tmp = tmp->parent)
6119 tmp->span_weight = cpumask_weight(sched_domain_span(tmp));
6121 /* Remove the sched domains which do not contribute to scheduling. */
6122 for (tmp = sd; tmp; ) {
6123 struct sched_domain *parent = tmp->parent;
6127 if (sd_parent_degenerate(tmp, parent)) {
6128 tmp->parent = parent->parent;
6130 parent->parent->child = tmp;
6135 if (sd && sd_degenerate(sd)) {
6141 sched_domain_debug(sd, cpu);
6143 rq_attach_root(rq, rd);
6144 rcu_assign_pointer(rq->sd, sd);
6147 /* cpus with isolated domains */
6148 static cpumask_var_t cpu_isolated_map;
6150 /* Setup the mask of cpus configured for isolated domains */
6151 static int __init isolated_cpu_setup(char *str)
6153 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6154 cpulist_parse(str, cpu_isolated_map);
6158 __setup("isolcpus=", isolated_cpu_setup);
6161 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6162 * to a function which identifies what group(along with sched group) a CPU
6163 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6164 * (due to the fact that we keep track of groups covered with a struct cpumask).
6166 * init_sched_build_groups will build a circular linked list of the groups
6167 * covered by the given span, and will set each group's ->cpumask correctly,
6168 * and ->cpu_power to 0.
6171 init_sched_build_groups(const struct cpumask *span,
6172 const struct cpumask *cpu_map,
6173 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
6174 struct sched_group **sg,
6175 struct cpumask *tmpmask),
6176 struct cpumask *covered, struct cpumask *tmpmask)
6178 struct sched_group *first = NULL, *last = NULL;
6181 cpumask_clear(covered);
6183 for_each_cpu(i, span) {
6184 struct sched_group *sg;
6185 int group = group_fn(i, cpu_map, &sg, tmpmask);
6188 if (cpumask_test_cpu(i, covered))
6191 cpumask_clear(sched_group_cpus(sg));
6194 for_each_cpu(j, span) {
6195 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6198 cpumask_set_cpu(j, covered);
6199 cpumask_set_cpu(j, sched_group_cpus(sg));
6210 #define SD_NODES_PER_DOMAIN 16
6215 * find_next_best_node - find the next node to include in a sched_domain
6216 * @node: node whose sched_domain we're building
6217 * @used_nodes: nodes already in the sched_domain
6219 * Find the next node to include in a given scheduling domain. Simply
6220 * finds the closest node not already in the @used_nodes map.
6222 * Should use nodemask_t.
6224 static int find_next_best_node(int node, nodemask_t *used_nodes)
6226 int i, n, val, min_val, best_node = 0;
6230 for (i = 0; i < nr_node_ids; i++) {
6231 /* Start at @node */
6232 n = (node + i) % nr_node_ids;
6234 if (!nr_cpus_node(n))
6237 /* Skip already used nodes */
6238 if (node_isset(n, *used_nodes))
6241 /* Simple min distance search */
6242 val = node_distance(node, n);
6244 if (val < min_val) {
6250 node_set(best_node, *used_nodes);
6255 * sched_domain_node_span - get a cpumask for a node's sched_domain
6256 * @node: node whose cpumask we're constructing
6257 * @span: resulting cpumask
6259 * Given a node, construct a good cpumask for its sched_domain to span. It
6260 * should be one that prevents unnecessary balancing, but also spreads tasks
6263 static void sched_domain_node_span(int node, struct cpumask *span)
6265 nodemask_t used_nodes;
6268 cpumask_clear(span);
6269 nodes_clear(used_nodes);
6271 cpumask_or(span, span, cpumask_of_node(node));
6272 node_set(node, used_nodes);
6274 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6275 int next_node = find_next_best_node(node, &used_nodes);
6277 cpumask_or(span, span, cpumask_of_node(next_node));
6280 #endif /* CONFIG_NUMA */
6282 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6285 * The cpus mask in sched_group and sched_domain hangs off the end.
6287 * ( See the the comments in include/linux/sched.h:struct sched_group
6288 * and struct sched_domain. )
6290 struct static_sched_group {
6291 struct sched_group sg;
6292 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
6295 struct static_sched_domain {
6296 struct sched_domain sd;
6297 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
6303 cpumask_var_t domainspan;
6304 cpumask_var_t covered;
6305 cpumask_var_t notcovered;
6307 cpumask_var_t nodemask;
6308 cpumask_var_t this_sibling_map;
6309 cpumask_var_t this_core_map;
6310 cpumask_var_t send_covered;
6311 cpumask_var_t tmpmask;
6312 struct sched_group **sched_group_nodes;
6313 struct root_domain *rd;
6317 sa_sched_groups = 0,
6322 sa_this_sibling_map,
6324 sa_sched_group_nodes,
6334 * SMT sched-domains:
6336 #ifdef CONFIG_SCHED_SMT
6337 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
6338 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
6341 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
6342 struct sched_group **sg, struct cpumask *unused)
6345 *sg = &per_cpu(sched_groups, cpu).sg;
6348 #endif /* CONFIG_SCHED_SMT */
6351 * multi-core sched-domains:
6353 #ifdef CONFIG_SCHED_MC
6354 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
6355 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
6356 #endif /* CONFIG_SCHED_MC */
6358 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6360 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6361 struct sched_group **sg, struct cpumask *mask)
6365 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6366 group = cpumask_first(mask);
6368 *sg = &per_cpu(sched_group_core, group).sg;
6371 #elif defined(CONFIG_SCHED_MC)
6373 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6374 struct sched_group **sg, struct cpumask *unused)
6377 *sg = &per_cpu(sched_group_core, cpu).sg;
6382 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
6383 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
6386 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
6387 struct sched_group **sg, struct cpumask *mask)
6390 #ifdef CONFIG_SCHED_MC
6391 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6392 group = cpumask_first(mask);
6393 #elif defined(CONFIG_SCHED_SMT)
6394 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6395 group = cpumask_first(mask);
6400 *sg = &per_cpu(sched_group_phys, group).sg;
6406 * The init_sched_build_groups can't handle what we want to do with node
6407 * groups, so roll our own. Now each node has its own list of groups which
6408 * gets dynamically allocated.
6410 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
6411 static struct sched_group ***sched_group_nodes_bycpu;
6413 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
6414 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
6416 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
6417 struct sched_group **sg,
6418 struct cpumask *nodemask)
6422 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
6423 group = cpumask_first(nodemask);
6426 *sg = &per_cpu(sched_group_allnodes, group).sg;
6430 static void init_numa_sched_groups_power(struct sched_group *group_head)
6432 struct sched_group *sg = group_head;
6438 for_each_cpu(j, sched_group_cpus(sg)) {
6439 struct sched_domain *sd;
6441 sd = &per_cpu(phys_domains, j).sd;
6442 if (j != group_first_cpu(sd->groups)) {
6444 * Only add "power" once for each
6450 sg->cpu_power += sd->groups->cpu_power;
6453 } while (sg != group_head);
6456 static int build_numa_sched_groups(struct s_data *d,
6457 const struct cpumask *cpu_map, int num)
6459 struct sched_domain *sd;
6460 struct sched_group *sg, *prev;
6463 cpumask_clear(d->covered);
6464 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
6465 if (cpumask_empty(d->nodemask)) {
6466 d->sched_group_nodes[num] = NULL;
6470 sched_domain_node_span(num, d->domainspan);
6471 cpumask_and(d->domainspan, d->domainspan, cpu_map);
6473 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6476 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
6480 d->sched_group_nodes[num] = sg;
6482 for_each_cpu(j, d->nodemask) {
6483 sd = &per_cpu(node_domains, j).sd;
6488 cpumask_copy(sched_group_cpus(sg), d->nodemask);
6490 cpumask_or(d->covered, d->covered, d->nodemask);
6493 for (j = 0; j < nr_node_ids; j++) {
6494 n = (num + j) % nr_node_ids;
6495 cpumask_complement(d->notcovered, d->covered);
6496 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
6497 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
6498 if (cpumask_empty(d->tmpmask))
6500 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
6501 if (cpumask_empty(d->tmpmask))
6503 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6507 "Can not alloc domain group for node %d\n", j);
6511 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
6512 sg->next = prev->next;
6513 cpumask_or(d->covered, d->covered, d->tmpmask);
6520 #endif /* CONFIG_NUMA */
6523 /* Free memory allocated for various sched_group structures */
6524 static void free_sched_groups(const struct cpumask *cpu_map,
6525 struct cpumask *nodemask)
6529 for_each_cpu(cpu, cpu_map) {
6530 struct sched_group **sched_group_nodes
6531 = sched_group_nodes_bycpu[cpu];
6533 if (!sched_group_nodes)
6536 for (i = 0; i < nr_node_ids; i++) {
6537 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6539 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
6540 if (cpumask_empty(nodemask))
6550 if (oldsg != sched_group_nodes[i])
6553 kfree(sched_group_nodes);
6554 sched_group_nodes_bycpu[cpu] = NULL;
6557 #else /* !CONFIG_NUMA */
6558 static void free_sched_groups(const struct cpumask *cpu_map,
6559 struct cpumask *nodemask)
6562 #endif /* CONFIG_NUMA */
6565 * Initialize sched groups cpu_power.
6567 * cpu_power indicates the capacity of sched group, which is used while
6568 * distributing the load between different sched groups in a sched domain.
6569 * Typically cpu_power for all the groups in a sched domain will be same unless
6570 * there are asymmetries in the topology. If there are asymmetries, group
6571 * having more cpu_power will pickup more load compared to the group having
6574 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6576 struct sched_domain *child;
6577 struct sched_group *group;
6581 WARN_ON(!sd || !sd->groups);
6583 if (cpu != group_first_cpu(sd->groups))
6588 sd->groups->cpu_power = 0;
6591 power = SCHED_LOAD_SCALE;
6592 weight = cpumask_weight(sched_domain_span(sd));
6594 * SMT siblings share the power of a single core.
6595 * Usually multiple threads get a better yield out of
6596 * that one core than a single thread would have,
6597 * reflect that in sd->smt_gain.
6599 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
6600 power *= sd->smt_gain;
6602 power >>= SCHED_LOAD_SHIFT;
6604 sd->groups->cpu_power += power;
6609 * Add cpu_power of each child group to this groups cpu_power.
6611 group = child->groups;
6613 sd->groups->cpu_power += group->cpu_power;
6614 group = group->next;
6615 } while (group != child->groups);
6619 * Initializers for schedule domains
6620 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6623 #ifdef CONFIG_SCHED_DEBUG
6624 # define SD_INIT_NAME(sd, type) sd->name = #type
6626 # define SD_INIT_NAME(sd, type) do { } while (0)
6629 #define SD_INIT(sd, type) sd_init_##type(sd)
6631 #define SD_INIT_FUNC(type) \
6632 static noinline void sd_init_##type(struct sched_domain *sd) \
6634 memset(sd, 0, sizeof(*sd)); \
6635 *sd = SD_##type##_INIT; \
6636 sd->level = SD_LV_##type; \
6637 SD_INIT_NAME(sd, type); \
6642 SD_INIT_FUNC(ALLNODES)
6645 #ifdef CONFIG_SCHED_SMT
6646 SD_INIT_FUNC(SIBLING)
6648 #ifdef CONFIG_SCHED_MC
6652 static int default_relax_domain_level = -1;
6654 static int __init setup_relax_domain_level(char *str)
6658 val = simple_strtoul(str, NULL, 0);
6659 if (val < SD_LV_MAX)
6660 default_relax_domain_level = val;
6664 __setup("relax_domain_level=", setup_relax_domain_level);
6666 static void set_domain_attribute(struct sched_domain *sd,
6667 struct sched_domain_attr *attr)
6671 if (!attr || attr->relax_domain_level < 0) {
6672 if (default_relax_domain_level < 0)
6675 request = default_relax_domain_level;
6677 request = attr->relax_domain_level;
6678 if (request < sd->level) {
6679 /* turn off idle balance on this domain */
6680 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6682 /* turn on idle balance on this domain */
6683 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6687 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6688 const struct cpumask *cpu_map)
6691 case sa_sched_groups:
6692 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
6693 d->sched_group_nodes = NULL;
6695 free_rootdomain(d->rd); /* fall through */
6697 free_cpumask_var(d->tmpmask); /* fall through */
6698 case sa_send_covered:
6699 free_cpumask_var(d->send_covered); /* fall through */
6700 case sa_this_core_map:
6701 free_cpumask_var(d->this_core_map); /* fall through */
6702 case sa_this_sibling_map:
6703 free_cpumask_var(d->this_sibling_map); /* fall through */
6705 free_cpumask_var(d->nodemask); /* fall through */
6706 case sa_sched_group_nodes:
6708 kfree(d->sched_group_nodes); /* fall through */
6710 free_cpumask_var(d->notcovered); /* fall through */
6712 free_cpumask_var(d->covered); /* fall through */
6714 free_cpumask_var(d->domainspan); /* fall through */
6721 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6722 const struct cpumask *cpu_map)
6725 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
6727 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
6728 return sa_domainspan;
6729 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
6731 /* Allocate the per-node list of sched groups */
6732 d->sched_group_nodes = kcalloc(nr_node_ids,
6733 sizeof(struct sched_group *), GFP_KERNEL);
6734 if (!d->sched_group_nodes) {
6735 printk(KERN_WARNING "Can not alloc sched group node list\n");
6736 return sa_notcovered;
6738 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
6740 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
6741 return sa_sched_group_nodes;
6742 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
6744 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
6745 return sa_this_sibling_map;
6746 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
6747 return sa_this_core_map;
6748 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
6749 return sa_send_covered;
6750 d->rd = alloc_rootdomain();
6752 printk(KERN_WARNING "Cannot alloc root domain\n");
6755 return sa_rootdomain;
6758 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
6759 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
6761 struct sched_domain *sd = NULL;
6763 struct sched_domain *parent;
6766 if (cpumask_weight(cpu_map) >
6767 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
6768 sd = &per_cpu(allnodes_domains, i).sd;
6769 SD_INIT(sd, ALLNODES);
6770 set_domain_attribute(sd, attr);
6771 cpumask_copy(sched_domain_span(sd), cpu_map);
6772 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
6777 sd = &per_cpu(node_domains, i).sd;
6779 set_domain_attribute(sd, attr);
6780 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
6781 sd->parent = parent;
6784 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
6789 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
6790 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6791 struct sched_domain *parent, int i)
6793 struct sched_domain *sd;
6794 sd = &per_cpu(phys_domains, i).sd;
6796 set_domain_attribute(sd, attr);
6797 cpumask_copy(sched_domain_span(sd), d->nodemask);
6798 sd->parent = parent;
6801 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
6805 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
6806 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6807 struct sched_domain *parent, int i)
6809 struct sched_domain *sd = parent;
6810 #ifdef CONFIG_SCHED_MC
6811 sd = &per_cpu(core_domains, i).sd;
6813 set_domain_attribute(sd, attr);
6814 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
6815 sd->parent = parent;
6817 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
6822 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
6823 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6824 struct sched_domain *parent, int i)
6826 struct sched_domain *sd = parent;
6827 #ifdef CONFIG_SCHED_SMT
6828 sd = &per_cpu(cpu_domains, i).sd;
6829 SD_INIT(sd, SIBLING);
6830 set_domain_attribute(sd, attr);
6831 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
6832 sd->parent = parent;
6834 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
6839 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
6840 const struct cpumask *cpu_map, int cpu)
6843 #ifdef CONFIG_SCHED_SMT
6844 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
6845 cpumask_and(d->this_sibling_map, cpu_map,
6846 topology_thread_cpumask(cpu));
6847 if (cpu == cpumask_first(d->this_sibling_map))
6848 init_sched_build_groups(d->this_sibling_map, cpu_map,
6850 d->send_covered, d->tmpmask);
6853 #ifdef CONFIG_SCHED_MC
6854 case SD_LV_MC: /* set up multi-core groups */
6855 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
6856 if (cpu == cpumask_first(d->this_core_map))
6857 init_sched_build_groups(d->this_core_map, cpu_map,
6859 d->send_covered, d->tmpmask);
6862 case SD_LV_CPU: /* set up physical groups */
6863 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
6864 if (!cpumask_empty(d->nodemask))
6865 init_sched_build_groups(d->nodemask, cpu_map,
6867 d->send_covered, d->tmpmask);
6870 case SD_LV_ALLNODES:
6871 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
6872 d->send_covered, d->tmpmask);
6881 * Build sched domains for a given set of cpus and attach the sched domains
6882 * to the individual cpus
6884 static int __build_sched_domains(const struct cpumask *cpu_map,
6885 struct sched_domain_attr *attr)
6887 enum s_alloc alloc_state = sa_none;
6889 struct sched_domain *sd;
6895 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6896 if (alloc_state != sa_rootdomain)
6898 alloc_state = sa_sched_groups;
6901 * Set up domains for cpus specified by the cpu_map.
6903 for_each_cpu(i, cpu_map) {
6904 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
6907 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
6908 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
6909 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
6910 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
6913 for_each_cpu(i, cpu_map) {
6914 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
6915 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
6918 /* Set up physical groups */
6919 for (i = 0; i < nr_node_ids; i++)
6920 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
6923 /* Set up node groups */
6925 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
6927 for (i = 0; i < nr_node_ids; i++)
6928 if (build_numa_sched_groups(&d, cpu_map, i))
6932 /* Calculate CPU power for physical packages and nodes */
6933 #ifdef CONFIG_SCHED_SMT
6934 for_each_cpu(i, cpu_map) {
6935 sd = &per_cpu(cpu_domains, i).sd;
6936 init_sched_groups_power(i, sd);
6939 #ifdef CONFIG_SCHED_MC
6940 for_each_cpu(i, cpu_map) {
6941 sd = &per_cpu(core_domains, i).sd;
6942 init_sched_groups_power(i, sd);
6946 for_each_cpu(i, cpu_map) {
6947 sd = &per_cpu(phys_domains, i).sd;
6948 init_sched_groups_power(i, sd);
6952 for (i = 0; i < nr_node_ids; i++)
6953 init_numa_sched_groups_power(d.sched_group_nodes[i]);
6955 if (d.sd_allnodes) {
6956 struct sched_group *sg;
6958 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
6960 init_numa_sched_groups_power(sg);
6964 /* Attach the domains */
6965 for_each_cpu(i, cpu_map) {
6966 #ifdef CONFIG_SCHED_SMT
6967 sd = &per_cpu(cpu_domains, i).sd;
6968 #elif defined(CONFIG_SCHED_MC)
6969 sd = &per_cpu(core_domains, i).sd;
6971 sd = &per_cpu(phys_domains, i).sd;
6973 cpu_attach_domain(sd, d.rd, i);
6976 d.sched_group_nodes = NULL; /* don't free this we still need it */
6977 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
6981 __free_domain_allocs(&d, alloc_state, cpu_map);
6985 static int build_sched_domains(const struct cpumask *cpu_map)
6987 return __build_sched_domains(cpu_map, NULL);
6990 static cpumask_var_t *doms_cur; /* current sched domains */
6991 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6992 static struct sched_domain_attr *dattr_cur;
6993 /* attribues of custom domains in 'doms_cur' */
6996 * Special case: If a kmalloc of a doms_cur partition (array of
6997 * cpumask) fails, then fallback to a single sched domain,
6998 * as determined by the single cpumask fallback_doms.
7000 static cpumask_var_t fallback_doms;
7003 * arch_update_cpu_topology lets virtualized architectures update the
7004 * cpu core maps. It is supposed to return 1 if the topology changed
7005 * or 0 if it stayed the same.
7007 int __attribute__((weak)) arch_update_cpu_topology(void)
7012 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7015 cpumask_var_t *doms;
7017 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7020 for (i = 0; i < ndoms; i++) {
7021 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7022 free_sched_domains(doms, i);
7029 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7032 for (i = 0; i < ndoms; i++)
7033 free_cpumask_var(doms[i]);
7038 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7039 * For now this just excludes isolated cpus, but could be used to
7040 * exclude other special cases in the future.
7042 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7046 arch_update_cpu_topology();
7048 doms_cur = alloc_sched_domains(ndoms_cur);
7050 doms_cur = &fallback_doms;
7051 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7053 err = build_sched_domains(doms_cur[0]);
7054 register_sched_domain_sysctl();
7059 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7060 struct cpumask *tmpmask)
7062 free_sched_groups(cpu_map, tmpmask);
7066 * Detach sched domains from a group of cpus specified in cpu_map
7067 * These cpus will now be attached to the NULL domain
7069 static void detach_destroy_domains(const struct cpumask *cpu_map)
7071 /* Save because hotplug lock held. */
7072 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7075 for_each_cpu(i, cpu_map)
7076 cpu_attach_domain(NULL, &def_root_domain, i);
7077 synchronize_sched();
7078 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7081 /* handle null as "default" */
7082 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7083 struct sched_domain_attr *new, int idx_new)
7085 struct sched_domain_attr tmp;
7092 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7093 new ? (new + idx_new) : &tmp,
7094 sizeof(struct sched_domain_attr));
7098 * Partition sched domains as specified by the 'ndoms_new'
7099 * cpumasks in the array doms_new[] of cpumasks. This compares
7100 * doms_new[] to the current sched domain partitioning, doms_cur[].
7101 * It destroys each deleted domain and builds each new domain.
7103 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7104 * The masks don't intersect (don't overlap.) We should setup one
7105 * sched domain for each mask. CPUs not in any of the cpumasks will
7106 * not be load balanced. If the same cpumask appears both in the
7107 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7110 * The passed in 'doms_new' should be allocated using
7111 * alloc_sched_domains. This routine takes ownership of it and will
7112 * free_sched_domains it when done with it. If the caller failed the
7113 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7114 * and partition_sched_domains() will fallback to the single partition
7115 * 'fallback_doms', it also forces the domains to be rebuilt.
7117 * If doms_new == NULL it will be replaced with cpu_online_mask.
7118 * ndoms_new == 0 is a special case for destroying existing domains,
7119 * and it will not create the default domain.
7121 * Call with hotplug lock held
7123 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7124 struct sched_domain_attr *dattr_new)
7129 mutex_lock(&sched_domains_mutex);
7131 /* always unregister in case we don't destroy any domains */
7132 unregister_sched_domain_sysctl();
7134 /* Let architecture update cpu core mappings. */
7135 new_topology = arch_update_cpu_topology();
7137 n = doms_new ? ndoms_new : 0;
7139 /* Destroy deleted domains */
7140 for (i = 0; i < ndoms_cur; i++) {
7141 for (j = 0; j < n && !new_topology; j++) {
7142 if (cpumask_equal(doms_cur[i], doms_new[j])
7143 && dattrs_equal(dattr_cur, i, dattr_new, j))
7146 /* no match - a current sched domain not in new doms_new[] */
7147 detach_destroy_domains(doms_cur[i]);
7152 if (doms_new == NULL) {
7154 doms_new = &fallback_doms;
7155 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7156 WARN_ON_ONCE(dattr_new);
7159 /* Build new domains */
7160 for (i = 0; i < ndoms_new; i++) {
7161 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7162 if (cpumask_equal(doms_new[i], doms_cur[j])
7163 && dattrs_equal(dattr_new, i, dattr_cur, j))
7166 /* no match - add a new doms_new */
7167 __build_sched_domains(doms_new[i],
7168 dattr_new ? dattr_new + i : NULL);
7173 /* Remember the new sched domains */
7174 if (doms_cur != &fallback_doms)
7175 free_sched_domains(doms_cur, ndoms_cur);
7176 kfree(dattr_cur); /* kfree(NULL) is safe */
7177 doms_cur = doms_new;
7178 dattr_cur = dattr_new;
7179 ndoms_cur = ndoms_new;
7181 register_sched_domain_sysctl();
7183 mutex_unlock(&sched_domains_mutex);
7186 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7187 static void arch_reinit_sched_domains(void)
7191 /* Destroy domains first to force the rebuild */
7192 partition_sched_domains(0, NULL, NULL);
7194 rebuild_sched_domains();
7198 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7200 unsigned int level = 0;
7202 if (sscanf(buf, "%u", &level) != 1)
7206 * level is always be positive so don't check for
7207 * level < POWERSAVINGS_BALANCE_NONE which is 0
7208 * What happens on 0 or 1 byte write,
7209 * need to check for count as well?
7212 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7216 sched_smt_power_savings = level;
7218 sched_mc_power_savings = level;
7220 arch_reinit_sched_domains();
7225 #ifdef CONFIG_SCHED_MC
7226 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7227 struct sysdev_class_attribute *attr,
7230 return sprintf(page, "%u\n", sched_mc_power_savings);
7232 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7233 struct sysdev_class_attribute *attr,
7234 const char *buf, size_t count)
7236 return sched_power_savings_store(buf, count, 0);
7238 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7239 sched_mc_power_savings_show,
7240 sched_mc_power_savings_store);
7243 #ifdef CONFIG_SCHED_SMT
7244 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7245 struct sysdev_class_attribute *attr,
7248 return sprintf(page, "%u\n", sched_smt_power_savings);
7250 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7251 struct sysdev_class_attribute *attr,
7252 const char *buf, size_t count)
7254 return sched_power_savings_store(buf, count, 1);
7256 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7257 sched_smt_power_savings_show,
7258 sched_smt_power_savings_store);
7261 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7265 #ifdef CONFIG_SCHED_SMT
7267 err = sysfs_create_file(&cls->kset.kobj,
7268 &attr_sched_smt_power_savings.attr);
7270 #ifdef CONFIG_SCHED_MC
7271 if (!err && mc_capable())
7272 err = sysfs_create_file(&cls->kset.kobj,
7273 &attr_sched_mc_power_savings.attr);
7277 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7279 #ifndef CONFIG_CPUSETS
7281 * Add online and remove offline CPUs from the scheduler domains.
7282 * When cpusets are enabled they take over this function.
7284 static int update_sched_domains(struct notifier_block *nfb,
7285 unsigned long action, void *hcpu)
7289 case CPU_ONLINE_FROZEN:
7290 case CPU_DOWN_PREPARE:
7291 case CPU_DOWN_PREPARE_FROZEN:
7292 case CPU_DOWN_FAILED:
7293 case CPU_DOWN_FAILED_FROZEN:
7294 partition_sched_domains(1, NULL, NULL);
7303 static int update_runtime(struct notifier_block *nfb,
7304 unsigned long action, void *hcpu)
7306 int cpu = (int)(long)hcpu;
7309 case CPU_DOWN_PREPARE:
7310 case CPU_DOWN_PREPARE_FROZEN:
7311 disable_runtime(cpu_rq(cpu));
7314 case CPU_DOWN_FAILED:
7315 case CPU_DOWN_FAILED_FROZEN:
7317 case CPU_ONLINE_FROZEN:
7318 enable_runtime(cpu_rq(cpu));
7326 void __init sched_init_smp(void)
7328 cpumask_var_t non_isolated_cpus;
7330 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7331 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7333 #if defined(CONFIG_NUMA)
7334 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7336 BUG_ON(sched_group_nodes_bycpu == NULL);
7339 mutex_lock(&sched_domains_mutex);
7340 arch_init_sched_domains(cpu_active_mask);
7341 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7342 if (cpumask_empty(non_isolated_cpus))
7343 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7344 mutex_unlock(&sched_domains_mutex);
7347 #ifndef CONFIG_CPUSETS
7348 /* XXX: Theoretical race here - CPU may be hotplugged now */
7349 hotcpu_notifier(update_sched_domains, 0);
7352 /* RT runtime code needs to handle some hotplug events */
7353 hotcpu_notifier(update_runtime, 0);
7357 /* Move init over to a non-isolated CPU */
7358 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7360 sched_init_granularity();
7361 free_cpumask_var(non_isolated_cpus);
7363 init_sched_rt_class();
7366 void __init sched_init_smp(void)
7368 sched_init_granularity();
7370 #endif /* CONFIG_SMP */
7372 const_debug unsigned int sysctl_timer_migration = 1;
7374 int in_sched_functions(unsigned long addr)
7376 return in_lock_functions(addr) ||
7377 (addr >= (unsigned long)__sched_text_start
7378 && addr < (unsigned long)__sched_text_end);
7381 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7383 cfs_rq->tasks_timeline = RB_ROOT;
7384 INIT_LIST_HEAD(&cfs_rq->tasks);
7385 #ifdef CONFIG_FAIR_GROUP_SCHED
7388 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7391 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7393 struct rt_prio_array *array;
7396 array = &rt_rq->active;
7397 for (i = 0; i < MAX_RT_PRIO; i++) {
7398 INIT_LIST_HEAD(array->queue + i);
7399 __clear_bit(i, array->bitmap);
7401 /* delimiter for bitsearch: */
7402 __set_bit(MAX_RT_PRIO, array->bitmap);
7404 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7405 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7407 rt_rq->highest_prio.next = MAX_RT_PRIO;
7411 rt_rq->rt_nr_migratory = 0;
7412 rt_rq->overloaded = 0;
7413 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
7417 rt_rq->rt_throttled = 0;
7418 rt_rq->rt_runtime = 0;
7419 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7421 #ifdef CONFIG_RT_GROUP_SCHED
7422 rt_rq->rt_nr_boosted = 0;
7427 #ifdef CONFIG_FAIR_GROUP_SCHED
7428 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7429 struct sched_entity *se, int cpu, int add,
7430 struct sched_entity *parent)
7432 struct rq *rq = cpu_rq(cpu);
7433 tg->cfs_rq[cpu] = cfs_rq;
7434 init_cfs_rq(cfs_rq, rq);
7437 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7440 /* se could be NULL for init_task_group */
7445 se->cfs_rq = &rq->cfs;
7447 se->cfs_rq = parent->my_q;
7450 se->load.weight = tg->shares;
7451 se->load.inv_weight = 0;
7452 se->parent = parent;
7456 #ifdef CONFIG_RT_GROUP_SCHED
7457 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7458 struct sched_rt_entity *rt_se, int cpu, int add,
7459 struct sched_rt_entity *parent)
7461 struct rq *rq = cpu_rq(cpu);
7463 tg->rt_rq[cpu] = rt_rq;
7464 init_rt_rq(rt_rq, rq);
7466 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7468 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7470 tg->rt_se[cpu] = rt_se;
7475 rt_se->rt_rq = &rq->rt;
7477 rt_se->rt_rq = parent->my_q;
7479 rt_se->my_q = rt_rq;
7480 rt_se->parent = parent;
7481 INIT_LIST_HEAD(&rt_se->run_list);
7485 void __init sched_init(void)
7488 unsigned long alloc_size = 0, ptr;
7490 #ifdef CONFIG_FAIR_GROUP_SCHED
7491 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7493 #ifdef CONFIG_RT_GROUP_SCHED
7494 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7496 #ifdef CONFIG_CPUMASK_OFFSTACK
7497 alloc_size += num_possible_cpus() * cpumask_size();
7500 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7502 #ifdef CONFIG_FAIR_GROUP_SCHED
7503 init_task_group.se = (struct sched_entity **)ptr;
7504 ptr += nr_cpu_ids * sizeof(void **);
7506 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7507 ptr += nr_cpu_ids * sizeof(void **);
7509 #endif /* CONFIG_FAIR_GROUP_SCHED */
7510 #ifdef CONFIG_RT_GROUP_SCHED
7511 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7512 ptr += nr_cpu_ids * sizeof(void **);
7514 init_task_group.rt_rq = (struct rt_rq **)ptr;
7515 ptr += nr_cpu_ids * sizeof(void **);
7517 #endif /* CONFIG_RT_GROUP_SCHED */
7518 #ifdef CONFIG_CPUMASK_OFFSTACK
7519 for_each_possible_cpu(i) {
7520 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
7521 ptr += cpumask_size();
7523 #endif /* CONFIG_CPUMASK_OFFSTACK */
7527 init_defrootdomain();
7530 init_rt_bandwidth(&def_rt_bandwidth,
7531 global_rt_period(), global_rt_runtime());
7533 #ifdef CONFIG_RT_GROUP_SCHED
7534 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7535 global_rt_period(), global_rt_runtime());
7536 #endif /* CONFIG_RT_GROUP_SCHED */
7538 #ifdef CONFIG_CGROUP_SCHED
7539 list_add(&init_task_group.list, &task_groups);
7540 INIT_LIST_HEAD(&init_task_group.children);
7542 #endif /* CONFIG_CGROUP_SCHED */
7544 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
7545 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
7546 __alignof__(unsigned long));
7548 for_each_possible_cpu(i) {
7552 raw_spin_lock_init(&rq->lock);
7554 rq->calc_load_active = 0;
7555 rq->calc_load_update = jiffies + LOAD_FREQ;
7556 init_cfs_rq(&rq->cfs, rq);
7557 init_rt_rq(&rq->rt, rq);
7558 #ifdef CONFIG_FAIR_GROUP_SCHED
7559 init_task_group.shares = init_task_group_load;
7560 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7561 #ifdef CONFIG_CGROUP_SCHED
7563 * How much cpu bandwidth does init_task_group get?
7565 * In case of task-groups formed thr' the cgroup filesystem, it
7566 * gets 100% of the cpu resources in the system. This overall
7567 * system cpu resource is divided among the tasks of
7568 * init_task_group and its child task-groups in a fair manner,
7569 * based on each entity's (task or task-group's) weight
7570 * (se->load.weight).
7572 * In other words, if init_task_group has 10 tasks of weight
7573 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7574 * then A0's share of the cpu resource is:
7576 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7578 * We achieve this by letting init_task_group's tasks sit
7579 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7581 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
7583 #endif /* CONFIG_FAIR_GROUP_SCHED */
7585 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7586 #ifdef CONFIG_RT_GROUP_SCHED
7587 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7588 #ifdef CONFIG_CGROUP_SCHED
7589 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
7593 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7594 rq->cpu_load[j] = 0;
7598 rq->cpu_power = SCHED_LOAD_SCALE;
7599 rq->post_schedule = 0;
7600 rq->active_balance = 0;
7601 rq->next_balance = jiffies;
7606 rq->avg_idle = 2*sysctl_sched_migration_cost;
7607 rq_attach_root(rq, &def_root_domain);
7610 atomic_set(&rq->nr_iowait, 0);
7613 set_load_weight(&init_task);
7615 #ifdef CONFIG_PREEMPT_NOTIFIERS
7616 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7620 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7623 #ifdef CONFIG_RT_MUTEXES
7624 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
7628 * The boot idle thread does lazy MMU switching as well:
7630 atomic_inc(&init_mm.mm_count);
7631 enter_lazy_tlb(&init_mm, current);
7634 * Make us the idle thread. Technically, schedule() should not be
7635 * called from this thread, however somewhere below it might be,
7636 * but because we are the idle thread, we just pick up running again
7637 * when this runqueue becomes "idle".
7639 init_idle(current, smp_processor_id());
7641 calc_load_update = jiffies + LOAD_FREQ;
7644 * During early bootup we pretend to be a normal task:
7646 current->sched_class = &fair_sched_class;
7648 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
7649 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
7652 zalloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
7653 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
7655 /* May be allocated at isolcpus cmdline parse time */
7656 if (cpu_isolated_map == NULL)
7657 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7662 scheduler_running = 1;
7665 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7666 static inline int preempt_count_equals(int preempt_offset)
7668 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7670 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
7673 void __might_sleep(const char *file, int line, int preempt_offset)
7676 static unsigned long prev_jiffy; /* ratelimiting */
7678 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7679 system_state != SYSTEM_RUNNING || oops_in_progress)
7681 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7683 prev_jiffy = jiffies;
7686 "BUG: sleeping function called from invalid context at %s:%d\n",
7689 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7690 in_atomic(), irqs_disabled(),
7691 current->pid, current->comm);
7693 debug_show_held_locks(current);
7694 if (irqs_disabled())
7695 print_irqtrace_events(current);
7699 EXPORT_SYMBOL(__might_sleep);
7702 #ifdef CONFIG_MAGIC_SYSRQ
7703 static void normalize_task(struct rq *rq, struct task_struct *p)
7707 on_rq = p->se.on_rq;
7709 deactivate_task(rq, p, 0);
7710 __setscheduler(rq, p, SCHED_NORMAL, 0);
7712 activate_task(rq, p, 0);
7713 resched_task(rq->curr);
7717 void normalize_rt_tasks(void)
7719 struct task_struct *g, *p;
7720 unsigned long flags;
7723 read_lock_irqsave(&tasklist_lock, flags);
7724 do_each_thread(g, p) {
7726 * Only normalize user tasks:
7731 p->se.exec_start = 0;
7732 #ifdef CONFIG_SCHEDSTATS
7733 p->se.statistics.wait_start = 0;
7734 p->se.statistics.sleep_start = 0;
7735 p->se.statistics.block_start = 0;
7740 * Renice negative nice level userspace
7743 if (TASK_NICE(p) < 0 && p->mm)
7744 set_user_nice(p, 0);
7748 raw_spin_lock(&p->pi_lock);
7749 rq = __task_rq_lock(p);
7751 normalize_task(rq, p);
7753 __task_rq_unlock(rq);
7754 raw_spin_unlock(&p->pi_lock);
7755 } while_each_thread(g, p);
7757 read_unlock_irqrestore(&tasklist_lock, flags);
7760 #endif /* CONFIG_MAGIC_SYSRQ */
7762 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7764 * These functions are only useful for the IA64 MCA handling, or kdb.
7766 * They can only be called when the whole system has been
7767 * stopped - every CPU needs to be quiescent, and no scheduling
7768 * activity can take place. Using them for anything else would
7769 * be a serious bug, and as a result, they aren't even visible
7770 * under any other configuration.
7774 * curr_task - return the current task for a given cpu.
7775 * @cpu: the processor in question.
7777 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7779 struct task_struct *curr_task(int cpu)
7781 return cpu_curr(cpu);
7784 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7788 * set_curr_task - set the current task for a given cpu.
7789 * @cpu: the processor in question.
7790 * @p: the task pointer to set.
7792 * Description: This function must only be used when non-maskable interrupts
7793 * are serviced on a separate stack. It allows the architecture to switch the
7794 * notion of the current task on a cpu in a non-blocking manner. This function
7795 * must be called with all CPU's synchronized, and interrupts disabled, the
7796 * and caller must save the original value of the current task (see
7797 * curr_task() above) and restore that value before reenabling interrupts and
7798 * re-starting the system.
7800 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7802 void set_curr_task(int cpu, struct task_struct *p)
7809 #ifdef CONFIG_FAIR_GROUP_SCHED
7810 static void free_fair_sched_group(struct task_group *tg)
7814 for_each_possible_cpu(i) {
7816 kfree(tg->cfs_rq[i]);
7826 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7828 struct cfs_rq *cfs_rq;
7829 struct sched_entity *se;
7833 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7836 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7840 tg->shares = NICE_0_LOAD;
7842 for_each_possible_cpu(i) {
7845 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
7846 GFP_KERNEL, cpu_to_node(i));
7850 se = kzalloc_node(sizeof(struct sched_entity),
7851 GFP_KERNEL, cpu_to_node(i));
7855 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
7866 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
7868 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
7869 &cpu_rq(cpu)->leaf_cfs_rq_list);
7872 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
7874 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
7876 #else /* !CONFG_FAIR_GROUP_SCHED */
7877 static inline void free_fair_sched_group(struct task_group *tg)
7882 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7887 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
7891 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
7894 #endif /* CONFIG_FAIR_GROUP_SCHED */
7896 #ifdef CONFIG_RT_GROUP_SCHED
7897 static void free_rt_sched_group(struct task_group *tg)
7901 destroy_rt_bandwidth(&tg->rt_bandwidth);
7903 for_each_possible_cpu(i) {
7905 kfree(tg->rt_rq[i]);
7907 kfree(tg->rt_se[i]);
7915 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
7917 struct rt_rq *rt_rq;
7918 struct sched_rt_entity *rt_se;
7922 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
7925 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
7929 init_rt_bandwidth(&tg->rt_bandwidth,
7930 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
7932 for_each_possible_cpu(i) {
7935 rt_rq = kzalloc_node(sizeof(struct rt_rq),
7936 GFP_KERNEL, cpu_to_node(i));
7940 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
7941 GFP_KERNEL, cpu_to_node(i));
7945 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
7956 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
7958 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
7959 &cpu_rq(cpu)->leaf_rt_rq_list);
7962 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
7964 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
7966 #else /* !CONFIG_RT_GROUP_SCHED */
7967 static inline void free_rt_sched_group(struct task_group *tg)
7972 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
7977 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
7981 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
7984 #endif /* CONFIG_RT_GROUP_SCHED */
7986 #ifdef CONFIG_CGROUP_SCHED
7987 static void free_sched_group(struct task_group *tg)
7989 free_fair_sched_group(tg);
7990 free_rt_sched_group(tg);
7994 /* allocate runqueue etc for a new task group */
7995 struct task_group *sched_create_group(struct task_group *parent)
7997 struct task_group *tg;
7998 unsigned long flags;
8001 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8003 return ERR_PTR(-ENOMEM);
8005 if (!alloc_fair_sched_group(tg, parent))
8008 if (!alloc_rt_sched_group(tg, parent))
8011 spin_lock_irqsave(&task_group_lock, flags);
8012 for_each_possible_cpu(i) {
8013 register_fair_sched_group(tg, i);
8014 register_rt_sched_group(tg, i);
8016 list_add_rcu(&tg->list, &task_groups);
8018 WARN_ON(!parent); /* root should already exist */
8020 tg->parent = parent;
8021 INIT_LIST_HEAD(&tg->children);
8022 list_add_rcu(&tg->siblings, &parent->children);
8023 spin_unlock_irqrestore(&task_group_lock, flags);
8028 free_sched_group(tg);
8029 return ERR_PTR(-ENOMEM);
8032 /* rcu callback to free various structures associated with a task group */
8033 static void free_sched_group_rcu(struct rcu_head *rhp)
8035 /* now it should be safe to free those cfs_rqs */
8036 free_sched_group(container_of(rhp, struct task_group, rcu));
8039 /* Destroy runqueue etc associated with a task group */
8040 void sched_destroy_group(struct task_group *tg)
8042 unsigned long flags;
8045 spin_lock_irqsave(&task_group_lock, flags);
8046 for_each_possible_cpu(i) {
8047 unregister_fair_sched_group(tg, i);
8048 unregister_rt_sched_group(tg, i);
8050 list_del_rcu(&tg->list);
8051 list_del_rcu(&tg->siblings);
8052 spin_unlock_irqrestore(&task_group_lock, flags);
8054 /* wait for possible concurrent references to cfs_rqs complete */
8055 call_rcu(&tg->rcu, free_sched_group_rcu);
8058 /* change task's runqueue when it moves between groups.
8059 * The caller of this function should have put the task in its new group
8060 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8061 * reflect its new group.
8063 void sched_move_task(struct task_struct *tsk)
8066 unsigned long flags;
8069 rq = task_rq_lock(tsk, &flags);
8071 running = task_current(rq, tsk);
8072 on_rq = tsk->se.on_rq;
8075 dequeue_task(rq, tsk, 0);
8076 if (unlikely(running))
8077 tsk->sched_class->put_prev_task(rq, tsk);
8079 set_task_rq(tsk, task_cpu(tsk));
8081 #ifdef CONFIG_FAIR_GROUP_SCHED
8082 if (tsk->sched_class->moved_group)
8083 tsk->sched_class->moved_group(tsk, on_rq);
8086 if (unlikely(running))
8087 tsk->sched_class->set_curr_task(rq);
8089 enqueue_task(rq, tsk, 0);
8091 task_rq_unlock(rq, &flags);
8093 #endif /* CONFIG_CGROUP_SCHED */
8095 #ifdef CONFIG_FAIR_GROUP_SCHED
8096 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8098 struct cfs_rq *cfs_rq = se->cfs_rq;
8103 dequeue_entity(cfs_rq, se, 0);
8105 se->load.weight = shares;
8106 se->load.inv_weight = 0;
8109 enqueue_entity(cfs_rq, se, 0);
8112 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8114 struct cfs_rq *cfs_rq = se->cfs_rq;
8115 struct rq *rq = cfs_rq->rq;
8116 unsigned long flags;
8118 raw_spin_lock_irqsave(&rq->lock, flags);
8119 __set_se_shares(se, shares);
8120 raw_spin_unlock_irqrestore(&rq->lock, flags);
8123 static DEFINE_MUTEX(shares_mutex);
8125 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8128 unsigned long flags;
8131 * We can't change the weight of the root cgroup.
8136 if (shares < MIN_SHARES)
8137 shares = MIN_SHARES;
8138 else if (shares > MAX_SHARES)
8139 shares = MAX_SHARES;
8141 mutex_lock(&shares_mutex);
8142 if (tg->shares == shares)
8145 spin_lock_irqsave(&task_group_lock, flags);
8146 for_each_possible_cpu(i)
8147 unregister_fair_sched_group(tg, i);
8148 list_del_rcu(&tg->siblings);
8149 spin_unlock_irqrestore(&task_group_lock, flags);
8151 /* wait for any ongoing reference to this group to finish */
8152 synchronize_sched();
8155 * Now we are free to modify the group's share on each cpu
8156 * w/o tripping rebalance_share or load_balance_fair.
8158 tg->shares = shares;
8159 for_each_possible_cpu(i) {
8163 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8164 set_se_shares(tg->se[i], shares);
8168 * Enable load balance activity on this group, by inserting it back on
8169 * each cpu's rq->leaf_cfs_rq_list.
8171 spin_lock_irqsave(&task_group_lock, flags);
8172 for_each_possible_cpu(i)
8173 register_fair_sched_group(tg, i);
8174 list_add_rcu(&tg->siblings, &tg->parent->children);
8175 spin_unlock_irqrestore(&task_group_lock, flags);
8177 mutex_unlock(&shares_mutex);
8181 unsigned long sched_group_shares(struct task_group *tg)
8187 #ifdef CONFIG_RT_GROUP_SCHED
8189 * Ensure that the real time constraints are schedulable.
8191 static DEFINE_MUTEX(rt_constraints_mutex);
8193 static unsigned long to_ratio(u64 period, u64 runtime)
8195 if (runtime == RUNTIME_INF)
8198 return div64_u64(runtime << 20, period);
8201 /* Must be called with tasklist_lock held */
8202 static inline int tg_has_rt_tasks(struct task_group *tg)
8204 struct task_struct *g, *p;
8206 do_each_thread(g, p) {
8207 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8209 } while_each_thread(g, p);
8214 struct rt_schedulable_data {
8215 struct task_group *tg;
8220 static int tg_schedulable(struct task_group *tg, void *data)
8222 struct rt_schedulable_data *d = data;
8223 struct task_group *child;
8224 unsigned long total, sum = 0;
8225 u64 period, runtime;
8227 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8228 runtime = tg->rt_bandwidth.rt_runtime;
8231 period = d->rt_period;
8232 runtime = d->rt_runtime;
8236 * Cannot have more runtime than the period.
8238 if (runtime > period && runtime != RUNTIME_INF)
8242 * Ensure we don't starve existing RT tasks.
8244 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8247 total = to_ratio(period, runtime);
8250 * Nobody can have more than the global setting allows.
8252 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8256 * The sum of our children's runtime should not exceed our own.
8258 list_for_each_entry_rcu(child, &tg->children, siblings) {
8259 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8260 runtime = child->rt_bandwidth.rt_runtime;
8262 if (child == d->tg) {
8263 period = d->rt_period;
8264 runtime = d->rt_runtime;
8267 sum += to_ratio(period, runtime);
8276 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8278 struct rt_schedulable_data data = {
8280 .rt_period = period,
8281 .rt_runtime = runtime,
8284 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8287 static int tg_set_bandwidth(struct task_group *tg,
8288 u64 rt_period, u64 rt_runtime)
8292 mutex_lock(&rt_constraints_mutex);
8293 read_lock(&tasklist_lock);
8294 err = __rt_schedulable(tg, rt_period, rt_runtime);
8298 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8299 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8300 tg->rt_bandwidth.rt_runtime = rt_runtime;
8302 for_each_possible_cpu(i) {
8303 struct rt_rq *rt_rq = tg->rt_rq[i];
8305 raw_spin_lock(&rt_rq->rt_runtime_lock);
8306 rt_rq->rt_runtime = rt_runtime;
8307 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8309 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8311 read_unlock(&tasklist_lock);
8312 mutex_unlock(&rt_constraints_mutex);
8317 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8319 u64 rt_runtime, rt_period;
8321 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8322 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8323 if (rt_runtime_us < 0)
8324 rt_runtime = RUNTIME_INF;
8326 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8329 long sched_group_rt_runtime(struct task_group *tg)
8333 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8336 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8337 do_div(rt_runtime_us, NSEC_PER_USEC);
8338 return rt_runtime_us;
8341 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8343 u64 rt_runtime, rt_period;
8345 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8346 rt_runtime = tg->rt_bandwidth.rt_runtime;
8351 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8354 long sched_group_rt_period(struct task_group *tg)
8358 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8359 do_div(rt_period_us, NSEC_PER_USEC);
8360 return rt_period_us;
8363 static int sched_rt_global_constraints(void)
8365 u64 runtime, period;
8368 if (sysctl_sched_rt_period <= 0)
8371 runtime = global_rt_runtime();
8372 period = global_rt_period();
8375 * Sanity check on the sysctl variables.
8377 if (runtime > period && runtime != RUNTIME_INF)
8380 mutex_lock(&rt_constraints_mutex);
8381 read_lock(&tasklist_lock);
8382 ret = __rt_schedulable(NULL, 0, 0);
8383 read_unlock(&tasklist_lock);
8384 mutex_unlock(&rt_constraints_mutex);
8389 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8391 /* Don't accept realtime tasks when there is no way for them to run */
8392 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8398 #else /* !CONFIG_RT_GROUP_SCHED */
8399 static int sched_rt_global_constraints(void)
8401 unsigned long flags;
8404 if (sysctl_sched_rt_period <= 0)
8408 * There's always some RT tasks in the root group
8409 * -- migration, kstopmachine etc..
8411 if (sysctl_sched_rt_runtime == 0)
8414 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8415 for_each_possible_cpu(i) {
8416 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8418 raw_spin_lock(&rt_rq->rt_runtime_lock);
8419 rt_rq->rt_runtime = global_rt_runtime();
8420 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8422 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8426 #endif /* CONFIG_RT_GROUP_SCHED */
8428 int sched_rt_handler(struct ctl_table *table, int write,
8429 void __user *buffer, size_t *lenp,
8433 int old_period, old_runtime;
8434 static DEFINE_MUTEX(mutex);
8437 old_period = sysctl_sched_rt_period;
8438 old_runtime = sysctl_sched_rt_runtime;
8440 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8442 if (!ret && write) {
8443 ret = sched_rt_global_constraints();
8445 sysctl_sched_rt_period = old_period;
8446 sysctl_sched_rt_runtime = old_runtime;
8448 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8449 def_rt_bandwidth.rt_period =
8450 ns_to_ktime(global_rt_period());
8453 mutex_unlock(&mutex);
8458 #ifdef CONFIG_CGROUP_SCHED
8460 /* return corresponding task_group object of a cgroup */
8461 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8463 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8464 struct task_group, css);
8467 static struct cgroup_subsys_state *
8468 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8470 struct task_group *tg, *parent;
8472 if (!cgrp->parent) {
8473 /* This is early initialization for the top cgroup */
8474 return &init_task_group.css;
8477 parent = cgroup_tg(cgrp->parent);
8478 tg = sched_create_group(parent);
8480 return ERR_PTR(-ENOMEM);
8486 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8488 struct task_group *tg = cgroup_tg(cgrp);
8490 sched_destroy_group(tg);
8494 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8496 #ifdef CONFIG_RT_GROUP_SCHED
8497 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
8500 /* We don't support RT-tasks being in separate groups */
8501 if (tsk->sched_class != &fair_sched_class)
8508 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8509 struct task_struct *tsk, bool threadgroup)
8511 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
8515 struct task_struct *c;
8517 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8518 retval = cpu_cgroup_can_attach_task(cgrp, c);
8530 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8531 struct cgroup *old_cont, struct task_struct *tsk,
8534 sched_move_task(tsk);
8536 struct task_struct *c;
8538 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8545 #ifdef CONFIG_FAIR_GROUP_SCHED
8546 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8549 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8552 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8554 struct task_group *tg = cgroup_tg(cgrp);
8556 return (u64) tg->shares;
8558 #endif /* CONFIG_FAIR_GROUP_SCHED */
8560 #ifdef CONFIG_RT_GROUP_SCHED
8561 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8564 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8567 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8569 return sched_group_rt_runtime(cgroup_tg(cgrp));
8572 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8575 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8578 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8580 return sched_group_rt_period(cgroup_tg(cgrp));
8582 #endif /* CONFIG_RT_GROUP_SCHED */
8584 static struct cftype cpu_files[] = {
8585 #ifdef CONFIG_FAIR_GROUP_SCHED
8588 .read_u64 = cpu_shares_read_u64,
8589 .write_u64 = cpu_shares_write_u64,
8592 #ifdef CONFIG_RT_GROUP_SCHED
8594 .name = "rt_runtime_us",
8595 .read_s64 = cpu_rt_runtime_read,
8596 .write_s64 = cpu_rt_runtime_write,
8599 .name = "rt_period_us",
8600 .read_u64 = cpu_rt_period_read_uint,
8601 .write_u64 = cpu_rt_period_write_uint,
8606 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8608 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8611 struct cgroup_subsys cpu_cgroup_subsys = {
8613 .create = cpu_cgroup_create,
8614 .destroy = cpu_cgroup_destroy,
8615 .can_attach = cpu_cgroup_can_attach,
8616 .attach = cpu_cgroup_attach,
8617 .populate = cpu_cgroup_populate,
8618 .subsys_id = cpu_cgroup_subsys_id,
8622 #endif /* CONFIG_CGROUP_SCHED */
8624 #ifdef CONFIG_CGROUP_CPUACCT
8627 * CPU accounting code for task groups.
8629 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8630 * (balbir@in.ibm.com).
8633 /* track cpu usage of a group of tasks and its child groups */
8635 struct cgroup_subsys_state css;
8636 /* cpuusage holds pointer to a u64-type object on every cpu */
8637 u64 __percpu *cpuusage;
8638 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
8639 struct cpuacct *parent;
8642 struct cgroup_subsys cpuacct_subsys;
8644 /* return cpu accounting group corresponding to this container */
8645 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8647 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8648 struct cpuacct, css);
8651 /* return cpu accounting group to which this task belongs */
8652 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8654 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8655 struct cpuacct, css);
8658 /* create a new cpu accounting group */
8659 static struct cgroup_subsys_state *cpuacct_create(
8660 struct cgroup_subsys *ss, struct cgroup *cgrp)
8662 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8668 ca->cpuusage = alloc_percpu(u64);
8672 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8673 if (percpu_counter_init(&ca->cpustat[i], 0))
8674 goto out_free_counters;
8677 ca->parent = cgroup_ca(cgrp->parent);
8683 percpu_counter_destroy(&ca->cpustat[i]);
8684 free_percpu(ca->cpuusage);
8688 return ERR_PTR(-ENOMEM);
8691 /* destroy an existing cpu accounting group */
8693 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8695 struct cpuacct *ca = cgroup_ca(cgrp);
8698 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8699 percpu_counter_destroy(&ca->cpustat[i]);
8700 free_percpu(ca->cpuusage);
8704 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
8706 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8709 #ifndef CONFIG_64BIT
8711 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8713 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8715 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8723 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
8725 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8727 #ifndef CONFIG_64BIT
8729 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8731 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8733 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8739 /* return total cpu usage (in nanoseconds) of a group */
8740 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8742 struct cpuacct *ca = cgroup_ca(cgrp);
8743 u64 totalcpuusage = 0;
8746 for_each_present_cpu(i)
8747 totalcpuusage += cpuacct_cpuusage_read(ca, i);
8749 return totalcpuusage;
8752 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8755 struct cpuacct *ca = cgroup_ca(cgrp);
8764 for_each_present_cpu(i)
8765 cpuacct_cpuusage_write(ca, i, 0);
8771 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
8774 struct cpuacct *ca = cgroup_ca(cgroup);
8778 for_each_present_cpu(i) {
8779 percpu = cpuacct_cpuusage_read(ca, i);
8780 seq_printf(m, "%llu ", (unsigned long long) percpu);
8782 seq_printf(m, "\n");
8786 static const char *cpuacct_stat_desc[] = {
8787 [CPUACCT_STAT_USER] = "user",
8788 [CPUACCT_STAT_SYSTEM] = "system",
8791 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
8792 struct cgroup_map_cb *cb)
8794 struct cpuacct *ca = cgroup_ca(cgrp);
8797 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
8798 s64 val = percpu_counter_read(&ca->cpustat[i]);
8799 val = cputime64_to_clock_t(val);
8800 cb->fill(cb, cpuacct_stat_desc[i], val);
8805 static struct cftype files[] = {
8808 .read_u64 = cpuusage_read,
8809 .write_u64 = cpuusage_write,
8812 .name = "usage_percpu",
8813 .read_seq_string = cpuacct_percpu_seq_read,
8817 .read_map = cpuacct_stats_show,
8821 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
8823 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
8827 * charge this task's execution time to its accounting group.
8829 * called with rq->lock held.
8831 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8836 if (unlikely(!cpuacct_subsys.active))
8839 cpu = task_cpu(tsk);
8845 for (; ca; ca = ca->parent) {
8846 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8847 *cpuusage += cputime;
8854 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
8855 * in cputime_t units. As a result, cpuacct_update_stats calls
8856 * percpu_counter_add with values large enough to always overflow the
8857 * per cpu batch limit causing bad SMP scalability.
8859 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
8860 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
8861 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
8864 #define CPUACCT_BATCH \
8865 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
8867 #define CPUACCT_BATCH 0
8871 * Charge the system/user time to the task's accounting group.
8873 static void cpuacct_update_stats(struct task_struct *tsk,
8874 enum cpuacct_stat_index idx, cputime_t val)
8877 int batch = CPUACCT_BATCH;
8879 if (unlikely(!cpuacct_subsys.active))
8886 __percpu_counter_add(&ca->cpustat[idx], val, batch);
8892 struct cgroup_subsys cpuacct_subsys = {
8894 .create = cpuacct_create,
8895 .destroy = cpuacct_destroy,
8896 .populate = cpuacct_populate,
8897 .subsys_id = cpuacct_subsys_id,
8899 #endif /* CONFIG_CGROUP_CPUACCT */
8903 void synchronize_sched_expedited(void)
8907 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
8909 #else /* #ifndef CONFIG_SMP */
8911 static atomic_t synchronize_sched_expedited_count = ATOMIC_INIT(0);
8913 static int synchronize_sched_expedited_cpu_stop(void *data)
8916 * There must be a full memory barrier on each affected CPU
8917 * between the time that try_stop_cpus() is called and the
8918 * time that it returns.
8920 * In the current initial implementation of cpu_stop, the
8921 * above condition is already met when the control reaches
8922 * this point and the following smp_mb() is not strictly
8923 * necessary. Do smp_mb() anyway for documentation and
8924 * robustness against future implementation changes.
8926 smp_mb(); /* See above comment block. */
8931 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
8932 * approach to force grace period to end quickly. This consumes
8933 * significant time on all CPUs, and is thus not recommended for
8934 * any sort of common-case code.
8936 * Note that it is illegal to call this function while holding any
8937 * lock that is acquired by a CPU-hotplug notifier. Failing to
8938 * observe this restriction will result in deadlock.
8940 void synchronize_sched_expedited(void)
8942 int snap, trycount = 0;
8944 smp_mb(); /* ensure prior mod happens before capturing snap. */
8945 snap = atomic_read(&synchronize_sched_expedited_count) + 1;
8947 while (try_stop_cpus(cpu_online_mask,
8948 synchronize_sched_expedited_cpu_stop,
8951 if (trycount++ < 10)
8952 udelay(trycount * num_online_cpus());
8954 synchronize_sched();
8957 if (atomic_read(&synchronize_sched_expedited_count) - snap > 0) {
8958 smp_mb(); /* ensure test happens before caller kfree */
8963 atomic_inc(&synchronize_sched_expedited_count);
8964 smp_mb__after_atomic_inc(); /* ensure post-GP actions seen after GP. */
8967 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
8969 #endif /* #else #ifndef CONFIG_SMP */