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"
80 #include "workqueue_sched.h"
82 #define CREATE_TRACE_POINTS
83 #include <trace/events/sched.h>
86 * Convert user-nice values [ -20 ... 0 ... 19 ]
87 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
90 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
91 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
92 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
95 * 'User priority' is the nice value converted to something we
96 * can work with better when scaling various scheduler parameters,
97 * it's a [ 0 ... 39 ] range.
99 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
100 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
101 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
104 * Helpers for converting nanosecond timing to jiffy resolution
106 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
108 #define NICE_0_LOAD SCHED_LOAD_SCALE
109 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
112 * These are the 'tuning knobs' of the scheduler:
114 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
115 * Timeslices get refilled after they expire.
117 #define DEF_TIMESLICE (100 * HZ / 1000)
120 * single value that denotes runtime == period, ie unlimited time.
122 #define RUNTIME_INF ((u64)~0ULL)
124 static inline int rt_policy(int policy)
126 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
131 static inline int task_has_rt_policy(struct task_struct *p)
133 return rt_policy(p->policy);
137 * This is the priority-queue data structure of the RT scheduling class:
139 struct rt_prio_array {
140 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
141 struct list_head queue[MAX_RT_PRIO];
144 struct rt_bandwidth {
145 /* nests inside the rq lock: */
146 raw_spinlock_t rt_runtime_lock;
149 struct hrtimer rt_period_timer;
152 static struct rt_bandwidth def_rt_bandwidth;
154 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
156 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
158 struct rt_bandwidth *rt_b =
159 container_of(timer, struct rt_bandwidth, rt_period_timer);
165 now = hrtimer_cb_get_time(timer);
166 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
171 idle = do_sched_rt_period_timer(rt_b, overrun);
174 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
178 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
180 rt_b->rt_period = ns_to_ktime(period);
181 rt_b->rt_runtime = runtime;
183 raw_spin_lock_init(&rt_b->rt_runtime_lock);
185 hrtimer_init(&rt_b->rt_period_timer,
186 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
187 rt_b->rt_period_timer.function = sched_rt_period_timer;
190 static inline int rt_bandwidth_enabled(void)
192 return sysctl_sched_rt_runtime >= 0;
195 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
199 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
202 if (hrtimer_active(&rt_b->rt_period_timer))
205 raw_spin_lock(&rt_b->rt_runtime_lock);
210 if (hrtimer_active(&rt_b->rt_period_timer))
213 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
214 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
216 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
217 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
218 delta = ktime_to_ns(ktime_sub(hard, soft));
219 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
220 HRTIMER_MODE_ABS_PINNED, 0);
222 raw_spin_unlock(&rt_b->rt_runtime_lock);
225 #ifdef CONFIG_RT_GROUP_SCHED
226 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
228 hrtimer_cancel(&rt_b->rt_period_timer);
233 * sched_domains_mutex serializes calls to arch_init_sched_domains,
234 * detach_destroy_domains and partition_sched_domains.
236 static DEFINE_MUTEX(sched_domains_mutex);
238 #ifdef CONFIG_CGROUP_SCHED
240 #include <linux/cgroup.h>
244 static LIST_HEAD(task_groups);
246 /* task group related information */
248 struct cgroup_subsys_state css;
250 #ifdef CONFIG_FAIR_GROUP_SCHED
251 /* schedulable entities of this group on each cpu */
252 struct sched_entity **se;
253 /* runqueue "owned" by this group on each cpu */
254 struct cfs_rq **cfs_rq;
255 unsigned long shares;
258 #ifdef CONFIG_RT_GROUP_SCHED
259 struct sched_rt_entity **rt_se;
260 struct rt_rq **rt_rq;
262 struct rt_bandwidth rt_bandwidth;
266 struct list_head list;
268 struct task_group *parent;
269 struct list_head siblings;
270 struct list_head children;
273 #define root_task_group init_task_group
275 /* task_group_lock serializes add/remove of task groups and also changes to
276 * a task group's cpu shares.
278 static DEFINE_SPINLOCK(task_group_lock);
280 #ifdef CONFIG_FAIR_GROUP_SCHED
283 static int root_task_group_empty(void)
285 return list_empty(&root_task_group.children);
289 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
292 * A weight of 0 or 1 can cause arithmetics problems.
293 * A weight of a cfs_rq is the sum of weights of which entities
294 * are queued on this cfs_rq, so a weight of a entity should not be
295 * too large, so as the shares value of a task group.
296 * (The default weight is 1024 - so there's no practical
297 * limitation from this.)
300 #define MAX_SHARES (1UL << 18)
302 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
305 /* Default task group.
306 * Every task in system belong to this group at bootup.
308 struct task_group init_task_group;
310 #endif /* CONFIG_CGROUP_SCHED */
312 /* CFS-related fields in a runqueue */
314 struct load_weight load;
315 unsigned long nr_running;
320 struct rb_root tasks_timeline;
321 struct rb_node *rb_leftmost;
323 struct list_head tasks;
324 struct list_head *balance_iterator;
327 * 'curr' points to currently running entity on this cfs_rq.
328 * It is set to NULL otherwise (i.e when none are currently running).
330 struct sched_entity *curr, *next, *last;
332 unsigned int nr_spread_over;
334 #ifdef CONFIG_FAIR_GROUP_SCHED
335 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
338 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
339 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
340 * (like users, containers etc.)
342 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
343 * list is used during load balance.
345 struct list_head leaf_cfs_rq_list;
346 struct task_group *tg; /* group that "owns" this runqueue */
350 * the part of load.weight contributed by tasks
352 unsigned long task_weight;
355 * h_load = weight * f(tg)
357 * Where f(tg) is the recursive weight fraction assigned to
360 unsigned long h_load;
363 * this cpu's part of tg->shares
365 unsigned long shares;
368 * load.weight at the time we set shares
370 unsigned long rq_weight;
375 /* Real-Time classes' related field in a runqueue: */
377 struct rt_prio_array active;
378 unsigned long rt_nr_running;
379 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
381 int curr; /* highest queued rt task prio */
383 int next; /* next highest */
388 unsigned long rt_nr_migratory;
389 unsigned long rt_nr_total;
391 struct plist_head pushable_tasks;
396 /* Nests inside the rq lock: */
397 raw_spinlock_t rt_runtime_lock;
399 #ifdef CONFIG_RT_GROUP_SCHED
400 unsigned long rt_nr_boosted;
403 struct list_head leaf_rt_rq_list;
404 struct task_group *tg;
411 * We add the notion of a root-domain which will be used to define per-domain
412 * variables. Each exclusive cpuset essentially defines an island domain by
413 * fully partitioning the member cpus from any other cpuset. Whenever a new
414 * exclusive cpuset is created, we also create and attach a new root-domain
421 cpumask_var_t online;
424 * The "RT overload" flag: it gets set if a CPU has more than
425 * one runnable RT task.
427 cpumask_var_t rto_mask;
429 struct cpupri cpupri;
433 * By default the system creates a single root-domain with all cpus as
434 * members (mimicking the global state we have today).
436 static struct root_domain def_root_domain;
438 #endif /* CONFIG_SMP */
441 * This is the main, per-CPU runqueue data structure.
443 * Locking rule: those places that want to lock multiple runqueues
444 * (such as the load balancing or the thread migration code), lock
445 * acquire operations must be ordered by ascending &runqueue.
452 * nr_running and cpu_load should be in the same cacheline because
453 * remote CPUs use both these fields when doing load calculation.
455 unsigned long nr_running;
456 #define CPU_LOAD_IDX_MAX 5
457 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
458 unsigned long last_load_update_tick;
461 unsigned char nohz_balance_kick;
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, *stop;
490 unsigned long next_balance;
491 struct mm_struct *prev_mm;
499 struct root_domain *rd;
500 struct sched_domain *sd;
502 unsigned long cpu_power;
504 unsigned char idle_at_tick;
505 /* For active balancing */
509 struct cpu_stop_work active_balance_work;
510 /* cpu of this runqueue: */
514 unsigned long avg_load_per_task;
522 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
526 /* calc_load related fields */
527 unsigned long calc_load_update;
528 long calc_load_active;
530 #ifdef CONFIG_SCHED_HRTICK
532 int hrtick_csd_pending;
533 struct call_single_data hrtick_csd;
535 struct hrtimer hrtick_timer;
538 #ifdef CONFIG_SCHEDSTATS
540 struct sched_info rq_sched_info;
541 unsigned long long rq_cpu_time;
542 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
544 /* sys_sched_yield() stats */
545 unsigned int yld_count;
547 /* schedule() stats */
548 unsigned int sched_switch;
549 unsigned int sched_count;
550 unsigned int sched_goidle;
552 /* try_to_wake_up() stats */
553 unsigned int ttwu_count;
554 unsigned int ttwu_local;
557 unsigned int bkl_count;
561 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
564 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags);
566 static inline int cpu_of(struct rq *rq)
575 #define rcu_dereference_check_sched_domain(p) \
576 rcu_dereference_check((p), \
577 rcu_read_lock_sched_held() || \
578 lockdep_is_held(&sched_domains_mutex))
581 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
582 * See detach_destroy_domains: synchronize_sched for details.
584 * The domain tree of any CPU may only be accessed from within
585 * preempt-disabled sections.
587 #define for_each_domain(cpu, __sd) \
588 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
590 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
591 #define this_rq() (&__get_cpu_var(runqueues))
592 #define task_rq(p) cpu_rq(task_cpu(p))
593 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
594 #define raw_rq() (&__raw_get_cpu_var(runqueues))
596 #ifdef CONFIG_CGROUP_SCHED
599 * Return the group to which this tasks belongs.
601 * We use task_subsys_state_check() and extend the RCU verification
602 * with lockdep_is_held(&task_rq(p)->lock) because cpu_cgroup_attach()
603 * holds that lock for each task it moves into the cgroup. Therefore
604 * by holding that lock, we pin the task to the current cgroup.
606 static inline struct task_group *task_group(struct task_struct *p)
608 struct cgroup_subsys_state *css;
610 css = task_subsys_state_check(p, cpu_cgroup_subsys_id,
611 lockdep_is_held(&task_rq(p)->lock));
612 return container_of(css, struct task_group, css);
615 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
616 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
618 #ifdef CONFIG_FAIR_GROUP_SCHED
619 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
620 p->se.parent = task_group(p)->se[cpu];
623 #ifdef CONFIG_RT_GROUP_SCHED
624 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
625 p->rt.parent = task_group(p)->rt_se[cpu];
629 #else /* CONFIG_CGROUP_SCHED */
631 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
632 static inline struct task_group *task_group(struct task_struct *p)
637 #endif /* CONFIG_CGROUP_SCHED */
639 static void update_rq_clock_task(struct rq *rq, s64 delta);
641 static void update_rq_clock(struct rq *rq)
645 if (rq->skip_clock_update)
648 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
650 update_rq_clock_task(rq, delta);
654 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
656 #ifdef CONFIG_SCHED_DEBUG
657 # define const_debug __read_mostly
659 # define const_debug static const
664 * @cpu: the processor in question.
666 * Returns true if the current cpu runqueue is locked.
667 * This interface allows printk to be called with the runqueue lock
668 * held and know whether or not it is OK to wake up the klogd.
670 int runqueue_is_locked(int cpu)
672 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
676 * Debugging: various feature bits
679 #define SCHED_FEAT(name, enabled) \
680 __SCHED_FEAT_##name ,
683 #include "sched_features.h"
688 #define SCHED_FEAT(name, enabled) \
689 (1UL << __SCHED_FEAT_##name) * enabled |
691 const_debug unsigned int sysctl_sched_features =
692 #include "sched_features.h"
697 #ifdef CONFIG_SCHED_DEBUG
698 #define SCHED_FEAT(name, enabled) \
701 static __read_mostly char *sched_feat_names[] = {
702 #include "sched_features.h"
708 static int sched_feat_show(struct seq_file *m, void *v)
712 for (i = 0; sched_feat_names[i]; i++) {
713 if (!(sysctl_sched_features & (1UL << i)))
715 seq_printf(m, "%s ", sched_feat_names[i]);
723 sched_feat_write(struct file *filp, const char __user *ubuf,
724 size_t cnt, loff_t *ppos)
734 if (copy_from_user(&buf, ubuf, cnt))
740 if (strncmp(buf, "NO_", 3) == 0) {
745 for (i = 0; sched_feat_names[i]; i++) {
746 if (strcmp(cmp, sched_feat_names[i]) == 0) {
748 sysctl_sched_features &= ~(1UL << i);
750 sysctl_sched_features |= (1UL << i);
755 if (!sched_feat_names[i])
763 static int sched_feat_open(struct inode *inode, struct file *filp)
765 return single_open(filp, sched_feat_show, NULL);
768 static const struct file_operations sched_feat_fops = {
769 .open = sched_feat_open,
770 .write = sched_feat_write,
773 .release = single_release,
776 static __init int sched_init_debug(void)
778 debugfs_create_file("sched_features", 0644, NULL, NULL,
783 late_initcall(sched_init_debug);
787 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
790 * Number of tasks to iterate in a single balance run.
791 * Limited because this is done with IRQs disabled.
793 const_debug unsigned int sysctl_sched_nr_migrate = 32;
796 * ratelimit for updating the group shares.
799 unsigned int sysctl_sched_shares_ratelimit = 250000;
800 unsigned int normalized_sysctl_sched_shares_ratelimit = 250000;
803 * Inject some fuzzyness into changing the per-cpu group shares
804 * this avoids remote rq-locks at the expense of fairness.
807 unsigned int sysctl_sched_shares_thresh = 4;
810 * period over which we average the RT time consumption, measured
815 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
818 * period over which we measure -rt task cpu usage in us.
821 unsigned int sysctl_sched_rt_period = 1000000;
823 static __read_mostly int scheduler_running;
826 * part of the period that we allow rt tasks to run in us.
829 int sysctl_sched_rt_runtime = 950000;
831 static inline u64 global_rt_period(void)
833 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
836 static inline u64 global_rt_runtime(void)
838 if (sysctl_sched_rt_runtime < 0)
841 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
844 #ifndef prepare_arch_switch
845 # define prepare_arch_switch(next) do { } while (0)
847 #ifndef finish_arch_switch
848 # define finish_arch_switch(prev) do { } while (0)
851 static inline int task_current(struct rq *rq, struct task_struct *p)
853 return rq->curr == p;
856 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
857 static inline int task_running(struct rq *rq, struct task_struct *p)
859 return task_current(rq, p);
862 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
866 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
868 #ifdef CONFIG_DEBUG_SPINLOCK
869 /* this is a valid case when another task releases the spinlock */
870 rq->lock.owner = current;
873 * If we are tracking spinlock dependencies then we have to
874 * fix up the runqueue lock - which gets 'carried over' from
877 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
879 raw_spin_unlock_irq(&rq->lock);
882 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
883 static inline int task_running(struct rq *rq, struct task_struct *p)
888 return task_current(rq, p);
892 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
896 * We can optimise this out completely for !SMP, because the
897 * SMP rebalancing from interrupt is the only thing that cares
902 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
903 raw_spin_unlock_irq(&rq->lock);
905 raw_spin_unlock(&rq->lock);
909 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
913 * After ->oncpu is cleared, the task can be moved to a different CPU.
914 * We must ensure this doesn't happen until the switch is completely
920 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
924 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
927 * Check whether the task is waking, we use this to synchronize ->cpus_allowed
930 static inline int task_is_waking(struct task_struct *p)
932 return unlikely(p->state == TASK_WAKING);
936 * __task_rq_lock - lock the runqueue a given task resides on.
937 * Must be called interrupts disabled.
939 static inline struct rq *__task_rq_lock(struct task_struct *p)
946 raw_spin_lock(&rq->lock);
947 if (likely(rq == task_rq(p)))
949 raw_spin_unlock(&rq->lock);
954 * task_rq_lock - lock the runqueue a given task resides on and disable
955 * interrupts. Note the ordering: we can safely lookup the task_rq without
956 * explicitly disabling preemption.
958 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
964 local_irq_save(*flags);
966 raw_spin_lock(&rq->lock);
967 if (likely(rq == task_rq(p)))
969 raw_spin_unlock_irqrestore(&rq->lock, *flags);
973 static void __task_rq_unlock(struct rq *rq)
976 raw_spin_unlock(&rq->lock);
979 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
982 raw_spin_unlock_irqrestore(&rq->lock, *flags);
986 * this_rq_lock - lock this runqueue and disable interrupts.
988 static struct rq *this_rq_lock(void)
995 raw_spin_lock(&rq->lock);
1000 #ifdef CONFIG_SCHED_HRTICK
1002 * Use HR-timers to deliver accurate preemption points.
1004 * Its all a bit involved since we cannot program an hrt while holding the
1005 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1008 * When we get rescheduled we reprogram the hrtick_timer outside of the
1014 * - enabled by features
1015 * - hrtimer is actually high res
1017 static inline int hrtick_enabled(struct rq *rq)
1019 if (!sched_feat(HRTICK))
1021 if (!cpu_active(cpu_of(rq)))
1023 return hrtimer_is_hres_active(&rq->hrtick_timer);
1026 static void hrtick_clear(struct rq *rq)
1028 if (hrtimer_active(&rq->hrtick_timer))
1029 hrtimer_cancel(&rq->hrtick_timer);
1033 * High-resolution timer tick.
1034 * Runs from hardirq context with interrupts disabled.
1036 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1038 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1040 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1042 raw_spin_lock(&rq->lock);
1043 update_rq_clock(rq);
1044 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1045 raw_spin_unlock(&rq->lock);
1047 return HRTIMER_NORESTART;
1052 * called from hardirq (IPI) context
1054 static void __hrtick_start(void *arg)
1056 struct rq *rq = arg;
1058 raw_spin_lock(&rq->lock);
1059 hrtimer_restart(&rq->hrtick_timer);
1060 rq->hrtick_csd_pending = 0;
1061 raw_spin_unlock(&rq->lock);
1065 * Called to set the hrtick timer state.
1067 * called with rq->lock held and irqs disabled
1069 static void hrtick_start(struct rq *rq, u64 delay)
1071 struct hrtimer *timer = &rq->hrtick_timer;
1072 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1074 hrtimer_set_expires(timer, time);
1076 if (rq == this_rq()) {
1077 hrtimer_restart(timer);
1078 } else if (!rq->hrtick_csd_pending) {
1079 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1080 rq->hrtick_csd_pending = 1;
1085 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1087 int cpu = (int)(long)hcpu;
1090 case CPU_UP_CANCELED:
1091 case CPU_UP_CANCELED_FROZEN:
1092 case CPU_DOWN_PREPARE:
1093 case CPU_DOWN_PREPARE_FROZEN:
1095 case CPU_DEAD_FROZEN:
1096 hrtick_clear(cpu_rq(cpu));
1103 static __init void init_hrtick(void)
1105 hotcpu_notifier(hotplug_hrtick, 0);
1109 * Called to set the hrtick timer state.
1111 * called with rq->lock held and irqs disabled
1113 static void hrtick_start(struct rq *rq, u64 delay)
1115 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1116 HRTIMER_MODE_REL_PINNED, 0);
1119 static inline void init_hrtick(void)
1122 #endif /* CONFIG_SMP */
1124 static void init_rq_hrtick(struct rq *rq)
1127 rq->hrtick_csd_pending = 0;
1129 rq->hrtick_csd.flags = 0;
1130 rq->hrtick_csd.func = __hrtick_start;
1131 rq->hrtick_csd.info = rq;
1134 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1135 rq->hrtick_timer.function = hrtick;
1137 #else /* CONFIG_SCHED_HRTICK */
1138 static inline void hrtick_clear(struct rq *rq)
1142 static inline void init_rq_hrtick(struct rq *rq)
1146 static inline void init_hrtick(void)
1149 #endif /* CONFIG_SCHED_HRTICK */
1152 * resched_task - mark a task 'to be rescheduled now'.
1154 * On UP this means the setting of the need_resched flag, on SMP it
1155 * might also involve a cross-CPU call to trigger the scheduler on
1160 #ifndef tsk_is_polling
1161 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1164 static void resched_task(struct task_struct *p)
1168 assert_raw_spin_locked(&task_rq(p)->lock);
1170 if (test_tsk_need_resched(p))
1173 set_tsk_need_resched(p);
1176 if (cpu == smp_processor_id())
1179 /* NEED_RESCHED must be visible before we test polling */
1181 if (!tsk_is_polling(p))
1182 smp_send_reschedule(cpu);
1185 static void resched_cpu(int cpu)
1187 struct rq *rq = cpu_rq(cpu);
1188 unsigned long flags;
1190 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1192 resched_task(cpu_curr(cpu));
1193 raw_spin_unlock_irqrestore(&rq->lock, flags);
1198 * In the semi idle case, use the nearest busy cpu for migrating timers
1199 * from an idle cpu. This is good for power-savings.
1201 * We don't do similar optimization for completely idle system, as
1202 * selecting an idle cpu will add more delays to the timers than intended
1203 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1205 int get_nohz_timer_target(void)
1207 int cpu = smp_processor_id();
1209 struct sched_domain *sd;
1211 for_each_domain(cpu, sd) {
1212 for_each_cpu(i, sched_domain_span(sd))
1219 * When add_timer_on() enqueues a timer into the timer wheel of an
1220 * idle CPU then this timer might expire before the next timer event
1221 * which is scheduled to wake up that CPU. In case of a completely
1222 * idle system the next event might even be infinite time into the
1223 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1224 * leaves the inner idle loop so the newly added timer is taken into
1225 * account when the CPU goes back to idle and evaluates the timer
1226 * wheel for the next timer event.
1228 void wake_up_idle_cpu(int cpu)
1230 struct rq *rq = cpu_rq(cpu);
1232 if (cpu == smp_processor_id())
1236 * This is safe, as this function is called with the timer
1237 * wheel base lock of (cpu) held. When the CPU is on the way
1238 * to idle and has not yet set rq->curr to idle then it will
1239 * be serialized on the timer wheel base lock and take the new
1240 * timer into account automatically.
1242 if (rq->curr != rq->idle)
1246 * We can set TIF_RESCHED on the idle task of the other CPU
1247 * lockless. The worst case is that the other CPU runs the
1248 * idle task through an additional NOOP schedule()
1250 set_tsk_need_resched(rq->idle);
1252 /* NEED_RESCHED must be visible before we test polling */
1254 if (!tsk_is_polling(rq->idle))
1255 smp_send_reschedule(cpu);
1258 #endif /* CONFIG_NO_HZ */
1260 static u64 sched_avg_period(void)
1262 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1265 static void sched_avg_update(struct rq *rq)
1267 s64 period = sched_avg_period();
1269 while ((s64)(rq->clock - rq->age_stamp) > period) {
1271 * Inline assembly required to prevent the compiler
1272 * optimising this loop into a divmod call.
1273 * See __iter_div_u64_rem() for another example of this.
1275 asm("" : "+rm" (rq->age_stamp));
1276 rq->age_stamp += period;
1281 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1283 rq->rt_avg += rt_delta;
1284 sched_avg_update(rq);
1287 #else /* !CONFIG_SMP */
1288 static void resched_task(struct task_struct *p)
1290 assert_raw_spin_locked(&task_rq(p)->lock);
1291 set_tsk_need_resched(p);
1294 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1298 static void sched_avg_update(struct rq *rq)
1301 #endif /* CONFIG_SMP */
1303 #if BITS_PER_LONG == 32
1304 # define WMULT_CONST (~0UL)
1306 # define WMULT_CONST (1UL << 32)
1309 #define WMULT_SHIFT 32
1312 * Shift right and round:
1314 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1317 * delta *= weight / lw
1319 static unsigned long
1320 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1321 struct load_weight *lw)
1325 if (!lw->inv_weight) {
1326 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1329 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1333 tmp = (u64)delta_exec * weight;
1335 * Check whether we'd overflow the 64-bit multiplication:
1337 if (unlikely(tmp > WMULT_CONST))
1338 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1341 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1343 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1346 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1352 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1359 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1360 * of tasks with abnormal "nice" values across CPUs the contribution that
1361 * each task makes to its run queue's load is weighted according to its
1362 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1363 * scaled version of the new time slice allocation that they receive on time
1367 #define WEIGHT_IDLEPRIO 3
1368 #define WMULT_IDLEPRIO 1431655765
1371 * Nice levels are multiplicative, with a gentle 10% change for every
1372 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1373 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1374 * that remained on nice 0.
1376 * The "10% effect" is relative and cumulative: from _any_ nice level,
1377 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1378 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1379 * If a task goes up by ~10% and another task goes down by ~10% then
1380 * the relative distance between them is ~25%.)
1382 static const int prio_to_weight[40] = {
1383 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1384 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1385 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1386 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1387 /* 0 */ 1024, 820, 655, 526, 423,
1388 /* 5 */ 335, 272, 215, 172, 137,
1389 /* 10 */ 110, 87, 70, 56, 45,
1390 /* 15 */ 36, 29, 23, 18, 15,
1394 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1396 * In cases where the weight does not change often, we can use the
1397 * precalculated inverse to speed up arithmetics by turning divisions
1398 * into multiplications:
1400 static const u32 prio_to_wmult[40] = {
1401 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1402 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1403 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1404 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1405 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1406 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1407 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1408 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1411 /* Time spent by the tasks of the cpu accounting group executing in ... */
1412 enum cpuacct_stat_index {
1413 CPUACCT_STAT_USER, /* ... user mode */
1414 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1416 CPUACCT_STAT_NSTATS,
1419 #ifdef CONFIG_CGROUP_CPUACCT
1420 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1421 static void cpuacct_update_stats(struct task_struct *tsk,
1422 enum cpuacct_stat_index idx, cputime_t val);
1424 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1425 static inline void cpuacct_update_stats(struct task_struct *tsk,
1426 enum cpuacct_stat_index idx, cputime_t val) {}
1429 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1431 update_load_add(&rq->load, load);
1434 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1436 update_load_sub(&rq->load, load);
1439 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1440 typedef int (*tg_visitor)(struct task_group *, void *);
1443 * Iterate the full tree, calling @down when first entering a node and @up when
1444 * leaving it for the final time.
1446 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1448 struct task_group *parent, *child;
1452 parent = &root_task_group;
1454 ret = (*down)(parent, data);
1457 list_for_each_entry_rcu(child, &parent->children, siblings) {
1464 ret = (*up)(parent, data);
1469 parent = parent->parent;
1478 static int tg_nop(struct task_group *tg, void *data)
1485 /* Used instead of source_load when we know the type == 0 */
1486 static unsigned long weighted_cpuload(const int cpu)
1488 return cpu_rq(cpu)->load.weight;
1492 * Return a low guess at the load of a migration-source cpu weighted
1493 * according to the scheduling class and "nice" value.
1495 * We want to under-estimate the load of migration sources, to
1496 * balance conservatively.
1498 static unsigned long source_load(int cpu, int type)
1500 struct rq *rq = cpu_rq(cpu);
1501 unsigned long total = weighted_cpuload(cpu);
1503 if (type == 0 || !sched_feat(LB_BIAS))
1506 return min(rq->cpu_load[type-1], total);
1510 * Return a high guess at the load of a migration-target cpu weighted
1511 * according to the scheduling class and "nice" value.
1513 static unsigned long target_load(int cpu, int type)
1515 struct rq *rq = cpu_rq(cpu);
1516 unsigned long total = weighted_cpuload(cpu);
1518 if (type == 0 || !sched_feat(LB_BIAS))
1521 return max(rq->cpu_load[type-1], total);
1524 static unsigned long power_of(int cpu)
1526 return cpu_rq(cpu)->cpu_power;
1529 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1531 static unsigned long cpu_avg_load_per_task(int cpu)
1533 struct rq *rq = cpu_rq(cpu);
1534 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1537 rq->avg_load_per_task = rq->load.weight / nr_running;
1539 rq->avg_load_per_task = 0;
1541 return rq->avg_load_per_task;
1544 #ifdef CONFIG_FAIR_GROUP_SCHED
1546 static __read_mostly unsigned long __percpu *update_shares_data;
1548 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1551 * Calculate and set the cpu's group shares.
1553 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1554 unsigned long sd_shares,
1555 unsigned long sd_rq_weight,
1556 unsigned long *usd_rq_weight)
1558 unsigned long shares, rq_weight;
1561 rq_weight = usd_rq_weight[cpu];
1564 rq_weight = NICE_0_LOAD;
1568 * \Sum_j shares_j * rq_weight_i
1569 * shares_i = -----------------------------
1570 * \Sum_j rq_weight_j
1572 shares = (sd_shares * rq_weight) / sd_rq_weight;
1573 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1575 if (abs(shares - tg->se[cpu]->load.weight) >
1576 sysctl_sched_shares_thresh) {
1577 struct rq *rq = cpu_rq(cpu);
1578 unsigned long flags;
1580 raw_spin_lock_irqsave(&rq->lock, flags);
1581 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1582 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1583 __set_se_shares(tg->se[cpu], shares);
1584 raw_spin_unlock_irqrestore(&rq->lock, flags);
1589 * Re-compute the task group their per cpu shares over the given domain.
1590 * This needs to be done in a bottom-up fashion because the rq weight of a
1591 * parent group depends on the shares of its child groups.
1593 static int tg_shares_up(struct task_group *tg, void *data)
1595 unsigned long weight, rq_weight = 0, sum_weight = 0, shares = 0;
1596 unsigned long *usd_rq_weight;
1597 struct sched_domain *sd = data;
1598 unsigned long flags;
1604 local_irq_save(flags);
1605 usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id());
1607 for_each_cpu(i, sched_domain_span(sd)) {
1608 weight = tg->cfs_rq[i]->load.weight;
1609 usd_rq_weight[i] = weight;
1611 rq_weight += weight;
1613 * If there are currently no tasks on the cpu pretend there
1614 * is one of average load so that when a new task gets to
1615 * run here it will not get delayed by group starvation.
1618 weight = NICE_0_LOAD;
1620 sum_weight += weight;
1621 shares += tg->cfs_rq[i]->shares;
1625 rq_weight = sum_weight;
1627 if ((!shares && rq_weight) || shares > tg->shares)
1628 shares = tg->shares;
1630 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1631 shares = tg->shares;
1633 for_each_cpu(i, sched_domain_span(sd))
1634 update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight);
1636 local_irq_restore(flags);
1642 * Compute the cpu's hierarchical load factor for each task group.
1643 * This needs to be done in a top-down fashion because the load of a child
1644 * group is a fraction of its parents load.
1646 static int tg_load_down(struct task_group *tg, void *data)
1649 long cpu = (long)data;
1652 load = cpu_rq(cpu)->load.weight;
1654 load = tg->parent->cfs_rq[cpu]->h_load;
1655 load *= tg->cfs_rq[cpu]->shares;
1656 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1659 tg->cfs_rq[cpu]->h_load = load;
1664 static void update_shares(struct sched_domain *sd)
1669 if (root_task_group_empty())
1672 now = local_clock();
1673 elapsed = now - sd->last_update;
1675 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1676 sd->last_update = now;
1677 walk_tg_tree(tg_nop, tg_shares_up, sd);
1681 static void update_h_load(long cpu)
1683 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1688 static inline void update_shares(struct sched_domain *sd)
1694 #ifdef CONFIG_PREEMPT
1696 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1699 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1700 * way at the expense of forcing extra atomic operations in all
1701 * invocations. This assures that the double_lock is acquired using the
1702 * same underlying policy as the spinlock_t on this architecture, which
1703 * reduces latency compared to the unfair variant below. However, it
1704 * also adds more overhead and therefore may reduce throughput.
1706 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1707 __releases(this_rq->lock)
1708 __acquires(busiest->lock)
1709 __acquires(this_rq->lock)
1711 raw_spin_unlock(&this_rq->lock);
1712 double_rq_lock(this_rq, busiest);
1719 * Unfair double_lock_balance: Optimizes throughput at the expense of
1720 * latency by eliminating extra atomic operations when the locks are
1721 * already in proper order on entry. This favors lower cpu-ids and will
1722 * grant the double lock to lower cpus over higher ids under contention,
1723 * regardless of entry order into the function.
1725 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1726 __releases(this_rq->lock)
1727 __acquires(busiest->lock)
1728 __acquires(this_rq->lock)
1732 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1733 if (busiest < this_rq) {
1734 raw_spin_unlock(&this_rq->lock);
1735 raw_spin_lock(&busiest->lock);
1736 raw_spin_lock_nested(&this_rq->lock,
1737 SINGLE_DEPTH_NESTING);
1740 raw_spin_lock_nested(&busiest->lock,
1741 SINGLE_DEPTH_NESTING);
1746 #endif /* CONFIG_PREEMPT */
1749 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1751 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1753 if (unlikely(!irqs_disabled())) {
1754 /* printk() doesn't work good under rq->lock */
1755 raw_spin_unlock(&this_rq->lock);
1759 return _double_lock_balance(this_rq, busiest);
1762 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1763 __releases(busiest->lock)
1765 raw_spin_unlock(&busiest->lock);
1766 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1770 * double_rq_lock - safely lock two runqueues
1772 * Note this does not disable interrupts like task_rq_lock,
1773 * you need to do so manually before calling.
1775 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1776 __acquires(rq1->lock)
1777 __acquires(rq2->lock)
1779 BUG_ON(!irqs_disabled());
1781 raw_spin_lock(&rq1->lock);
1782 __acquire(rq2->lock); /* Fake it out ;) */
1785 raw_spin_lock(&rq1->lock);
1786 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1788 raw_spin_lock(&rq2->lock);
1789 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1795 * double_rq_unlock - safely unlock two runqueues
1797 * Note this does not restore interrupts like task_rq_unlock,
1798 * you need to do so manually after calling.
1800 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1801 __releases(rq1->lock)
1802 __releases(rq2->lock)
1804 raw_spin_unlock(&rq1->lock);
1806 raw_spin_unlock(&rq2->lock);
1808 __release(rq2->lock);
1813 #ifdef CONFIG_FAIR_GROUP_SCHED
1814 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1817 cfs_rq->shares = shares;
1822 static void calc_load_account_idle(struct rq *this_rq);
1823 static void update_sysctl(void);
1824 static int get_update_sysctl_factor(void);
1825 static void update_cpu_load(struct rq *this_rq);
1827 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1829 set_task_rq(p, cpu);
1832 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1833 * successfuly executed on another CPU. We must ensure that updates of
1834 * per-task data have been completed by this moment.
1837 task_thread_info(p)->cpu = cpu;
1841 static const struct sched_class rt_sched_class;
1843 #define sched_class_highest (&stop_sched_class)
1844 #define for_each_class(class) \
1845 for (class = sched_class_highest; class; class = class->next)
1847 #include "sched_stats.h"
1849 static void inc_nr_running(struct rq *rq)
1854 static void dec_nr_running(struct rq *rq)
1859 static void set_load_weight(struct task_struct *p)
1862 * SCHED_IDLE tasks get minimal weight:
1864 if (p->policy == SCHED_IDLE) {
1865 p->se.load.weight = WEIGHT_IDLEPRIO;
1866 p->se.load.inv_weight = WMULT_IDLEPRIO;
1870 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1871 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1874 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1876 update_rq_clock(rq);
1877 sched_info_queued(p);
1878 p->sched_class->enqueue_task(rq, p, flags);
1882 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1884 update_rq_clock(rq);
1885 sched_info_dequeued(p);
1886 p->sched_class->dequeue_task(rq, p, flags);
1891 * activate_task - move a task to the runqueue.
1893 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1895 if (task_contributes_to_load(p))
1896 rq->nr_uninterruptible--;
1898 enqueue_task(rq, p, flags);
1903 * deactivate_task - remove a task from the runqueue.
1905 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1907 if (task_contributes_to_load(p))
1908 rq->nr_uninterruptible++;
1910 dequeue_task(rq, p, flags);
1914 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1917 * There are no locks covering percpu hardirq/softirq time.
1918 * They are only modified in account_system_vtime, on corresponding CPU
1919 * with interrupts disabled. So, writes are safe.
1920 * They are read and saved off onto struct rq in update_rq_clock().
1921 * This may result in other CPU reading this CPU's irq time and can
1922 * race with irq/account_system_vtime on this CPU. We would either get old
1923 * or new value (or semi updated value on 32 bit) with a side effect of
1924 * accounting a slice of irq time to wrong task when irq is in progress
1925 * while we read rq->clock. That is a worthy compromise in place of having
1926 * locks on each irq in account_system_time.
1928 static DEFINE_PER_CPU(u64, cpu_hardirq_time);
1929 static DEFINE_PER_CPU(u64, cpu_softirq_time);
1931 static DEFINE_PER_CPU(u64, irq_start_time);
1932 static int sched_clock_irqtime;
1934 void enable_sched_clock_irqtime(void)
1936 sched_clock_irqtime = 1;
1939 void disable_sched_clock_irqtime(void)
1941 sched_clock_irqtime = 0;
1944 static inline u64 irq_time_cpu(int cpu)
1946 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
1950 * Called before incrementing preempt_count on {soft,}irq_enter
1951 * and before decrementing preempt_count on {soft,}irq_exit.
1953 void account_system_vtime(struct task_struct *curr)
1955 unsigned long flags;
1959 if (!sched_clock_irqtime)
1962 local_irq_save(flags);
1964 cpu = smp_processor_id();
1965 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
1966 __this_cpu_add(irq_start_time, delta);
1969 * We do not account for softirq time from ksoftirqd here.
1970 * We want to continue accounting softirq time to ksoftirqd thread
1971 * in that case, so as not to confuse scheduler with a special task
1972 * that do not consume any time, but still wants to run.
1974 if (hardirq_count())
1975 __this_cpu_add(cpu_hardirq_time, delta);
1976 else if (in_serving_softirq() && !(curr->flags & PF_KSOFTIRQD))
1977 __this_cpu_add(cpu_softirq_time, delta);
1979 local_irq_restore(flags);
1981 EXPORT_SYMBOL_GPL(account_system_vtime);
1983 static void update_rq_clock_task(struct rq *rq, s64 delta)
1987 irq_delta = irq_time_cpu(cpu_of(rq)) - rq->prev_irq_time;
1990 * Since irq_time is only updated on {soft,}irq_exit, we might run into
1991 * this case when a previous update_rq_clock() happened inside a
1992 * {soft,}irq region.
1994 * When this happens, we stop ->clock_task and only update the
1995 * prev_irq_time stamp to account for the part that fit, so that a next
1996 * update will consume the rest. This ensures ->clock_task is
1999 * It does however cause some slight miss-attribution of {soft,}irq
2000 * time, a more accurate solution would be to update the irq_time using
2001 * the current rq->clock timestamp, except that would require using
2004 if (irq_delta > delta)
2007 rq->prev_irq_time += irq_delta;
2009 rq->clock_task += delta;
2011 if (irq_delta && sched_feat(NONIRQ_POWER))
2012 sched_rt_avg_update(rq, irq_delta);
2015 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
2017 static void update_rq_clock_task(struct rq *rq, s64 delta)
2019 rq->clock_task += delta;
2022 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2024 #include "sched_idletask.c"
2025 #include "sched_fair.c"
2026 #include "sched_rt.c"
2027 #include "sched_stoptask.c"
2028 #ifdef CONFIG_SCHED_DEBUG
2029 # include "sched_debug.c"
2032 void sched_set_stop_task(int cpu, struct task_struct *stop)
2034 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
2035 struct task_struct *old_stop = cpu_rq(cpu)->stop;
2039 * Make it appear like a SCHED_FIFO task, its something
2040 * userspace knows about and won't get confused about.
2042 * Also, it will make PI more or less work without too
2043 * much confusion -- but then, stop work should not
2044 * rely on PI working anyway.
2046 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
2048 stop->sched_class = &stop_sched_class;
2051 cpu_rq(cpu)->stop = stop;
2055 * Reset it back to a normal scheduling class so that
2056 * it can die in pieces.
2058 old_stop->sched_class = &rt_sched_class;
2063 * __normal_prio - return the priority that is based on the static prio
2065 static inline int __normal_prio(struct task_struct *p)
2067 return p->static_prio;
2071 * Calculate the expected normal priority: i.e. priority
2072 * without taking RT-inheritance into account. Might be
2073 * boosted by interactivity modifiers. Changes upon fork,
2074 * setprio syscalls, and whenever the interactivity
2075 * estimator recalculates.
2077 static inline int normal_prio(struct task_struct *p)
2081 if (task_has_rt_policy(p))
2082 prio = MAX_RT_PRIO-1 - p->rt_priority;
2084 prio = __normal_prio(p);
2089 * Calculate the current priority, i.e. the priority
2090 * taken into account by the scheduler. This value might
2091 * be boosted by RT tasks, or might be boosted by
2092 * interactivity modifiers. Will be RT if the task got
2093 * RT-boosted. If not then it returns p->normal_prio.
2095 static int effective_prio(struct task_struct *p)
2097 p->normal_prio = normal_prio(p);
2099 * If we are RT tasks or we were boosted to RT priority,
2100 * keep the priority unchanged. Otherwise, update priority
2101 * to the normal priority:
2103 if (!rt_prio(p->prio))
2104 return p->normal_prio;
2109 * task_curr - is this task currently executing on a CPU?
2110 * @p: the task in question.
2112 inline int task_curr(const struct task_struct *p)
2114 return cpu_curr(task_cpu(p)) == p;
2117 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2118 const struct sched_class *prev_class,
2119 int oldprio, int running)
2121 if (prev_class != p->sched_class) {
2122 if (prev_class->switched_from)
2123 prev_class->switched_from(rq, p, running);
2124 p->sched_class->switched_to(rq, p, running);
2126 p->sched_class->prio_changed(rq, p, oldprio, running);
2129 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2131 const struct sched_class *class;
2133 if (p->sched_class == rq->curr->sched_class) {
2134 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2136 for_each_class(class) {
2137 if (class == rq->curr->sched_class)
2139 if (class == p->sched_class) {
2140 resched_task(rq->curr);
2147 * A queue event has occurred, and we're going to schedule. In
2148 * this case, we can save a useless back to back clock update.
2150 if (rq->curr->se.on_rq && test_tsk_need_resched(rq->curr))
2151 rq->skip_clock_update = 1;
2156 * Is this task likely cache-hot:
2159 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2163 if (p->sched_class != &fair_sched_class)
2166 if (unlikely(p->policy == SCHED_IDLE))
2170 * Buddy candidates are cache hot:
2172 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2173 (&p->se == cfs_rq_of(&p->se)->next ||
2174 &p->se == cfs_rq_of(&p->se)->last))
2177 if (sysctl_sched_migration_cost == -1)
2179 if (sysctl_sched_migration_cost == 0)
2182 delta = now - p->se.exec_start;
2184 return delta < (s64)sysctl_sched_migration_cost;
2187 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2189 #ifdef CONFIG_SCHED_DEBUG
2191 * We should never call set_task_cpu() on a blocked task,
2192 * ttwu() will sort out the placement.
2194 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2195 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2198 trace_sched_migrate_task(p, new_cpu);
2200 if (task_cpu(p) != new_cpu) {
2201 p->se.nr_migrations++;
2202 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2205 __set_task_cpu(p, new_cpu);
2208 struct migration_arg {
2209 struct task_struct *task;
2213 static int migration_cpu_stop(void *data);
2216 * The task's runqueue lock must be held.
2217 * Returns true if you have to wait for migration thread.
2219 static bool migrate_task(struct task_struct *p, int dest_cpu)
2221 struct rq *rq = task_rq(p);
2224 * If the task is not on a runqueue (and not running), then
2225 * the next wake-up will properly place the task.
2227 return p->se.on_rq || task_running(rq, p);
2231 * wait_task_inactive - wait for a thread to unschedule.
2233 * If @match_state is nonzero, it's the @p->state value just checked and
2234 * not expected to change. If it changes, i.e. @p might have woken up,
2235 * then return zero. When we succeed in waiting for @p to be off its CPU,
2236 * we return a positive number (its total switch count). If a second call
2237 * a short while later returns the same number, the caller can be sure that
2238 * @p has remained unscheduled the whole time.
2240 * The caller must ensure that the task *will* unschedule sometime soon,
2241 * else this function might spin for a *long* time. This function can't
2242 * be called with interrupts off, or it may introduce deadlock with
2243 * smp_call_function() if an IPI is sent by the same process we are
2244 * waiting to become inactive.
2246 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2248 unsigned long flags;
2255 * We do the initial early heuristics without holding
2256 * any task-queue locks at all. We'll only try to get
2257 * the runqueue lock when things look like they will
2263 * If the task is actively running on another CPU
2264 * still, just relax and busy-wait without holding
2267 * NOTE! Since we don't hold any locks, it's not
2268 * even sure that "rq" stays as the right runqueue!
2269 * But we don't care, since "task_running()" will
2270 * return false if the runqueue has changed and p
2271 * is actually now running somewhere else!
2273 while (task_running(rq, p)) {
2274 if (match_state && unlikely(p->state != match_state))
2280 * Ok, time to look more closely! We need the rq
2281 * lock now, to be *sure*. If we're wrong, we'll
2282 * just go back and repeat.
2284 rq = task_rq_lock(p, &flags);
2285 trace_sched_wait_task(p);
2286 running = task_running(rq, p);
2287 on_rq = p->se.on_rq;
2289 if (!match_state || p->state == match_state)
2290 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2291 task_rq_unlock(rq, &flags);
2294 * If it changed from the expected state, bail out now.
2296 if (unlikely(!ncsw))
2300 * Was it really running after all now that we
2301 * checked with the proper locks actually held?
2303 * Oops. Go back and try again..
2305 if (unlikely(running)) {
2311 * It's not enough that it's not actively running,
2312 * it must be off the runqueue _entirely_, and not
2315 * So if it was still runnable (but just not actively
2316 * running right now), it's preempted, and we should
2317 * yield - it could be a while.
2319 if (unlikely(on_rq)) {
2320 schedule_timeout_uninterruptible(1);
2325 * Ahh, all good. It wasn't running, and it wasn't
2326 * runnable, which means that it will never become
2327 * running in the future either. We're all done!
2336 * kick_process - kick a running thread to enter/exit the kernel
2337 * @p: the to-be-kicked thread
2339 * Cause a process which is running on another CPU to enter
2340 * kernel-mode, without any delay. (to get signals handled.)
2342 * NOTE: this function doesnt have to take the runqueue lock,
2343 * because all it wants to ensure is that the remote task enters
2344 * the kernel. If the IPI races and the task has been migrated
2345 * to another CPU then no harm is done and the purpose has been
2348 void kick_process(struct task_struct *p)
2354 if ((cpu != smp_processor_id()) && task_curr(p))
2355 smp_send_reschedule(cpu);
2358 EXPORT_SYMBOL_GPL(kick_process);
2359 #endif /* CONFIG_SMP */
2362 * task_oncpu_function_call - call a function on the cpu on which a task runs
2363 * @p: the task to evaluate
2364 * @func: the function to be called
2365 * @info: the function call argument
2367 * Calls the function @func when the task is currently running. This might
2368 * be on the current CPU, which just calls the function directly
2370 void task_oncpu_function_call(struct task_struct *p,
2371 void (*func) (void *info), void *info)
2378 smp_call_function_single(cpu, func, info, 1);
2384 * ->cpus_allowed is protected by either TASK_WAKING or rq->lock held.
2386 static int select_fallback_rq(int cpu, struct task_struct *p)
2389 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2391 /* Look for allowed, online CPU in same node. */
2392 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2393 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2396 /* Any allowed, online CPU? */
2397 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2398 if (dest_cpu < nr_cpu_ids)
2401 /* No more Mr. Nice Guy. */
2402 if (unlikely(dest_cpu >= nr_cpu_ids)) {
2403 dest_cpu = cpuset_cpus_allowed_fallback(p);
2405 * Don't tell them about moving exiting tasks or
2406 * kernel threads (both mm NULL), since they never
2409 if (p->mm && printk_ratelimit()) {
2410 printk(KERN_INFO "process %d (%s) no "
2411 "longer affine to cpu%d\n",
2412 task_pid_nr(p), p->comm, cpu);
2420 * The caller (fork, wakeup) owns TASK_WAKING, ->cpus_allowed is stable.
2423 int select_task_rq(struct rq *rq, struct task_struct *p, int sd_flags, int wake_flags)
2425 int cpu = p->sched_class->select_task_rq(rq, p, sd_flags, wake_flags);
2428 * In order not to call set_task_cpu() on a blocking task we need
2429 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2432 * Since this is common to all placement strategies, this lives here.
2434 * [ this allows ->select_task() to simply return task_cpu(p) and
2435 * not worry about this generic constraint ]
2437 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2439 cpu = select_fallback_rq(task_cpu(p), p);
2444 static void update_avg(u64 *avg, u64 sample)
2446 s64 diff = sample - *avg;
2451 static inline void ttwu_activate(struct task_struct *p, struct rq *rq,
2452 bool is_sync, bool is_migrate, bool is_local,
2453 unsigned long en_flags)
2455 schedstat_inc(p, se.statistics.nr_wakeups);
2457 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2459 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2461 schedstat_inc(p, se.statistics.nr_wakeups_local);
2463 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2465 activate_task(rq, p, en_flags);
2468 static inline void ttwu_post_activation(struct task_struct *p, struct rq *rq,
2469 int wake_flags, bool success)
2471 trace_sched_wakeup(p, success);
2472 check_preempt_curr(rq, p, wake_flags);
2474 p->state = TASK_RUNNING;
2476 if (p->sched_class->task_woken)
2477 p->sched_class->task_woken(rq, p);
2479 if (unlikely(rq->idle_stamp)) {
2480 u64 delta = rq->clock - rq->idle_stamp;
2481 u64 max = 2*sysctl_sched_migration_cost;
2486 update_avg(&rq->avg_idle, delta);
2490 /* if a worker is waking up, notify workqueue */
2491 if ((p->flags & PF_WQ_WORKER) && success)
2492 wq_worker_waking_up(p, cpu_of(rq));
2496 * try_to_wake_up - wake up a thread
2497 * @p: the thread to be awakened
2498 * @state: the mask of task states that can be woken
2499 * @wake_flags: wake modifier flags (WF_*)
2501 * Put it on the run-queue if it's not already there. The "current"
2502 * thread is always on the run-queue (except when the actual
2503 * re-schedule is in progress), and as such you're allowed to do
2504 * the simpler "current->state = TASK_RUNNING" to mark yourself
2505 * runnable without the overhead of this.
2507 * Returns %true if @p was woken up, %false if it was already running
2508 * or @state didn't match @p's state.
2510 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2513 int cpu, orig_cpu, this_cpu, success = 0;
2514 unsigned long flags;
2515 unsigned long en_flags = ENQUEUE_WAKEUP;
2518 this_cpu = get_cpu();
2521 rq = task_rq_lock(p, &flags);
2522 if (!(p->state & state))
2532 if (unlikely(task_running(rq, p)))
2536 * In order to handle concurrent wakeups and release the rq->lock
2537 * we put the task in TASK_WAKING state.
2539 * First fix up the nr_uninterruptible count:
2541 if (task_contributes_to_load(p)) {
2542 if (likely(cpu_online(orig_cpu)))
2543 rq->nr_uninterruptible--;
2545 this_rq()->nr_uninterruptible--;
2547 p->state = TASK_WAKING;
2549 if (p->sched_class->task_waking) {
2550 p->sched_class->task_waking(rq, p);
2551 en_flags |= ENQUEUE_WAKING;
2554 cpu = select_task_rq(rq, p, SD_BALANCE_WAKE, wake_flags);
2555 if (cpu != orig_cpu)
2556 set_task_cpu(p, cpu);
2557 __task_rq_unlock(rq);
2560 raw_spin_lock(&rq->lock);
2563 * We migrated the task without holding either rq->lock, however
2564 * since the task is not on the task list itself, nobody else
2565 * will try and migrate the task, hence the rq should match the
2566 * cpu we just moved it to.
2568 WARN_ON(task_cpu(p) != cpu);
2569 WARN_ON(p->state != TASK_WAKING);
2571 #ifdef CONFIG_SCHEDSTATS
2572 schedstat_inc(rq, ttwu_count);
2573 if (cpu == this_cpu)
2574 schedstat_inc(rq, ttwu_local);
2576 struct sched_domain *sd;
2577 for_each_domain(this_cpu, sd) {
2578 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2579 schedstat_inc(sd, ttwu_wake_remote);
2584 #endif /* CONFIG_SCHEDSTATS */
2587 #endif /* CONFIG_SMP */
2588 ttwu_activate(p, rq, wake_flags & WF_SYNC, orig_cpu != cpu,
2589 cpu == this_cpu, en_flags);
2592 ttwu_post_activation(p, rq, wake_flags, success);
2594 task_rq_unlock(rq, &flags);
2601 * try_to_wake_up_local - try to wake up a local task with rq lock held
2602 * @p: the thread to be awakened
2604 * Put @p on the run-queue if it's not alredy there. The caller must
2605 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2606 * the current task. this_rq() stays locked over invocation.
2608 static void try_to_wake_up_local(struct task_struct *p)
2610 struct rq *rq = task_rq(p);
2611 bool success = false;
2613 BUG_ON(rq != this_rq());
2614 BUG_ON(p == current);
2615 lockdep_assert_held(&rq->lock);
2617 if (!(p->state & TASK_NORMAL))
2621 if (likely(!task_running(rq, p))) {
2622 schedstat_inc(rq, ttwu_count);
2623 schedstat_inc(rq, ttwu_local);
2625 ttwu_activate(p, rq, false, false, true, ENQUEUE_WAKEUP);
2628 ttwu_post_activation(p, rq, 0, success);
2632 * wake_up_process - Wake up a specific process
2633 * @p: The process to be woken up.
2635 * Attempt to wake up the nominated process and move it to the set of runnable
2636 * processes. Returns 1 if the process was woken up, 0 if it was already
2639 * It may be assumed that this function implies a write memory barrier before
2640 * changing the task state if and only if any tasks are woken up.
2642 int wake_up_process(struct task_struct *p)
2644 return try_to_wake_up(p, TASK_ALL, 0);
2646 EXPORT_SYMBOL(wake_up_process);
2648 int wake_up_state(struct task_struct *p, unsigned int state)
2650 return try_to_wake_up(p, state, 0);
2654 * Perform scheduler related setup for a newly forked process p.
2655 * p is forked by current.
2657 * __sched_fork() is basic setup used by init_idle() too:
2659 static void __sched_fork(struct task_struct *p)
2661 p->se.exec_start = 0;
2662 p->se.sum_exec_runtime = 0;
2663 p->se.prev_sum_exec_runtime = 0;
2664 p->se.nr_migrations = 0;
2666 #ifdef CONFIG_SCHEDSTATS
2667 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2670 INIT_LIST_HEAD(&p->rt.run_list);
2672 INIT_LIST_HEAD(&p->se.group_node);
2674 #ifdef CONFIG_PREEMPT_NOTIFIERS
2675 INIT_HLIST_HEAD(&p->preempt_notifiers);
2680 * fork()/clone()-time setup:
2682 void sched_fork(struct task_struct *p, int clone_flags)
2684 int cpu = get_cpu();
2688 * We mark the process as running here. This guarantees that
2689 * nobody will actually run it, and a signal or other external
2690 * event cannot wake it up and insert it on the runqueue either.
2692 p->state = TASK_RUNNING;
2695 * Revert to default priority/policy on fork if requested.
2697 if (unlikely(p->sched_reset_on_fork)) {
2698 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2699 p->policy = SCHED_NORMAL;
2700 p->normal_prio = p->static_prio;
2703 if (PRIO_TO_NICE(p->static_prio) < 0) {
2704 p->static_prio = NICE_TO_PRIO(0);
2705 p->normal_prio = p->static_prio;
2710 * We don't need the reset flag anymore after the fork. It has
2711 * fulfilled its duty:
2713 p->sched_reset_on_fork = 0;
2717 * Make sure we do not leak PI boosting priority to the child.
2719 p->prio = current->normal_prio;
2721 if (!rt_prio(p->prio))
2722 p->sched_class = &fair_sched_class;
2724 if (p->sched_class->task_fork)
2725 p->sched_class->task_fork(p);
2728 * The child is not yet in the pid-hash so no cgroup attach races,
2729 * and the cgroup is pinned to this child due to cgroup_fork()
2730 * is ran before sched_fork().
2732 * Silence PROVE_RCU.
2735 set_task_cpu(p, cpu);
2738 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2739 if (likely(sched_info_on()))
2740 memset(&p->sched_info, 0, sizeof(p->sched_info));
2742 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2745 #ifdef CONFIG_PREEMPT
2746 /* Want to start with kernel preemption disabled. */
2747 task_thread_info(p)->preempt_count = 1;
2749 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2755 * wake_up_new_task - wake up a newly created task for the first time.
2757 * This function will do some initial scheduler statistics housekeeping
2758 * that must be done for every newly created context, then puts the task
2759 * on the runqueue and wakes it.
2761 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2763 unsigned long flags;
2765 int cpu __maybe_unused = get_cpu();
2768 rq = task_rq_lock(p, &flags);
2769 p->state = TASK_WAKING;
2772 * Fork balancing, do it here and not earlier because:
2773 * - cpus_allowed can change in the fork path
2774 * - any previously selected cpu might disappear through hotplug
2776 * We set TASK_WAKING so that select_task_rq() can drop rq->lock
2777 * without people poking at ->cpus_allowed.
2779 cpu = select_task_rq(rq, p, SD_BALANCE_FORK, 0);
2780 set_task_cpu(p, cpu);
2782 p->state = TASK_RUNNING;
2783 task_rq_unlock(rq, &flags);
2786 rq = task_rq_lock(p, &flags);
2787 activate_task(rq, p, 0);
2788 trace_sched_wakeup_new(p, 1);
2789 check_preempt_curr(rq, p, WF_FORK);
2791 if (p->sched_class->task_woken)
2792 p->sched_class->task_woken(rq, p);
2794 task_rq_unlock(rq, &flags);
2798 #ifdef CONFIG_PREEMPT_NOTIFIERS
2801 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2802 * @notifier: notifier struct to register
2804 void preempt_notifier_register(struct preempt_notifier *notifier)
2806 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2808 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2811 * preempt_notifier_unregister - no longer interested in preemption notifications
2812 * @notifier: notifier struct to unregister
2814 * This is safe to call from within a preemption notifier.
2816 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2818 hlist_del(¬ifier->link);
2820 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2822 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2824 struct preempt_notifier *notifier;
2825 struct hlist_node *node;
2827 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2828 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2832 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2833 struct task_struct *next)
2835 struct preempt_notifier *notifier;
2836 struct hlist_node *node;
2838 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2839 notifier->ops->sched_out(notifier, next);
2842 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2844 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2849 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2850 struct task_struct *next)
2854 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2857 * prepare_task_switch - prepare to switch tasks
2858 * @rq: the runqueue preparing to switch
2859 * @prev: the current task that is being switched out
2860 * @next: the task we are going to switch to.
2862 * This is called with the rq lock held and interrupts off. It must
2863 * be paired with a subsequent finish_task_switch after the context
2866 * prepare_task_switch sets up locking and calls architecture specific
2870 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2871 struct task_struct *next)
2873 fire_sched_out_preempt_notifiers(prev, next);
2874 prepare_lock_switch(rq, next);
2875 prepare_arch_switch(next);
2879 * finish_task_switch - clean up after a task-switch
2880 * @rq: runqueue associated with task-switch
2881 * @prev: the thread we just switched away from.
2883 * finish_task_switch must be called after the context switch, paired
2884 * with a prepare_task_switch call before the context switch.
2885 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2886 * and do any other architecture-specific cleanup actions.
2888 * Note that we may have delayed dropping an mm in context_switch(). If
2889 * so, we finish that here outside of the runqueue lock. (Doing it
2890 * with the lock held can cause deadlocks; see schedule() for
2893 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2894 __releases(rq->lock)
2896 struct mm_struct *mm = rq->prev_mm;
2902 * A task struct has one reference for the use as "current".
2903 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2904 * schedule one last time. The schedule call will never return, and
2905 * the scheduled task must drop that reference.
2906 * The test for TASK_DEAD must occur while the runqueue locks are
2907 * still held, otherwise prev could be scheduled on another cpu, die
2908 * there before we look at prev->state, and then the reference would
2910 * Manfred Spraul <manfred@colorfullife.com>
2912 prev_state = prev->state;
2913 finish_arch_switch(prev);
2914 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2915 local_irq_disable();
2916 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2917 perf_event_task_sched_in(current);
2918 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2920 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2921 finish_lock_switch(rq, prev);
2923 fire_sched_in_preempt_notifiers(current);
2926 if (unlikely(prev_state == TASK_DEAD)) {
2928 * Remove function-return probe instances associated with this
2929 * task and put them back on the free list.
2931 kprobe_flush_task(prev);
2932 put_task_struct(prev);
2938 /* assumes rq->lock is held */
2939 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2941 if (prev->sched_class->pre_schedule)
2942 prev->sched_class->pre_schedule(rq, prev);
2945 /* rq->lock is NOT held, but preemption is disabled */
2946 static inline void post_schedule(struct rq *rq)
2948 if (rq->post_schedule) {
2949 unsigned long flags;
2951 raw_spin_lock_irqsave(&rq->lock, flags);
2952 if (rq->curr->sched_class->post_schedule)
2953 rq->curr->sched_class->post_schedule(rq);
2954 raw_spin_unlock_irqrestore(&rq->lock, flags);
2956 rq->post_schedule = 0;
2962 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2966 static inline void post_schedule(struct rq *rq)
2973 * schedule_tail - first thing a freshly forked thread must call.
2974 * @prev: the thread we just switched away from.
2976 asmlinkage void schedule_tail(struct task_struct *prev)
2977 __releases(rq->lock)
2979 struct rq *rq = this_rq();
2981 finish_task_switch(rq, prev);
2984 * FIXME: do we need to worry about rq being invalidated by the
2989 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2990 /* In this case, finish_task_switch does not reenable preemption */
2993 if (current->set_child_tid)
2994 put_user(task_pid_vnr(current), current->set_child_tid);
2998 * context_switch - switch to the new MM and the new
2999 * thread's register state.
3002 context_switch(struct rq *rq, struct task_struct *prev,
3003 struct task_struct *next)
3005 struct mm_struct *mm, *oldmm;
3007 prepare_task_switch(rq, prev, next);
3008 trace_sched_switch(prev, next);
3010 oldmm = prev->active_mm;
3012 * For paravirt, this is coupled with an exit in switch_to to
3013 * combine the page table reload and the switch backend into
3016 arch_start_context_switch(prev);
3019 next->active_mm = oldmm;
3020 atomic_inc(&oldmm->mm_count);
3021 enter_lazy_tlb(oldmm, next);
3023 switch_mm(oldmm, mm, next);
3026 prev->active_mm = NULL;
3027 rq->prev_mm = oldmm;
3030 * Since the runqueue lock will be released by the next
3031 * task (which is an invalid locking op but in the case
3032 * of the scheduler it's an obvious special-case), so we
3033 * do an early lockdep release here:
3035 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
3036 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3039 /* Here we just switch the register state and the stack. */
3040 switch_to(prev, next, prev);
3044 * this_rq must be evaluated again because prev may have moved
3045 * CPUs since it called schedule(), thus the 'rq' on its stack
3046 * frame will be invalid.
3048 finish_task_switch(this_rq(), prev);
3052 * nr_running, nr_uninterruptible and nr_context_switches:
3054 * externally visible scheduler statistics: current number of runnable
3055 * threads, current number of uninterruptible-sleeping threads, total
3056 * number of context switches performed since bootup.
3058 unsigned long nr_running(void)
3060 unsigned long i, sum = 0;
3062 for_each_online_cpu(i)
3063 sum += cpu_rq(i)->nr_running;
3068 unsigned long nr_uninterruptible(void)
3070 unsigned long i, sum = 0;
3072 for_each_possible_cpu(i)
3073 sum += cpu_rq(i)->nr_uninterruptible;
3076 * Since we read the counters lockless, it might be slightly
3077 * inaccurate. Do not allow it to go below zero though:
3079 if (unlikely((long)sum < 0))
3085 unsigned long long nr_context_switches(void)
3088 unsigned long long sum = 0;
3090 for_each_possible_cpu(i)
3091 sum += cpu_rq(i)->nr_switches;
3096 unsigned long nr_iowait(void)
3098 unsigned long i, sum = 0;
3100 for_each_possible_cpu(i)
3101 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3106 unsigned long nr_iowait_cpu(int cpu)
3108 struct rq *this = cpu_rq(cpu);
3109 return atomic_read(&this->nr_iowait);
3112 unsigned long this_cpu_load(void)
3114 struct rq *this = this_rq();
3115 return this->cpu_load[0];
3119 /* Variables and functions for calc_load */
3120 static atomic_long_t calc_load_tasks;
3121 static unsigned long calc_load_update;
3122 unsigned long avenrun[3];
3123 EXPORT_SYMBOL(avenrun);
3125 static long calc_load_fold_active(struct rq *this_rq)
3127 long nr_active, delta = 0;
3129 nr_active = this_rq->nr_running;
3130 nr_active += (long) this_rq->nr_uninterruptible;
3132 if (nr_active != this_rq->calc_load_active) {
3133 delta = nr_active - this_rq->calc_load_active;
3134 this_rq->calc_load_active = nr_active;
3140 static unsigned long
3141 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3144 load += active * (FIXED_1 - exp);
3145 load += 1UL << (FSHIFT - 1);
3146 return load >> FSHIFT;
3151 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
3153 * When making the ILB scale, we should try to pull this in as well.
3155 static atomic_long_t calc_load_tasks_idle;
3157 static void calc_load_account_idle(struct rq *this_rq)
3161 delta = calc_load_fold_active(this_rq);
3163 atomic_long_add(delta, &calc_load_tasks_idle);
3166 static long calc_load_fold_idle(void)
3171 * Its got a race, we don't care...
3173 if (atomic_long_read(&calc_load_tasks_idle))
3174 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
3180 * fixed_power_int - compute: x^n, in O(log n) time
3182 * @x: base of the power
3183 * @frac_bits: fractional bits of @x
3184 * @n: power to raise @x to.
3186 * By exploiting the relation between the definition of the natural power
3187 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
3188 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
3189 * (where: n_i \elem {0, 1}, the binary vector representing n),
3190 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
3191 * of course trivially computable in O(log_2 n), the length of our binary
3194 static unsigned long
3195 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
3197 unsigned long result = 1UL << frac_bits;
3202 result += 1UL << (frac_bits - 1);
3203 result >>= frac_bits;
3209 x += 1UL << (frac_bits - 1);
3217 * a1 = a0 * e + a * (1 - e)
3219 * a2 = a1 * e + a * (1 - e)
3220 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
3221 * = a0 * e^2 + a * (1 - e) * (1 + e)
3223 * a3 = a2 * e + a * (1 - e)
3224 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
3225 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
3229 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
3230 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
3231 * = a0 * e^n + a * (1 - e^n)
3233 * [1] application of the geometric series:
3236 * S_n := \Sum x^i = -------------
3239 static unsigned long
3240 calc_load_n(unsigned long load, unsigned long exp,
3241 unsigned long active, unsigned int n)
3244 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
3248 * NO_HZ can leave us missing all per-cpu ticks calling
3249 * calc_load_account_active(), but since an idle CPU folds its delta into
3250 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
3251 * in the pending idle delta if our idle period crossed a load cycle boundary.
3253 * Once we've updated the global active value, we need to apply the exponential
3254 * weights adjusted to the number of cycles missed.
3256 static void calc_global_nohz(unsigned long ticks)
3258 long delta, active, n;
3260 if (time_before(jiffies, calc_load_update))
3264 * If we crossed a calc_load_update boundary, make sure to fold
3265 * any pending idle changes, the respective CPUs might have
3266 * missed the tick driven calc_load_account_active() update
3269 delta = calc_load_fold_idle();
3271 atomic_long_add(delta, &calc_load_tasks);
3274 * If we were idle for multiple load cycles, apply them.
3276 if (ticks >= LOAD_FREQ) {
3277 n = ticks / LOAD_FREQ;
3279 active = atomic_long_read(&calc_load_tasks);
3280 active = active > 0 ? active * FIXED_1 : 0;
3282 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
3283 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
3284 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
3286 calc_load_update += n * LOAD_FREQ;
3290 * Its possible the remainder of the above division also crosses
3291 * a LOAD_FREQ period, the regular check in calc_global_load()
3292 * which comes after this will take care of that.
3294 * Consider us being 11 ticks before a cycle completion, and us
3295 * sleeping for 4*LOAD_FREQ + 22 ticks, then the above code will
3296 * age us 4 cycles, and the test in calc_global_load() will
3297 * pick up the final one.
3301 static void calc_load_account_idle(struct rq *this_rq)
3305 static inline long calc_load_fold_idle(void)
3310 static void calc_global_nohz(unsigned long ticks)
3316 * get_avenrun - get the load average array
3317 * @loads: pointer to dest load array
3318 * @offset: offset to add
3319 * @shift: shift count to shift the result left
3321 * These values are estimates at best, so no need for locking.
3323 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3325 loads[0] = (avenrun[0] + offset) << shift;
3326 loads[1] = (avenrun[1] + offset) << shift;
3327 loads[2] = (avenrun[2] + offset) << shift;
3331 * calc_load - update the avenrun load estimates 10 ticks after the
3332 * CPUs have updated calc_load_tasks.
3334 void calc_global_load(unsigned long ticks)
3338 calc_global_nohz(ticks);
3340 if (time_before(jiffies, calc_load_update + 10))
3343 active = atomic_long_read(&calc_load_tasks);
3344 active = active > 0 ? active * FIXED_1 : 0;
3346 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3347 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3348 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3350 calc_load_update += LOAD_FREQ;
3354 * Called from update_cpu_load() to periodically update this CPU's
3357 static void calc_load_account_active(struct rq *this_rq)
3361 if (time_before(jiffies, this_rq->calc_load_update))
3364 delta = calc_load_fold_active(this_rq);
3365 delta += calc_load_fold_idle();
3367 atomic_long_add(delta, &calc_load_tasks);
3369 this_rq->calc_load_update += LOAD_FREQ;
3373 * The exact cpuload at various idx values, calculated at every tick would be
3374 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3376 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3377 * on nth tick when cpu may be busy, then we have:
3378 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3379 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3381 * decay_load_missed() below does efficient calculation of
3382 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3383 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3385 * The calculation is approximated on a 128 point scale.
3386 * degrade_zero_ticks is the number of ticks after which load at any
3387 * particular idx is approximated to be zero.
3388 * degrade_factor is a precomputed table, a row for each load idx.
3389 * Each column corresponds to degradation factor for a power of two ticks,
3390 * based on 128 point scale.
3392 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3393 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3395 * With this power of 2 load factors, we can degrade the load n times
3396 * by looking at 1 bits in n and doing as many mult/shift instead of
3397 * n mult/shifts needed by the exact degradation.
3399 #define DEGRADE_SHIFT 7
3400 static const unsigned char
3401 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3402 static const unsigned char
3403 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3404 {0, 0, 0, 0, 0, 0, 0, 0},
3405 {64, 32, 8, 0, 0, 0, 0, 0},
3406 {96, 72, 40, 12, 1, 0, 0},
3407 {112, 98, 75, 43, 15, 1, 0},
3408 {120, 112, 98, 76, 45, 16, 2} };
3411 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3412 * would be when CPU is idle and so we just decay the old load without
3413 * adding any new load.
3415 static unsigned long
3416 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3420 if (!missed_updates)
3423 if (missed_updates >= degrade_zero_ticks[idx])
3427 return load >> missed_updates;
3429 while (missed_updates) {
3430 if (missed_updates % 2)
3431 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3433 missed_updates >>= 1;
3440 * Update rq->cpu_load[] statistics. This function is usually called every
3441 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3442 * every tick. We fix it up based on jiffies.
3444 static void update_cpu_load(struct rq *this_rq)
3446 unsigned long this_load = this_rq->load.weight;
3447 unsigned long curr_jiffies = jiffies;
3448 unsigned long pending_updates;
3451 this_rq->nr_load_updates++;
3453 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3454 if (curr_jiffies == this_rq->last_load_update_tick)
3457 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3458 this_rq->last_load_update_tick = curr_jiffies;
3460 /* Update our load: */
3461 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3462 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3463 unsigned long old_load, new_load;
3465 /* scale is effectively 1 << i now, and >> i divides by scale */
3467 old_load = this_rq->cpu_load[i];
3468 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3469 new_load = this_load;
3471 * Round up the averaging division if load is increasing. This
3472 * prevents us from getting stuck on 9 if the load is 10, for
3475 if (new_load > old_load)
3476 new_load += scale - 1;
3478 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3481 sched_avg_update(this_rq);
3484 static void update_cpu_load_active(struct rq *this_rq)
3486 update_cpu_load(this_rq);
3488 calc_load_account_active(this_rq);
3494 * sched_exec - execve() is a valuable balancing opportunity, because at
3495 * this point the task has the smallest effective memory and cache footprint.
3497 void sched_exec(void)
3499 struct task_struct *p = current;
3500 unsigned long flags;
3504 rq = task_rq_lock(p, &flags);
3505 dest_cpu = p->sched_class->select_task_rq(rq, p, SD_BALANCE_EXEC, 0);
3506 if (dest_cpu == smp_processor_id())
3510 * select_task_rq() can race against ->cpus_allowed
3512 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed) &&
3513 likely(cpu_active(dest_cpu)) && migrate_task(p, dest_cpu)) {
3514 struct migration_arg arg = { p, dest_cpu };
3516 task_rq_unlock(rq, &flags);
3517 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
3521 task_rq_unlock(rq, &flags);
3526 DEFINE_PER_CPU(struct kernel_stat, kstat);
3528 EXPORT_PER_CPU_SYMBOL(kstat);
3531 * Return any ns on the sched_clock that have not yet been accounted in
3532 * @p in case that task is currently running.
3534 * Called with task_rq_lock() held on @rq.
3536 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3540 if (task_current(rq, p)) {
3541 update_rq_clock(rq);
3542 ns = rq->clock_task - p->se.exec_start;
3550 unsigned long long task_delta_exec(struct task_struct *p)
3552 unsigned long flags;
3556 rq = task_rq_lock(p, &flags);
3557 ns = do_task_delta_exec(p, rq);
3558 task_rq_unlock(rq, &flags);
3564 * Return accounted runtime for the task.
3565 * In case the task is currently running, return the runtime plus current's
3566 * pending runtime that have not been accounted yet.
3568 unsigned long long task_sched_runtime(struct task_struct *p)
3570 unsigned long flags;
3574 rq = task_rq_lock(p, &flags);
3575 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3576 task_rq_unlock(rq, &flags);
3582 * Return sum_exec_runtime for the thread group.
3583 * In case the task is currently running, return the sum plus current's
3584 * pending runtime that have not been accounted yet.
3586 * Note that the thread group might have other running tasks as well,
3587 * so the return value not includes other pending runtime that other
3588 * running tasks might have.
3590 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3592 struct task_cputime totals;
3593 unsigned long flags;
3597 rq = task_rq_lock(p, &flags);
3598 thread_group_cputime(p, &totals);
3599 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3600 task_rq_unlock(rq, &flags);
3606 * Account user cpu time to a process.
3607 * @p: the process that the cpu time gets accounted to
3608 * @cputime: the cpu time spent in user space since the last update
3609 * @cputime_scaled: cputime scaled by cpu frequency
3611 void account_user_time(struct task_struct *p, cputime_t cputime,
3612 cputime_t cputime_scaled)
3614 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3617 /* Add user time to process. */
3618 p->utime = cputime_add(p->utime, cputime);
3619 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3620 account_group_user_time(p, cputime);
3622 /* Add user time to cpustat. */
3623 tmp = cputime_to_cputime64(cputime);
3624 if (TASK_NICE(p) > 0)
3625 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3627 cpustat->user = cputime64_add(cpustat->user, tmp);
3629 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3630 /* Account for user time used */
3631 acct_update_integrals(p);
3635 * Account guest cpu time to a process.
3636 * @p: the process that the cpu time gets accounted to
3637 * @cputime: the cpu time spent in virtual machine since the last update
3638 * @cputime_scaled: cputime scaled by cpu frequency
3640 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3641 cputime_t cputime_scaled)
3644 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3646 tmp = cputime_to_cputime64(cputime);
3648 /* Add guest time to process. */
3649 p->utime = cputime_add(p->utime, cputime);
3650 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3651 account_group_user_time(p, cputime);
3652 p->gtime = cputime_add(p->gtime, cputime);
3654 /* Add guest time to cpustat. */
3655 if (TASK_NICE(p) > 0) {
3656 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3657 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3659 cpustat->user = cputime64_add(cpustat->user, tmp);
3660 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3665 * Account system cpu time to a process.
3666 * @p: the process that the cpu time gets accounted to
3667 * @hardirq_offset: the offset to subtract from hardirq_count()
3668 * @cputime: the cpu time spent in kernel space since the last update
3669 * @cputime_scaled: cputime scaled by cpu frequency
3671 void account_system_time(struct task_struct *p, int hardirq_offset,
3672 cputime_t cputime, cputime_t cputime_scaled)
3674 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3677 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3678 account_guest_time(p, cputime, cputime_scaled);
3682 /* Add system time to process. */
3683 p->stime = cputime_add(p->stime, cputime);
3684 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3685 account_group_system_time(p, cputime);
3687 /* Add system time to cpustat. */
3688 tmp = cputime_to_cputime64(cputime);
3689 if (hardirq_count() - hardirq_offset)
3690 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3691 else if (in_serving_softirq())
3692 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3694 cpustat->system = cputime64_add(cpustat->system, tmp);
3696 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3698 /* Account for system time used */
3699 acct_update_integrals(p);
3703 * Account for involuntary wait time.
3704 * @steal: the cpu time spent in involuntary wait
3706 void account_steal_time(cputime_t cputime)
3708 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3709 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3711 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3715 * Account for idle time.
3716 * @cputime: the cpu time spent in idle wait
3718 void account_idle_time(cputime_t cputime)
3720 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3721 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3722 struct rq *rq = this_rq();
3724 if (atomic_read(&rq->nr_iowait) > 0)
3725 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3727 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3730 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3733 * Account a single tick of cpu time.
3734 * @p: the process that the cpu time gets accounted to
3735 * @user_tick: indicates if the tick is a user or a system tick
3737 void account_process_tick(struct task_struct *p, int user_tick)
3739 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3740 struct rq *rq = this_rq();
3743 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3744 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3745 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3748 account_idle_time(cputime_one_jiffy);
3752 * Account multiple ticks of steal time.
3753 * @p: the process from which the cpu time has been stolen
3754 * @ticks: number of stolen ticks
3756 void account_steal_ticks(unsigned long ticks)
3758 account_steal_time(jiffies_to_cputime(ticks));
3762 * Account multiple ticks of idle time.
3763 * @ticks: number of stolen ticks
3765 void account_idle_ticks(unsigned long ticks)
3767 account_idle_time(jiffies_to_cputime(ticks));
3773 * Use precise platform statistics if available:
3775 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3776 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3782 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3784 struct task_cputime cputime;
3786 thread_group_cputime(p, &cputime);
3788 *ut = cputime.utime;
3789 *st = cputime.stime;
3793 #ifndef nsecs_to_cputime
3794 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3797 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3799 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
3802 * Use CFS's precise accounting:
3804 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3810 do_div(temp, total);
3811 utime = (cputime_t)temp;
3816 * Compare with previous values, to keep monotonicity:
3818 p->prev_utime = max(p->prev_utime, utime);
3819 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
3821 *ut = p->prev_utime;
3822 *st = p->prev_stime;
3826 * Must be called with siglock held.
3828 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3830 struct signal_struct *sig = p->signal;
3831 struct task_cputime cputime;
3832 cputime_t rtime, utime, total;
3834 thread_group_cputime(p, &cputime);
3836 total = cputime_add(cputime.utime, cputime.stime);
3837 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3842 temp *= cputime.utime;
3843 do_div(temp, total);
3844 utime = (cputime_t)temp;
3848 sig->prev_utime = max(sig->prev_utime, utime);
3849 sig->prev_stime = max(sig->prev_stime,
3850 cputime_sub(rtime, sig->prev_utime));
3852 *ut = sig->prev_utime;
3853 *st = sig->prev_stime;
3858 * This function gets called by the timer code, with HZ frequency.
3859 * We call it with interrupts disabled.
3861 * It also gets called by the fork code, when changing the parent's
3864 void scheduler_tick(void)
3866 int cpu = smp_processor_id();
3867 struct rq *rq = cpu_rq(cpu);
3868 struct task_struct *curr = rq->curr;
3872 raw_spin_lock(&rq->lock);
3873 update_rq_clock(rq);
3874 update_cpu_load_active(rq);
3875 curr->sched_class->task_tick(rq, curr, 0);
3876 raw_spin_unlock(&rq->lock);
3878 perf_event_task_tick();
3881 rq->idle_at_tick = idle_cpu(cpu);
3882 trigger_load_balance(rq, cpu);
3886 notrace unsigned long get_parent_ip(unsigned long addr)
3888 if (in_lock_functions(addr)) {
3889 addr = CALLER_ADDR2;
3890 if (in_lock_functions(addr))
3891 addr = CALLER_ADDR3;
3896 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3897 defined(CONFIG_PREEMPT_TRACER))
3899 void __kprobes add_preempt_count(int val)
3901 #ifdef CONFIG_DEBUG_PREEMPT
3905 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3908 preempt_count() += val;
3909 #ifdef CONFIG_DEBUG_PREEMPT
3911 * Spinlock count overflowing soon?
3913 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3916 if (preempt_count() == val)
3917 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3919 EXPORT_SYMBOL(add_preempt_count);
3921 void __kprobes sub_preempt_count(int val)
3923 #ifdef CONFIG_DEBUG_PREEMPT
3927 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3930 * Is the spinlock portion underflowing?
3932 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3933 !(preempt_count() & PREEMPT_MASK)))
3937 if (preempt_count() == val)
3938 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3939 preempt_count() -= val;
3941 EXPORT_SYMBOL(sub_preempt_count);
3946 * Print scheduling while atomic bug:
3948 static noinline void __schedule_bug(struct task_struct *prev)
3950 struct pt_regs *regs = get_irq_regs();
3952 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3953 prev->comm, prev->pid, preempt_count());
3955 debug_show_held_locks(prev);
3957 if (irqs_disabled())
3958 print_irqtrace_events(prev);
3967 * Various schedule()-time debugging checks and statistics:
3969 static inline void schedule_debug(struct task_struct *prev)
3972 * Test if we are atomic. Since do_exit() needs to call into
3973 * schedule() atomically, we ignore that path for now.
3974 * Otherwise, whine if we are scheduling when we should not be.
3976 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3977 __schedule_bug(prev);
3979 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3981 schedstat_inc(this_rq(), sched_count);
3982 #ifdef CONFIG_SCHEDSTATS
3983 if (unlikely(prev->lock_depth >= 0)) {
3984 schedstat_inc(this_rq(), bkl_count);
3985 schedstat_inc(prev, sched_info.bkl_count);
3990 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3993 update_rq_clock(rq);
3994 prev->sched_class->put_prev_task(rq, prev);
3998 * Pick up the highest-prio task:
4000 static inline struct task_struct *
4001 pick_next_task(struct rq *rq)
4003 const struct sched_class *class;
4004 struct task_struct *p;
4007 * Optimization: we know that if all tasks are in
4008 * the fair class we can call that function directly:
4010 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4011 p = fair_sched_class.pick_next_task(rq);
4016 for_each_class(class) {
4017 p = class->pick_next_task(rq);
4022 BUG(); /* the idle class will always have a runnable task */
4026 * schedule() is the main scheduler function.
4028 asmlinkage void __sched schedule(void)
4030 struct task_struct *prev, *next;
4031 unsigned long *switch_count;
4037 cpu = smp_processor_id();
4039 rcu_note_context_switch(cpu);
4042 release_kernel_lock(prev);
4043 need_resched_nonpreemptible:
4045 schedule_debug(prev);
4047 if (sched_feat(HRTICK))
4050 raw_spin_lock_irq(&rq->lock);
4052 switch_count = &prev->nivcsw;
4053 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4054 if (unlikely(signal_pending_state(prev->state, prev))) {
4055 prev->state = TASK_RUNNING;
4058 * If a worker is going to sleep, notify and
4059 * ask workqueue whether it wants to wake up a
4060 * task to maintain concurrency. If so, wake
4063 if (prev->flags & PF_WQ_WORKER) {
4064 struct task_struct *to_wakeup;
4066 to_wakeup = wq_worker_sleeping(prev, cpu);
4068 try_to_wake_up_local(to_wakeup);
4070 deactivate_task(rq, prev, DEQUEUE_SLEEP);
4072 switch_count = &prev->nvcsw;
4075 pre_schedule(rq, prev);
4077 if (unlikely(!rq->nr_running))
4078 idle_balance(cpu, rq);
4080 put_prev_task(rq, prev);
4081 next = pick_next_task(rq);
4082 clear_tsk_need_resched(prev);
4083 rq->skip_clock_update = 0;
4085 if (likely(prev != next)) {
4086 sched_info_switch(prev, next);
4087 perf_event_task_sched_out(prev, next);
4092 WARN_ON_ONCE(test_tsk_need_resched(next));
4094 context_switch(rq, prev, next); /* unlocks the rq */
4096 * The context switch have flipped the stack from under us
4097 * and restored the local variables which were saved when
4098 * this task called schedule() in the past. prev == current
4099 * is still correct, but it can be moved to another cpu/rq.
4101 cpu = smp_processor_id();
4104 raw_spin_unlock_irq(&rq->lock);
4108 if (unlikely(reacquire_kernel_lock(prev)))
4109 goto need_resched_nonpreemptible;
4111 preempt_enable_no_resched();
4115 EXPORT_SYMBOL(schedule);
4117 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
4119 * Look out! "owner" is an entirely speculative pointer
4120 * access and not reliable.
4122 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
4127 if (!sched_feat(OWNER_SPIN))
4130 #ifdef CONFIG_DEBUG_PAGEALLOC
4132 * Need to access the cpu field knowing that
4133 * DEBUG_PAGEALLOC could have unmapped it if
4134 * the mutex owner just released it and exited.
4136 if (probe_kernel_address(&owner->cpu, cpu))
4143 * Even if the access succeeded (likely case),
4144 * the cpu field may no longer be valid.
4146 if (cpu >= nr_cpumask_bits)
4150 * We need to validate that we can do a
4151 * get_cpu() and that we have the percpu area.
4153 if (!cpu_online(cpu))
4160 * Owner changed, break to re-assess state.
4162 if (lock->owner != owner) {
4164 * If the lock has switched to a different owner,
4165 * we likely have heavy contention. Return 0 to quit
4166 * optimistic spinning and not contend further:
4174 * Is that owner really running on that cpu?
4176 if (task_thread_info(rq->curr) != owner || need_resched())
4186 #ifdef CONFIG_PREEMPT
4188 * this is the entry point to schedule() from in-kernel preemption
4189 * off of preempt_enable. Kernel preemptions off return from interrupt
4190 * occur there and call schedule directly.
4192 asmlinkage void __sched notrace preempt_schedule(void)
4194 struct thread_info *ti = current_thread_info();
4197 * If there is a non-zero preempt_count or interrupts are disabled,
4198 * we do not want to preempt the current task. Just return..
4200 if (likely(ti->preempt_count || irqs_disabled()))
4204 add_preempt_count_notrace(PREEMPT_ACTIVE);
4206 sub_preempt_count_notrace(PREEMPT_ACTIVE);
4209 * Check again in case we missed a preemption opportunity
4210 * between schedule and now.
4213 } while (need_resched());
4215 EXPORT_SYMBOL(preempt_schedule);
4218 * this is the entry point to schedule() from kernel preemption
4219 * off of irq context.
4220 * Note, that this is called and return with irqs disabled. This will
4221 * protect us against recursive calling from irq.
4223 asmlinkage void __sched preempt_schedule_irq(void)
4225 struct thread_info *ti = current_thread_info();
4227 /* Catch callers which need to be fixed */
4228 BUG_ON(ti->preempt_count || !irqs_disabled());
4231 add_preempt_count(PREEMPT_ACTIVE);
4234 local_irq_disable();
4235 sub_preempt_count(PREEMPT_ACTIVE);
4238 * Check again in case we missed a preemption opportunity
4239 * between schedule and now.
4242 } while (need_resched());
4245 #endif /* CONFIG_PREEMPT */
4247 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
4250 return try_to_wake_up(curr->private, mode, wake_flags);
4252 EXPORT_SYMBOL(default_wake_function);
4255 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4256 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4257 * number) then we wake all the non-exclusive tasks and one exclusive task.
4259 * There are circumstances in which we can try to wake a task which has already
4260 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4261 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4263 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4264 int nr_exclusive, int wake_flags, void *key)
4266 wait_queue_t *curr, *next;
4268 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4269 unsigned flags = curr->flags;
4271 if (curr->func(curr, mode, wake_flags, key) &&
4272 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4278 * __wake_up - wake up threads blocked on a waitqueue.
4280 * @mode: which threads
4281 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4282 * @key: is directly passed to the wakeup function
4284 * It may be assumed that this function implies a write memory barrier before
4285 * changing the task state if and only if any tasks are woken up.
4287 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4288 int nr_exclusive, void *key)
4290 unsigned long flags;
4292 spin_lock_irqsave(&q->lock, flags);
4293 __wake_up_common(q, mode, nr_exclusive, 0, key);
4294 spin_unlock_irqrestore(&q->lock, flags);
4296 EXPORT_SYMBOL(__wake_up);
4299 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4301 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4303 __wake_up_common(q, mode, 1, 0, NULL);
4305 EXPORT_SYMBOL_GPL(__wake_up_locked);
4307 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
4309 __wake_up_common(q, mode, 1, 0, key);
4313 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4315 * @mode: which threads
4316 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4317 * @key: opaque value to be passed to wakeup targets
4319 * The sync wakeup differs that the waker knows that it will schedule
4320 * away soon, so while the target thread will be woken up, it will not
4321 * be migrated to another CPU - ie. the two threads are 'synchronized'
4322 * with each other. This can prevent needless bouncing between CPUs.
4324 * On UP it can prevent extra preemption.
4326 * It may be assumed that this function implies a write memory barrier before
4327 * changing the task state if and only if any tasks are woken up.
4329 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4330 int nr_exclusive, void *key)
4332 unsigned long flags;
4333 int wake_flags = WF_SYNC;
4338 if (unlikely(!nr_exclusive))
4341 spin_lock_irqsave(&q->lock, flags);
4342 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4343 spin_unlock_irqrestore(&q->lock, flags);
4345 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4348 * __wake_up_sync - see __wake_up_sync_key()
4350 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4352 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4354 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4357 * complete: - signals a single thread waiting on this completion
4358 * @x: holds the state of this particular completion
4360 * This will wake up a single thread waiting on this completion. Threads will be
4361 * awakened in the same order in which they were queued.
4363 * See also complete_all(), wait_for_completion() and related routines.
4365 * It may be assumed that this function implies a write memory barrier before
4366 * changing the task state if and only if any tasks are woken up.
4368 void complete(struct completion *x)
4370 unsigned long flags;
4372 spin_lock_irqsave(&x->wait.lock, flags);
4374 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4375 spin_unlock_irqrestore(&x->wait.lock, flags);
4377 EXPORT_SYMBOL(complete);
4380 * complete_all: - signals all threads waiting on this completion
4381 * @x: holds the state of this particular completion
4383 * This will wake up all threads waiting on this particular completion event.
4385 * It may be assumed that this function implies a write memory barrier before
4386 * changing the task state if and only if any tasks are woken up.
4388 void complete_all(struct completion *x)
4390 unsigned long flags;
4392 spin_lock_irqsave(&x->wait.lock, flags);
4393 x->done += UINT_MAX/2;
4394 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4395 spin_unlock_irqrestore(&x->wait.lock, flags);
4397 EXPORT_SYMBOL(complete_all);
4399 static inline long __sched
4400 do_wait_for_common(struct completion *x, long timeout, int state)
4403 DECLARE_WAITQUEUE(wait, current);
4405 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4407 if (signal_pending_state(state, current)) {
4408 timeout = -ERESTARTSYS;
4411 __set_current_state(state);
4412 spin_unlock_irq(&x->wait.lock);
4413 timeout = schedule_timeout(timeout);
4414 spin_lock_irq(&x->wait.lock);
4415 } while (!x->done && timeout);
4416 __remove_wait_queue(&x->wait, &wait);
4421 return timeout ?: 1;
4425 wait_for_common(struct completion *x, long timeout, int state)
4429 spin_lock_irq(&x->wait.lock);
4430 timeout = do_wait_for_common(x, timeout, state);
4431 spin_unlock_irq(&x->wait.lock);
4436 * wait_for_completion: - waits for completion of a task
4437 * @x: holds the state of this particular completion
4439 * This waits to be signaled for completion of a specific task. It is NOT
4440 * interruptible and there is no timeout.
4442 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4443 * and interrupt capability. Also see complete().
4445 void __sched wait_for_completion(struct completion *x)
4447 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4449 EXPORT_SYMBOL(wait_for_completion);
4452 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4453 * @x: holds the state of this particular completion
4454 * @timeout: timeout value in jiffies
4456 * This waits for either a completion of a specific task to be signaled or for a
4457 * specified timeout to expire. The timeout is in jiffies. It is not
4460 unsigned long __sched
4461 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4463 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4465 EXPORT_SYMBOL(wait_for_completion_timeout);
4468 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4469 * @x: holds the state of this particular completion
4471 * This waits for completion of a specific task to be signaled. It is
4474 int __sched wait_for_completion_interruptible(struct completion *x)
4476 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4477 if (t == -ERESTARTSYS)
4481 EXPORT_SYMBOL(wait_for_completion_interruptible);
4484 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4485 * @x: holds the state of this particular completion
4486 * @timeout: timeout value in jiffies
4488 * This waits for either a completion of a specific task to be signaled or for a
4489 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4491 unsigned long __sched
4492 wait_for_completion_interruptible_timeout(struct completion *x,
4493 unsigned long timeout)
4495 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4497 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4500 * wait_for_completion_killable: - waits for completion of a task (killable)
4501 * @x: holds the state of this particular completion
4503 * This waits to be signaled for completion of a specific task. It can be
4504 * interrupted by a kill signal.
4506 int __sched wait_for_completion_killable(struct completion *x)
4508 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4509 if (t == -ERESTARTSYS)
4513 EXPORT_SYMBOL(wait_for_completion_killable);
4516 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4517 * @x: holds the state of this particular completion
4518 * @timeout: timeout value in jiffies
4520 * This waits for either a completion of a specific task to be
4521 * signaled or for a specified timeout to expire. It can be
4522 * interrupted by a kill signal. The timeout is in jiffies.
4524 unsigned long __sched
4525 wait_for_completion_killable_timeout(struct completion *x,
4526 unsigned long timeout)
4528 return wait_for_common(x, timeout, TASK_KILLABLE);
4530 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
4533 * try_wait_for_completion - try to decrement a completion without blocking
4534 * @x: completion structure
4536 * Returns: 0 if a decrement cannot be done without blocking
4537 * 1 if a decrement succeeded.
4539 * If a completion is being used as a counting completion,
4540 * attempt to decrement the counter without blocking. This
4541 * enables us to avoid waiting if the resource the completion
4542 * is protecting is not available.
4544 bool try_wait_for_completion(struct completion *x)
4546 unsigned long flags;
4549 spin_lock_irqsave(&x->wait.lock, flags);
4554 spin_unlock_irqrestore(&x->wait.lock, flags);
4557 EXPORT_SYMBOL(try_wait_for_completion);
4560 * completion_done - Test to see if a completion has any waiters
4561 * @x: completion structure
4563 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4564 * 1 if there are no waiters.
4567 bool completion_done(struct completion *x)
4569 unsigned long flags;
4572 spin_lock_irqsave(&x->wait.lock, flags);
4575 spin_unlock_irqrestore(&x->wait.lock, flags);
4578 EXPORT_SYMBOL(completion_done);
4581 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4583 unsigned long flags;
4586 init_waitqueue_entry(&wait, current);
4588 __set_current_state(state);
4590 spin_lock_irqsave(&q->lock, flags);
4591 __add_wait_queue(q, &wait);
4592 spin_unlock(&q->lock);
4593 timeout = schedule_timeout(timeout);
4594 spin_lock_irq(&q->lock);
4595 __remove_wait_queue(q, &wait);
4596 spin_unlock_irqrestore(&q->lock, flags);
4601 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4603 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4605 EXPORT_SYMBOL(interruptible_sleep_on);
4608 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4610 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4612 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4614 void __sched sleep_on(wait_queue_head_t *q)
4616 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4618 EXPORT_SYMBOL(sleep_on);
4620 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4622 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4624 EXPORT_SYMBOL(sleep_on_timeout);
4626 #ifdef CONFIG_RT_MUTEXES
4629 * rt_mutex_setprio - set the current priority of a task
4631 * @prio: prio value (kernel-internal form)
4633 * This function changes the 'effective' priority of a task. It does
4634 * not touch ->normal_prio like __setscheduler().
4636 * Used by the rt_mutex code to implement priority inheritance logic.
4638 void rt_mutex_setprio(struct task_struct *p, int prio)
4640 unsigned long flags;
4641 int oldprio, on_rq, running;
4643 const struct sched_class *prev_class;
4645 BUG_ON(prio < 0 || prio > MAX_PRIO);
4647 rq = task_rq_lock(p, &flags);
4649 trace_sched_pi_setprio(p, prio);
4651 prev_class = p->sched_class;
4652 on_rq = p->se.on_rq;
4653 running = task_current(rq, p);
4655 dequeue_task(rq, p, 0);
4657 p->sched_class->put_prev_task(rq, p);
4660 p->sched_class = &rt_sched_class;
4662 p->sched_class = &fair_sched_class;
4667 p->sched_class->set_curr_task(rq);
4669 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4671 check_class_changed(rq, p, prev_class, oldprio, running);
4673 task_rq_unlock(rq, &flags);
4678 void set_user_nice(struct task_struct *p, long nice)
4680 int old_prio, delta, on_rq;
4681 unsigned long flags;
4684 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4687 * We have to be careful, if called from sys_setpriority(),
4688 * the task might be in the middle of scheduling on another CPU.
4690 rq = task_rq_lock(p, &flags);
4692 * The RT priorities are set via sched_setscheduler(), but we still
4693 * allow the 'normal' nice value to be set - but as expected
4694 * it wont have any effect on scheduling until the task is
4695 * SCHED_FIFO/SCHED_RR:
4697 if (task_has_rt_policy(p)) {
4698 p->static_prio = NICE_TO_PRIO(nice);
4701 on_rq = p->se.on_rq;
4703 dequeue_task(rq, p, 0);
4705 p->static_prio = NICE_TO_PRIO(nice);
4708 p->prio = effective_prio(p);
4709 delta = p->prio - old_prio;
4712 enqueue_task(rq, p, 0);
4714 * If the task increased its priority or is running and
4715 * lowered its priority, then reschedule its CPU:
4717 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4718 resched_task(rq->curr);
4721 task_rq_unlock(rq, &flags);
4723 EXPORT_SYMBOL(set_user_nice);
4726 * can_nice - check if a task can reduce its nice value
4730 int can_nice(const struct task_struct *p, const int nice)
4732 /* convert nice value [19,-20] to rlimit style value [1,40] */
4733 int nice_rlim = 20 - nice;
4735 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4736 capable(CAP_SYS_NICE));
4739 #ifdef __ARCH_WANT_SYS_NICE
4742 * sys_nice - change the priority of the current process.
4743 * @increment: priority increment
4745 * sys_setpriority is a more generic, but much slower function that
4746 * does similar things.
4748 SYSCALL_DEFINE1(nice, int, increment)
4753 * Setpriority might change our priority at the same moment.
4754 * We don't have to worry. Conceptually one call occurs first
4755 * and we have a single winner.
4757 if (increment < -40)
4762 nice = TASK_NICE(current) + increment;
4768 if (increment < 0 && !can_nice(current, nice))
4771 retval = security_task_setnice(current, nice);
4775 set_user_nice(current, nice);
4782 * task_prio - return the priority value of a given task.
4783 * @p: the task in question.
4785 * This is the priority value as seen by users in /proc.
4786 * RT tasks are offset by -200. Normal tasks are centered
4787 * around 0, value goes from -16 to +15.
4789 int task_prio(const struct task_struct *p)
4791 return p->prio - MAX_RT_PRIO;
4795 * task_nice - return the nice value of a given task.
4796 * @p: the task in question.
4798 int task_nice(const struct task_struct *p)
4800 return TASK_NICE(p);
4802 EXPORT_SYMBOL(task_nice);
4805 * idle_cpu - is a given cpu idle currently?
4806 * @cpu: the processor in question.
4808 int idle_cpu(int cpu)
4810 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4814 * idle_task - return the idle task for a given cpu.
4815 * @cpu: the processor in question.
4817 struct task_struct *idle_task(int cpu)
4819 return cpu_rq(cpu)->idle;
4823 * find_process_by_pid - find a process with a matching PID value.
4824 * @pid: the pid in question.
4826 static struct task_struct *find_process_by_pid(pid_t pid)
4828 return pid ? find_task_by_vpid(pid) : current;
4831 /* Actually do priority change: must hold rq lock. */
4833 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4835 BUG_ON(p->se.on_rq);
4838 p->rt_priority = prio;
4839 p->normal_prio = normal_prio(p);
4840 /* we are holding p->pi_lock already */
4841 p->prio = rt_mutex_getprio(p);
4842 if (rt_prio(p->prio))
4843 p->sched_class = &rt_sched_class;
4845 p->sched_class = &fair_sched_class;
4850 * check the target process has a UID that matches the current process's
4852 static bool check_same_owner(struct task_struct *p)
4854 const struct cred *cred = current_cred(), *pcred;
4858 pcred = __task_cred(p);
4859 match = (cred->euid == pcred->euid ||
4860 cred->euid == pcred->uid);
4865 static int __sched_setscheduler(struct task_struct *p, int policy,
4866 struct sched_param *param, bool user)
4868 int retval, oldprio, oldpolicy = -1, on_rq, running;
4869 unsigned long flags;
4870 const struct sched_class *prev_class;
4874 /* may grab non-irq protected spin_locks */
4875 BUG_ON(in_interrupt());
4877 /* double check policy once rq lock held */
4879 reset_on_fork = p->sched_reset_on_fork;
4880 policy = oldpolicy = p->policy;
4882 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4883 policy &= ~SCHED_RESET_ON_FORK;
4885 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4886 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4887 policy != SCHED_IDLE)
4892 * Valid priorities for SCHED_FIFO and SCHED_RR are
4893 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4894 * SCHED_BATCH and SCHED_IDLE is 0.
4896 if (param->sched_priority < 0 ||
4897 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4898 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4900 if (rt_policy(policy) != (param->sched_priority != 0))
4904 * Allow unprivileged RT tasks to decrease priority:
4906 if (user && !capable(CAP_SYS_NICE)) {
4907 if (rt_policy(policy)) {
4908 unsigned long rlim_rtprio =
4909 task_rlimit(p, RLIMIT_RTPRIO);
4911 /* can't set/change the rt policy */
4912 if (policy != p->policy && !rlim_rtprio)
4915 /* can't increase priority */
4916 if (param->sched_priority > p->rt_priority &&
4917 param->sched_priority > rlim_rtprio)
4921 * Like positive nice levels, dont allow tasks to
4922 * move out of SCHED_IDLE either:
4924 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4927 /* can't change other user's priorities */
4928 if (!check_same_owner(p))
4931 /* Normal users shall not reset the sched_reset_on_fork flag */
4932 if (p->sched_reset_on_fork && !reset_on_fork)
4937 retval = security_task_setscheduler(p);
4943 * make sure no PI-waiters arrive (or leave) while we are
4944 * changing the priority of the task:
4946 raw_spin_lock_irqsave(&p->pi_lock, flags);
4948 * To be able to change p->policy safely, the apropriate
4949 * runqueue lock must be held.
4951 rq = __task_rq_lock(p);
4954 * Changing the policy of the stop threads its a very bad idea
4956 if (p == rq->stop) {
4957 __task_rq_unlock(rq);
4958 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4962 #ifdef CONFIG_RT_GROUP_SCHED
4965 * Do not allow realtime tasks into groups that have no runtime
4968 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4969 task_group(p)->rt_bandwidth.rt_runtime == 0) {
4970 __task_rq_unlock(rq);
4971 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4977 /* recheck policy now with rq lock held */
4978 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4979 policy = oldpolicy = -1;
4980 __task_rq_unlock(rq);
4981 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4984 on_rq = p->se.on_rq;
4985 running = task_current(rq, p);
4987 deactivate_task(rq, p, 0);
4989 p->sched_class->put_prev_task(rq, p);
4991 p->sched_reset_on_fork = reset_on_fork;
4994 prev_class = p->sched_class;
4995 __setscheduler(rq, p, policy, param->sched_priority);
4998 p->sched_class->set_curr_task(rq);
5000 activate_task(rq, p, 0);
5002 check_class_changed(rq, p, prev_class, oldprio, running);
5004 __task_rq_unlock(rq);
5005 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5007 rt_mutex_adjust_pi(p);
5013 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5014 * @p: the task in question.
5015 * @policy: new policy.
5016 * @param: structure containing the new RT priority.
5018 * NOTE that the task may be already dead.
5020 int sched_setscheduler(struct task_struct *p, int policy,
5021 struct sched_param *param)
5023 return __sched_setscheduler(p, policy, param, true);
5025 EXPORT_SYMBOL_GPL(sched_setscheduler);
5028 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5029 * @p: the task in question.
5030 * @policy: new policy.
5031 * @param: structure containing the new RT priority.
5033 * Just like sched_setscheduler, only don't bother checking if the
5034 * current context has permission. For example, this is needed in
5035 * stop_machine(): we create temporary high priority worker threads,
5036 * but our caller might not have that capability.
5038 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5039 struct sched_param *param)
5041 return __sched_setscheduler(p, policy, param, false);
5045 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5047 struct sched_param lparam;
5048 struct task_struct *p;
5051 if (!param || pid < 0)
5053 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5058 p = find_process_by_pid(pid);
5060 retval = sched_setscheduler(p, policy, &lparam);
5067 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5068 * @pid: the pid in question.
5069 * @policy: new policy.
5070 * @param: structure containing the new RT priority.
5072 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
5073 struct sched_param __user *, param)
5075 /* negative values for policy are not valid */
5079 return do_sched_setscheduler(pid, policy, param);
5083 * sys_sched_setparam - set/change the RT priority of a thread
5084 * @pid: the pid in question.
5085 * @param: structure containing the new RT priority.
5087 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5089 return do_sched_setscheduler(pid, -1, param);
5093 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5094 * @pid: the pid in question.
5096 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5098 struct task_struct *p;
5106 p = find_process_by_pid(pid);
5108 retval = security_task_getscheduler(p);
5111 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
5118 * sys_sched_getparam - get the RT priority of a thread
5119 * @pid: the pid in question.
5120 * @param: structure containing the RT priority.
5122 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5124 struct sched_param lp;
5125 struct task_struct *p;
5128 if (!param || pid < 0)
5132 p = find_process_by_pid(pid);
5137 retval = security_task_getscheduler(p);
5141 lp.sched_priority = p->rt_priority;
5145 * This one might sleep, we cannot do it with a spinlock held ...
5147 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5156 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5158 cpumask_var_t cpus_allowed, new_mask;
5159 struct task_struct *p;
5165 p = find_process_by_pid(pid);
5172 /* Prevent p going away */
5176 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5180 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5182 goto out_free_cpus_allowed;
5185 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
5188 retval = security_task_setscheduler(p);
5192 cpuset_cpus_allowed(p, cpus_allowed);
5193 cpumask_and(new_mask, in_mask, cpus_allowed);
5195 retval = set_cpus_allowed_ptr(p, new_mask);
5198 cpuset_cpus_allowed(p, cpus_allowed);
5199 if (!cpumask_subset(new_mask, cpus_allowed)) {
5201 * We must have raced with a concurrent cpuset
5202 * update. Just reset the cpus_allowed to the
5203 * cpuset's cpus_allowed
5205 cpumask_copy(new_mask, cpus_allowed);
5210 free_cpumask_var(new_mask);
5211 out_free_cpus_allowed:
5212 free_cpumask_var(cpus_allowed);
5219 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5220 struct cpumask *new_mask)
5222 if (len < cpumask_size())
5223 cpumask_clear(new_mask);
5224 else if (len > cpumask_size())
5225 len = cpumask_size();
5227 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5231 * sys_sched_setaffinity - set the cpu affinity of a process
5232 * @pid: pid of the process
5233 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5234 * @user_mask_ptr: user-space pointer to the new cpu mask
5236 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5237 unsigned long __user *, user_mask_ptr)
5239 cpumask_var_t new_mask;
5242 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5245 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5247 retval = sched_setaffinity(pid, new_mask);
5248 free_cpumask_var(new_mask);
5252 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5254 struct task_struct *p;
5255 unsigned long flags;
5263 p = find_process_by_pid(pid);
5267 retval = security_task_getscheduler(p);
5271 rq = task_rq_lock(p, &flags);
5272 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5273 task_rq_unlock(rq, &flags);
5283 * sys_sched_getaffinity - get the cpu affinity of a process
5284 * @pid: pid of the process
5285 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5286 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5288 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5289 unsigned long __user *, user_mask_ptr)
5294 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5296 if (len & (sizeof(unsigned long)-1))
5299 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5302 ret = sched_getaffinity(pid, mask);
5304 size_t retlen = min_t(size_t, len, cpumask_size());
5306 if (copy_to_user(user_mask_ptr, mask, retlen))
5311 free_cpumask_var(mask);
5317 * sys_sched_yield - yield the current processor to other threads.
5319 * This function yields the current CPU to other tasks. If there are no
5320 * other threads running on this CPU then this function will return.
5322 SYSCALL_DEFINE0(sched_yield)
5324 struct rq *rq = this_rq_lock();
5326 schedstat_inc(rq, yld_count);
5327 current->sched_class->yield_task(rq);
5330 * Since we are going to call schedule() anyway, there's
5331 * no need to preempt or enable interrupts:
5333 __release(rq->lock);
5334 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5335 do_raw_spin_unlock(&rq->lock);
5336 preempt_enable_no_resched();
5343 static inline int should_resched(void)
5345 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5348 static void __cond_resched(void)
5350 add_preempt_count(PREEMPT_ACTIVE);
5352 sub_preempt_count(PREEMPT_ACTIVE);
5355 int __sched _cond_resched(void)
5357 if (should_resched()) {
5363 EXPORT_SYMBOL(_cond_resched);
5366 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5367 * call schedule, and on return reacquire the lock.
5369 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5370 * operations here to prevent schedule() from being called twice (once via
5371 * spin_unlock(), once by hand).
5373 int __cond_resched_lock(spinlock_t *lock)
5375 int resched = should_resched();
5378 lockdep_assert_held(lock);
5380 if (spin_needbreak(lock) || resched) {
5391 EXPORT_SYMBOL(__cond_resched_lock);
5393 int __sched __cond_resched_softirq(void)
5395 BUG_ON(!in_softirq());
5397 if (should_resched()) {
5405 EXPORT_SYMBOL(__cond_resched_softirq);
5408 * yield - yield the current processor to other threads.
5410 * This is a shortcut for kernel-space yielding - it marks the
5411 * thread runnable and calls sys_sched_yield().
5413 void __sched yield(void)
5415 set_current_state(TASK_RUNNING);
5418 EXPORT_SYMBOL(yield);
5421 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5422 * that process accounting knows that this is a task in IO wait state.
5424 void __sched io_schedule(void)
5426 struct rq *rq = raw_rq();
5428 delayacct_blkio_start();
5429 atomic_inc(&rq->nr_iowait);
5430 current->in_iowait = 1;
5432 current->in_iowait = 0;
5433 atomic_dec(&rq->nr_iowait);
5434 delayacct_blkio_end();
5436 EXPORT_SYMBOL(io_schedule);
5438 long __sched io_schedule_timeout(long timeout)
5440 struct rq *rq = raw_rq();
5443 delayacct_blkio_start();
5444 atomic_inc(&rq->nr_iowait);
5445 current->in_iowait = 1;
5446 ret = schedule_timeout(timeout);
5447 current->in_iowait = 0;
5448 atomic_dec(&rq->nr_iowait);
5449 delayacct_blkio_end();
5454 * sys_sched_get_priority_max - return maximum RT priority.
5455 * @policy: scheduling class.
5457 * this syscall returns the maximum rt_priority that can be used
5458 * by a given scheduling class.
5460 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5467 ret = MAX_USER_RT_PRIO-1;
5479 * sys_sched_get_priority_min - return minimum RT priority.
5480 * @policy: scheduling class.
5482 * this syscall returns the minimum rt_priority that can be used
5483 * by a given scheduling class.
5485 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5503 * sys_sched_rr_get_interval - return the default timeslice of a process.
5504 * @pid: pid of the process.
5505 * @interval: userspace pointer to the timeslice value.
5507 * this syscall writes the default timeslice value of a given process
5508 * into the user-space timespec buffer. A value of '0' means infinity.
5510 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5511 struct timespec __user *, interval)
5513 struct task_struct *p;
5514 unsigned int time_slice;
5515 unsigned long flags;
5525 p = find_process_by_pid(pid);
5529 retval = security_task_getscheduler(p);
5533 rq = task_rq_lock(p, &flags);
5534 time_slice = p->sched_class->get_rr_interval(rq, p);
5535 task_rq_unlock(rq, &flags);
5538 jiffies_to_timespec(time_slice, &t);
5539 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5547 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5549 void sched_show_task(struct task_struct *p)
5551 unsigned long free = 0;
5554 state = p->state ? __ffs(p->state) + 1 : 0;
5555 printk(KERN_INFO "%-13.13s %c", p->comm,
5556 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5557 #if BITS_PER_LONG == 32
5558 if (state == TASK_RUNNING)
5559 printk(KERN_CONT " running ");
5561 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5563 if (state == TASK_RUNNING)
5564 printk(KERN_CONT " running task ");
5566 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5568 #ifdef CONFIG_DEBUG_STACK_USAGE
5569 free = stack_not_used(p);
5571 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5572 task_pid_nr(p), task_pid_nr(p->real_parent),
5573 (unsigned long)task_thread_info(p)->flags);
5575 show_stack(p, NULL);
5578 void show_state_filter(unsigned long state_filter)
5580 struct task_struct *g, *p;
5582 #if BITS_PER_LONG == 32
5584 " task PC stack pid father\n");
5587 " task PC stack pid father\n");
5589 read_lock(&tasklist_lock);
5590 do_each_thread(g, p) {
5592 * reset the NMI-timeout, listing all files on a slow
5593 * console might take alot of time:
5595 touch_nmi_watchdog();
5596 if (!state_filter || (p->state & state_filter))
5598 } while_each_thread(g, p);
5600 touch_all_softlockup_watchdogs();
5602 #ifdef CONFIG_SCHED_DEBUG
5603 sysrq_sched_debug_show();
5605 read_unlock(&tasklist_lock);
5607 * Only show locks if all tasks are dumped:
5610 debug_show_all_locks();
5613 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5615 idle->sched_class = &idle_sched_class;
5619 * init_idle - set up an idle thread for a given CPU
5620 * @idle: task in question
5621 * @cpu: cpu the idle task belongs to
5623 * NOTE: this function does not set the idle thread's NEED_RESCHED
5624 * flag, to make booting more robust.
5626 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5628 struct rq *rq = cpu_rq(cpu);
5629 unsigned long flags;
5631 raw_spin_lock_irqsave(&rq->lock, flags);
5634 idle->state = TASK_RUNNING;
5635 idle->se.exec_start = sched_clock();
5637 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
5639 * We're having a chicken and egg problem, even though we are
5640 * holding rq->lock, the cpu isn't yet set to this cpu so the
5641 * lockdep check in task_group() will fail.
5643 * Similar case to sched_fork(). / Alternatively we could
5644 * use task_rq_lock() here and obtain the other rq->lock.
5649 __set_task_cpu(idle, cpu);
5652 rq->curr = rq->idle = idle;
5653 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5656 raw_spin_unlock_irqrestore(&rq->lock, flags);
5658 /* Set the preempt count _outside_ the spinlocks! */
5659 #if defined(CONFIG_PREEMPT)
5660 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5662 task_thread_info(idle)->preempt_count = 0;
5665 * The idle tasks have their own, simple scheduling class:
5667 idle->sched_class = &idle_sched_class;
5668 ftrace_graph_init_task(idle);
5672 * In a system that switches off the HZ timer nohz_cpu_mask
5673 * indicates which cpus entered this state. This is used
5674 * in the rcu update to wait only for active cpus. For system
5675 * which do not switch off the HZ timer nohz_cpu_mask should
5676 * always be CPU_BITS_NONE.
5678 cpumask_var_t nohz_cpu_mask;
5681 * Increase the granularity value when there are more CPUs,
5682 * because with more CPUs the 'effective latency' as visible
5683 * to users decreases. But the relationship is not linear,
5684 * so pick a second-best guess by going with the log2 of the
5687 * This idea comes from the SD scheduler of Con Kolivas:
5689 static int get_update_sysctl_factor(void)
5691 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5692 unsigned int factor;
5694 switch (sysctl_sched_tunable_scaling) {
5695 case SCHED_TUNABLESCALING_NONE:
5698 case SCHED_TUNABLESCALING_LINEAR:
5701 case SCHED_TUNABLESCALING_LOG:
5703 factor = 1 + ilog2(cpus);
5710 static void update_sysctl(void)
5712 unsigned int factor = get_update_sysctl_factor();
5714 #define SET_SYSCTL(name) \
5715 (sysctl_##name = (factor) * normalized_sysctl_##name)
5716 SET_SYSCTL(sched_min_granularity);
5717 SET_SYSCTL(sched_latency);
5718 SET_SYSCTL(sched_wakeup_granularity);
5719 SET_SYSCTL(sched_shares_ratelimit);
5723 static inline void sched_init_granularity(void)
5730 * This is how migration works:
5732 * 1) we invoke migration_cpu_stop() on the target CPU using
5734 * 2) stopper starts to run (implicitly forcing the migrated thread
5736 * 3) it checks whether the migrated task is still in the wrong runqueue.
5737 * 4) if it's in the wrong runqueue then the migration thread removes
5738 * it and puts it into the right queue.
5739 * 5) stopper completes and stop_one_cpu() returns and the migration
5744 * Change a given task's CPU affinity. Migrate the thread to a
5745 * proper CPU and schedule it away if the CPU it's executing on
5746 * is removed from the allowed bitmask.
5748 * NOTE: the caller must have a valid reference to the task, the
5749 * task must not exit() & deallocate itself prematurely. The
5750 * call is not atomic; no spinlocks may be held.
5752 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5754 unsigned long flags;
5756 unsigned int dest_cpu;
5760 * Serialize against TASK_WAKING so that ttwu() and wunt() can
5761 * drop the rq->lock and still rely on ->cpus_allowed.
5764 while (task_is_waking(p))
5766 rq = task_rq_lock(p, &flags);
5767 if (task_is_waking(p)) {
5768 task_rq_unlock(rq, &flags);
5772 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5777 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5778 !cpumask_equal(&p->cpus_allowed, new_mask))) {
5783 if (p->sched_class->set_cpus_allowed)
5784 p->sched_class->set_cpus_allowed(p, new_mask);
5786 cpumask_copy(&p->cpus_allowed, new_mask);
5787 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5790 /* Can the task run on the task's current CPU? If so, we're done */
5791 if (cpumask_test_cpu(task_cpu(p), new_mask))
5794 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
5795 if (migrate_task(p, dest_cpu)) {
5796 struct migration_arg arg = { p, dest_cpu };
5797 /* Need help from migration thread: drop lock and wait. */
5798 task_rq_unlock(rq, &flags);
5799 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
5800 tlb_migrate_finish(p->mm);
5804 task_rq_unlock(rq, &flags);
5808 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5811 * Move (not current) task off this cpu, onto dest cpu. We're doing
5812 * this because either it can't run here any more (set_cpus_allowed()
5813 * away from this CPU, or CPU going down), or because we're
5814 * attempting to rebalance this task on exec (sched_exec).
5816 * So we race with normal scheduler movements, but that's OK, as long
5817 * as the task is no longer on this CPU.
5819 * Returns non-zero if task was successfully migrated.
5821 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5823 struct rq *rq_dest, *rq_src;
5826 if (unlikely(!cpu_active(dest_cpu)))
5829 rq_src = cpu_rq(src_cpu);
5830 rq_dest = cpu_rq(dest_cpu);
5832 double_rq_lock(rq_src, rq_dest);
5833 /* Already moved. */
5834 if (task_cpu(p) != src_cpu)
5836 /* Affinity changed (again). */
5837 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
5841 * If we're not on a rq, the next wake-up will ensure we're
5845 deactivate_task(rq_src, p, 0);
5846 set_task_cpu(p, dest_cpu);
5847 activate_task(rq_dest, p, 0);
5848 check_preempt_curr(rq_dest, p, 0);
5853 double_rq_unlock(rq_src, rq_dest);
5858 * migration_cpu_stop - this will be executed by a highprio stopper thread
5859 * and performs thread migration by bumping thread off CPU then
5860 * 'pushing' onto another runqueue.
5862 static int migration_cpu_stop(void *data)
5864 struct migration_arg *arg = data;
5867 * The original target cpu might have gone down and we might
5868 * be on another cpu but it doesn't matter.
5870 local_irq_disable();
5871 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
5876 #ifdef CONFIG_HOTPLUG_CPU
5878 * Figure out where task on dead CPU should go, use force if necessary.
5880 void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5882 struct rq *rq = cpu_rq(dead_cpu);
5883 int needs_cpu, uninitialized_var(dest_cpu);
5884 unsigned long flags;
5886 local_irq_save(flags);
5888 raw_spin_lock(&rq->lock);
5889 needs_cpu = (task_cpu(p) == dead_cpu) && (p->state != TASK_WAKING);
5891 dest_cpu = select_fallback_rq(dead_cpu, p);
5892 raw_spin_unlock(&rq->lock);
5894 * It can only fail if we race with set_cpus_allowed(),
5895 * in the racer should migrate the task anyway.
5898 __migrate_task(p, dead_cpu, dest_cpu);
5899 local_irq_restore(flags);
5903 * While a dead CPU has no uninterruptible tasks queued at this point,
5904 * it might still have a nonzero ->nr_uninterruptible counter, because
5905 * for performance reasons the counter is not stricly tracking tasks to
5906 * their home CPUs. So we just add the counter to another CPU's counter,
5907 * to keep the global sum constant after CPU-down:
5909 static void migrate_nr_uninterruptible(struct rq *rq_src)
5911 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5912 unsigned long flags;
5914 local_irq_save(flags);
5915 double_rq_lock(rq_src, rq_dest);
5916 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5917 rq_src->nr_uninterruptible = 0;
5918 double_rq_unlock(rq_src, rq_dest);
5919 local_irq_restore(flags);
5922 /* Run through task list and migrate tasks from the dead cpu. */
5923 static void migrate_live_tasks(int src_cpu)
5925 struct task_struct *p, *t;
5927 read_lock(&tasklist_lock);
5929 do_each_thread(t, p) {
5933 if (task_cpu(p) == src_cpu)
5934 move_task_off_dead_cpu(src_cpu, p);
5935 } while_each_thread(t, p);
5937 read_unlock(&tasklist_lock);
5941 * Schedules idle task to be the next runnable task on current CPU.
5942 * It does so by boosting its priority to highest possible.
5943 * Used by CPU offline code.
5945 void sched_idle_next(void)
5947 int this_cpu = smp_processor_id();
5948 struct rq *rq = cpu_rq(this_cpu);
5949 struct task_struct *p = rq->idle;
5950 unsigned long flags;
5952 /* cpu has to be offline */
5953 BUG_ON(cpu_online(this_cpu));
5956 * Strictly not necessary since rest of the CPUs are stopped by now
5957 * and interrupts disabled on the current cpu.
5959 raw_spin_lock_irqsave(&rq->lock, flags);
5961 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5963 activate_task(rq, p, 0);
5965 raw_spin_unlock_irqrestore(&rq->lock, flags);
5969 * Ensures that the idle task is using init_mm right before its cpu goes
5972 void idle_task_exit(void)
5974 struct mm_struct *mm = current->active_mm;
5976 BUG_ON(cpu_online(smp_processor_id()));
5979 switch_mm(mm, &init_mm, current);
5983 /* called under rq->lock with disabled interrupts */
5984 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5986 struct rq *rq = cpu_rq(dead_cpu);
5988 /* Must be exiting, otherwise would be on tasklist. */
5989 BUG_ON(!p->exit_state);
5991 /* Cannot have done final schedule yet: would have vanished. */
5992 BUG_ON(p->state == TASK_DEAD);
5997 * Drop lock around migration; if someone else moves it,
5998 * that's OK. No task can be added to this CPU, so iteration is
6001 raw_spin_unlock_irq(&rq->lock);
6002 move_task_off_dead_cpu(dead_cpu, p);
6003 raw_spin_lock_irq(&rq->lock);
6008 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6009 static void migrate_dead_tasks(unsigned int dead_cpu)
6011 struct rq *rq = cpu_rq(dead_cpu);
6012 struct task_struct *next;
6015 if (!rq->nr_running)
6017 next = pick_next_task(rq);
6020 next->sched_class->put_prev_task(rq, next);
6021 migrate_dead(dead_cpu, next);
6027 * remove the tasks which were accounted by rq from calc_load_tasks.
6029 static void calc_global_load_remove(struct rq *rq)
6031 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
6032 rq->calc_load_active = 0;
6034 #endif /* CONFIG_HOTPLUG_CPU */
6036 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6038 static struct ctl_table sd_ctl_dir[] = {
6040 .procname = "sched_domain",
6046 static struct ctl_table sd_ctl_root[] = {
6048 .procname = "kernel",
6050 .child = sd_ctl_dir,
6055 static struct ctl_table *sd_alloc_ctl_entry(int n)
6057 struct ctl_table *entry =
6058 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6063 static void sd_free_ctl_entry(struct ctl_table **tablep)
6065 struct ctl_table *entry;
6068 * In the intermediate directories, both the child directory and
6069 * procname are dynamically allocated and could fail but the mode
6070 * will always be set. In the lowest directory the names are
6071 * static strings and all have proc handlers.
6073 for (entry = *tablep; entry->mode; entry++) {
6075 sd_free_ctl_entry(&entry->child);
6076 if (entry->proc_handler == NULL)
6077 kfree(entry->procname);
6085 set_table_entry(struct ctl_table *entry,
6086 const char *procname, void *data, int maxlen,
6087 mode_t mode, proc_handler *proc_handler)
6089 entry->procname = procname;
6091 entry->maxlen = maxlen;
6093 entry->proc_handler = proc_handler;
6096 static struct ctl_table *
6097 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6099 struct ctl_table *table = sd_alloc_ctl_entry(13);
6104 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6105 sizeof(long), 0644, proc_doulongvec_minmax);
6106 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6107 sizeof(long), 0644, proc_doulongvec_minmax);
6108 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6109 sizeof(int), 0644, proc_dointvec_minmax);
6110 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6111 sizeof(int), 0644, proc_dointvec_minmax);
6112 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6113 sizeof(int), 0644, proc_dointvec_minmax);
6114 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6115 sizeof(int), 0644, proc_dointvec_minmax);
6116 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6117 sizeof(int), 0644, proc_dointvec_minmax);
6118 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6119 sizeof(int), 0644, proc_dointvec_minmax);
6120 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6121 sizeof(int), 0644, proc_dointvec_minmax);
6122 set_table_entry(&table[9], "cache_nice_tries",
6123 &sd->cache_nice_tries,
6124 sizeof(int), 0644, proc_dointvec_minmax);
6125 set_table_entry(&table[10], "flags", &sd->flags,
6126 sizeof(int), 0644, proc_dointvec_minmax);
6127 set_table_entry(&table[11], "name", sd->name,
6128 CORENAME_MAX_SIZE, 0444, proc_dostring);
6129 /* &table[12] is terminator */
6134 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6136 struct ctl_table *entry, *table;
6137 struct sched_domain *sd;
6138 int domain_num = 0, i;
6141 for_each_domain(cpu, sd)
6143 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6148 for_each_domain(cpu, sd) {
6149 snprintf(buf, 32, "domain%d", i);
6150 entry->procname = kstrdup(buf, GFP_KERNEL);
6152 entry->child = sd_alloc_ctl_domain_table(sd);
6159 static struct ctl_table_header *sd_sysctl_header;
6160 static void register_sched_domain_sysctl(void)
6162 int i, cpu_num = num_possible_cpus();
6163 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6166 WARN_ON(sd_ctl_dir[0].child);
6167 sd_ctl_dir[0].child = entry;
6172 for_each_possible_cpu(i) {
6173 snprintf(buf, 32, "cpu%d", i);
6174 entry->procname = kstrdup(buf, GFP_KERNEL);
6176 entry->child = sd_alloc_ctl_cpu_table(i);
6180 WARN_ON(sd_sysctl_header);
6181 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6184 /* may be called multiple times per register */
6185 static void unregister_sched_domain_sysctl(void)
6187 if (sd_sysctl_header)
6188 unregister_sysctl_table(sd_sysctl_header);
6189 sd_sysctl_header = NULL;
6190 if (sd_ctl_dir[0].child)
6191 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6194 static void register_sched_domain_sysctl(void)
6197 static void unregister_sched_domain_sysctl(void)
6202 static void set_rq_online(struct rq *rq)
6205 const struct sched_class *class;
6207 cpumask_set_cpu(rq->cpu, rq->rd->online);
6210 for_each_class(class) {
6211 if (class->rq_online)
6212 class->rq_online(rq);
6217 static void set_rq_offline(struct rq *rq)
6220 const struct sched_class *class;
6222 for_each_class(class) {
6223 if (class->rq_offline)
6224 class->rq_offline(rq);
6227 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6233 * migration_call - callback that gets triggered when a CPU is added.
6234 * Here we can start up the necessary migration thread for the new CPU.
6236 static int __cpuinit
6237 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6239 int cpu = (long)hcpu;
6240 unsigned long flags;
6241 struct rq *rq = cpu_rq(cpu);
6245 case CPU_UP_PREPARE:
6246 case CPU_UP_PREPARE_FROZEN:
6247 rq->calc_load_update = calc_load_update;
6251 case CPU_ONLINE_FROZEN:
6252 /* Update our root-domain */
6253 raw_spin_lock_irqsave(&rq->lock, flags);
6255 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6259 raw_spin_unlock_irqrestore(&rq->lock, flags);
6262 #ifdef CONFIG_HOTPLUG_CPU
6264 case CPU_DEAD_FROZEN:
6265 migrate_live_tasks(cpu);
6266 /* Idle task back to normal (off runqueue, low prio) */
6267 raw_spin_lock_irq(&rq->lock);
6268 deactivate_task(rq, rq->idle, 0);
6269 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6270 rq->idle->sched_class = &idle_sched_class;
6271 migrate_dead_tasks(cpu);
6272 raw_spin_unlock_irq(&rq->lock);
6273 migrate_nr_uninterruptible(rq);
6274 BUG_ON(rq->nr_running != 0);
6275 calc_global_load_remove(rq);
6279 case CPU_DYING_FROZEN:
6280 /* Update our root-domain */
6281 raw_spin_lock_irqsave(&rq->lock, flags);
6283 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6286 raw_spin_unlock_irqrestore(&rq->lock, flags);
6294 * Register at high priority so that task migration (migrate_all_tasks)
6295 * happens before everything else. This has to be lower priority than
6296 * the notifier in the perf_event subsystem, though.
6298 static struct notifier_block __cpuinitdata migration_notifier = {
6299 .notifier_call = migration_call,
6300 .priority = CPU_PRI_MIGRATION,
6303 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
6304 unsigned long action, void *hcpu)
6306 switch (action & ~CPU_TASKS_FROZEN) {
6308 case CPU_DOWN_FAILED:
6309 set_cpu_active((long)hcpu, true);
6316 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
6317 unsigned long action, void *hcpu)
6319 switch (action & ~CPU_TASKS_FROZEN) {
6320 case CPU_DOWN_PREPARE:
6321 set_cpu_active((long)hcpu, false);
6328 static int __init migration_init(void)
6330 void *cpu = (void *)(long)smp_processor_id();
6333 /* Initialize migration for the boot CPU */
6334 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6335 BUG_ON(err == NOTIFY_BAD);
6336 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6337 register_cpu_notifier(&migration_notifier);
6339 /* Register cpu active notifiers */
6340 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
6341 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
6345 early_initcall(migration_init);
6350 #ifdef CONFIG_SCHED_DEBUG
6352 static __read_mostly int sched_domain_debug_enabled;
6354 static int __init sched_domain_debug_setup(char *str)
6356 sched_domain_debug_enabled = 1;
6360 early_param("sched_debug", sched_domain_debug_setup);
6362 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6363 struct cpumask *groupmask)
6365 struct sched_group *group = sd->groups;
6368 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6369 cpumask_clear(groupmask);
6371 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6373 if (!(sd->flags & SD_LOAD_BALANCE)) {
6374 printk("does not load-balance\n");
6376 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6381 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6383 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6384 printk(KERN_ERR "ERROR: domain->span does not contain "
6387 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6388 printk(KERN_ERR "ERROR: domain->groups does not contain"
6392 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6396 printk(KERN_ERR "ERROR: group is NULL\n");
6400 if (!group->cpu_power) {
6401 printk(KERN_CONT "\n");
6402 printk(KERN_ERR "ERROR: domain->cpu_power not "
6407 if (!cpumask_weight(sched_group_cpus(group))) {
6408 printk(KERN_CONT "\n");
6409 printk(KERN_ERR "ERROR: empty group\n");
6413 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6414 printk(KERN_CONT "\n");
6415 printk(KERN_ERR "ERROR: repeated CPUs\n");
6419 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6421 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6423 printk(KERN_CONT " %s", str);
6424 if (group->cpu_power != SCHED_LOAD_SCALE) {
6425 printk(KERN_CONT " (cpu_power = %d)",
6429 group = group->next;
6430 } while (group != sd->groups);
6431 printk(KERN_CONT "\n");
6433 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6434 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6437 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6438 printk(KERN_ERR "ERROR: parent span is not a superset "
6439 "of domain->span\n");
6443 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6445 cpumask_var_t groupmask;
6448 if (!sched_domain_debug_enabled)
6452 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6456 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6458 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6459 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6464 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6471 free_cpumask_var(groupmask);
6473 #else /* !CONFIG_SCHED_DEBUG */
6474 # define sched_domain_debug(sd, cpu) do { } while (0)
6475 #endif /* CONFIG_SCHED_DEBUG */
6477 static int sd_degenerate(struct sched_domain *sd)
6479 if (cpumask_weight(sched_domain_span(sd)) == 1)
6482 /* Following flags need at least 2 groups */
6483 if (sd->flags & (SD_LOAD_BALANCE |
6484 SD_BALANCE_NEWIDLE |
6488 SD_SHARE_PKG_RESOURCES)) {
6489 if (sd->groups != sd->groups->next)
6493 /* Following flags don't use groups */
6494 if (sd->flags & (SD_WAKE_AFFINE))
6501 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6503 unsigned long cflags = sd->flags, pflags = parent->flags;
6505 if (sd_degenerate(parent))
6508 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6511 /* Flags needing groups don't count if only 1 group in parent */
6512 if (parent->groups == parent->groups->next) {
6513 pflags &= ~(SD_LOAD_BALANCE |
6514 SD_BALANCE_NEWIDLE |
6518 SD_SHARE_PKG_RESOURCES);
6519 if (nr_node_ids == 1)
6520 pflags &= ~SD_SERIALIZE;
6522 if (~cflags & pflags)
6528 static void free_rootdomain(struct root_domain *rd)
6530 synchronize_sched();
6532 cpupri_cleanup(&rd->cpupri);
6534 free_cpumask_var(rd->rto_mask);
6535 free_cpumask_var(rd->online);
6536 free_cpumask_var(rd->span);
6540 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6542 struct root_domain *old_rd = NULL;
6543 unsigned long flags;
6545 raw_spin_lock_irqsave(&rq->lock, flags);
6550 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6553 cpumask_clear_cpu(rq->cpu, old_rd->span);
6556 * If we dont want to free the old_rt yet then
6557 * set old_rd to NULL to skip the freeing later
6560 if (!atomic_dec_and_test(&old_rd->refcount))
6564 atomic_inc(&rd->refcount);
6567 cpumask_set_cpu(rq->cpu, rd->span);
6568 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6571 raw_spin_unlock_irqrestore(&rq->lock, flags);
6574 free_rootdomain(old_rd);
6577 static int init_rootdomain(struct root_domain *rd)
6579 memset(rd, 0, sizeof(*rd));
6581 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
6583 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
6585 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6588 if (cpupri_init(&rd->cpupri) != 0)
6593 free_cpumask_var(rd->rto_mask);
6595 free_cpumask_var(rd->online);
6597 free_cpumask_var(rd->span);
6602 static void init_defrootdomain(void)
6604 init_rootdomain(&def_root_domain);
6606 atomic_set(&def_root_domain.refcount, 1);
6609 static struct root_domain *alloc_rootdomain(void)
6611 struct root_domain *rd;
6613 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6617 if (init_rootdomain(rd) != 0) {
6626 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6627 * hold the hotplug lock.
6630 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6632 struct rq *rq = cpu_rq(cpu);
6633 struct sched_domain *tmp;
6635 for (tmp = sd; tmp; tmp = tmp->parent)
6636 tmp->span_weight = cpumask_weight(sched_domain_span(tmp));
6638 /* Remove the sched domains which do not contribute to scheduling. */
6639 for (tmp = sd; tmp; ) {
6640 struct sched_domain *parent = tmp->parent;
6644 if (sd_parent_degenerate(tmp, parent)) {
6645 tmp->parent = parent->parent;
6647 parent->parent->child = tmp;
6652 if (sd && sd_degenerate(sd)) {
6658 sched_domain_debug(sd, cpu);
6660 rq_attach_root(rq, rd);
6661 rcu_assign_pointer(rq->sd, sd);
6664 /* cpus with isolated domains */
6665 static cpumask_var_t cpu_isolated_map;
6667 /* Setup the mask of cpus configured for isolated domains */
6668 static int __init isolated_cpu_setup(char *str)
6670 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6671 cpulist_parse(str, cpu_isolated_map);
6675 __setup("isolcpus=", isolated_cpu_setup);
6678 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6679 * to a function which identifies what group(along with sched group) a CPU
6680 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6681 * (due to the fact that we keep track of groups covered with a struct cpumask).
6683 * init_sched_build_groups will build a circular linked list of the groups
6684 * covered by the given span, and will set each group's ->cpumask correctly,
6685 * and ->cpu_power to 0.
6688 init_sched_build_groups(const struct cpumask *span,
6689 const struct cpumask *cpu_map,
6690 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
6691 struct sched_group **sg,
6692 struct cpumask *tmpmask),
6693 struct cpumask *covered, struct cpumask *tmpmask)
6695 struct sched_group *first = NULL, *last = NULL;
6698 cpumask_clear(covered);
6700 for_each_cpu(i, span) {
6701 struct sched_group *sg;
6702 int group = group_fn(i, cpu_map, &sg, tmpmask);
6705 if (cpumask_test_cpu(i, covered))
6708 cpumask_clear(sched_group_cpus(sg));
6711 for_each_cpu(j, span) {
6712 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6715 cpumask_set_cpu(j, covered);
6716 cpumask_set_cpu(j, sched_group_cpus(sg));
6727 #define SD_NODES_PER_DOMAIN 16
6732 * find_next_best_node - find the next node to include in a sched_domain
6733 * @node: node whose sched_domain we're building
6734 * @used_nodes: nodes already in the sched_domain
6736 * Find the next node to include in a given scheduling domain. Simply
6737 * finds the closest node not already in the @used_nodes map.
6739 * Should use nodemask_t.
6741 static int find_next_best_node(int node, nodemask_t *used_nodes)
6743 int i, n, val, min_val, best_node = 0;
6747 for (i = 0; i < nr_node_ids; i++) {
6748 /* Start at @node */
6749 n = (node + i) % nr_node_ids;
6751 if (!nr_cpus_node(n))
6754 /* Skip already used nodes */
6755 if (node_isset(n, *used_nodes))
6758 /* Simple min distance search */
6759 val = node_distance(node, n);
6761 if (val < min_val) {
6767 node_set(best_node, *used_nodes);
6772 * sched_domain_node_span - get a cpumask for a node's sched_domain
6773 * @node: node whose cpumask we're constructing
6774 * @span: resulting cpumask
6776 * Given a node, construct a good cpumask for its sched_domain to span. It
6777 * should be one that prevents unnecessary balancing, but also spreads tasks
6780 static void sched_domain_node_span(int node, struct cpumask *span)
6782 nodemask_t used_nodes;
6785 cpumask_clear(span);
6786 nodes_clear(used_nodes);
6788 cpumask_or(span, span, cpumask_of_node(node));
6789 node_set(node, used_nodes);
6791 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6792 int next_node = find_next_best_node(node, &used_nodes);
6794 cpumask_or(span, span, cpumask_of_node(next_node));
6797 #endif /* CONFIG_NUMA */
6799 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6802 * The cpus mask in sched_group and sched_domain hangs off the end.
6804 * ( See the the comments in include/linux/sched.h:struct sched_group
6805 * and struct sched_domain. )
6807 struct static_sched_group {
6808 struct sched_group sg;
6809 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
6812 struct static_sched_domain {
6813 struct sched_domain sd;
6814 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
6820 cpumask_var_t domainspan;
6821 cpumask_var_t covered;
6822 cpumask_var_t notcovered;
6824 cpumask_var_t nodemask;
6825 cpumask_var_t this_sibling_map;
6826 cpumask_var_t this_core_map;
6827 cpumask_var_t this_book_map;
6828 cpumask_var_t send_covered;
6829 cpumask_var_t tmpmask;
6830 struct sched_group **sched_group_nodes;
6831 struct root_domain *rd;
6835 sa_sched_groups = 0,
6841 sa_this_sibling_map,
6843 sa_sched_group_nodes,
6853 * SMT sched-domains:
6855 #ifdef CONFIG_SCHED_SMT
6856 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
6857 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
6860 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
6861 struct sched_group **sg, struct cpumask *unused)
6864 *sg = &per_cpu(sched_groups, cpu).sg;
6867 #endif /* CONFIG_SCHED_SMT */
6870 * multi-core sched-domains:
6872 #ifdef CONFIG_SCHED_MC
6873 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
6874 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
6877 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6878 struct sched_group **sg, struct cpumask *mask)
6881 #ifdef CONFIG_SCHED_SMT
6882 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6883 group = cpumask_first(mask);
6888 *sg = &per_cpu(sched_group_core, group).sg;
6891 #endif /* CONFIG_SCHED_MC */
6894 * book sched-domains:
6896 #ifdef CONFIG_SCHED_BOOK
6897 static DEFINE_PER_CPU(struct static_sched_domain, book_domains);
6898 static DEFINE_PER_CPU(struct static_sched_group, sched_group_book);
6901 cpu_to_book_group(int cpu, const struct cpumask *cpu_map,
6902 struct sched_group **sg, struct cpumask *mask)
6905 #ifdef CONFIG_SCHED_MC
6906 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6907 group = cpumask_first(mask);
6908 #elif defined(CONFIG_SCHED_SMT)
6909 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6910 group = cpumask_first(mask);
6913 *sg = &per_cpu(sched_group_book, group).sg;
6916 #endif /* CONFIG_SCHED_BOOK */
6918 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
6919 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
6922 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
6923 struct sched_group **sg, struct cpumask *mask)
6926 #ifdef CONFIG_SCHED_BOOK
6927 cpumask_and(mask, cpu_book_mask(cpu), cpu_map);
6928 group = cpumask_first(mask);
6929 #elif defined(CONFIG_SCHED_MC)
6930 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6931 group = cpumask_first(mask);
6932 #elif defined(CONFIG_SCHED_SMT)
6933 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6934 group = cpumask_first(mask);
6939 *sg = &per_cpu(sched_group_phys, group).sg;
6945 * The init_sched_build_groups can't handle what we want to do with node
6946 * groups, so roll our own. Now each node has its own list of groups which
6947 * gets dynamically allocated.
6949 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
6950 static struct sched_group ***sched_group_nodes_bycpu;
6952 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
6953 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
6955 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
6956 struct sched_group **sg,
6957 struct cpumask *nodemask)
6961 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
6962 group = cpumask_first(nodemask);
6965 *sg = &per_cpu(sched_group_allnodes, group).sg;
6969 static void init_numa_sched_groups_power(struct sched_group *group_head)
6971 struct sched_group *sg = group_head;
6977 for_each_cpu(j, sched_group_cpus(sg)) {
6978 struct sched_domain *sd;
6980 sd = &per_cpu(phys_domains, j).sd;
6981 if (j != group_first_cpu(sd->groups)) {
6983 * Only add "power" once for each
6989 sg->cpu_power += sd->groups->cpu_power;
6992 } while (sg != group_head);
6995 static int build_numa_sched_groups(struct s_data *d,
6996 const struct cpumask *cpu_map, int num)
6998 struct sched_domain *sd;
6999 struct sched_group *sg, *prev;
7002 cpumask_clear(d->covered);
7003 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
7004 if (cpumask_empty(d->nodemask)) {
7005 d->sched_group_nodes[num] = NULL;
7009 sched_domain_node_span(num, d->domainspan);
7010 cpumask_and(d->domainspan, d->domainspan, cpu_map);
7012 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
7015 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
7019 d->sched_group_nodes[num] = sg;
7021 for_each_cpu(j, d->nodemask) {
7022 sd = &per_cpu(node_domains, j).sd;
7027 cpumask_copy(sched_group_cpus(sg), d->nodemask);
7029 cpumask_or(d->covered, d->covered, d->nodemask);
7032 for (j = 0; j < nr_node_ids; j++) {
7033 n = (num + j) % nr_node_ids;
7034 cpumask_complement(d->notcovered, d->covered);
7035 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
7036 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
7037 if (cpumask_empty(d->tmpmask))
7039 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
7040 if (cpumask_empty(d->tmpmask))
7042 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
7046 "Can not alloc domain group for node %d\n", j);
7050 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
7051 sg->next = prev->next;
7052 cpumask_or(d->covered, d->covered, d->tmpmask);
7059 #endif /* CONFIG_NUMA */
7062 /* Free memory allocated for various sched_group structures */
7063 static void free_sched_groups(const struct cpumask *cpu_map,
7064 struct cpumask *nodemask)
7068 for_each_cpu(cpu, cpu_map) {
7069 struct sched_group **sched_group_nodes
7070 = sched_group_nodes_bycpu[cpu];
7072 if (!sched_group_nodes)
7075 for (i = 0; i < nr_node_ids; i++) {
7076 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7078 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
7079 if (cpumask_empty(nodemask))
7089 if (oldsg != sched_group_nodes[i])
7092 kfree(sched_group_nodes);
7093 sched_group_nodes_bycpu[cpu] = NULL;
7096 #else /* !CONFIG_NUMA */
7097 static void free_sched_groups(const struct cpumask *cpu_map,
7098 struct cpumask *nodemask)
7101 #endif /* CONFIG_NUMA */
7104 * Initialize sched groups cpu_power.
7106 * cpu_power indicates the capacity of sched group, which is used while
7107 * distributing the load between different sched groups in a sched domain.
7108 * Typically cpu_power for all the groups in a sched domain will be same unless
7109 * there are asymmetries in the topology. If there are asymmetries, group
7110 * having more cpu_power will pickup more load compared to the group having
7113 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7115 struct sched_domain *child;
7116 struct sched_group *group;
7120 WARN_ON(!sd || !sd->groups);
7122 if (cpu != group_first_cpu(sd->groups))
7125 sd->groups->group_weight = cpumask_weight(sched_group_cpus(sd->groups));
7129 sd->groups->cpu_power = 0;
7132 power = SCHED_LOAD_SCALE;
7133 weight = cpumask_weight(sched_domain_span(sd));
7135 * SMT siblings share the power of a single core.
7136 * Usually multiple threads get a better yield out of
7137 * that one core than a single thread would have,
7138 * reflect that in sd->smt_gain.
7140 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
7141 power *= sd->smt_gain;
7143 power >>= SCHED_LOAD_SHIFT;
7145 sd->groups->cpu_power += power;
7150 * Add cpu_power of each child group to this groups cpu_power.
7152 group = child->groups;
7154 sd->groups->cpu_power += group->cpu_power;
7155 group = group->next;
7156 } while (group != child->groups);
7160 * Initializers for schedule domains
7161 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7164 #ifdef CONFIG_SCHED_DEBUG
7165 # define SD_INIT_NAME(sd, type) sd->name = #type
7167 # define SD_INIT_NAME(sd, type) do { } while (0)
7170 #define SD_INIT(sd, type) sd_init_##type(sd)
7172 #define SD_INIT_FUNC(type) \
7173 static noinline void sd_init_##type(struct sched_domain *sd) \
7175 memset(sd, 0, sizeof(*sd)); \
7176 *sd = SD_##type##_INIT; \
7177 sd->level = SD_LV_##type; \
7178 SD_INIT_NAME(sd, type); \
7183 SD_INIT_FUNC(ALLNODES)
7186 #ifdef CONFIG_SCHED_SMT
7187 SD_INIT_FUNC(SIBLING)
7189 #ifdef CONFIG_SCHED_MC
7192 #ifdef CONFIG_SCHED_BOOK
7196 static int default_relax_domain_level = -1;
7198 static int __init setup_relax_domain_level(char *str)
7202 val = simple_strtoul(str, NULL, 0);
7203 if (val < SD_LV_MAX)
7204 default_relax_domain_level = val;
7208 __setup("relax_domain_level=", setup_relax_domain_level);
7210 static void set_domain_attribute(struct sched_domain *sd,
7211 struct sched_domain_attr *attr)
7215 if (!attr || attr->relax_domain_level < 0) {
7216 if (default_relax_domain_level < 0)
7219 request = default_relax_domain_level;
7221 request = attr->relax_domain_level;
7222 if (request < sd->level) {
7223 /* turn off idle balance on this domain */
7224 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7226 /* turn on idle balance on this domain */
7227 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7231 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
7232 const struct cpumask *cpu_map)
7235 case sa_sched_groups:
7236 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
7237 d->sched_group_nodes = NULL;
7239 free_rootdomain(d->rd); /* fall through */
7241 free_cpumask_var(d->tmpmask); /* fall through */
7242 case sa_send_covered:
7243 free_cpumask_var(d->send_covered); /* fall through */
7244 case sa_this_book_map:
7245 free_cpumask_var(d->this_book_map); /* fall through */
7246 case sa_this_core_map:
7247 free_cpumask_var(d->this_core_map); /* fall through */
7248 case sa_this_sibling_map:
7249 free_cpumask_var(d->this_sibling_map); /* fall through */
7251 free_cpumask_var(d->nodemask); /* fall through */
7252 case sa_sched_group_nodes:
7254 kfree(d->sched_group_nodes); /* fall through */
7256 free_cpumask_var(d->notcovered); /* fall through */
7258 free_cpumask_var(d->covered); /* fall through */
7260 free_cpumask_var(d->domainspan); /* fall through */
7267 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
7268 const struct cpumask *cpu_map)
7271 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
7273 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
7274 return sa_domainspan;
7275 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
7277 /* Allocate the per-node list of sched groups */
7278 d->sched_group_nodes = kcalloc(nr_node_ids,
7279 sizeof(struct sched_group *), GFP_KERNEL);
7280 if (!d->sched_group_nodes) {
7281 printk(KERN_WARNING "Can not alloc sched group node list\n");
7282 return sa_notcovered;
7284 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
7286 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
7287 return sa_sched_group_nodes;
7288 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
7290 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
7291 return sa_this_sibling_map;
7292 if (!alloc_cpumask_var(&d->this_book_map, GFP_KERNEL))
7293 return sa_this_core_map;
7294 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
7295 return sa_this_book_map;
7296 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
7297 return sa_send_covered;
7298 d->rd = alloc_rootdomain();
7300 printk(KERN_WARNING "Cannot alloc root domain\n");
7303 return sa_rootdomain;
7306 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
7307 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
7309 struct sched_domain *sd = NULL;
7311 struct sched_domain *parent;
7314 if (cpumask_weight(cpu_map) >
7315 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
7316 sd = &per_cpu(allnodes_domains, i).sd;
7317 SD_INIT(sd, ALLNODES);
7318 set_domain_attribute(sd, attr);
7319 cpumask_copy(sched_domain_span(sd), cpu_map);
7320 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
7325 sd = &per_cpu(node_domains, i).sd;
7327 set_domain_attribute(sd, attr);
7328 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
7329 sd->parent = parent;
7332 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
7337 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
7338 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7339 struct sched_domain *parent, int i)
7341 struct sched_domain *sd;
7342 sd = &per_cpu(phys_domains, i).sd;
7344 set_domain_attribute(sd, attr);
7345 cpumask_copy(sched_domain_span(sd), d->nodemask);
7346 sd->parent = parent;
7349 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
7353 static struct sched_domain *__build_book_sched_domain(struct s_data *d,
7354 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7355 struct sched_domain *parent, int i)
7357 struct sched_domain *sd = parent;
7358 #ifdef CONFIG_SCHED_BOOK
7359 sd = &per_cpu(book_domains, i).sd;
7361 set_domain_attribute(sd, attr);
7362 cpumask_and(sched_domain_span(sd), cpu_map, cpu_book_mask(i));
7363 sd->parent = parent;
7365 cpu_to_book_group(i, cpu_map, &sd->groups, d->tmpmask);
7370 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
7371 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7372 struct sched_domain *parent, int i)
7374 struct sched_domain *sd = parent;
7375 #ifdef CONFIG_SCHED_MC
7376 sd = &per_cpu(core_domains, i).sd;
7378 set_domain_attribute(sd, attr);
7379 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
7380 sd->parent = parent;
7382 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
7387 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
7388 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7389 struct sched_domain *parent, int i)
7391 struct sched_domain *sd = parent;
7392 #ifdef CONFIG_SCHED_SMT
7393 sd = &per_cpu(cpu_domains, i).sd;
7394 SD_INIT(sd, SIBLING);
7395 set_domain_attribute(sd, attr);
7396 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
7397 sd->parent = parent;
7399 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
7404 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
7405 const struct cpumask *cpu_map, int cpu)
7408 #ifdef CONFIG_SCHED_SMT
7409 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
7410 cpumask_and(d->this_sibling_map, cpu_map,
7411 topology_thread_cpumask(cpu));
7412 if (cpu == cpumask_first(d->this_sibling_map))
7413 init_sched_build_groups(d->this_sibling_map, cpu_map,
7415 d->send_covered, d->tmpmask);
7418 #ifdef CONFIG_SCHED_MC
7419 case SD_LV_MC: /* set up multi-core groups */
7420 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
7421 if (cpu == cpumask_first(d->this_core_map))
7422 init_sched_build_groups(d->this_core_map, cpu_map,
7424 d->send_covered, d->tmpmask);
7427 #ifdef CONFIG_SCHED_BOOK
7428 case SD_LV_BOOK: /* set up book groups */
7429 cpumask_and(d->this_book_map, cpu_map, cpu_book_mask(cpu));
7430 if (cpu == cpumask_first(d->this_book_map))
7431 init_sched_build_groups(d->this_book_map, cpu_map,
7433 d->send_covered, d->tmpmask);
7436 case SD_LV_CPU: /* set up physical groups */
7437 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
7438 if (!cpumask_empty(d->nodemask))
7439 init_sched_build_groups(d->nodemask, cpu_map,
7441 d->send_covered, d->tmpmask);
7444 case SD_LV_ALLNODES:
7445 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
7446 d->send_covered, d->tmpmask);
7455 * Build sched domains for a given set of cpus and attach the sched domains
7456 * to the individual cpus
7458 static int __build_sched_domains(const struct cpumask *cpu_map,
7459 struct sched_domain_attr *attr)
7461 enum s_alloc alloc_state = sa_none;
7463 struct sched_domain *sd;
7469 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7470 if (alloc_state != sa_rootdomain)
7472 alloc_state = sa_sched_groups;
7475 * Set up domains for cpus specified by the cpu_map.
7477 for_each_cpu(i, cpu_map) {
7478 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
7481 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
7482 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
7483 sd = __build_book_sched_domain(&d, cpu_map, attr, sd, i);
7484 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
7485 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
7488 for_each_cpu(i, cpu_map) {
7489 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
7490 build_sched_groups(&d, SD_LV_BOOK, cpu_map, i);
7491 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
7494 /* Set up physical groups */
7495 for (i = 0; i < nr_node_ids; i++)
7496 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
7499 /* Set up node groups */
7501 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
7503 for (i = 0; i < nr_node_ids; i++)
7504 if (build_numa_sched_groups(&d, cpu_map, i))
7508 /* Calculate CPU power for physical packages and nodes */
7509 #ifdef CONFIG_SCHED_SMT
7510 for_each_cpu(i, cpu_map) {
7511 sd = &per_cpu(cpu_domains, i).sd;
7512 init_sched_groups_power(i, sd);
7515 #ifdef CONFIG_SCHED_MC
7516 for_each_cpu(i, cpu_map) {
7517 sd = &per_cpu(core_domains, i).sd;
7518 init_sched_groups_power(i, sd);
7521 #ifdef CONFIG_SCHED_BOOK
7522 for_each_cpu(i, cpu_map) {
7523 sd = &per_cpu(book_domains, i).sd;
7524 init_sched_groups_power(i, sd);
7528 for_each_cpu(i, cpu_map) {
7529 sd = &per_cpu(phys_domains, i).sd;
7530 init_sched_groups_power(i, sd);
7534 for (i = 0; i < nr_node_ids; i++)
7535 init_numa_sched_groups_power(d.sched_group_nodes[i]);
7537 if (d.sd_allnodes) {
7538 struct sched_group *sg;
7540 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7542 init_numa_sched_groups_power(sg);
7546 /* Attach the domains */
7547 for_each_cpu(i, cpu_map) {
7548 #ifdef CONFIG_SCHED_SMT
7549 sd = &per_cpu(cpu_domains, i).sd;
7550 #elif defined(CONFIG_SCHED_MC)
7551 sd = &per_cpu(core_domains, i).sd;
7552 #elif defined(CONFIG_SCHED_BOOK)
7553 sd = &per_cpu(book_domains, i).sd;
7555 sd = &per_cpu(phys_domains, i).sd;
7557 cpu_attach_domain(sd, d.rd, i);
7560 d.sched_group_nodes = NULL; /* don't free this we still need it */
7561 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
7565 __free_domain_allocs(&d, alloc_state, cpu_map);
7569 static int build_sched_domains(const struct cpumask *cpu_map)
7571 return __build_sched_domains(cpu_map, NULL);
7574 static cpumask_var_t *doms_cur; /* current sched domains */
7575 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7576 static struct sched_domain_attr *dattr_cur;
7577 /* attribues of custom domains in 'doms_cur' */
7580 * Special case: If a kmalloc of a doms_cur partition (array of
7581 * cpumask) fails, then fallback to a single sched domain,
7582 * as determined by the single cpumask fallback_doms.
7584 static cpumask_var_t fallback_doms;
7587 * arch_update_cpu_topology lets virtualized architectures update the
7588 * cpu core maps. It is supposed to return 1 if the topology changed
7589 * or 0 if it stayed the same.
7591 int __attribute__((weak)) arch_update_cpu_topology(void)
7596 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7599 cpumask_var_t *doms;
7601 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7604 for (i = 0; i < ndoms; i++) {
7605 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7606 free_sched_domains(doms, i);
7613 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7616 for (i = 0; i < ndoms; i++)
7617 free_cpumask_var(doms[i]);
7622 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7623 * For now this just excludes isolated cpus, but could be used to
7624 * exclude other special cases in the future.
7626 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7630 arch_update_cpu_topology();
7632 doms_cur = alloc_sched_domains(ndoms_cur);
7634 doms_cur = &fallback_doms;
7635 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7637 err = build_sched_domains(doms_cur[0]);
7638 register_sched_domain_sysctl();
7643 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7644 struct cpumask *tmpmask)
7646 free_sched_groups(cpu_map, tmpmask);
7650 * Detach sched domains from a group of cpus specified in cpu_map
7651 * These cpus will now be attached to the NULL domain
7653 static void detach_destroy_domains(const struct cpumask *cpu_map)
7655 /* Save because hotplug lock held. */
7656 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7659 for_each_cpu(i, cpu_map)
7660 cpu_attach_domain(NULL, &def_root_domain, i);
7661 synchronize_sched();
7662 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7665 /* handle null as "default" */
7666 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7667 struct sched_domain_attr *new, int idx_new)
7669 struct sched_domain_attr tmp;
7676 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7677 new ? (new + idx_new) : &tmp,
7678 sizeof(struct sched_domain_attr));
7682 * Partition sched domains as specified by the 'ndoms_new'
7683 * cpumasks in the array doms_new[] of cpumasks. This compares
7684 * doms_new[] to the current sched domain partitioning, doms_cur[].
7685 * It destroys each deleted domain and builds each new domain.
7687 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7688 * The masks don't intersect (don't overlap.) We should setup one
7689 * sched domain for each mask. CPUs not in any of the cpumasks will
7690 * not be load balanced. If the same cpumask appears both in the
7691 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7694 * The passed in 'doms_new' should be allocated using
7695 * alloc_sched_domains. This routine takes ownership of it and will
7696 * free_sched_domains it when done with it. If the caller failed the
7697 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7698 * and partition_sched_domains() will fallback to the single partition
7699 * 'fallback_doms', it also forces the domains to be rebuilt.
7701 * If doms_new == NULL it will be replaced with cpu_online_mask.
7702 * ndoms_new == 0 is a special case for destroying existing domains,
7703 * and it will not create the default domain.
7705 * Call with hotplug lock held
7707 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7708 struct sched_domain_attr *dattr_new)
7713 mutex_lock(&sched_domains_mutex);
7715 /* always unregister in case we don't destroy any domains */
7716 unregister_sched_domain_sysctl();
7718 /* Let architecture update cpu core mappings. */
7719 new_topology = arch_update_cpu_topology();
7721 n = doms_new ? ndoms_new : 0;
7723 /* Destroy deleted domains */
7724 for (i = 0; i < ndoms_cur; i++) {
7725 for (j = 0; j < n && !new_topology; j++) {
7726 if (cpumask_equal(doms_cur[i], doms_new[j])
7727 && dattrs_equal(dattr_cur, i, dattr_new, j))
7730 /* no match - a current sched domain not in new doms_new[] */
7731 detach_destroy_domains(doms_cur[i]);
7736 if (doms_new == NULL) {
7738 doms_new = &fallback_doms;
7739 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7740 WARN_ON_ONCE(dattr_new);
7743 /* Build new domains */
7744 for (i = 0; i < ndoms_new; i++) {
7745 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7746 if (cpumask_equal(doms_new[i], doms_cur[j])
7747 && dattrs_equal(dattr_new, i, dattr_cur, j))
7750 /* no match - add a new doms_new */
7751 __build_sched_domains(doms_new[i],
7752 dattr_new ? dattr_new + i : NULL);
7757 /* Remember the new sched domains */
7758 if (doms_cur != &fallback_doms)
7759 free_sched_domains(doms_cur, ndoms_cur);
7760 kfree(dattr_cur); /* kfree(NULL) is safe */
7761 doms_cur = doms_new;
7762 dattr_cur = dattr_new;
7763 ndoms_cur = ndoms_new;
7765 register_sched_domain_sysctl();
7767 mutex_unlock(&sched_domains_mutex);
7770 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7771 static void arch_reinit_sched_domains(void)
7775 /* Destroy domains first to force the rebuild */
7776 partition_sched_domains(0, NULL, NULL);
7778 rebuild_sched_domains();
7782 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7784 unsigned int level = 0;
7786 if (sscanf(buf, "%u", &level) != 1)
7790 * level is always be positive so don't check for
7791 * level < POWERSAVINGS_BALANCE_NONE which is 0
7792 * What happens on 0 or 1 byte write,
7793 * need to check for count as well?
7796 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7800 sched_smt_power_savings = level;
7802 sched_mc_power_savings = level;
7804 arch_reinit_sched_domains();
7809 #ifdef CONFIG_SCHED_MC
7810 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7811 struct sysdev_class_attribute *attr,
7814 return sprintf(page, "%u\n", sched_mc_power_savings);
7816 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7817 struct sysdev_class_attribute *attr,
7818 const char *buf, size_t count)
7820 return sched_power_savings_store(buf, count, 0);
7822 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7823 sched_mc_power_savings_show,
7824 sched_mc_power_savings_store);
7827 #ifdef CONFIG_SCHED_SMT
7828 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7829 struct sysdev_class_attribute *attr,
7832 return sprintf(page, "%u\n", sched_smt_power_savings);
7834 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7835 struct sysdev_class_attribute *attr,
7836 const char *buf, size_t count)
7838 return sched_power_savings_store(buf, count, 1);
7840 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7841 sched_smt_power_savings_show,
7842 sched_smt_power_savings_store);
7845 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7849 #ifdef CONFIG_SCHED_SMT
7851 err = sysfs_create_file(&cls->kset.kobj,
7852 &attr_sched_smt_power_savings.attr);
7854 #ifdef CONFIG_SCHED_MC
7855 if (!err && mc_capable())
7856 err = sysfs_create_file(&cls->kset.kobj,
7857 &attr_sched_mc_power_savings.attr);
7861 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7864 * Update cpusets according to cpu_active mask. If cpusets are
7865 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7866 * around partition_sched_domains().
7868 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7871 switch (action & ~CPU_TASKS_FROZEN) {
7873 case CPU_DOWN_FAILED:
7874 cpuset_update_active_cpus();
7881 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7884 switch (action & ~CPU_TASKS_FROZEN) {
7885 case CPU_DOWN_PREPARE:
7886 cpuset_update_active_cpus();
7893 static int update_runtime(struct notifier_block *nfb,
7894 unsigned long action, void *hcpu)
7896 int cpu = (int)(long)hcpu;
7899 case CPU_DOWN_PREPARE:
7900 case CPU_DOWN_PREPARE_FROZEN:
7901 disable_runtime(cpu_rq(cpu));
7904 case CPU_DOWN_FAILED:
7905 case CPU_DOWN_FAILED_FROZEN:
7907 case CPU_ONLINE_FROZEN:
7908 enable_runtime(cpu_rq(cpu));
7916 void __init sched_init_smp(void)
7918 cpumask_var_t non_isolated_cpus;
7920 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7921 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7923 #if defined(CONFIG_NUMA)
7924 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7926 BUG_ON(sched_group_nodes_bycpu == NULL);
7929 mutex_lock(&sched_domains_mutex);
7930 arch_init_sched_domains(cpu_active_mask);
7931 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7932 if (cpumask_empty(non_isolated_cpus))
7933 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7934 mutex_unlock(&sched_domains_mutex);
7937 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7938 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7940 /* RT runtime code needs to handle some hotplug events */
7941 hotcpu_notifier(update_runtime, 0);
7945 /* Move init over to a non-isolated CPU */
7946 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7948 sched_init_granularity();
7949 free_cpumask_var(non_isolated_cpus);
7951 init_sched_rt_class();
7954 void __init sched_init_smp(void)
7956 sched_init_granularity();
7958 #endif /* CONFIG_SMP */
7960 const_debug unsigned int sysctl_timer_migration = 1;
7962 int in_sched_functions(unsigned long addr)
7964 return in_lock_functions(addr) ||
7965 (addr >= (unsigned long)__sched_text_start
7966 && addr < (unsigned long)__sched_text_end);
7969 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7971 cfs_rq->tasks_timeline = RB_ROOT;
7972 INIT_LIST_HEAD(&cfs_rq->tasks);
7973 #ifdef CONFIG_FAIR_GROUP_SCHED
7976 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7979 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7981 struct rt_prio_array *array;
7984 array = &rt_rq->active;
7985 for (i = 0; i < MAX_RT_PRIO; i++) {
7986 INIT_LIST_HEAD(array->queue + i);
7987 __clear_bit(i, array->bitmap);
7989 /* delimiter for bitsearch: */
7990 __set_bit(MAX_RT_PRIO, array->bitmap);
7992 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7993 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7995 rt_rq->highest_prio.next = MAX_RT_PRIO;
7999 rt_rq->rt_nr_migratory = 0;
8000 rt_rq->overloaded = 0;
8001 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
8005 rt_rq->rt_throttled = 0;
8006 rt_rq->rt_runtime = 0;
8007 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
8009 #ifdef CONFIG_RT_GROUP_SCHED
8010 rt_rq->rt_nr_boosted = 0;
8015 #ifdef CONFIG_FAIR_GROUP_SCHED
8016 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8017 struct sched_entity *se, int cpu, int add,
8018 struct sched_entity *parent)
8020 struct rq *rq = cpu_rq(cpu);
8021 tg->cfs_rq[cpu] = cfs_rq;
8022 init_cfs_rq(cfs_rq, rq);
8025 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
8028 /* se could be NULL for init_task_group */
8033 se->cfs_rq = &rq->cfs;
8035 se->cfs_rq = parent->my_q;
8038 se->load.weight = tg->shares;
8039 se->load.inv_weight = 0;
8040 se->parent = parent;
8044 #ifdef CONFIG_RT_GROUP_SCHED
8045 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8046 struct sched_rt_entity *rt_se, int cpu, int add,
8047 struct sched_rt_entity *parent)
8049 struct rq *rq = cpu_rq(cpu);
8051 tg->rt_rq[cpu] = rt_rq;
8052 init_rt_rq(rt_rq, rq);
8054 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8056 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
8058 tg->rt_se[cpu] = rt_se;
8063 rt_se->rt_rq = &rq->rt;
8065 rt_se->rt_rq = parent->my_q;
8067 rt_se->my_q = rt_rq;
8068 rt_se->parent = parent;
8069 INIT_LIST_HEAD(&rt_se->run_list);
8073 void __init sched_init(void)
8076 unsigned long alloc_size = 0, ptr;
8078 #ifdef CONFIG_FAIR_GROUP_SCHED
8079 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8081 #ifdef CONFIG_RT_GROUP_SCHED
8082 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8084 #ifdef CONFIG_CPUMASK_OFFSTACK
8085 alloc_size += num_possible_cpus() * cpumask_size();
8088 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
8090 #ifdef CONFIG_FAIR_GROUP_SCHED
8091 init_task_group.se = (struct sched_entity **)ptr;
8092 ptr += nr_cpu_ids * sizeof(void **);
8094 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
8095 ptr += nr_cpu_ids * sizeof(void **);
8097 #endif /* CONFIG_FAIR_GROUP_SCHED */
8098 #ifdef CONFIG_RT_GROUP_SCHED
8099 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
8100 ptr += nr_cpu_ids * sizeof(void **);
8102 init_task_group.rt_rq = (struct rt_rq **)ptr;
8103 ptr += nr_cpu_ids * sizeof(void **);
8105 #endif /* CONFIG_RT_GROUP_SCHED */
8106 #ifdef CONFIG_CPUMASK_OFFSTACK
8107 for_each_possible_cpu(i) {
8108 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
8109 ptr += cpumask_size();
8111 #endif /* CONFIG_CPUMASK_OFFSTACK */
8115 init_defrootdomain();
8118 init_rt_bandwidth(&def_rt_bandwidth,
8119 global_rt_period(), global_rt_runtime());
8121 #ifdef CONFIG_RT_GROUP_SCHED
8122 init_rt_bandwidth(&init_task_group.rt_bandwidth,
8123 global_rt_period(), global_rt_runtime());
8124 #endif /* CONFIG_RT_GROUP_SCHED */
8126 #ifdef CONFIG_CGROUP_SCHED
8127 list_add(&init_task_group.list, &task_groups);
8128 INIT_LIST_HEAD(&init_task_group.children);
8130 #endif /* CONFIG_CGROUP_SCHED */
8132 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
8133 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
8134 __alignof__(unsigned long));
8136 for_each_possible_cpu(i) {
8140 raw_spin_lock_init(&rq->lock);
8142 rq->calc_load_active = 0;
8143 rq->calc_load_update = jiffies + LOAD_FREQ;
8144 init_cfs_rq(&rq->cfs, rq);
8145 init_rt_rq(&rq->rt, rq);
8146 #ifdef CONFIG_FAIR_GROUP_SCHED
8147 init_task_group.shares = init_task_group_load;
8148 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8149 #ifdef CONFIG_CGROUP_SCHED
8151 * How much cpu bandwidth does init_task_group get?
8153 * In case of task-groups formed thr' the cgroup filesystem, it
8154 * gets 100% of the cpu resources in the system. This overall
8155 * system cpu resource is divided among the tasks of
8156 * init_task_group and its child task-groups in a fair manner,
8157 * based on each entity's (task or task-group's) weight
8158 * (se->load.weight).
8160 * In other words, if init_task_group has 10 tasks of weight
8161 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8162 * then A0's share of the cpu resource is:
8164 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8166 * We achieve this by letting init_task_group's tasks sit
8167 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8169 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
8171 #endif /* CONFIG_FAIR_GROUP_SCHED */
8173 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8174 #ifdef CONFIG_RT_GROUP_SCHED
8175 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8176 #ifdef CONFIG_CGROUP_SCHED
8177 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
8181 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8182 rq->cpu_load[j] = 0;
8184 rq->last_load_update_tick = jiffies;
8189 rq->cpu_power = SCHED_LOAD_SCALE;
8190 rq->post_schedule = 0;
8191 rq->active_balance = 0;
8192 rq->next_balance = jiffies;
8197 rq->avg_idle = 2*sysctl_sched_migration_cost;
8198 rq_attach_root(rq, &def_root_domain);
8200 rq->nohz_balance_kick = 0;
8201 init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb, i));
8205 atomic_set(&rq->nr_iowait, 0);
8208 set_load_weight(&init_task);
8210 #ifdef CONFIG_PREEMPT_NOTIFIERS
8211 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8215 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8218 #ifdef CONFIG_RT_MUTEXES
8219 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
8223 * The boot idle thread does lazy MMU switching as well:
8225 atomic_inc(&init_mm.mm_count);
8226 enter_lazy_tlb(&init_mm, current);
8229 * Make us the idle thread. Technically, schedule() should not be
8230 * called from this thread, however somewhere below it might be,
8231 * but because we are the idle thread, we just pick up running again
8232 * when this runqueue becomes "idle".
8234 init_idle(current, smp_processor_id());
8236 calc_load_update = jiffies + LOAD_FREQ;
8239 * During early bootup we pretend to be a normal task:
8241 current->sched_class = &fair_sched_class;
8243 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8244 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
8247 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8248 alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT);
8249 atomic_set(&nohz.load_balancer, nr_cpu_ids);
8250 atomic_set(&nohz.first_pick_cpu, nr_cpu_ids);
8251 atomic_set(&nohz.second_pick_cpu, nr_cpu_ids);
8253 /* May be allocated at isolcpus cmdline parse time */
8254 if (cpu_isolated_map == NULL)
8255 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
8260 scheduler_running = 1;
8263 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8264 static inline int preempt_count_equals(int preempt_offset)
8266 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
8268 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
8271 void __might_sleep(const char *file, int line, int preempt_offset)
8274 static unsigned long prev_jiffy; /* ratelimiting */
8276 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
8277 system_state != SYSTEM_RUNNING || oops_in_progress)
8279 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8281 prev_jiffy = jiffies;
8284 "BUG: sleeping function called from invalid context at %s:%d\n",
8287 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8288 in_atomic(), irqs_disabled(),
8289 current->pid, current->comm);
8291 debug_show_held_locks(current);
8292 if (irqs_disabled())
8293 print_irqtrace_events(current);
8297 EXPORT_SYMBOL(__might_sleep);
8300 #ifdef CONFIG_MAGIC_SYSRQ
8301 static void normalize_task(struct rq *rq, struct task_struct *p)
8305 on_rq = p->se.on_rq;
8307 deactivate_task(rq, p, 0);
8308 __setscheduler(rq, p, SCHED_NORMAL, 0);
8310 activate_task(rq, p, 0);
8311 resched_task(rq->curr);
8315 void normalize_rt_tasks(void)
8317 struct task_struct *g, *p;
8318 unsigned long flags;
8321 read_lock_irqsave(&tasklist_lock, flags);
8322 do_each_thread(g, p) {
8324 * Only normalize user tasks:
8329 p->se.exec_start = 0;
8330 #ifdef CONFIG_SCHEDSTATS
8331 p->se.statistics.wait_start = 0;
8332 p->se.statistics.sleep_start = 0;
8333 p->se.statistics.block_start = 0;
8338 * Renice negative nice level userspace
8341 if (TASK_NICE(p) < 0 && p->mm)
8342 set_user_nice(p, 0);
8346 raw_spin_lock(&p->pi_lock);
8347 rq = __task_rq_lock(p);
8349 normalize_task(rq, p);
8351 __task_rq_unlock(rq);
8352 raw_spin_unlock(&p->pi_lock);
8353 } while_each_thread(g, p);
8355 read_unlock_irqrestore(&tasklist_lock, flags);
8358 #endif /* CONFIG_MAGIC_SYSRQ */
8360 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8362 * These functions are only useful for the IA64 MCA handling, or kdb.
8364 * They can only be called when the whole system has been
8365 * stopped - every CPU needs to be quiescent, and no scheduling
8366 * activity can take place. Using them for anything else would
8367 * be a serious bug, and as a result, they aren't even visible
8368 * under any other configuration.
8372 * curr_task - return the current task for a given cpu.
8373 * @cpu: the processor in question.
8375 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8377 struct task_struct *curr_task(int cpu)
8379 return cpu_curr(cpu);
8382 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8386 * set_curr_task - set the current task for a given cpu.
8387 * @cpu: the processor in question.
8388 * @p: the task pointer to set.
8390 * Description: This function must only be used when non-maskable interrupts
8391 * are serviced on a separate stack. It allows the architecture to switch the
8392 * notion of the current task on a cpu in a non-blocking manner. This function
8393 * must be called with all CPU's synchronized, and interrupts disabled, the
8394 * and caller must save the original value of the current task (see
8395 * curr_task() above) and restore that value before reenabling interrupts and
8396 * re-starting the system.
8398 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8400 void set_curr_task(int cpu, struct task_struct *p)
8407 #ifdef CONFIG_FAIR_GROUP_SCHED
8408 static void free_fair_sched_group(struct task_group *tg)
8412 for_each_possible_cpu(i) {
8414 kfree(tg->cfs_rq[i]);
8424 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8426 struct cfs_rq *cfs_rq;
8427 struct sched_entity *se;
8431 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8434 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8438 tg->shares = NICE_0_LOAD;
8440 for_each_possible_cpu(i) {
8443 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8444 GFP_KERNEL, cpu_to_node(i));
8448 se = kzalloc_node(sizeof(struct sched_entity),
8449 GFP_KERNEL, cpu_to_node(i));
8453 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
8464 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8466 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8467 &cpu_rq(cpu)->leaf_cfs_rq_list);
8470 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8472 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8474 #else /* !CONFG_FAIR_GROUP_SCHED */
8475 static inline void free_fair_sched_group(struct task_group *tg)
8480 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8485 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8489 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8492 #endif /* CONFIG_FAIR_GROUP_SCHED */
8494 #ifdef CONFIG_RT_GROUP_SCHED
8495 static void free_rt_sched_group(struct task_group *tg)
8499 destroy_rt_bandwidth(&tg->rt_bandwidth);
8501 for_each_possible_cpu(i) {
8503 kfree(tg->rt_rq[i]);
8505 kfree(tg->rt_se[i]);
8513 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8515 struct rt_rq *rt_rq;
8516 struct sched_rt_entity *rt_se;
8520 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8523 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8527 init_rt_bandwidth(&tg->rt_bandwidth,
8528 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8530 for_each_possible_cpu(i) {
8533 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8534 GFP_KERNEL, cpu_to_node(i));
8538 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8539 GFP_KERNEL, cpu_to_node(i));
8543 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
8554 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8556 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8557 &cpu_rq(cpu)->leaf_rt_rq_list);
8560 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8562 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8564 #else /* !CONFIG_RT_GROUP_SCHED */
8565 static inline void free_rt_sched_group(struct task_group *tg)
8570 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8575 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8579 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8582 #endif /* CONFIG_RT_GROUP_SCHED */
8584 #ifdef CONFIG_CGROUP_SCHED
8585 static void free_sched_group(struct task_group *tg)
8587 free_fair_sched_group(tg);
8588 free_rt_sched_group(tg);
8592 /* allocate runqueue etc for a new task group */
8593 struct task_group *sched_create_group(struct task_group *parent)
8595 struct task_group *tg;
8596 unsigned long flags;
8599 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8601 return ERR_PTR(-ENOMEM);
8603 if (!alloc_fair_sched_group(tg, parent))
8606 if (!alloc_rt_sched_group(tg, parent))
8609 spin_lock_irqsave(&task_group_lock, flags);
8610 for_each_possible_cpu(i) {
8611 register_fair_sched_group(tg, i);
8612 register_rt_sched_group(tg, i);
8614 list_add_rcu(&tg->list, &task_groups);
8616 WARN_ON(!parent); /* root should already exist */
8618 tg->parent = parent;
8619 INIT_LIST_HEAD(&tg->children);
8620 list_add_rcu(&tg->siblings, &parent->children);
8621 spin_unlock_irqrestore(&task_group_lock, flags);
8626 free_sched_group(tg);
8627 return ERR_PTR(-ENOMEM);
8630 /* rcu callback to free various structures associated with a task group */
8631 static void free_sched_group_rcu(struct rcu_head *rhp)
8633 /* now it should be safe to free those cfs_rqs */
8634 free_sched_group(container_of(rhp, struct task_group, rcu));
8637 /* Destroy runqueue etc associated with a task group */
8638 void sched_destroy_group(struct task_group *tg)
8640 unsigned long flags;
8643 spin_lock_irqsave(&task_group_lock, flags);
8644 for_each_possible_cpu(i) {
8645 unregister_fair_sched_group(tg, i);
8646 unregister_rt_sched_group(tg, i);
8648 list_del_rcu(&tg->list);
8649 list_del_rcu(&tg->siblings);
8650 spin_unlock_irqrestore(&task_group_lock, flags);
8652 /* wait for possible concurrent references to cfs_rqs complete */
8653 call_rcu(&tg->rcu, free_sched_group_rcu);
8656 /* change task's runqueue when it moves between groups.
8657 * The caller of this function should have put the task in its new group
8658 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8659 * reflect its new group.
8661 void sched_move_task(struct task_struct *tsk)
8664 unsigned long flags;
8667 rq = task_rq_lock(tsk, &flags);
8669 running = task_current(rq, tsk);
8670 on_rq = tsk->se.on_rq;
8673 dequeue_task(rq, tsk, 0);
8674 if (unlikely(running))
8675 tsk->sched_class->put_prev_task(rq, tsk);
8677 #ifdef CONFIG_FAIR_GROUP_SCHED
8678 if (tsk->sched_class->task_move_group)
8679 tsk->sched_class->task_move_group(tsk, on_rq);
8682 set_task_rq(tsk, task_cpu(tsk));
8684 if (unlikely(running))
8685 tsk->sched_class->set_curr_task(rq);
8687 enqueue_task(rq, tsk, 0);
8689 task_rq_unlock(rq, &flags);
8691 #endif /* CONFIG_CGROUP_SCHED */
8693 #ifdef CONFIG_FAIR_GROUP_SCHED
8694 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8696 struct cfs_rq *cfs_rq = se->cfs_rq;
8701 dequeue_entity(cfs_rq, se, 0);
8703 se->load.weight = shares;
8704 se->load.inv_weight = 0;
8707 enqueue_entity(cfs_rq, se, 0);
8710 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8712 struct cfs_rq *cfs_rq = se->cfs_rq;
8713 struct rq *rq = cfs_rq->rq;
8714 unsigned long flags;
8716 raw_spin_lock_irqsave(&rq->lock, flags);
8717 __set_se_shares(se, shares);
8718 raw_spin_unlock_irqrestore(&rq->lock, flags);
8721 static DEFINE_MUTEX(shares_mutex);
8723 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8726 unsigned long flags;
8729 * We can't change the weight of the root cgroup.
8734 if (shares < MIN_SHARES)
8735 shares = MIN_SHARES;
8736 else if (shares > MAX_SHARES)
8737 shares = MAX_SHARES;
8739 mutex_lock(&shares_mutex);
8740 if (tg->shares == shares)
8743 spin_lock_irqsave(&task_group_lock, flags);
8744 for_each_possible_cpu(i)
8745 unregister_fair_sched_group(tg, i);
8746 list_del_rcu(&tg->siblings);
8747 spin_unlock_irqrestore(&task_group_lock, flags);
8749 /* wait for any ongoing reference to this group to finish */
8750 synchronize_sched();
8753 * Now we are free to modify the group's share on each cpu
8754 * w/o tripping rebalance_share or load_balance_fair.
8756 tg->shares = shares;
8757 for_each_possible_cpu(i) {
8761 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8762 set_se_shares(tg->se[i], shares);
8766 * Enable load balance activity on this group, by inserting it back on
8767 * each cpu's rq->leaf_cfs_rq_list.
8769 spin_lock_irqsave(&task_group_lock, flags);
8770 for_each_possible_cpu(i)
8771 register_fair_sched_group(tg, i);
8772 list_add_rcu(&tg->siblings, &tg->parent->children);
8773 spin_unlock_irqrestore(&task_group_lock, flags);
8775 mutex_unlock(&shares_mutex);
8779 unsigned long sched_group_shares(struct task_group *tg)
8785 #ifdef CONFIG_RT_GROUP_SCHED
8787 * Ensure that the real time constraints are schedulable.
8789 static DEFINE_MUTEX(rt_constraints_mutex);
8791 static unsigned long to_ratio(u64 period, u64 runtime)
8793 if (runtime == RUNTIME_INF)
8796 return div64_u64(runtime << 20, period);
8799 /* Must be called with tasklist_lock held */
8800 static inline int tg_has_rt_tasks(struct task_group *tg)
8802 struct task_struct *g, *p;
8804 do_each_thread(g, p) {
8805 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8807 } while_each_thread(g, p);
8812 struct rt_schedulable_data {
8813 struct task_group *tg;
8818 static int tg_schedulable(struct task_group *tg, void *data)
8820 struct rt_schedulable_data *d = data;
8821 struct task_group *child;
8822 unsigned long total, sum = 0;
8823 u64 period, runtime;
8825 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8826 runtime = tg->rt_bandwidth.rt_runtime;
8829 period = d->rt_period;
8830 runtime = d->rt_runtime;
8834 * Cannot have more runtime than the period.
8836 if (runtime > period && runtime != RUNTIME_INF)
8840 * Ensure we don't starve existing RT tasks.
8842 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8845 total = to_ratio(period, runtime);
8848 * Nobody can have more than the global setting allows.
8850 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8854 * The sum of our children's runtime should not exceed our own.
8856 list_for_each_entry_rcu(child, &tg->children, siblings) {
8857 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8858 runtime = child->rt_bandwidth.rt_runtime;
8860 if (child == d->tg) {
8861 period = d->rt_period;
8862 runtime = d->rt_runtime;
8865 sum += to_ratio(period, runtime);
8874 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8876 struct rt_schedulable_data data = {
8878 .rt_period = period,
8879 .rt_runtime = runtime,
8882 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8885 static int tg_set_bandwidth(struct task_group *tg,
8886 u64 rt_period, u64 rt_runtime)
8890 mutex_lock(&rt_constraints_mutex);
8891 read_lock(&tasklist_lock);
8892 err = __rt_schedulable(tg, rt_period, rt_runtime);
8896 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8897 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8898 tg->rt_bandwidth.rt_runtime = rt_runtime;
8900 for_each_possible_cpu(i) {
8901 struct rt_rq *rt_rq = tg->rt_rq[i];
8903 raw_spin_lock(&rt_rq->rt_runtime_lock);
8904 rt_rq->rt_runtime = rt_runtime;
8905 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8907 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8909 read_unlock(&tasklist_lock);
8910 mutex_unlock(&rt_constraints_mutex);
8915 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8917 u64 rt_runtime, rt_period;
8919 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8920 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8921 if (rt_runtime_us < 0)
8922 rt_runtime = RUNTIME_INF;
8924 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8927 long sched_group_rt_runtime(struct task_group *tg)
8931 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8934 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8935 do_div(rt_runtime_us, NSEC_PER_USEC);
8936 return rt_runtime_us;
8939 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8941 u64 rt_runtime, rt_period;
8943 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8944 rt_runtime = tg->rt_bandwidth.rt_runtime;
8949 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8952 long sched_group_rt_period(struct task_group *tg)
8956 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8957 do_div(rt_period_us, NSEC_PER_USEC);
8958 return rt_period_us;
8961 static int sched_rt_global_constraints(void)
8963 u64 runtime, period;
8966 if (sysctl_sched_rt_period <= 0)
8969 runtime = global_rt_runtime();
8970 period = global_rt_period();
8973 * Sanity check on the sysctl variables.
8975 if (runtime > period && runtime != RUNTIME_INF)
8978 mutex_lock(&rt_constraints_mutex);
8979 read_lock(&tasklist_lock);
8980 ret = __rt_schedulable(NULL, 0, 0);
8981 read_unlock(&tasklist_lock);
8982 mutex_unlock(&rt_constraints_mutex);
8987 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8989 /* Don't accept realtime tasks when there is no way for them to run */
8990 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8996 #else /* !CONFIG_RT_GROUP_SCHED */
8997 static int sched_rt_global_constraints(void)
8999 unsigned long flags;
9002 if (sysctl_sched_rt_period <= 0)
9006 * There's always some RT tasks in the root group
9007 * -- migration, kstopmachine etc..
9009 if (sysctl_sched_rt_runtime == 0)
9012 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
9013 for_each_possible_cpu(i) {
9014 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
9016 raw_spin_lock(&rt_rq->rt_runtime_lock);
9017 rt_rq->rt_runtime = global_rt_runtime();
9018 raw_spin_unlock(&rt_rq->rt_runtime_lock);
9020 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
9024 #endif /* CONFIG_RT_GROUP_SCHED */
9026 int sched_rt_handler(struct ctl_table *table, int write,
9027 void __user *buffer, size_t *lenp,
9031 int old_period, old_runtime;
9032 static DEFINE_MUTEX(mutex);
9035 old_period = sysctl_sched_rt_period;
9036 old_runtime = sysctl_sched_rt_runtime;
9038 ret = proc_dointvec(table, write, buffer, lenp, ppos);
9040 if (!ret && write) {
9041 ret = sched_rt_global_constraints();
9043 sysctl_sched_rt_period = old_period;
9044 sysctl_sched_rt_runtime = old_runtime;
9046 def_rt_bandwidth.rt_runtime = global_rt_runtime();
9047 def_rt_bandwidth.rt_period =
9048 ns_to_ktime(global_rt_period());
9051 mutex_unlock(&mutex);
9056 #ifdef CONFIG_CGROUP_SCHED
9058 /* return corresponding task_group object of a cgroup */
9059 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9061 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9062 struct task_group, css);
9065 static struct cgroup_subsys_state *
9066 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9068 struct task_group *tg, *parent;
9070 if (!cgrp->parent) {
9071 /* This is early initialization for the top cgroup */
9072 return &init_task_group.css;
9075 parent = cgroup_tg(cgrp->parent);
9076 tg = sched_create_group(parent);
9078 return ERR_PTR(-ENOMEM);
9084 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9086 struct task_group *tg = cgroup_tg(cgrp);
9088 sched_destroy_group(tg);
9092 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
9094 #ifdef CONFIG_RT_GROUP_SCHED
9095 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
9098 /* We don't support RT-tasks being in separate groups */
9099 if (tsk->sched_class != &fair_sched_class)
9106 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9107 struct task_struct *tsk, bool threadgroup)
9109 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
9113 struct task_struct *c;
9115 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
9116 retval = cpu_cgroup_can_attach_task(cgrp, c);
9128 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9129 struct cgroup *old_cont, struct task_struct *tsk,
9132 sched_move_task(tsk);
9134 struct task_struct *c;
9136 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
9143 #ifdef CONFIG_FAIR_GROUP_SCHED
9144 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9147 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
9150 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9152 struct task_group *tg = cgroup_tg(cgrp);
9154 return (u64) tg->shares;
9156 #endif /* CONFIG_FAIR_GROUP_SCHED */
9158 #ifdef CONFIG_RT_GROUP_SCHED
9159 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9162 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9165 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9167 return sched_group_rt_runtime(cgroup_tg(cgrp));
9170 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9173 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9176 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9178 return sched_group_rt_period(cgroup_tg(cgrp));
9180 #endif /* CONFIG_RT_GROUP_SCHED */
9182 static struct cftype cpu_files[] = {
9183 #ifdef CONFIG_FAIR_GROUP_SCHED
9186 .read_u64 = cpu_shares_read_u64,
9187 .write_u64 = cpu_shares_write_u64,
9190 #ifdef CONFIG_RT_GROUP_SCHED
9192 .name = "rt_runtime_us",
9193 .read_s64 = cpu_rt_runtime_read,
9194 .write_s64 = cpu_rt_runtime_write,
9197 .name = "rt_period_us",
9198 .read_u64 = cpu_rt_period_read_uint,
9199 .write_u64 = cpu_rt_period_write_uint,
9204 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9206 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9209 struct cgroup_subsys cpu_cgroup_subsys = {
9211 .create = cpu_cgroup_create,
9212 .destroy = cpu_cgroup_destroy,
9213 .can_attach = cpu_cgroup_can_attach,
9214 .attach = cpu_cgroup_attach,
9215 .populate = cpu_cgroup_populate,
9216 .subsys_id = cpu_cgroup_subsys_id,
9220 #endif /* CONFIG_CGROUP_SCHED */
9222 #ifdef CONFIG_CGROUP_CPUACCT
9225 * CPU accounting code for task groups.
9227 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9228 * (balbir@in.ibm.com).
9231 /* track cpu usage of a group of tasks and its child groups */
9233 struct cgroup_subsys_state css;
9234 /* cpuusage holds pointer to a u64-type object on every cpu */
9235 u64 __percpu *cpuusage;
9236 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
9237 struct cpuacct *parent;
9240 struct cgroup_subsys cpuacct_subsys;
9242 /* return cpu accounting group corresponding to this container */
9243 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9245 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9246 struct cpuacct, css);
9249 /* return cpu accounting group to which this task belongs */
9250 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9252 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9253 struct cpuacct, css);
9256 /* create a new cpu accounting group */
9257 static struct cgroup_subsys_state *cpuacct_create(
9258 struct cgroup_subsys *ss, struct cgroup *cgrp)
9260 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9266 ca->cpuusage = alloc_percpu(u64);
9270 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9271 if (percpu_counter_init(&ca->cpustat[i], 0))
9272 goto out_free_counters;
9275 ca->parent = cgroup_ca(cgrp->parent);
9281 percpu_counter_destroy(&ca->cpustat[i]);
9282 free_percpu(ca->cpuusage);
9286 return ERR_PTR(-ENOMEM);
9289 /* destroy an existing cpu accounting group */
9291 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9293 struct cpuacct *ca = cgroup_ca(cgrp);
9296 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9297 percpu_counter_destroy(&ca->cpustat[i]);
9298 free_percpu(ca->cpuusage);
9302 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
9304 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9307 #ifndef CONFIG_64BIT
9309 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9311 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9313 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9321 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9323 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9325 #ifndef CONFIG_64BIT
9327 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9329 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9331 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9337 /* return total cpu usage (in nanoseconds) of a group */
9338 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9340 struct cpuacct *ca = cgroup_ca(cgrp);
9341 u64 totalcpuusage = 0;
9344 for_each_present_cpu(i)
9345 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9347 return totalcpuusage;
9350 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9353 struct cpuacct *ca = cgroup_ca(cgrp);
9362 for_each_present_cpu(i)
9363 cpuacct_cpuusage_write(ca, i, 0);
9369 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9372 struct cpuacct *ca = cgroup_ca(cgroup);
9376 for_each_present_cpu(i) {
9377 percpu = cpuacct_cpuusage_read(ca, i);
9378 seq_printf(m, "%llu ", (unsigned long long) percpu);
9380 seq_printf(m, "\n");
9384 static const char *cpuacct_stat_desc[] = {
9385 [CPUACCT_STAT_USER] = "user",
9386 [CPUACCT_STAT_SYSTEM] = "system",
9389 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
9390 struct cgroup_map_cb *cb)
9392 struct cpuacct *ca = cgroup_ca(cgrp);
9395 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
9396 s64 val = percpu_counter_read(&ca->cpustat[i]);
9397 val = cputime64_to_clock_t(val);
9398 cb->fill(cb, cpuacct_stat_desc[i], val);
9403 static struct cftype files[] = {
9406 .read_u64 = cpuusage_read,
9407 .write_u64 = cpuusage_write,
9410 .name = "usage_percpu",
9411 .read_seq_string = cpuacct_percpu_seq_read,
9415 .read_map = cpuacct_stats_show,
9419 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9421 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9425 * charge this task's execution time to its accounting group.
9427 * called with rq->lock held.
9429 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9434 if (unlikely(!cpuacct_subsys.active))
9437 cpu = task_cpu(tsk);
9443 for (; ca; ca = ca->parent) {
9444 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9445 *cpuusage += cputime;
9452 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9453 * in cputime_t units. As a result, cpuacct_update_stats calls
9454 * percpu_counter_add with values large enough to always overflow the
9455 * per cpu batch limit causing bad SMP scalability.
9457 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9458 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9459 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9462 #define CPUACCT_BATCH \
9463 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9465 #define CPUACCT_BATCH 0
9469 * Charge the system/user time to the task's accounting group.
9471 static void cpuacct_update_stats(struct task_struct *tsk,
9472 enum cpuacct_stat_index idx, cputime_t val)
9475 int batch = CPUACCT_BATCH;
9477 if (unlikely(!cpuacct_subsys.active))
9484 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9490 struct cgroup_subsys cpuacct_subsys = {
9492 .create = cpuacct_create,
9493 .destroy = cpuacct_destroy,
9494 .populate = cpuacct_populate,
9495 .subsys_id = cpuacct_subsys_id,
9497 #endif /* CONFIG_CGROUP_CPUACCT */
9501 void synchronize_sched_expedited(void)
9505 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9507 #else /* #ifndef CONFIG_SMP */
9509 static atomic_t synchronize_sched_expedited_count = ATOMIC_INIT(0);
9511 static int synchronize_sched_expedited_cpu_stop(void *data)
9514 * There must be a full memory barrier on each affected CPU
9515 * between the time that try_stop_cpus() is called and the
9516 * time that it returns.
9518 * In the current initial implementation of cpu_stop, the
9519 * above condition is already met when the control reaches
9520 * this point and the following smp_mb() is not strictly
9521 * necessary. Do smp_mb() anyway for documentation and
9522 * robustness against future implementation changes.
9524 smp_mb(); /* See above comment block. */
9529 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
9530 * approach to force grace period to end quickly. This consumes
9531 * significant time on all CPUs, and is thus not recommended for
9532 * any sort of common-case code.
9534 * Note that it is illegal to call this function while holding any
9535 * lock that is acquired by a CPU-hotplug notifier. Failing to
9536 * observe this restriction will result in deadlock.
9538 void synchronize_sched_expedited(void)
9540 int snap, trycount = 0;
9542 smp_mb(); /* ensure prior mod happens before capturing snap. */
9543 snap = atomic_read(&synchronize_sched_expedited_count) + 1;
9545 while (try_stop_cpus(cpu_online_mask,
9546 synchronize_sched_expedited_cpu_stop,
9549 if (trycount++ < 10)
9550 udelay(trycount * num_online_cpus());
9552 synchronize_sched();
9555 if (atomic_read(&synchronize_sched_expedited_count) - snap > 0) {
9556 smp_mb(); /* ensure test happens before caller kfree */
9561 atomic_inc(&synchronize_sched_expedited_count);
9562 smp_mb__after_atomic_inc(); /* ensure post-GP actions seen after GP. */
9565 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9567 #endif /* #else #ifndef CONFIG_SMP */