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;
430 struct cpupri cpupri;
435 * By default the system creates a single root-domain with all cpus as
436 * members (mimicking the global state we have today).
438 static struct root_domain def_root_domain;
443 * This is the main, per-CPU runqueue data structure.
445 * Locking rule: those places that want to lock multiple runqueues
446 * (such as the load balancing or the thread migration code), lock
447 * acquire operations must be ordered by ascending &runqueue.
454 * nr_running and cpu_load should be in the same cacheline because
455 * remote CPUs use both these fields when doing load calculation.
457 unsigned long nr_running;
458 #define CPU_LOAD_IDX_MAX 5
459 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
460 unsigned long last_load_update_tick;
463 unsigned char nohz_balance_kick;
465 unsigned int skip_clock_update;
467 /* capture load from *all* tasks on this cpu: */
468 struct load_weight load;
469 unsigned long nr_load_updates;
475 #ifdef CONFIG_FAIR_GROUP_SCHED
476 /* list of leaf cfs_rq on this cpu: */
477 struct list_head leaf_cfs_rq_list;
479 #ifdef CONFIG_RT_GROUP_SCHED
480 struct list_head leaf_rt_rq_list;
484 * This is part of a global counter where only the total sum
485 * over all CPUs matters. A task can increase this counter on
486 * one CPU and if it got migrated afterwards it may decrease
487 * it on another CPU. Always updated under the runqueue lock:
489 unsigned long nr_uninterruptible;
491 struct task_struct *curr, *idle;
492 unsigned long next_balance;
493 struct mm_struct *prev_mm;
500 struct root_domain *rd;
501 struct sched_domain *sd;
503 unsigned long cpu_power;
505 unsigned char idle_at_tick;
506 /* For active balancing */
510 struct cpu_stop_work active_balance_work;
511 /* cpu of this runqueue: */
515 unsigned long avg_load_per_task;
523 /* calc_load related fields */
524 unsigned long calc_load_update;
525 long calc_load_active;
527 #ifdef CONFIG_SCHED_HRTICK
529 int hrtick_csd_pending;
530 struct call_single_data hrtick_csd;
532 struct hrtimer hrtick_timer;
535 #ifdef CONFIG_SCHEDSTATS
537 struct sched_info rq_sched_info;
538 unsigned long long rq_cpu_time;
539 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
541 /* sys_sched_yield() stats */
542 unsigned int yld_count;
544 /* schedule() stats */
545 unsigned int sched_switch;
546 unsigned int sched_count;
547 unsigned int sched_goidle;
549 /* try_to_wake_up() stats */
550 unsigned int ttwu_count;
551 unsigned int ttwu_local;
554 unsigned int bkl_count;
558 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
561 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
563 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
566 * A queue event has occurred, and we're going to schedule. In
567 * this case, we can save a useless back to back clock update.
569 if (rq->curr->se.on_rq && test_tsk_need_resched(rq->curr))
570 rq->skip_clock_update = 1;
573 static inline int cpu_of(struct rq *rq)
582 #define rcu_dereference_check_sched_domain(p) \
583 rcu_dereference_check((p), \
584 rcu_read_lock_sched_held() || \
585 lockdep_is_held(&sched_domains_mutex))
588 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
589 * See detach_destroy_domains: synchronize_sched for details.
591 * The domain tree of any CPU may only be accessed from within
592 * preempt-disabled sections.
594 #define for_each_domain(cpu, __sd) \
595 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
597 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
598 #define this_rq() (&__get_cpu_var(runqueues))
599 #define task_rq(p) cpu_rq(task_cpu(p))
600 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
601 #define raw_rq() (&__raw_get_cpu_var(runqueues))
603 #ifdef CONFIG_CGROUP_SCHED
606 * Return the group to which this tasks belongs.
608 * We use task_subsys_state_check() and extend the RCU verification
609 * with lockdep_is_held(&task_rq(p)->lock) because cpu_cgroup_attach()
610 * holds that lock for each task it moves into the cgroup. Therefore
611 * by holding that lock, we pin the task to the current cgroup.
613 static inline struct task_group *task_group(struct task_struct *p)
615 struct cgroup_subsys_state *css;
617 css = task_subsys_state_check(p, cpu_cgroup_subsys_id,
618 lockdep_is_held(&task_rq(p)->lock));
619 return container_of(css, struct task_group, css);
622 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
623 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
625 #ifdef CONFIG_FAIR_GROUP_SCHED
626 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
627 p->se.parent = task_group(p)->se[cpu];
630 #ifdef CONFIG_RT_GROUP_SCHED
631 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
632 p->rt.parent = task_group(p)->rt_se[cpu];
636 #else /* CONFIG_CGROUP_SCHED */
638 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
639 static inline struct task_group *task_group(struct task_struct *p)
644 #endif /* CONFIG_CGROUP_SCHED */
646 inline void update_rq_clock(struct rq *rq)
648 if (!rq->skip_clock_update)
649 rq->clock = sched_clock_cpu(cpu_of(rq));
653 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
655 #ifdef CONFIG_SCHED_DEBUG
656 # define const_debug __read_mostly
658 # define const_debug static const
663 * @cpu: the processor in question.
665 * Returns true if the current cpu runqueue is locked.
666 * This interface allows printk to be called with the runqueue lock
667 * held and know whether or not it is OK to wake up the klogd.
669 int runqueue_is_locked(int cpu)
671 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
675 * Debugging: various feature bits
678 #define SCHED_FEAT(name, enabled) \
679 __SCHED_FEAT_##name ,
682 #include "sched_features.h"
687 #define SCHED_FEAT(name, enabled) \
688 (1UL << __SCHED_FEAT_##name) * enabled |
690 const_debug unsigned int sysctl_sched_features =
691 #include "sched_features.h"
696 #ifdef CONFIG_SCHED_DEBUG
697 #define SCHED_FEAT(name, enabled) \
700 static __read_mostly char *sched_feat_names[] = {
701 #include "sched_features.h"
707 static int sched_feat_show(struct seq_file *m, void *v)
711 for (i = 0; sched_feat_names[i]; i++) {
712 if (!(sysctl_sched_features & (1UL << i)))
714 seq_printf(m, "%s ", sched_feat_names[i]);
722 sched_feat_write(struct file *filp, const char __user *ubuf,
723 size_t cnt, loff_t *ppos)
733 if (copy_from_user(&buf, ubuf, cnt))
739 if (strncmp(buf, "NO_", 3) == 0) {
744 for (i = 0; sched_feat_names[i]; i++) {
745 if (strcmp(cmp, sched_feat_names[i]) == 0) {
747 sysctl_sched_features &= ~(1UL << i);
749 sysctl_sched_features |= (1UL << i);
754 if (!sched_feat_names[i])
762 static int sched_feat_open(struct inode *inode, struct file *filp)
764 return single_open(filp, sched_feat_show, NULL);
767 static const struct file_operations sched_feat_fops = {
768 .open = sched_feat_open,
769 .write = sched_feat_write,
772 .release = single_release,
775 static __init int sched_init_debug(void)
777 debugfs_create_file("sched_features", 0644, NULL, NULL,
782 late_initcall(sched_init_debug);
786 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
789 * Number of tasks to iterate in a single balance run.
790 * Limited because this is done with IRQs disabled.
792 const_debug unsigned int sysctl_sched_nr_migrate = 32;
795 * ratelimit for updating the group shares.
798 unsigned int sysctl_sched_shares_ratelimit = 250000;
799 unsigned int normalized_sysctl_sched_shares_ratelimit = 250000;
802 * Inject some fuzzyness into changing the per-cpu group shares
803 * this avoids remote rq-locks at the expense of fairness.
806 unsigned int sysctl_sched_shares_thresh = 4;
809 * period over which we average the RT time consumption, measured
814 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
817 * period over which we measure -rt task cpu usage in us.
820 unsigned int sysctl_sched_rt_period = 1000000;
822 static __read_mostly int scheduler_running;
825 * part of the period that we allow rt tasks to run in us.
828 int sysctl_sched_rt_runtime = 950000;
830 static inline u64 global_rt_period(void)
832 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
835 static inline u64 global_rt_runtime(void)
837 if (sysctl_sched_rt_runtime < 0)
840 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
843 #ifndef prepare_arch_switch
844 # define prepare_arch_switch(next) do { } while (0)
846 #ifndef finish_arch_switch
847 # define finish_arch_switch(prev) do { } while (0)
850 static inline int task_current(struct rq *rq, struct task_struct *p)
852 return rq->curr == p;
855 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
856 static inline int task_running(struct rq *rq, struct task_struct *p)
858 return task_current(rq, p);
861 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
865 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
867 #ifdef CONFIG_DEBUG_SPINLOCK
868 /* this is a valid case when another task releases the spinlock */
869 rq->lock.owner = current;
872 * If we are tracking spinlock dependencies then we have to
873 * fix up the runqueue lock - which gets 'carried over' from
876 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
878 raw_spin_unlock_irq(&rq->lock);
881 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
882 static inline int task_running(struct rq *rq, struct task_struct *p)
887 return task_current(rq, p);
891 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
895 * We can optimise this out completely for !SMP, because the
896 * SMP rebalancing from interrupt is the only thing that cares
901 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
902 raw_spin_unlock_irq(&rq->lock);
904 raw_spin_unlock(&rq->lock);
908 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
912 * After ->oncpu is cleared, the task can be moved to a different CPU.
913 * We must ensure this doesn't happen until the switch is completely
919 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
923 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
926 * Check whether the task is waking, we use this to synchronize ->cpus_allowed
929 static inline int task_is_waking(struct task_struct *p)
931 return unlikely(p->state == TASK_WAKING);
935 * __task_rq_lock - lock the runqueue a given task resides on.
936 * Must be called interrupts disabled.
938 static inline struct rq *__task_rq_lock(struct task_struct *p)
945 raw_spin_lock(&rq->lock);
946 if (likely(rq == task_rq(p)))
948 raw_spin_unlock(&rq->lock);
953 * task_rq_lock - lock the runqueue a given task resides on and disable
954 * interrupts. Note the ordering: we can safely lookup the task_rq without
955 * explicitly disabling preemption.
957 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
963 local_irq_save(*flags);
965 raw_spin_lock(&rq->lock);
966 if (likely(rq == task_rq(p)))
968 raw_spin_unlock_irqrestore(&rq->lock, *flags);
972 static void __task_rq_unlock(struct rq *rq)
975 raw_spin_unlock(&rq->lock);
978 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
981 raw_spin_unlock_irqrestore(&rq->lock, *flags);
985 * this_rq_lock - lock this runqueue and disable interrupts.
987 static struct rq *this_rq_lock(void)
994 raw_spin_lock(&rq->lock);
999 #ifdef CONFIG_SCHED_HRTICK
1001 * Use HR-timers to deliver accurate preemption points.
1003 * Its all a bit involved since we cannot program an hrt while holding the
1004 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1007 * When we get rescheduled we reprogram the hrtick_timer outside of the
1013 * - enabled by features
1014 * - hrtimer is actually high res
1016 static inline int hrtick_enabled(struct rq *rq)
1018 if (!sched_feat(HRTICK))
1020 if (!cpu_active(cpu_of(rq)))
1022 return hrtimer_is_hres_active(&rq->hrtick_timer);
1025 static void hrtick_clear(struct rq *rq)
1027 if (hrtimer_active(&rq->hrtick_timer))
1028 hrtimer_cancel(&rq->hrtick_timer);
1032 * High-resolution timer tick.
1033 * Runs from hardirq context with interrupts disabled.
1035 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1037 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1039 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1041 raw_spin_lock(&rq->lock);
1042 update_rq_clock(rq);
1043 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1044 raw_spin_unlock(&rq->lock);
1046 return HRTIMER_NORESTART;
1051 * called from hardirq (IPI) context
1053 static void __hrtick_start(void *arg)
1055 struct rq *rq = arg;
1057 raw_spin_lock(&rq->lock);
1058 hrtimer_restart(&rq->hrtick_timer);
1059 rq->hrtick_csd_pending = 0;
1060 raw_spin_unlock(&rq->lock);
1064 * Called to set the hrtick timer state.
1066 * called with rq->lock held and irqs disabled
1068 static void hrtick_start(struct rq *rq, u64 delay)
1070 struct hrtimer *timer = &rq->hrtick_timer;
1071 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1073 hrtimer_set_expires(timer, time);
1075 if (rq == this_rq()) {
1076 hrtimer_restart(timer);
1077 } else if (!rq->hrtick_csd_pending) {
1078 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1079 rq->hrtick_csd_pending = 1;
1084 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1086 int cpu = (int)(long)hcpu;
1089 case CPU_UP_CANCELED:
1090 case CPU_UP_CANCELED_FROZEN:
1091 case CPU_DOWN_PREPARE:
1092 case CPU_DOWN_PREPARE_FROZEN:
1094 case CPU_DEAD_FROZEN:
1095 hrtick_clear(cpu_rq(cpu));
1102 static __init void init_hrtick(void)
1104 hotcpu_notifier(hotplug_hrtick, 0);
1108 * Called to set the hrtick timer state.
1110 * called with rq->lock held and irqs disabled
1112 static void hrtick_start(struct rq *rq, u64 delay)
1114 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1115 HRTIMER_MODE_REL_PINNED, 0);
1118 static inline void init_hrtick(void)
1121 #endif /* CONFIG_SMP */
1123 static void init_rq_hrtick(struct rq *rq)
1126 rq->hrtick_csd_pending = 0;
1128 rq->hrtick_csd.flags = 0;
1129 rq->hrtick_csd.func = __hrtick_start;
1130 rq->hrtick_csd.info = rq;
1133 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1134 rq->hrtick_timer.function = hrtick;
1136 #else /* CONFIG_SCHED_HRTICK */
1137 static inline void hrtick_clear(struct rq *rq)
1141 static inline void init_rq_hrtick(struct rq *rq)
1145 static inline void init_hrtick(void)
1148 #endif /* CONFIG_SCHED_HRTICK */
1151 * resched_task - mark a task 'to be rescheduled now'.
1153 * On UP this means the setting of the need_resched flag, on SMP it
1154 * might also involve a cross-CPU call to trigger the scheduler on
1159 #ifndef tsk_is_polling
1160 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1163 static void resched_task(struct task_struct *p)
1167 assert_raw_spin_locked(&task_rq(p)->lock);
1169 if (test_tsk_need_resched(p))
1172 set_tsk_need_resched(p);
1175 if (cpu == smp_processor_id())
1178 /* NEED_RESCHED must be visible before we test polling */
1180 if (!tsk_is_polling(p))
1181 smp_send_reschedule(cpu);
1184 static void resched_cpu(int cpu)
1186 struct rq *rq = cpu_rq(cpu);
1187 unsigned long flags;
1189 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1191 resched_task(cpu_curr(cpu));
1192 raw_spin_unlock_irqrestore(&rq->lock, flags);
1197 * In the semi idle case, use the nearest busy cpu for migrating timers
1198 * from an idle cpu. This is good for power-savings.
1200 * We don't do similar optimization for completely idle system, as
1201 * selecting an idle cpu will add more delays to the timers than intended
1202 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1204 int get_nohz_timer_target(void)
1206 int cpu = smp_processor_id();
1208 struct sched_domain *sd;
1210 for_each_domain(cpu, sd) {
1211 for_each_cpu(i, sched_domain_span(sd))
1218 * When add_timer_on() enqueues a timer into the timer wheel of an
1219 * idle CPU then this timer might expire before the next timer event
1220 * which is scheduled to wake up that CPU. In case of a completely
1221 * idle system the next event might even be infinite time into the
1222 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1223 * leaves the inner idle loop so the newly added timer is taken into
1224 * account when the CPU goes back to idle and evaluates the timer
1225 * wheel for the next timer event.
1227 void wake_up_idle_cpu(int cpu)
1229 struct rq *rq = cpu_rq(cpu);
1231 if (cpu == smp_processor_id())
1235 * This is safe, as this function is called with the timer
1236 * wheel base lock of (cpu) held. When the CPU is on the way
1237 * to idle and has not yet set rq->curr to idle then it will
1238 * be serialized on the timer wheel base lock and take the new
1239 * timer into account automatically.
1241 if (rq->curr != rq->idle)
1245 * We can set TIF_RESCHED on the idle task of the other CPU
1246 * lockless. The worst case is that the other CPU runs the
1247 * idle task through an additional NOOP schedule()
1249 set_tsk_need_resched(rq->idle);
1251 /* NEED_RESCHED must be visible before we test polling */
1253 if (!tsk_is_polling(rq->idle))
1254 smp_send_reschedule(cpu);
1257 #endif /* CONFIG_NO_HZ */
1259 static u64 sched_avg_period(void)
1261 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1264 static void sched_avg_update(struct rq *rq)
1266 s64 period = sched_avg_period();
1268 while ((s64)(rq->clock - rq->age_stamp) > period) {
1270 * Inline assembly required to prevent the compiler
1271 * optimising this loop into a divmod call.
1272 * See __iter_div_u64_rem() for another example of this.
1274 asm("" : "+rm" (rq->age_stamp));
1275 rq->age_stamp += period;
1280 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1282 rq->rt_avg += rt_delta;
1283 sched_avg_update(rq);
1286 #else /* !CONFIG_SMP */
1287 static void resched_task(struct task_struct *p)
1289 assert_raw_spin_locked(&task_rq(p)->lock);
1290 set_tsk_need_resched(p);
1293 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1297 static void sched_avg_update(struct rq *rq)
1300 #endif /* CONFIG_SMP */
1302 #if BITS_PER_LONG == 32
1303 # define WMULT_CONST (~0UL)
1305 # define WMULT_CONST (1UL << 32)
1308 #define WMULT_SHIFT 32
1311 * Shift right and round:
1313 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1316 * delta *= weight / lw
1318 static unsigned long
1319 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1320 struct load_weight *lw)
1324 if (!lw->inv_weight) {
1325 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1328 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1332 tmp = (u64)delta_exec * weight;
1334 * Check whether we'd overflow the 64-bit multiplication:
1336 if (unlikely(tmp > WMULT_CONST))
1337 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1340 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1342 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1345 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1351 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1358 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1359 * of tasks with abnormal "nice" values across CPUs the contribution that
1360 * each task makes to its run queue's load is weighted according to its
1361 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1362 * scaled version of the new time slice allocation that they receive on time
1366 #define WEIGHT_IDLEPRIO 3
1367 #define WMULT_IDLEPRIO 1431655765
1370 * Nice levels are multiplicative, with a gentle 10% change for every
1371 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1372 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1373 * that remained on nice 0.
1375 * The "10% effect" is relative and cumulative: from _any_ nice level,
1376 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1377 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1378 * If a task goes up by ~10% and another task goes down by ~10% then
1379 * the relative distance between them is ~25%.)
1381 static const int prio_to_weight[40] = {
1382 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1383 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1384 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1385 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1386 /* 0 */ 1024, 820, 655, 526, 423,
1387 /* 5 */ 335, 272, 215, 172, 137,
1388 /* 10 */ 110, 87, 70, 56, 45,
1389 /* 15 */ 36, 29, 23, 18, 15,
1393 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1395 * In cases where the weight does not change often, we can use the
1396 * precalculated inverse to speed up arithmetics by turning divisions
1397 * into multiplications:
1399 static const u32 prio_to_wmult[40] = {
1400 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1401 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1402 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1403 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1404 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1405 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1406 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1407 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1410 /* Time spent by the tasks of the cpu accounting group executing in ... */
1411 enum cpuacct_stat_index {
1412 CPUACCT_STAT_USER, /* ... user mode */
1413 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1415 CPUACCT_STAT_NSTATS,
1418 #ifdef CONFIG_CGROUP_CPUACCT
1419 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1420 static void cpuacct_update_stats(struct task_struct *tsk,
1421 enum cpuacct_stat_index idx, cputime_t val);
1423 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1424 static inline void cpuacct_update_stats(struct task_struct *tsk,
1425 enum cpuacct_stat_index idx, cputime_t val) {}
1428 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1430 update_load_add(&rq->load, load);
1433 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1435 update_load_sub(&rq->load, load);
1438 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1439 typedef int (*tg_visitor)(struct task_group *, void *);
1442 * Iterate the full tree, calling @down when first entering a node and @up when
1443 * leaving it for the final time.
1445 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1447 struct task_group *parent, *child;
1451 parent = &root_task_group;
1453 ret = (*down)(parent, data);
1456 list_for_each_entry_rcu(child, &parent->children, siblings) {
1463 ret = (*up)(parent, data);
1468 parent = parent->parent;
1477 static int tg_nop(struct task_group *tg, void *data)
1484 /* Used instead of source_load when we know the type == 0 */
1485 static unsigned long weighted_cpuload(const int cpu)
1487 return cpu_rq(cpu)->load.weight;
1491 * Return a low guess at the load of a migration-source cpu weighted
1492 * according to the scheduling class and "nice" value.
1494 * We want to under-estimate the load of migration sources, to
1495 * balance conservatively.
1497 static unsigned long source_load(int cpu, int type)
1499 struct rq *rq = cpu_rq(cpu);
1500 unsigned long total = weighted_cpuload(cpu);
1502 if (type == 0 || !sched_feat(LB_BIAS))
1505 return min(rq->cpu_load[type-1], total);
1509 * Return a high guess at the load of a migration-target cpu weighted
1510 * according to the scheduling class and "nice" value.
1512 static unsigned long target_load(int cpu, int type)
1514 struct rq *rq = cpu_rq(cpu);
1515 unsigned long total = weighted_cpuload(cpu);
1517 if (type == 0 || !sched_feat(LB_BIAS))
1520 return max(rq->cpu_load[type-1], total);
1523 static unsigned long power_of(int cpu)
1525 return cpu_rq(cpu)->cpu_power;
1528 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1530 static unsigned long cpu_avg_load_per_task(int cpu)
1532 struct rq *rq = cpu_rq(cpu);
1533 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1536 rq->avg_load_per_task = rq->load.weight / nr_running;
1538 rq->avg_load_per_task = 0;
1540 return rq->avg_load_per_task;
1543 #ifdef CONFIG_FAIR_GROUP_SCHED
1545 static __read_mostly unsigned long __percpu *update_shares_data;
1547 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1550 * Calculate and set the cpu's group shares.
1552 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1553 unsigned long sd_shares,
1554 unsigned long sd_rq_weight,
1555 unsigned long *usd_rq_weight)
1557 unsigned long shares, rq_weight;
1560 rq_weight = usd_rq_weight[cpu];
1563 rq_weight = NICE_0_LOAD;
1567 * \Sum_j shares_j * rq_weight_i
1568 * shares_i = -----------------------------
1569 * \Sum_j rq_weight_j
1571 shares = (sd_shares * rq_weight) / sd_rq_weight;
1572 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1574 if (abs(shares - tg->se[cpu]->load.weight) >
1575 sysctl_sched_shares_thresh) {
1576 struct rq *rq = cpu_rq(cpu);
1577 unsigned long flags;
1579 raw_spin_lock_irqsave(&rq->lock, flags);
1580 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1581 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1582 __set_se_shares(tg->se[cpu], shares);
1583 raw_spin_unlock_irqrestore(&rq->lock, flags);
1588 * Re-compute the task group their per cpu shares over the given domain.
1589 * This needs to be done in a bottom-up fashion because the rq weight of a
1590 * parent group depends on the shares of its child groups.
1592 static int tg_shares_up(struct task_group *tg, void *data)
1594 unsigned long weight, rq_weight = 0, sum_weight = 0, shares = 0;
1595 unsigned long *usd_rq_weight;
1596 struct sched_domain *sd = data;
1597 unsigned long flags;
1603 local_irq_save(flags);
1604 usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id());
1606 for_each_cpu(i, sched_domain_span(sd)) {
1607 weight = tg->cfs_rq[i]->load.weight;
1608 usd_rq_weight[i] = weight;
1610 rq_weight += weight;
1612 * If there are currently no tasks on the cpu pretend there
1613 * is one of average load so that when a new task gets to
1614 * run here it will not get delayed by group starvation.
1617 weight = NICE_0_LOAD;
1619 sum_weight += weight;
1620 shares += tg->cfs_rq[i]->shares;
1624 rq_weight = sum_weight;
1626 if ((!shares && rq_weight) || shares > tg->shares)
1627 shares = tg->shares;
1629 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1630 shares = tg->shares;
1632 for_each_cpu(i, sched_domain_span(sd))
1633 update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight);
1635 local_irq_restore(flags);
1641 * Compute the cpu's hierarchical load factor for each task group.
1642 * This needs to be done in a top-down fashion because the load of a child
1643 * group is a fraction of its parents load.
1645 static int tg_load_down(struct task_group *tg, void *data)
1648 long cpu = (long)data;
1651 load = cpu_rq(cpu)->load.weight;
1653 load = tg->parent->cfs_rq[cpu]->h_load;
1654 load *= tg->cfs_rq[cpu]->shares;
1655 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1658 tg->cfs_rq[cpu]->h_load = load;
1663 static void update_shares(struct sched_domain *sd)
1668 if (root_task_group_empty())
1671 now = local_clock();
1672 elapsed = now - sd->last_update;
1674 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1675 sd->last_update = now;
1676 walk_tg_tree(tg_nop, tg_shares_up, sd);
1680 static void update_h_load(long cpu)
1682 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1687 static inline void update_shares(struct sched_domain *sd)
1693 #ifdef CONFIG_PREEMPT
1695 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1698 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1699 * way at the expense of forcing extra atomic operations in all
1700 * invocations. This assures that the double_lock is acquired using the
1701 * same underlying policy as the spinlock_t on this architecture, which
1702 * reduces latency compared to the unfair variant below. However, it
1703 * also adds more overhead and therefore may reduce throughput.
1705 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1706 __releases(this_rq->lock)
1707 __acquires(busiest->lock)
1708 __acquires(this_rq->lock)
1710 raw_spin_unlock(&this_rq->lock);
1711 double_rq_lock(this_rq, busiest);
1718 * Unfair double_lock_balance: Optimizes throughput at the expense of
1719 * latency by eliminating extra atomic operations when the locks are
1720 * already in proper order on entry. This favors lower cpu-ids and will
1721 * grant the double lock to lower cpus over higher ids under contention,
1722 * regardless of entry order into the function.
1724 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1725 __releases(this_rq->lock)
1726 __acquires(busiest->lock)
1727 __acquires(this_rq->lock)
1731 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1732 if (busiest < this_rq) {
1733 raw_spin_unlock(&this_rq->lock);
1734 raw_spin_lock(&busiest->lock);
1735 raw_spin_lock_nested(&this_rq->lock,
1736 SINGLE_DEPTH_NESTING);
1739 raw_spin_lock_nested(&busiest->lock,
1740 SINGLE_DEPTH_NESTING);
1745 #endif /* CONFIG_PREEMPT */
1748 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1750 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1752 if (unlikely(!irqs_disabled())) {
1753 /* printk() doesn't work good under rq->lock */
1754 raw_spin_unlock(&this_rq->lock);
1758 return _double_lock_balance(this_rq, busiest);
1761 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1762 __releases(busiest->lock)
1764 raw_spin_unlock(&busiest->lock);
1765 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1769 * double_rq_lock - safely lock two runqueues
1771 * Note this does not disable interrupts like task_rq_lock,
1772 * you need to do so manually before calling.
1774 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1775 __acquires(rq1->lock)
1776 __acquires(rq2->lock)
1778 BUG_ON(!irqs_disabled());
1780 raw_spin_lock(&rq1->lock);
1781 __acquire(rq2->lock); /* Fake it out ;) */
1784 raw_spin_lock(&rq1->lock);
1785 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1787 raw_spin_lock(&rq2->lock);
1788 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1794 * double_rq_unlock - safely unlock two runqueues
1796 * Note this does not restore interrupts like task_rq_unlock,
1797 * you need to do so manually after calling.
1799 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1800 __releases(rq1->lock)
1801 __releases(rq2->lock)
1803 raw_spin_unlock(&rq1->lock);
1805 raw_spin_unlock(&rq2->lock);
1807 __release(rq2->lock);
1812 #ifdef CONFIG_FAIR_GROUP_SCHED
1813 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1816 cfs_rq->shares = shares;
1821 static void calc_load_account_idle(struct rq *this_rq);
1822 static void update_sysctl(void);
1823 static int get_update_sysctl_factor(void);
1824 static void update_cpu_load(struct rq *this_rq);
1826 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1828 set_task_rq(p, cpu);
1831 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1832 * successfuly executed on another CPU. We must ensure that updates of
1833 * per-task data have been completed by this moment.
1836 task_thread_info(p)->cpu = cpu;
1840 static const struct sched_class rt_sched_class;
1842 #define sched_class_highest (&rt_sched_class)
1843 #define for_each_class(class) \
1844 for (class = sched_class_highest; class; class = class->next)
1846 #include "sched_stats.h"
1848 static void inc_nr_running(struct rq *rq)
1853 static void dec_nr_running(struct rq *rq)
1858 static void set_load_weight(struct task_struct *p)
1861 * SCHED_IDLE tasks get minimal weight:
1863 if (p->policy == SCHED_IDLE) {
1864 p->se.load.weight = WEIGHT_IDLEPRIO;
1865 p->se.load.inv_weight = WMULT_IDLEPRIO;
1869 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1870 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1873 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1875 update_rq_clock(rq);
1876 sched_info_queued(p);
1877 p->sched_class->enqueue_task(rq, p, flags);
1881 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1883 update_rq_clock(rq);
1884 sched_info_dequeued(p);
1885 p->sched_class->dequeue_task(rq, p, flags);
1890 * activate_task - move a task to the runqueue.
1892 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1894 if (task_contributes_to_load(p))
1895 rq->nr_uninterruptible--;
1897 enqueue_task(rq, p, flags);
1902 * deactivate_task - remove a task from the runqueue.
1904 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1906 if (task_contributes_to_load(p))
1907 rq->nr_uninterruptible++;
1909 dequeue_task(rq, p, flags);
1913 #include "sched_idletask.c"
1914 #include "sched_fair.c"
1915 #include "sched_rt.c"
1916 #ifdef CONFIG_SCHED_DEBUG
1917 # include "sched_debug.c"
1921 * __normal_prio - return the priority that is based on the static prio
1923 static inline int __normal_prio(struct task_struct *p)
1925 return p->static_prio;
1929 * Calculate the expected normal priority: i.e. priority
1930 * without taking RT-inheritance into account. Might be
1931 * boosted by interactivity modifiers. Changes upon fork,
1932 * setprio syscalls, and whenever the interactivity
1933 * estimator recalculates.
1935 static inline int normal_prio(struct task_struct *p)
1939 if (task_has_rt_policy(p))
1940 prio = MAX_RT_PRIO-1 - p->rt_priority;
1942 prio = __normal_prio(p);
1947 * Calculate the current priority, i.e. the priority
1948 * taken into account by the scheduler. This value might
1949 * be boosted by RT tasks, or might be boosted by
1950 * interactivity modifiers. Will be RT if the task got
1951 * RT-boosted. If not then it returns p->normal_prio.
1953 static int effective_prio(struct task_struct *p)
1955 p->normal_prio = normal_prio(p);
1957 * If we are RT tasks or we were boosted to RT priority,
1958 * keep the priority unchanged. Otherwise, update priority
1959 * to the normal priority:
1961 if (!rt_prio(p->prio))
1962 return p->normal_prio;
1967 * task_curr - is this task currently executing on a CPU?
1968 * @p: the task in question.
1970 inline int task_curr(const struct task_struct *p)
1972 return cpu_curr(task_cpu(p)) == p;
1975 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1976 const struct sched_class *prev_class,
1977 int oldprio, int running)
1979 if (prev_class != p->sched_class) {
1980 if (prev_class->switched_from)
1981 prev_class->switched_from(rq, p, running);
1982 p->sched_class->switched_to(rq, p, running);
1984 p->sched_class->prio_changed(rq, p, oldprio, running);
1989 * Is this task likely cache-hot:
1992 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1996 if (p->sched_class != &fair_sched_class)
2000 * Buddy candidates are cache hot:
2002 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2003 (&p->se == cfs_rq_of(&p->se)->next ||
2004 &p->se == cfs_rq_of(&p->se)->last))
2007 if (sysctl_sched_migration_cost == -1)
2009 if (sysctl_sched_migration_cost == 0)
2012 delta = now - p->se.exec_start;
2014 return delta < (s64)sysctl_sched_migration_cost;
2017 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2019 #ifdef CONFIG_SCHED_DEBUG
2021 * We should never call set_task_cpu() on a blocked task,
2022 * ttwu() will sort out the placement.
2024 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2025 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2028 trace_sched_migrate_task(p, new_cpu);
2030 if (task_cpu(p) != new_cpu) {
2031 p->se.nr_migrations++;
2032 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2035 __set_task_cpu(p, new_cpu);
2038 struct migration_arg {
2039 struct task_struct *task;
2043 static int migration_cpu_stop(void *data);
2046 * The task's runqueue lock must be held.
2047 * Returns true if you have to wait for migration thread.
2049 static bool migrate_task(struct task_struct *p, int dest_cpu)
2051 struct rq *rq = task_rq(p);
2054 * If the task is not on a runqueue (and not running), then
2055 * the next wake-up will properly place the task.
2057 return p->se.on_rq || task_running(rq, p);
2061 * wait_task_inactive - wait for a thread to unschedule.
2063 * If @match_state is nonzero, it's the @p->state value just checked and
2064 * not expected to change. If it changes, i.e. @p might have woken up,
2065 * then return zero. When we succeed in waiting for @p to be off its CPU,
2066 * we return a positive number (its total switch count). If a second call
2067 * a short while later returns the same number, the caller can be sure that
2068 * @p has remained unscheduled the whole time.
2070 * The caller must ensure that the task *will* unschedule sometime soon,
2071 * else this function might spin for a *long* time. This function can't
2072 * be called with interrupts off, or it may introduce deadlock with
2073 * smp_call_function() if an IPI is sent by the same process we are
2074 * waiting to become inactive.
2076 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2078 unsigned long flags;
2085 * We do the initial early heuristics without holding
2086 * any task-queue locks at all. We'll only try to get
2087 * the runqueue lock when things look like they will
2093 * If the task is actively running on another CPU
2094 * still, just relax and busy-wait without holding
2097 * NOTE! Since we don't hold any locks, it's not
2098 * even sure that "rq" stays as the right runqueue!
2099 * But we don't care, since "task_running()" will
2100 * return false if the runqueue has changed and p
2101 * is actually now running somewhere else!
2103 while (task_running(rq, p)) {
2104 if (match_state && unlikely(p->state != match_state))
2110 * Ok, time to look more closely! We need the rq
2111 * lock now, to be *sure*. If we're wrong, we'll
2112 * just go back and repeat.
2114 rq = task_rq_lock(p, &flags);
2115 trace_sched_wait_task(p);
2116 running = task_running(rq, p);
2117 on_rq = p->se.on_rq;
2119 if (!match_state || p->state == match_state)
2120 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2121 task_rq_unlock(rq, &flags);
2124 * If it changed from the expected state, bail out now.
2126 if (unlikely(!ncsw))
2130 * Was it really running after all now that we
2131 * checked with the proper locks actually held?
2133 * Oops. Go back and try again..
2135 if (unlikely(running)) {
2141 * It's not enough that it's not actively running,
2142 * it must be off the runqueue _entirely_, and not
2145 * So if it was still runnable (but just not actively
2146 * running right now), it's preempted, and we should
2147 * yield - it could be a while.
2149 if (unlikely(on_rq)) {
2150 schedule_timeout_uninterruptible(1);
2155 * Ahh, all good. It wasn't running, and it wasn't
2156 * runnable, which means that it will never become
2157 * running in the future either. We're all done!
2166 * kick_process - kick a running thread to enter/exit the kernel
2167 * @p: the to-be-kicked thread
2169 * Cause a process which is running on another CPU to enter
2170 * kernel-mode, without any delay. (to get signals handled.)
2172 * NOTE: this function doesnt have to take the runqueue lock,
2173 * because all it wants to ensure is that the remote task enters
2174 * the kernel. If the IPI races and the task has been migrated
2175 * to another CPU then no harm is done and the purpose has been
2178 void kick_process(struct task_struct *p)
2184 if ((cpu != smp_processor_id()) && task_curr(p))
2185 smp_send_reschedule(cpu);
2188 EXPORT_SYMBOL_GPL(kick_process);
2189 #endif /* CONFIG_SMP */
2192 * task_oncpu_function_call - call a function on the cpu on which a task runs
2193 * @p: the task to evaluate
2194 * @func: the function to be called
2195 * @info: the function call argument
2197 * Calls the function @func when the task is currently running. This might
2198 * be on the current CPU, which just calls the function directly
2200 void task_oncpu_function_call(struct task_struct *p,
2201 void (*func) (void *info), void *info)
2208 smp_call_function_single(cpu, func, info, 1);
2214 * ->cpus_allowed is protected by either TASK_WAKING or rq->lock held.
2216 static int select_fallback_rq(int cpu, struct task_struct *p)
2219 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2221 /* Look for allowed, online CPU in same node. */
2222 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2223 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2226 /* Any allowed, online CPU? */
2227 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2228 if (dest_cpu < nr_cpu_ids)
2231 /* No more Mr. Nice Guy. */
2232 if (unlikely(dest_cpu >= nr_cpu_ids)) {
2233 dest_cpu = cpuset_cpus_allowed_fallback(p);
2235 * Don't tell them about moving exiting tasks or
2236 * kernel threads (both mm NULL), since they never
2239 if (p->mm && printk_ratelimit()) {
2240 printk(KERN_INFO "process %d (%s) no "
2241 "longer affine to cpu%d\n",
2242 task_pid_nr(p), p->comm, cpu);
2250 * The caller (fork, wakeup) owns TASK_WAKING, ->cpus_allowed is stable.
2253 int select_task_rq(struct rq *rq, struct task_struct *p, int sd_flags, int wake_flags)
2255 int cpu = p->sched_class->select_task_rq(rq, p, sd_flags, wake_flags);
2258 * In order not to call set_task_cpu() on a blocking task we need
2259 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2262 * Since this is common to all placement strategies, this lives here.
2264 * [ this allows ->select_task() to simply return task_cpu(p) and
2265 * not worry about this generic constraint ]
2267 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2269 cpu = select_fallback_rq(task_cpu(p), p);
2274 static void update_avg(u64 *avg, u64 sample)
2276 s64 diff = sample - *avg;
2281 static inline void ttwu_activate(struct task_struct *p, struct rq *rq,
2282 bool is_sync, bool is_migrate, bool is_local,
2283 unsigned long en_flags)
2285 schedstat_inc(p, se.statistics.nr_wakeups);
2287 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2289 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2291 schedstat_inc(p, se.statistics.nr_wakeups_local);
2293 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2295 activate_task(rq, p, en_flags);
2298 static inline void ttwu_post_activation(struct task_struct *p, struct rq *rq,
2299 int wake_flags, bool success)
2301 trace_sched_wakeup(p, success);
2302 check_preempt_curr(rq, p, wake_flags);
2304 p->state = TASK_RUNNING;
2306 if (p->sched_class->task_woken)
2307 p->sched_class->task_woken(rq, p);
2309 if (unlikely(rq->idle_stamp)) {
2310 u64 delta = rq->clock - rq->idle_stamp;
2311 u64 max = 2*sysctl_sched_migration_cost;
2316 update_avg(&rq->avg_idle, delta);
2320 /* if a worker is waking up, notify workqueue */
2321 if ((p->flags & PF_WQ_WORKER) && success)
2322 wq_worker_waking_up(p, cpu_of(rq));
2326 * try_to_wake_up - wake up a thread
2327 * @p: the thread to be awakened
2328 * @state: the mask of task states that can be woken
2329 * @wake_flags: wake modifier flags (WF_*)
2331 * Put it on the run-queue if it's not already there. The "current"
2332 * thread is always on the run-queue (except when the actual
2333 * re-schedule is in progress), and as such you're allowed to do
2334 * the simpler "current->state = TASK_RUNNING" to mark yourself
2335 * runnable without the overhead of this.
2337 * Returns %true if @p was woken up, %false if it was already running
2338 * or @state didn't match @p's state.
2340 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2343 int cpu, orig_cpu, this_cpu, success = 0;
2344 unsigned long flags;
2345 unsigned long en_flags = ENQUEUE_WAKEUP;
2348 this_cpu = get_cpu();
2351 rq = task_rq_lock(p, &flags);
2352 if (!(p->state & state))
2362 if (unlikely(task_running(rq, p)))
2366 * In order to handle concurrent wakeups and release the rq->lock
2367 * we put the task in TASK_WAKING state.
2369 * First fix up the nr_uninterruptible count:
2371 if (task_contributes_to_load(p)) {
2372 if (likely(cpu_online(orig_cpu)))
2373 rq->nr_uninterruptible--;
2375 this_rq()->nr_uninterruptible--;
2377 p->state = TASK_WAKING;
2379 if (p->sched_class->task_waking) {
2380 p->sched_class->task_waking(rq, p);
2381 en_flags |= ENQUEUE_WAKING;
2384 cpu = select_task_rq(rq, p, SD_BALANCE_WAKE, wake_flags);
2385 if (cpu != orig_cpu)
2386 set_task_cpu(p, cpu);
2387 __task_rq_unlock(rq);
2390 raw_spin_lock(&rq->lock);
2393 * We migrated the task without holding either rq->lock, however
2394 * since the task is not on the task list itself, nobody else
2395 * will try and migrate the task, hence the rq should match the
2396 * cpu we just moved it to.
2398 WARN_ON(task_cpu(p) != cpu);
2399 WARN_ON(p->state != TASK_WAKING);
2401 #ifdef CONFIG_SCHEDSTATS
2402 schedstat_inc(rq, ttwu_count);
2403 if (cpu == this_cpu)
2404 schedstat_inc(rq, ttwu_local);
2406 struct sched_domain *sd;
2407 for_each_domain(this_cpu, sd) {
2408 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2409 schedstat_inc(sd, ttwu_wake_remote);
2414 #endif /* CONFIG_SCHEDSTATS */
2417 #endif /* CONFIG_SMP */
2418 ttwu_activate(p, rq, wake_flags & WF_SYNC, orig_cpu != cpu,
2419 cpu == this_cpu, en_flags);
2422 ttwu_post_activation(p, rq, wake_flags, success);
2424 task_rq_unlock(rq, &flags);
2431 * try_to_wake_up_local - try to wake up a local task with rq lock held
2432 * @p: the thread to be awakened
2434 * Put @p on the run-queue if it's not alredy there. The caller must
2435 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2436 * the current task. this_rq() stays locked over invocation.
2438 static void try_to_wake_up_local(struct task_struct *p)
2440 struct rq *rq = task_rq(p);
2441 bool success = false;
2443 BUG_ON(rq != this_rq());
2444 BUG_ON(p == current);
2445 lockdep_assert_held(&rq->lock);
2447 if (!(p->state & TASK_NORMAL))
2451 if (likely(!task_running(rq, p))) {
2452 schedstat_inc(rq, ttwu_count);
2453 schedstat_inc(rq, ttwu_local);
2455 ttwu_activate(p, rq, false, false, true, ENQUEUE_WAKEUP);
2458 ttwu_post_activation(p, rq, 0, success);
2462 * wake_up_process - Wake up a specific process
2463 * @p: The process to be woken up.
2465 * Attempt to wake up the nominated process and move it to the set of runnable
2466 * processes. Returns 1 if the process was woken up, 0 if it was already
2469 * It may be assumed that this function implies a write memory barrier before
2470 * changing the task state if and only if any tasks are woken up.
2472 int wake_up_process(struct task_struct *p)
2474 return try_to_wake_up(p, TASK_ALL, 0);
2476 EXPORT_SYMBOL(wake_up_process);
2478 int wake_up_state(struct task_struct *p, unsigned int state)
2480 return try_to_wake_up(p, state, 0);
2484 * Perform scheduler related setup for a newly forked process p.
2485 * p is forked by current.
2487 * __sched_fork() is basic setup used by init_idle() too:
2489 static void __sched_fork(struct task_struct *p)
2491 p->se.exec_start = 0;
2492 p->se.sum_exec_runtime = 0;
2493 p->se.prev_sum_exec_runtime = 0;
2494 p->se.nr_migrations = 0;
2496 #ifdef CONFIG_SCHEDSTATS
2497 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2500 INIT_LIST_HEAD(&p->rt.run_list);
2502 INIT_LIST_HEAD(&p->se.group_node);
2504 #ifdef CONFIG_PREEMPT_NOTIFIERS
2505 INIT_HLIST_HEAD(&p->preempt_notifiers);
2510 * fork()/clone()-time setup:
2512 void sched_fork(struct task_struct *p, int clone_flags)
2514 int cpu = get_cpu();
2518 * We mark the process as running here. This guarantees that
2519 * nobody will actually run it, and a signal or other external
2520 * event cannot wake it up and insert it on the runqueue either.
2522 p->state = TASK_RUNNING;
2525 * Revert to default priority/policy on fork if requested.
2527 if (unlikely(p->sched_reset_on_fork)) {
2528 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2529 p->policy = SCHED_NORMAL;
2530 p->normal_prio = p->static_prio;
2533 if (PRIO_TO_NICE(p->static_prio) < 0) {
2534 p->static_prio = NICE_TO_PRIO(0);
2535 p->normal_prio = p->static_prio;
2540 * We don't need the reset flag anymore after the fork. It has
2541 * fulfilled its duty:
2543 p->sched_reset_on_fork = 0;
2547 * Make sure we do not leak PI boosting priority to the child.
2549 p->prio = current->normal_prio;
2551 if (!rt_prio(p->prio))
2552 p->sched_class = &fair_sched_class;
2554 if (p->sched_class->task_fork)
2555 p->sched_class->task_fork(p);
2558 * The child is not yet in the pid-hash so no cgroup attach races,
2559 * and the cgroup is pinned to this child due to cgroup_fork()
2560 * is ran before sched_fork().
2562 * Silence PROVE_RCU.
2565 set_task_cpu(p, cpu);
2568 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2569 if (likely(sched_info_on()))
2570 memset(&p->sched_info, 0, sizeof(p->sched_info));
2572 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2575 #ifdef CONFIG_PREEMPT
2576 /* Want to start with kernel preemption disabled. */
2577 task_thread_info(p)->preempt_count = 1;
2579 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2585 * wake_up_new_task - wake up a newly created task for the first time.
2587 * This function will do some initial scheduler statistics housekeeping
2588 * that must be done for every newly created context, then puts the task
2589 * on the runqueue and wakes it.
2591 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2593 unsigned long flags;
2595 int cpu __maybe_unused = get_cpu();
2598 rq = task_rq_lock(p, &flags);
2599 p->state = TASK_WAKING;
2602 * Fork balancing, do it here and not earlier because:
2603 * - cpus_allowed can change in the fork path
2604 * - any previously selected cpu might disappear through hotplug
2606 * We set TASK_WAKING so that select_task_rq() can drop rq->lock
2607 * without people poking at ->cpus_allowed.
2609 cpu = select_task_rq(rq, p, SD_BALANCE_FORK, 0);
2610 set_task_cpu(p, cpu);
2612 p->state = TASK_RUNNING;
2613 task_rq_unlock(rq, &flags);
2616 rq = task_rq_lock(p, &flags);
2617 activate_task(rq, p, 0);
2618 trace_sched_wakeup_new(p, 1);
2619 check_preempt_curr(rq, p, WF_FORK);
2621 if (p->sched_class->task_woken)
2622 p->sched_class->task_woken(rq, p);
2624 task_rq_unlock(rq, &flags);
2628 #ifdef CONFIG_PREEMPT_NOTIFIERS
2631 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2632 * @notifier: notifier struct to register
2634 void preempt_notifier_register(struct preempt_notifier *notifier)
2636 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2638 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2641 * preempt_notifier_unregister - no longer interested in preemption notifications
2642 * @notifier: notifier struct to unregister
2644 * This is safe to call from within a preemption notifier.
2646 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2648 hlist_del(¬ifier->link);
2650 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2652 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2654 struct preempt_notifier *notifier;
2655 struct hlist_node *node;
2657 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2658 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2662 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2663 struct task_struct *next)
2665 struct preempt_notifier *notifier;
2666 struct hlist_node *node;
2668 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2669 notifier->ops->sched_out(notifier, next);
2672 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2674 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2679 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2680 struct task_struct *next)
2684 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2687 * prepare_task_switch - prepare to switch tasks
2688 * @rq: the runqueue preparing to switch
2689 * @prev: the current task that is being switched out
2690 * @next: the task we are going to switch to.
2692 * This is called with the rq lock held and interrupts off. It must
2693 * be paired with a subsequent finish_task_switch after the context
2696 * prepare_task_switch sets up locking and calls architecture specific
2700 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2701 struct task_struct *next)
2703 fire_sched_out_preempt_notifiers(prev, next);
2704 prepare_lock_switch(rq, next);
2705 prepare_arch_switch(next);
2709 * finish_task_switch - clean up after a task-switch
2710 * @rq: runqueue associated with task-switch
2711 * @prev: the thread we just switched away from.
2713 * finish_task_switch must be called after the context switch, paired
2714 * with a prepare_task_switch call before the context switch.
2715 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2716 * and do any other architecture-specific cleanup actions.
2718 * Note that we may have delayed dropping an mm in context_switch(). If
2719 * so, we finish that here outside of the runqueue lock. (Doing it
2720 * with the lock held can cause deadlocks; see schedule() for
2723 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2724 __releases(rq->lock)
2726 struct mm_struct *mm = rq->prev_mm;
2732 * A task struct has one reference for the use as "current".
2733 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2734 * schedule one last time. The schedule call will never return, and
2735 * the scheduled task must drop that reference.
2736 * The test for TASK_DEAD must occur while the runqueue locks are
2737 * still held, otherwise prev could be scheduled on another cpu, die
2738 * there before we look at prev->state, and then the reference would
2740 * Manfred Spraul <manfred@colorfullife.com>
2742 prev_state = prev->state;
2743 finish_arch_switch(prev);
2744 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2745 local_irq_disable();
2746 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2747 perf_event_task_sched_in(current);
2748 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2750 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2751 finish_lock_switch(rq, prev);
2753 fire_sched_in_preempt_notifiers(current);
2756 if (unlikely(prev_state == TASK_DEAD)) {
2758 * Remove function-return probe instances associated with this
2759 * task and put them back on the free list.
2761 kprobe_flush_task(prev);
2762 put_task_struct(prev);
2768 /* assumes rq->lock is held */
2769 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2771 if (prev->sched_class->pre_schedule)
2772 prev->sched_class->pre_schedule(rq, prev);
2775 /* rq->lock is NOT held, but preemption is disabled */
2776 static inline void post_schedule(struct rq *rq)
2778 if (rq->post_schedule) {
2779 unsigned long flags;
2781 raw_spin_lock_irqsave(&rq->lock, flags);
2782 if (rq->curr->sched_class->post_schedule)
2783 rq->curr->sched_class->post_schedule(rq);
2784 raw_spin_unlock_irqrestore(&rq->lock, flags);
2786 rq->post_schedule = 0;
2792 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2796 static inline void post_schedule(struct rq *rq)
2803 * schedule_tail - first thing a freshly forked thread must call.
2804 * @prev: the thread we just switched away from.
2806 asmlinkage void schedule_tail(struct task_struct *prev)
2807 __releases(rq->lock)
2809 struct rq *rq = this_rq();
2811 finish_task_switch(rq, prev);
2814 * FIXME: do we need to worry about rq being invalidated by the
2819 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2820 /* In this case, finish_task_switch does not reenable preemption */
2823 if (current->set_child_tid)
2824 put_user(task_pid_vnr(current), current->set_child_tid);
2828 * context_switch - switch to the new MM and the new
2829 * thread's register state.
2832 context_switch(struct rq *rq, struct task_struct *prev,
2833 struct task_struct *next)
2835 struct mm_struct *mm, *oldmm;
2837 prepare_task_switch(rq, prev, next);
2838 trace_sched_switch(prev, next);
2840 oldmm = prev->active_mm;
2842 * For paravirt, this is coupled with an exit in switch_to to
2843 * combine the page table reload and the switch backend into
2846 arch_start_context_switch(prev);
2849 next->active_mm = oldmm;
2850 atomic_inc(&oldmm->mm_count);
2851 enter_lazy_tlb(oldmm, next);
2853 switch_mm(oldmm, mm, next);
2855 if (likely(!prev->mm)) {
2856 prev->active_mm = NULL;
2857 rq->prev_mm = oldmm;
2860 * Since the runqueue lock will be released by the next
2861 * task (which is an invalid locking op but in the case
2862 * of the scheduler it's an obvious special-case), so we
2863 * do an early lockdep release here:
2865 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2866 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2869 /* Here we just switch the register state and the stack. */
2870 switch_to(prev, next, prev);
2874 * this_rq must be evaluated again because prev may have moved
2875 * CPUs since it called schedule(), thus the 'rq' on its stack
2876 * frame will be invalid.
2878 finish_task_switch(this_rq(), prev);
2882 * nr_running, nr_uninterruptible and nr_context_switches:
2884 * externally visible scheduler statistics: current number of runnable
2885 * threads, current number of uninterruptible-sleeping threads, total
2886 * number of context switches performed since bootup.
2888 unsigned long nr_running(void)
2890 unsigned long i, sum = 0;
2892 for_each_online_cpu(i)
2893 sum += cpu_rq(i)->nr_running;
2898 unsigned long nr_uninterruptible(void)
2900 unsigned long i, sum = 0;
2902 for_each_possible_cpu(i)
2903 sum += cpu_rq(i)->nr_uninterruptible;
2906 * Since we read the counters lockless, it might be slightly
2907 * inaccurate. Do not allow it to go below zero though:
2909 if (unlikely((long)sum < 0))
2915 unsigned long long nr_context_switches(void)
2918 unsigned long long sum = 0;
2920 for_each_possible_cpu(i)
2921 sum += cpu_rq(i)->nr_switches;
2926 unsigned long nr_iowait(void)
2928 unsigned long i, sum = 0;
2930 for_each_possible_cpu(i)
2931 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2936 unsigned long nr_iowait_cpu(int cpu)
2938 struct rq *this = cpu_rq(cpu);
2939 return atomic_read(&this->nr_iowait);
2942 unsigned long this_cpu_load(void)
2944 struct rq *this = this_rq();
2945 return this->cpu_load[0];
2949 /* Variables and functions for calc_load */
2950 static atomic_long_t calc_load_tasks;
2951 static unsigned long calc_load_update;
2952 unsigned long avenrun[3];
2953 EXPORT_SYMBOL(avenrun);
2955 static long calc_load_fold_active(struct rq *this_rq)
2957 long nr_active, delta = 0;
2959 nr_active = this_rq->nr_running;
2960 nr_active += (long) this_rq->nr_uninterruptible;
2962 if (nr_active != this_rq->calc_load_active) {
2963 delta = nr_active - this_rq->calc_load_active;
2964 this_rq->calc_load_active = nr_active;
2970 static unsigned long
2971 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2974 load += active * (FIXED_1 - exp);
2975 load += 1UL << (FSHIFT - 1);
2976 return load >> FSHIFT;
2981 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
2983 * When making the ILB scale, we should try to pull this in as well.
2985 static atomic_long_t calc_load_tasks_idle;
2987 static void calc_load_account_idle(struct rq *this_rq)
2991 delta = calc_load_fold_active(this_rq);
2993 atomic_long_add(delta, &calc_load_tasks_idle);
2996 static long calc_load_fold_idle(void)
3001 * Its got a race, we don't care...
3003 if (atomic_long_read(&calc_load_tasks_idle))
3004 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
3010 * fixed_power_int - compute: x^n, in O(log n) time
3012 * @x: base of the power
3013 * @frac_bits: fractional bits of @x
3014 * @n: power to raise @x to.
3016 * By exploiting the relation between the definition of the natural power
3017 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
3018 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
3019 * (where: n_i \elem {0, 1}, the binary vector representing n),
3020 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
3021 * of course trivially computable in O(log_2 n), the length of our binary
3024 static unsigned long
3025 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
3027 unsigned long result = 1UL << frac_bits;
3032 result += 1UL << (frac_bits - 1);
3033 result >>= frac_bits;
3039 x += 1UL << (frac_bits - 1);
3047 * a1 = a0 * e + a * (1 - e)
3049 * a2 = a1 * e + a * (1 - e)
3050 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
3051 * = a0 * e^2 + a * (1 - e) * (1 + e)
3053 * a3 = a2 * e + a * (1 - e)
3054 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
3055 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
3059 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
3060 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
3061 * = a0 * e^n + a * (1 - e^n)
3063 * [1] application of the geometric series:
3066 * S_n := \Sum x^i = -------------
3069 static unsigned long
3070 calc_load_n(unsigned long load, unsigned long exp,
3071 unsigned long active, unsigned int n)
3074 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
3078 * NO_HZ can leave us missing all per-cpu ticks calling
3079 * calc_load_account_active(), but since an idle CPU folds its delta into
3080 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
3081 * in the pending idle delta if our idle period crossed a load cycle boundary.
3083 * Once we've updated the global active value, we need to apply the exponential
3084 * weights adjusted to the number of cycles missed.
3086 static void calc_global_nohz(unsigned long ticks)
3088 long delta, active, n;
3090 if (time_before(jiffies, calc_load_update))
3094 * If we crossed a calc_load_update boundary, make sure to fold
3095 * any pending idle changes, the respective CPUs might have
3096 * missed the tick driven calc_load_account_active() update
3099 delta = calc_load_fold_idle();
3101 atomic_long_add(delta, &calc_load_tasks);
3104 * If we were idle for multiple load cycles, apply them.
3106 if (ticks >= LOAD_FREQ) {
3107 n = ticks / LOAD_FREQ;
3109 active = atomic_long_read(&calc_load_tasks);
3110 active = active > 0 ? active * FIXED_1 : 0;
3112 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
3113 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
3114 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
3116 calc_load_update += n * LOAD_FREQ;
3120 * Its possible the remainder of the above division also crosses
3121 * a LOAD_FREQ period, the regular check in calc_global_load()
3122 * which comes after this will take care of that.
3124 * Consider us being 11 ticks before a cycle completion, and us
3125 * sleeping for 4*LOAD_FREQ + 22 ticks, then the above code will
3126 * age us 4 cycles, and the test in calc_global_load() will
3127 * pick up the final one.
3131 static void calc_load_account_idle(struct rq *this_rq)
3135 static inline long calc_load_fold_idle(void)
3140 static void calc_global_nohz(unsigned long ticks)
3146 * get_avenrun - get the load average array
3147 * @loads: pointer to dest load array
3148 * @offset: offset to add
3149 * @shift: shift count to shift the result left
3151 * These values are estimates at best, so no need for locking.
3153 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3155 loads[0] = (avenrun[0] + offset) << shift;
3156 loads[1] = (avenrun[1] + offset) << shift;
3157 loads[2] = (avenrun[2] + offset) << shift;
3161 * calc_load - update the avenrun load estimates 10 ticks after the
3162 * CPUs have updated calc_load_tasks.
3164 void calc_global_load(unsigned long ticks)
3168 calc_global_nohz(ticks);
3170 if (time_before(jiffies, calc_load_update + 10))
3173 active = atomic_long_read(&calc_load_tasks);
3174 active = active > 0 ? active * FIXED_1 : 0;
3176 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3177 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3178 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3180 calc_load_update += LOAD_FREQ;
3184 * Called from update_cpu_load() to periodically update this CPU's
3187 static void calc_load_account_active(struct rq *this_rq)
3191 if (time_before(jiffies, this_rq->calc_load_update))
3194 delta = calc_load_fold_active(this_rq);
3195 delta += calc_load_fold_idle();
3197 atomic_long_add(delta, &calc_load_tasks);
3199 this_rq->calc_load_update += LOAD_FREQ;
3203 * The exact cpuload at various idx values, calculated at every tick would be
3204 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3206 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3207 * on nth tick when cpu may be busy, then we have:
3208 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3209 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3211 * decay_load_missed() below does efficient calculation of
3212 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3213 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3215 * The calculation is approximated on a 128 point scale.
3216 * degrade_zero_ticks is the number of ticks after which load at any
3217 * particular idx is approximated to be zero.
3218 * degrade_factor is a precomputed table, a row for each load idx.
3219 * Each column corresponds to degradation factor for a power of two ticks,
3220 * based on 128 point scale.
3222 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3223 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3225 * With this power of 2 load factors, we can degrade the load n times
3226 * by looking at 1 bits in n and doing as many mult/shift instead of
3227 * n mult/shifts needed by the exact degradation.
3229 #define DEGRADE_SHIFT 7
3230 static const unsigned char
3231 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3232 static const unsigned char
3233 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3234 {0, 0, 0, 0, 0, 0, 0, 0},
3235 {64, 32, 8, 0, 0, 0, 0, 0},
3236 {96, 72, 40, 12, 1, 0, 0},
3237 {112, 98, 75, 43, 15, 1, 0},
3238 {120, 112, 98, 76, 45, 16, 2} };
3241 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3242 * would be when CPU is idle and so we just decay the old load without
3243 * adding any new load.
3245 static unsigned long
3246 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3250 if (!missed_updates)
3253 if (missed_updates >= degrade_zero_ticks[idx])
3257 return load >> missed_updates;
3259 while (missed_updates) {
3260 if (missed_updates % 2)
3261 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3263 missed_updates >>= 1;
3270 * Update rq->cpu_load[] statistics. This function is usually called every
3271 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3272 * every tick. We fix it up based on jiffies.
3274 static void update_cpu_load(struct rq *this_rq)
3276 unsigned long this_load = this_rq->load.weight;
3277 unsigned long curr_jiffies = jiffies;
3278 unsigned long pending_updates;
3281 this_rq->nr_load_updates++;
3283 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3284 if (curr_jiffies == this_rq->last_load_update_tick)
3287 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3288 this_rq->last_load_update_tick = curr_jiffies;
3290 /* Update our load: */
3291 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3292 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3293 unsigned long old_load, new_load;
3295 /* scale is effectively 1 << i now, and >> i divides by scale */
3297 old_load = this_rq->cpu_load[i];
3298 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3299 new_load = this_load;
3301 * Round up the averaging division if load is increasing. This
3302 * prevents us from getting stuck on 9 if the load is 10, for
3305 if (new_load > old_load)
3306 new_load += scale - 1;
3308 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3311 sched_avg_update(this_rq);
3314 static void update_cpu_load_active(struct rq *this_rq)
3316 update_cpu_load(this_rq);
3318 calc_load_account_active(this_rq);
3324 * sched_exec - execve() is a valuable balancing opportunity, because at
3325 * this point the task has the smallest effective memory and cache footprint.
3327 void sched_exec(void)
3329 struct task_struct *p = current;
3330 unsigned long flags;
3334 rq = task_rq_lock(p, &flags);
3335 dest_cpu = p->sched_class->select_task_rq(rq, p, SD_BALANCE_EXEC, 0);
3336 if (dest_cpu == smp_processor_id())
3340 * select_task_rq() can race against ->cpus_allowed
3342 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed) &&
3343 likely(cpu_active(dest_cpu)) && migrate_task(p, dest_cpu)) {
3344 struct migration_arg arg = { p, dest_cpu };
3346 task_rq_unlock(rq, &flags);
3347 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
3351 task_rq_unlock(rq, &flags);
3356 DEFINE_PER_CPU(struct kernel_stat, kstat);
3358 EXPORT_PER_CPU_SYMBOL(kstat);
3361 * Return any ns on the sched_clock that have not yet been accounted in
3362 * @p in case that task is currently running.
3364 * Called with task_rq_lock() held on @rq.
3366 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3370 if (task_current(rq, p)) {
3371 update_rq_clock(rq);
3372 ns = rq->clock - p->se.exec_start;
3380 unsigned long long task_delta_exec(struct task_struct *p)
3382 unsigned long flags;
3386 rq = task_rq_lock(p, &flags);
3387 ns = do_task_delta_exec(p, rq);
3388 task_rq_unlock(rq, &flags);
3394 * Return accounted runtime for the task.
3395 * In case the task is currently running, return the runtime plus current's
3396 * pending runtime that have not been accounted yet.
3398 unsigned long long task_sched_runtime(struct task_struct *p)
3400 unsigned long flags;
3404 rq = task_rq_lock(p, &flags);
3405 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3406 task_rq_unlock(rq, &flags);
3412 * Return sum_exec_runtime for the thread group.
3413 * In case the task is currently running, return the sum plus current's
3414 * pending runtime that have not been accounted yet.
3416 * Note that the thread group might have other running tasks as well,
3417 * so the return value not includes other pending runtime that other
3418 * running tasks might have.
3420 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3422 struct task_cputime totals;
3423 unsigned long flags;
3427 rq = task_rq_lock(p, &flags);
3428 thread_group_cputime(p, &totals);
3429 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3430 task_rq_unlock(rq, &flags);
3436 * Account user cpu time to a process.
3437 * @p: the process that the cpu time gets accounted to
3438 * @cputime: the cpu time spent in user space since the last update
3439 * @cputime_scaled: cputime scaled by cpu frequency
3441 void account_user_time(struct task_struct *p, cputime_t cputime,
3442 cputime_t cputime_scaled)
3444 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3447 /* Add user time to process. */
3448 p->utime = cputime_add(p->utime, cputime);
3449 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3450 account_group_user_time(p, cputime);
3452 /* Add user time to cpustat. */
3453 tmp = cputime_to_cputime64(cputime);
3454 if (TASK_NICE(p) > 0)
3455 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3457 cpustat->user = cputime64_add(cpustat->user, tmp);
3459 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3460 /* Account for user time used */
3461 acct_update_integrals(p);
3465 * Account guest cpu time to a process.
3466 * @p: the process that the cpu time gets accounted to
3467 * @cputime: the cpu time spent in virtual machine since the last update
3468 * @cputime_scaled: cputime scaled by cpu frequency
3470 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3471 cputime_t cputime_scaled)
3474 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3476 tmp = cputime_to_cputime64(cputime);
3478 /* Add guest time to process. */
3479 p->utime = cputime_add(p->utime, cputime);
3480 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3481 account_group_user_time(p, cputime);
3482 p->gtime = cputime_add(p->gtime, cputime);
3484 /* Add guest time to cpustat. */
3485 if (TASK_NICE(p) > 0) {
3486 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3487 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3489 cpustat->user = cputime64_add(cpustat->user, tmp);
3490 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3495 * Account system cpu time to a process.
3496 * @p: the process that the cpu time gets accounted to
3497 * @hardirq_offset: the offset to subtract from hardirq_count()
3498 * @cputime: the cpu time spent in kernel space since the last update
3499 * @cputime_scaled: cputime scaled by cpu frequency
3501 void account_system_time(struct task_struct *p, int hardirq_offset,
3502 cputime_t cputime, cputime_t cputime_scaled)
3504 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3507 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3508 account_guest_time(p, cputime, cputime_scaled);
3512 /* Add system time to process. */
3513 p->stime = cputime_add(p->stime, cputime);
3514 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3515 account_group_system_time(p, cputime);
3517 /* Add system time to cpustat. */
3518 tmp = cputime_to_cputime64(cputime);
3519 if (hardirq_count() - hardirq_offset)
3520 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3521 else if (softirq_count())
3522 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3524 cpustat->system = cputime64_add(cpustat->system, tmp);
3526 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3528 /* Account for system time used */
3529 acct_update_integrals(p);
3533 * Account for involuntary wait time.
3534 * @steal: the cpu time spent in involuntary wait
3536 void account_steal_time(cputime_t cputime)
3538 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3539 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3541 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3545 * Account for idle time.
3546 * @cputime: the cpu time spent in idle wait
3548 void account_idle_time(cputime_t cputime)
3550 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3551 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3552 struct rq *rq = this_rq();
3554 if (atomic_read(&rq->nr_iowait) > 0)
3555 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3557 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3560 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3563 * Account a single tick of cpu time.
3564 * @p: the process that the cpu time gets accounted to
3565 * @user_tick: indicates if the tick is a user or a system tick
3567 void account_process_tick(struct task_struct *p, int user_tick)
3569 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3570 struct rq *rq = this_rq();
3573 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3574 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3575 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3578 account_idle_time(cputime_one_jiffy);
3582 * Account multiple ticks of steal time.
3583 * @p: the process from which the cpu time has been stolen
3584 * @ticks: number of stolen ticks
3586 void account_steal_ticks(unsigned long ticks)
3588 account_steal_time(jiffies_to_cputime(ticks));
3592 * Account multiple ticks of idle time.
3593 * @ticks: number of stolen ticks
3595 void account_idle_ticks(unsigned long ticks)
3597 account_idle_time(jiffies_to_cputime(ticks));
3603 * Use precise platform statistics if available:
3605 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3606 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3612 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3614 struct task_cputime cputime;
3616 thread_group_cputime(p, &cputime);
3618 *ut = cputime.utime;
3619 *st = cputime.stime;
3623 #ifndef nsecs_to_cputime
3624 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3627 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3629 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
3632 * Use CFS's precise accounting:
3634 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3640 do_div(temp, total);
3641 utime = (cputime_t)temp;
3646 * Compare with previous values, to keep monotonicity:
3648 p->prev_utime = max(p->prev_utime, utime);
3649 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
3651 *ut = p->prev_utime;
3652 *st = p->prev_stime;
3656 * Must be called with siglock held.
3658 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3660 struct signal_struct *sig = p->signal;
3661 struct task_cputime cputime;
3662 cputime_t rtime, utime, total;
3664 thread_group_cputime(p, &cputime);
3666 total = cputime_add(cputime.utime, cputime.stime);
3667 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3672 temp *= cputime.utime;
3673 do_div(temp, total);
3674 utime = (cputime_t)temp;
3678 sig->prev_utime = max(sig->prev_utime, utime);
3679 sig->prev_stime = max(sig->prev_stime,
3680 cputime_sub(rtime, sig->prev_utime));
3682 *ut = sig->prev_utime;
3683 *st = sig->prev_stime;
3688 * This function gets called by the timer code, with HZ frequency.
3689 * We call it with interrupts disabled.
3691 * It also gets called by the fork code, when changing the parent's
3694 void scheduler_tick(void)
3696 int cpu = smp_processor_id();
3697 struct rq *rq = cpu_rq(cpu);
3698 struct task_struct *curr = rq->curr;
3702 raw_spin_lock(&rq->lock);
3703 update_rq_clock(rq);
3704 update_cpu_load_active(rq);
3705 curr->sched_class->task_tick(rq, curr, 0);
3706 raw_spin_unlock(&rq->lock);
3708 perf_event_task_tick(curr);
3711 rq->idle_at_tick = idle_cpu(cpu);
3712 trigger_load_balance(rq, cpu);
3716 notrace unsigned long get_parent_ip(unsigned long addr)
3718 if (in_lock_functions(addr)) {
3719 addr = CALLER_ADDR2;
3720 if (in_lock_functions(addr))
3721 addr = CALLER_ADDR3;
3726 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3727 defined(CONFIG_PREEMPT_TRACER))
3729 void __kprobes add_preempt_count(int val)
3731 #ifdef CONFIG_DEBUG_PREEMPT
3735 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3738 preempt_count() += val;
3739 #ifdef CONFIG_DEBUG_PREEMPT
3741 * Spinlock count overflowing soon?
3743 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3746 if (preempt_count() == val)
3747 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3749 EXPORT_SYMBOL(add_preempt_count);
3751 void __kprobes sub_preempt_count(int val)
3753 #ifdef CONFIG_DEBUG_PREEMPT
3757 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3760 * Is the spinlock portion underflowing?
3762 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3763 !(preempt_count() & PREEMPT_MASK)))
3767 if (preempt_count() == val)
3768 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3769 preempt_count() -= val;
3771 EXPORT_SYMBOL(sub_preempt_count);
3776 * Print scheduling while atomic bug:
3778 static noinline void __schedule_bug(struct task_struct *prev)
3780 struct pt_regs *regs = get_irq_regs();
3782 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3783 prev->comm, prev->pid, preempt_count());
3785 debug_show_held_locks(prev);
3787 if (irqs_disabled())
3788 print_irqtrace_events(prev);
3797 * Various schedule()-time debugging checks and statistics:
3799 static inline void schedule_debug(struct task_struct *prev)
3802 * Test if we are atomic. Since do_exit() needs to call into
3803 * schedule() atomically, we ignore that path for now.
3804 * Otherwise, whine if we are scheduling when we should not be.
3806 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3807 __schedule_bug(prev);
3809 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3811 schedstat_inc(this_rq(), sched_count);
3812 #ifdef CONFIG_SCHEDSTATS
3813 if (unlikely(prev->lock_depth >= 0)) {
3814 schedstat_inc(this_rq(), bkl_count);
3815 schedstat_inc(prev, sched_info.bkl_count);
3820 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3823 update_rq_clock(rq);
3824 prev->sched_class->put_prev_task(rq, prev);
3828 * Pick up the highest-prio task:
3830 static inline struct task_struct *
3831 pick_next_task(struct rq *rq)
3833 const struct sched_class *class;
3834 struct task_struct *p;
3837 * Optimization: we know that if all tasks are in
3838 * the fair class we can call that function directly:
3840 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3841 p = fair_sched_class.pick_next_task(rq);
3846 class = sched_class_highest;
3848 p = class->pick_next_task(rq);
3852 * Will never be NULL as the idle class always
3853 * returns a non-NULL p:
3855 class = class->next;
3860 * schedule() is the main scheduler function.
3862 asmlinkage void __sched schedule(void)
3864 struct task_struct *prev, *next;
3865 unsigned long *switch_count;
3871 cpu = smp_processor_id();
3873 rcu_note_context_switch(cpu);
3876 release_kernel_lock(prev);
3877 need_resched_nonpreemptible:
3879 schedule_debug(prev);
3881 if (sched_feat(HRTICK))
3884 raw_spin_lock_irq(&rq->lock);
3886 switch_count = &prev->nivcsw;
3887 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3888 if (unlikely(signal_pending_state(prev->state, prev))) {
3889 prev->state = TASK_RUNNING;
3892 * If a worker is going to sleep, notify and
3893 * ask workqueue whether it wants to wake up a
3894 * task to maintain concurrency. If so, wake
3897 if (prev->flags & PF_WQ_WORKER) {
3898 struct task_struct *to_wakeup;
3900 to_wakeup = wq_worker_sleeping(prev, cpu);
3902 try_to_wake_up_local(to_wakeup);
3904 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3906 switch_count = &prev->nvcsw;
3909 pre_schedule(rq, prev);
3911 if (unlikely(!rq->nr_running))
3912 idle_balance(cpu, rq);
3914 put_prev_task(rq, prev);
3915 next = pick_next_task(rq);
3916 clear_tsk_need_resched(prev);
3917 rq->skip_clock_update = 0;
3919 if (likely(prev != next)) {
3920 sched_info_switch(prev, next);
3921 perf_event_task_sched_out(prev, next);
3927 context_switch(rq, prev, next); /* unlocks the rq */
3929 * The context switch have flipped the stack from under us
3930 * and restored the local variables which were saved when
3931 * this task called schedule() in the past. prev == current
3932 * is still correct, but it can be moved to another cpu/rq.
3934 cpu = smp_processor_id();
3937 raw_spin_unlock_irq(&rq->lock);
3941 if (unlikely(reacquire_kernel_lock(prev)))
3942 goto need_resched_nonpreemptible;
3944 preempt_enable_no_resched();
3948 EXPORT_SYMBOL(schedule);
3950 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3952 * Look out! "owner" is an entirely speculative pointer
3953 * access and not reliable.
3955 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
3960 if (!sched_feat(OWNER_SPIN))
3963 #ifdef CONFIG_DEBUG_PAGEALLOC
3965 * Need to access the cpu field knowing that
3966 * DEBUG_PAGEALLOC could have unmapped it if
3967 * the mutex owner just released it and exited.
3969 if (probe_kernel_address(&owner->cpu, cpu))
3976 * Even if the access succeeded (likely case),
3977 * the cpu field may no longer be valid.
3979 if (cpu >= nr_cpumask_bits)
3983 * We need to validate that we can do a
3984 * get_cpu() and that we have the percpu area.
3986 if (!cpu_online(cpu))
3993 * Owner changed, break to re-assess state.
3995 if (lock->owner != owner) {
3997 * If the lock has switched to a different owner,
3998 * we likely have heavy contention. Return 0 to quit
3999 * optimistic spinning and not contend further:
4007 * Is that owner really running on that cpu?
4009 if (task_thread_info(rq->curr) != owner || need_resched())
4019 #ifdef CONFIG_PREEMPT
4021 * this is the entry point to schedule() from in-kernel preemption
4022 * off of preempt_enable. Kernel preemptions off return from interrupt
4023 * occur there and call schedule directly.
4025 asmlinkage void __sched notrace preempt_schedule(void)
4027 struct thread_info *ti = current_thread_info();
4030 * If there is a non-zero preempt_count or interrupts are disabled,
4031 * we do not want to preempt the current task. Just return..
4033 if (likely(ti->preempt_count || irqs_disabled()))
4037 add_preempt_count_notrace(PREEMPT_ACTIVE);
4039 sub_preempt_count_notrace(PREEMPT_ACTIVE);
4042 * Check again in case we missed a preemption opportunity
4043 * between schedule and now.
4046 } while (need_resched());
4048 EXPORT_SYMBOL(preempt_schedule);
4051 * this is the entry point to schedule() from kernel preemption
4052 * off of irq context.
4053 * Note, that this is called and return with irqs disabled. This will
4054 * protect us against recursive calling from irq.
4056 asmlinkage void __sched preempt_schedule_irq(void)
4058 struct thread_info *ti = current_thread_info();
4060 /* Catch callers which need to be fixed */
4061 BUG_ON(ti->preempt_count || !irqs_disabled());
4064 add_preempt_count(PREEMPT_ACTIVE);
4067 local_irq_disable();
4068 sub_preempt_count(PREEMPT_ACTIVE);
4071 * Check again in case we missed a preemption opportunity
4072 * between schedule and now.
4075 } while (need_resched());
4078 #endif /* CONFIG_PREEMPT */
4080 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
4083 return try_to_wake_up(curr->private, mode, wake_flags);
4085 EXPORT_SYMBOL(default_wake_function);
4088 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4089 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4090 * number) then we wake all the non-exclusive tasks and one exclusive task.
4092 * There are circumstances in which we can try to wake a task which has already
4093 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4094 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4096 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4097 int nr_exclusive, int wake_flags, void *key)
4099 wait_queue_t *curr, *next;
4101 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4102 unsigned flags = curr->flags;
4104 if (curr->func(curr, mode, wake_flags, key) &&
4105 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4111 * __wake_up - wake up threads blocked on a waitqueue.
4113 * @mode: which threads
4114 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4115 * @key: is directly passed to the wakeup function
4117 * It may be assumed that this function implies a write memory barrier before
4118 * changing the task state if and only if any tasks are woken up.
4120 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4121 int nr_exclusive, void *key)
4123 unsigned long flags;
4125 spin_lock_irqsave(&q->lock, flags);
4126 __wake_up_common(q, mode, nr_exclusive, 0, key);
4127 spin_unlock_irqrestore(&q->lock, flags);
4129 EXPORT_SYMBOL(__wake_up);
4132 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4134 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4136 __wake_up_common(q, mode, 1, 0, NULL);
4138 EXPORT_SYMBOL_GPL(__wake_up_locked);
4140 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
4142 __wake_up_common(q, mode, 1, 0, key);
4146 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4148 * @mode: which threads
4149 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4150 * @key: opaque value to be passed to wakeup targets
4152 * The sync wakeup differs that the waker knows that it will schedule
4153 * away soon, so while the target thread will be woken up, it will not
4154 * be migrated to another CPU - ie. the two threads are 'synchronized'
4155 * with each other. This can prevent needless bouncing between CPUs.
4157 * On UP it can prevent extra preemption.
4159 * It may be assumed that this function implies a write memory barrier before
4160 * changing the task state if and only if any tasks are woken up.
4162 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4163 int nr_exclusive, void *key)
4165 unsigned long flags;
4166 int wake_flags = WF_SYNC;
4171 if (unlikely(!nr_exclusive))
4174 spin_lock_irqsave(&q->lock, flags);
4175 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4176 spin_unlock_irqrestore(&q->lock, flags);
4178 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4181 * __wake_up_sync - see __wake_up_sync_key()
4183 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4185 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4187 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4190 * complete: - signals a single thread waiting on this completion
4191 * @x: holds the state of this particular completion
4193 * This will wake up a single thread waiting on this completion. Threads will be
4194 * awakened in the same order in which they were queued.
4196 * See also complete_all(), wait_for_completion() and related routines.
4198 * It may be assumed that this function implies a write memory barrier before
4199 * changing the task state if and only if any tasks are woken up.
4201 void complete(struct completion *x)
4203 unsigned long flags;
4205 spin_lock_irqsave(&x->wait.lock, flags);
4207 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4208 spin_unlock_irqrestore(&x->wait.lock, flags);
4210 EXPORT_SYMBOL(complete);
4213 * complete_all: - signals all threads waiting on this completion
4214 * @x: holds the state of this particular completion
4216 * This will wake up all threads waiting on this particular completion event.
4218 * It may be assumed that this function implies a write memory barrier before
4219 * changing the task state if and only if any tasks are woken up.
4221 void complete_all(struct completion *x)
4223 unsigned long flags;
4225 spin_lock_irqsave(&x->wait.lock, flags);
4226 x->done += UINT_MAX/2;
4227 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4228 spin_unlock_irqrestore(&x->wait.lock, flags);
4230 EXPORT_SYMBOL(complete_all);
4232 static inline long __sched
4233 do_wait_for_common(struct completion *x, long timeout, int state)
4236 DECLARE_WAITQUEUE(wait, current);
4238 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4240 if (signal_pending_state(state, current)) {
4241 timeout = -ERESTARTSYS;
4244 __set_current_state(state);
4245 spin_unlock_irq(&x->wait.lock);
4246 timeout = schedule_timeout(timeout);
4247 spin_lock_irq(&x->wait.lock);
4248 } while (!x->done && timeout);
4249 __remove_wait_queue(&x->wait, &wait);
4254 return timeout ?: 1;
4258 wait_for_common(struct completion *x, long timeout, int state)
4262 spin_lock_irq(&x->wait.lock);
4263 timeout = do_wait_for_common(x, timeout, state);
4264 spin_unlock_irq(&x->wait.lock);
4269 * wait_for_completion: - waits for completion of a task
4270 * @x: holds the state of this particular completion
4272 * This waits to be signaled for completion of a specific task. It is NOT
4273 * interruptible and there is no timeout.
4275 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4276 * and interrupt capability. Also see complete().
4278 void __sched wait_for_completion(struct completion *x)
4280 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4282 EXPORT_SYMBOL(wait_for_completion);
4285 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4286 * @x: holds the state of this particular completion
4287 * @timeout: timeout value in jiffies
4289 * This waits for either a completion of a specific task to be signaled or for a
4290 * specified timeout to expire. The timeout is in jiffies. It is not
4293 unsigned long __sched
4294 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4296 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4298 EXPORT_SYMBOL(wait_for_completion_timeout);
4301 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4302 * @x: holds the state of this particular completion
4304 * This waits for completion of a specific task to be signaled. It is
4307 int __sched wait_for_completion_interruptible(struct completion *x)
4309 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4310 if (t == -ERESTARTSYS)
4314 EXPORT_SYMBOL(wait_for_completion_interruptible);
4317 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4318 * @x: holds the state of this particular completion
4319 * @timeout: timeout value in jiffies
4321 * This waits for either a completion of a specific task to be signaled or for a
4322 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4324 unsigned long __sched
4325 wait_for_completion_interruptible_timeout(struct completion *x,
4326 unsigned long timeout)
4328 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4330 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4333 * wait_for_completion_killable: - waits for completion of a task (killable)
4334 * @x: holds the state of this particular completion
4336 * This waits to be signaled for completion of a specific task. It can be
4337 * interrupted by a kill signal.
4339 int __sched wait_for_completion_killable(struct completion *x)
4341 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4342 if (t == -ERESTARTSYS)
4346 EXPORT_SYMBOL(wait_for_completion_killable);
4349 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4350 * @x: holds the state of this particular completion
4351 * @timeout: timeout value in jiffies
4353 * This waits for either a completion of a specific task to be
4354 * signaled or for a specified timeout to expire. It can be
4355 * interrupted by a kill signal. The timeout is in jiffies.
4357 unsigned long __sched
4358 wait_for_completion_killable_timeout(struct completion *x,
4359 unsigned long timeout)
4361 return wait_for_common(x, timeout, TASK_KILLABLE);
4363 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
4366 * try_wait_for_completion - try to decrement a completion without blocking
4367 * @x: completion structure
4369 * Returns: 0 if a decrement cannot be done without blocking
4370 * 1 if a decrement succeeded.
4372 * If a completion is being used as a counting completion,
4373 * attempt to decrement the counter without blocking. This
4374 * enables us to avoid waiting if the resource the completion
4375 * is protecting is not available.
4377 bool try_wait_for_completion(struct completion *x)
4379 unsigned long flags;
4382 spin_lock_irqsave(&x->wait.lock, flags);
4387 spin_unlock_irqrestore(&x->wait.lock, flags);
4390 EXPORT_SYMBOL(try_wait_for_completion);
4393 * completion_done - Test to see if a completion has any waiters
4394 * @x: completion structure
4396 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4397 * 1 if there are no waiters.
4400 bool completion_done(struct completion *x)
4402 unsigned long flags;
4405 spin_lock_irqsave(&x->wait.lock, flags);
4408 spin_unlock_irqrestore(&x->wait.lock, flags);
4411 EXPORT_SYMBOL(completion_done);
4414 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4416 unsigned long flags;
4419 init_waitqueue_entry(&wait, current);
4421 __set_current_state(state);
4423 spin_lock_irqsave(&q->lock, flags);
4424 __add_wait_queue(q, &wait);
4425 spin_unlock(&q->lock);
4426 timeout = schedule_timeout(timeout);
4427 spin_lock_irq(&q->lock);
4428 __remove_wait_queue(q, &wait);
4429 spin_unlock_irqrestore(&q->lock, flags);
4434 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4436 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4438 EXPORT_SYMBOL(interruptible_sleep_on);
4441 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4443 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4445 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4447 void __sched sleep_on(wait_queue_head_t *q)
4449 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4451 EXPORT_SYMBOL(sleep_on);
4453 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4455 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4457 EXPORT_SYMBOL(sleep_on_timeout);
4459 #ifdef CONFIG_RT_MUTEXES
4462 * rt_mutex_setprio - set the current priority of a task
4464 * @prio: prio value (kernel-internal form)
4466 * This function changes the 'effective' priority of a task. It does
4467 * not touch ->normal_prio like __setscheduler().
4469 * Used by the rt_mutex code to implement priority inheritance logic.
4471 void rt_mutex_setprio(struct task_struct *p, int prio)
4473 unsigned long flags;
4474 int oldprio, on_rq, running;
4476 const struct sched_class *prev_class;
4478 BUG_ON(prio < 0 || prio > MAX_PRIO);
4480 rq = task_rq_lock(p, &flags);
4483 prev_class = p->sched_class;
4484 on_rq = p->se.on_rq;
4485 running = task_current(rq, p);
4487 dequeue_task(rq, p, 0);
4489 p->sched_class->put_prev_task(rq, p);
4492 p->sched_class = &rt_sched_class;
4494 p->sched_class = &fair_sched_class;
4499 p->sched_class->set_curr_task(rq);
4501 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4503 check_class_changed(rq, p, prev_class, oldprio, running);
4505 task_rq_unlock(rq, &flags);
4510 void set_user_nice(struct task_struct *p, long nice)
4512 int old_prio, delta, on_rq;
4513 unsigned long flags;
4516 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4519 * We have to be careful, if called from sys_setpriority(),
4520 * the task might be in the middle of scheduling on another CPU.
4522 rq = task_rq_lock(p, &flags);
4524 * The RT priorities are set via sched_setscheduler(), but we still
4525 * allow the 'normal' nice value to be set - but as expected
4526 * it wont have any effect on scheduling until the task is
4527 * SCHED_FIFO/SCHED_RR:
4529 if (task_has_rt_policy(p)) {
4530 p->static_prio = NICE_TO_PRIO(nice);
4533 on_rq = p->se.on_rq;
4535 dequeue_task(rq, p, 0);
4537 p->static_prio = NICE_TO_PRIO(nice);
4540 p->prio = effective_prio(p);
4541 delta = p->prio - old_prio;
4544 enqueue_task(rq, p, 0);
4546 * If the task increased its priority or is running and
4547 * lowered its priority, then reschedule its CPU:
4549 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4550 resched_task(rq->curr);
4553 task_rq_unlock(rq, &flags);
4555 EXPORT_SYMBOL(set_user_nice);
4558 * can_nice - check if a task can reduce its nice value
4562 int can_nice(const struct task_struct *p, const int nice)
4564 /* convert nice value [19,-20] to rlimit style value [1,40] */
4565 int nice_rlim = 20 - nice;
4567 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4568 capable(CAP_SYS_NICE));
4571 #ifdef __ARCH_WANT_SYS_NICE
4574 * sys_nice - change the priority of the current process.
4575 * @increment: priority increment
4577 * sys_setpriority is a more generic, but much slower function that
4578 * does similar things.
4580 SYSCALL_DEFINE1(nice, int, increment)
4585 * Setpriority might change our priority at the same moment.
4586 * We don't have to worry. Conceptually one call occurs first
4587 * and we have a single winner.
4589 if (increment < -40)
4594 nice = TASK_NICE(current) + increment;
4600 if (increment < 0 && !can_nice(current, nice))
4603 retval = security_task_setnice(current, nice);
4607 set_user_nice(current, nice);
4614 * task_prio - return the priority value of a given task.
4615 * @p: the task in question.
4617 * This is the priority value as seen by users in /proc.
4618 * RT tasks are offset by -200. Normal tasks are centered
4619 * around 0, value goes from -16 to +15.
4621 int task_prio(const struct task_struct *p)
4623 return p->prio - MAX_RT_PRIO;
4627 * task_nice - return the nice value of a given task.
4628 * @p: the task in question.
4630 int task_nice(const struct task_struct *p)
4632 return TASK_NICE(p);
4634 EXPORT_SYMBOL(task_nice);
4637 * idle_cpu - is a given cpu idle currently?
4638 * @cpu: the processor in question.
4640 int idle_cpu(int cpu)
4642 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4646 * idle_task - return the idle task for a given cpu.
4647 * @cpu: the processor in question.
4649 struct task_struct *idle_task(int cpu)
4651 return cpu_rq(cpu)->idle;
4655 * find_process_by_pid - find a process with a matching PID value.
4656 * @pid: the pid in question.
4658 static struct task_struct *find_process_by_pid(pid_t pid)
4660 return pid ? find_task_by_vpid(pid) : current;
4663 /* Actually do priority change: must hold rq lock. */
4665 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4667 BUG_ON(p->se.on_rq);
4670 p->rt_priority = prio;
4671 p->normal_prio = normal_prio(p);
4672 /* we are holding p->pi_lock already */
4673 p->prio = rt_mutex_getprio(p);
4674 if (rt_prio(p->prio))
4675 p->sched_class = &rt_sched_class;
4677 p->sched_class = &fair_sched_class;
4682 * check the target process has a UID that matches the current process's
4684 static bool check_same_owner(struct task_struct *p)
4686 const struct cred *cred = current_cred(), *pcred;
4690 pcred = __task_cred(p);
4691 match = (cred->euid == pcred->euid ||
4692 cred->euid == pcred->uid);
4697 static int __sched_setscheduler(struct task_struct *p, int policy,
4698 struct sched_param *param, bool user)
4700 int retval, oldprio, oldpolicy = -1, on_rq, running;
4701 unsigned long flags;
4702 const struct sched_class *prev_class;
4706 /* may grab non-irq protected spin_locks */
4707 BUG_ON(in_interrupt());
4709 /* double check policy once rq lock held */
4711 reset_on_fork = p->sched_reset_on_fork;
4712 policy = oldpolicy = p->policy;
4714 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4715 policy &= ~SCHED_RESET_ON_FORK;
4717 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4718 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4719 policy != SCHED_IDLE)
4724 * Valid priorities for SCHED_FIFO and SCHED_RR are
4725 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4726 * SCHED_BATCH and SCHED_IDLE is 0.
4728 if (param->sched_priority < 0 ||
4729 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4730 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4732 if (rt_policy(policy) != (param->sched_priority != 0))
4736 * Allow unprivileged RT tasks to decrease priority:
4738 if (user && !capable(CAP_SYS_NICE)) {
4739 if (rt_policy(policy)) {
4740 unsigned long rlim_rtprio =
4741 task_rlimit(p, RLIMIT_RTPRIO);
4743 /* can't set/change the rt policy */
4744 if (policy != p->policy && !rlim_rtprio)
4747 /* can't increase priority */
4748 if (param->sched_priority > p->rt_priority &&
4749 param->sched_priority > rlim_rtprio)
4753 * Like positive nice levels, dont allow tasks to
4754 * move out of SCHED_IDLE either:
4756 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4759 /* can't change other user's priorities */
4760 if (!check_same_owner(p))
4763 /* Normal users shall not reset the sched_reset_on_fork flag */
4764 if (p->sched_reset_on_fork && !reset_on_fork)
4769 retval = security_task_setscheduler(p, policy, param);
4775 * make sure no PI-waiters arrive (or leave) while we are
4776 * changing the priority of the task:
4778 raw_spin_lock_irqsave(&p->pi_lock, flags);
4780 * To be able to change p->policy safely, the apropriate
4781 * runqueue lock must be held.
4783 rq = __task_rq_lock(p);
4785 #ifdef CONFIG_RT_GROUP_SCHED
4788 * Do not allow realtime tasks into groups that have no runtime
4791 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4792 task_group(p)->rt_bandwidth.rt_runtime == 0) {
4793 __task_rq_unlock(rq);
4794 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4800 /* recheck policy now with rq lock held */
4801 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4802 policy = oldpolicy = -1;
4803 __task_rq_unlock(rq);
4804 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4807 on_rq = p->se.on_rq;
4808 running = task_current(rq, p);
4810 deactivate_task(rq, p, 0);
4812 p->sched_class->put_prev_task(rq, p);
4814 p->sched_reset_on_fork = reset_on_fork;
4817 prev_class = p->sched_class;
4818 __setscheduler(rq, p, policy, param->sched_priority);
4821 p->sched_class->set_curr_task(rq);
4823 activate_task(rq, p, 0);
4825 check_class_changed(rq, p, prev_class, oldprio, running);
4827 __task_rq_unlock(rq);
4828 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4830 rt_mutex_adjust_pi(p);
4836 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4837 * @p: the task in question.
4838 * @policy: new policy.
4839 * @param: structure containing the new RT priority.
4841 * NOTE that the task may be already dead.
4843 int sched_setscheduler(struct task_struct *p, int policy,
4844 struct sched_param *param)
4846 return __sched_setscheduler(p, policy, param, true);
4848 EXPORT_SYMBOL_GPL(sched_setscheduler);
4851 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4852 * @p: the task in question.
4853 * @policy: new policy.
4854 * @param: structure containing the new RT priority.
4856 * Just like sched_setscheduler, only don't bother checking if the
4857 * current context has permission. For example, this is needed in
4858 * stop_machine(): we create temporary high priority worker threads,
4859 * but our caller might not have that capability.
4861 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4862 struct sched_param *param)
4864 return __sched_setscheduler(p, policy, param, false);
4868 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4870 struct sched_param lparam;
4871 struct task_struct *p;
4874 if (!param || pid < 0)
4876 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4881 p = find_process_by_pid(pid);
4883 retval = sched_setscheduler(p, policy, &lparam);
4890 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4891 * @pid: the pid in question.
4892 * @policy: new policy.
4893 * @param: structure containing the new RT priority.
4895 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4896 struct sched_param __user *, param)
4898 /* negative values for policy are not valid */
4902 return do_sched_setscheduler(pid, policy, param);
4906 * sys_sched_setparam - set/change the RT priority of a thread
4907 * @pid: the pid in question.
4908 * @param: structure containing the new RT priority.
4910 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4912 return do_sched_setscheduler(pid, -1, param);
4916 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4917 * @pid: the pid in question.
4919 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4921 struct task_struct *p;
4929 p = find_process_by_pid(pid);
4931 retval = security_task_getscheduler(p);
4934 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4941 * sys_sched_getparam - get the RT priority of a thread
4942 * @pid: the pid in question.
4943 * @param: structure containing the RT priority.
4945 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4947 struct sched_param lp;
4948 struct task_struct *p;
4951 if (!param || pid < 0)
4955 p = find_process_by_pid(pid);
4960 retval = security_task_getscheduler(p);
4964 lp.sched_priority = p->rt_priority;
4968 * This one might sleep, we cannot do it with a spinlock held ...
4970 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4979 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4981 cpumask_var_t cpus_allowed, new_mask;
4982 struct task_struct *p;
4988 p = find_process_by_pid(pid);
4995 /* Prevent p going away */
4999 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5003 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5005 goto out_free_cpus_allowed;
5008 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
5011 retval = security_task_setscheduler(p, 0, NULL);
5015 cpuset_cpus_allowed(p, cpus_allowed);
5016 cpumask_and(new_mask, in_mask, cpus_allowed);
5018 retval = set_cpus_allowed_ptr(p, new_mask);
5021 cpuset_cpus_allowed(p, cpus_allowed);
5022 if (!cpumask_subset(new_mask, cpus_allowed)) {
5024 * We must have raced with a concurrent cpuset
5025 * update. Just reset the cpus_allowed to the
5026 * cpuset's cpus_allowed
5028 cpumask_copy(new_mask, cpus_allowed);
5033 free_cpumask_var(new_mask);
5034 out_free_cpus_allowed:
5035 free_cpumask_var(cpus_allowed);
5042 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5043 struct cpumask *new_mask)
5045 if (len < cpumask_size())
5046 cpumask_clear(new_mask);
5047 else if (len > cpumask_size())
5048 len = cpumask_size();
5050 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5054 * sys_sched_setaffinity - set the cpu affinity of a process
5055 * @pid: pid of the process
5056 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5057 * @user_mask_ptr: user-space pointer to the new cpu mask
5059 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5060 unsigned long __user *, user_mask_ptr)
5062 cpumask_var_t new_mask;
5065 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5068 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5070 retval = sched_setaffinity(pid, new_mask);
5071 free_cpumask_var(new_mask);
5075 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5077 struct task_struct *p;
5078 unsigned long flags;
5086 p = find_process_by_pid(pid);
5090 retval = security_task_getscheduler(p);
5094 rq = task_rq_lock(p, &flags);
5095 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5096 task_rq_unlock(rq, &flags);
5106 * sys_sched_getaffinity - get the cpu affinity of a process
5107 * @pid: pid of the process
5108 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5109 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5111 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5112 unsigned long __user *, user_mask_ptr)
5117 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5119 if (len & (sizeof(unsigned long)-1))
5122 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5125 ret = sched_getaffinity(pid, mask);
5127 size_t retlen = min_t(size_t, len, cpumask_size());
5129 if (copy_to_user(user_mask_ptr, mask, retlen))
5134 free_cpumask_var(mask);
5140 * sys_sched_yield - yield the current processor to other threads.
5142 * This function yields the current CPU to other tasks. If there are no
5143 * other threads running on this CPU then this function will return.
5145 SYSCALL_DEFINE0(sched_yield)
5147 struct rq *rq = this_rq_lock();
5149 schedstat_inc(rq, yld_count);
5150 current->sched_class->yield_task(rq);
5153 * Since we are going to call schedule() anyway, there's
5154 * no need to preempt or enable interrupts:
5156 __release(rq->lock);
5157 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5158 do_raw_spin_unlock(&rq->lock);
5159 preempt_enable_no_resched();
5166 static inline int should_resched(void)
5168 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5171 static void __cond_resched(void)
5173 add_preempt_count(PREEMPT_ACTIVE);
5175 sub_preempt_count(PREEMPT_ACTIVE);
5178 int __sched _cond_resched(void)
5180 if (should_resched()) {
5186 EXPORT_SYMBOL(_cond_resched);
5189 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5190 * call schedule, and on return reacquire the lock.
5192 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5193 * operations here to prevent schedule() from being called twice (once via
5194 * spin_unlock(), once by hand).
5196 int __cond_resched_lock(spinlock_t *lock)
5198 int resched = should_resched();
5201 lockdep_assert_held(lock);
5203 if (spin_needbreak(lock) || resched) {
5214 EXPORT_SYMBOL(__cond_resched_lock);
5216 int __sched __cond_resched_softirq(void)
5218 BUG_ON(!in_softirq());
5220 if (should_resched()) {
5228 EXPORT_SYMBOL(__cond_resched_softirq);
5231 * yield - yield the current processor to other threads.
5233 * This is a shortcut for kernel-space yielding - it marks the
5234 * thread runnable and calls sys_sched_yield().
5236 void __sched yield(void)
5238 set_current_state(TASK_RUNNING);
5241 EXPORT_SYMBOL(yield);
5244 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5245 * that process accounting knows that this is a task in IO wait state.
5247 void __sched io_schedule(void)
5249 struct rq *rq = raw_rq();
5251 delayacct_blkio_start();
5252 atomic_inc(&rq->nr_iowait);
5253 current->in_iowait = 1;
5255 current->in_iowait = 0;
5256 atomic_dec(&rq->nr_iowait);
5257 delayacct_blkio_end();
5259 EXPORT_SYMBOL(io_schedule);
5261 long __sched io_schedule_timeout(long timeout)
5263 struct rq *rq = raw_rq();
5266 delayacct_blkio_start();
5267 atomic_inc(&rq->nr_iowait);
5268 current->in_iowait = 1;
5269 ret = schedule_timeout(timeout);
5270 current->in_iowait = 0;
5271 atomic_dec(&rq->nr_iowait);
5272 delayacct_blkio_end();
5277 * sys_sched_get_priority_max - return maximum RT priority.
5278 * @policy: scheduling class.
5280 * this syscall returns the maximum rt_priority that can be used
5281 * by a given scheduling class.
5283 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5290 ret = MAX_USER_RT_PRIO-1;
5302 * sys_sched_get_priority_min - return minimum RT priority.
5303 * @policy: scheduling class.
5305 * this syscall returns the minimum rt_priority that can be used
5306 * by a given scheduling class.
5308 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5326 * sys_sched_rr_get_interval - return the default timeslice of a process.
5327 * @pid: pid of the process.
5328 * @interval: userspace pointer to the timeslice value.
5330 * this syscall writes the default timeslice value of a given process
5331 * into the user-space timespec buffer. A value of '0' means infinity.
5333 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5334 struct timespec __user *, interval)
5336 struct task_struct *p;
5337 unsigned int time_slice;
5338 unsigned long flags;
5348 p = find_process_by_pid(pid);
5352 retval = security_task_getscheduler(p);
5356 rq = task_rq_lock(p, &flags);
5357 time_slice = p->sched_class->get_rr_interval(rq, p);
5358 task_rq_unlock(rq, &flags);
5361 jiffies_to_timespec(time_slice, &t);
5362 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5370 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5372 void sched_show_task(struct task_struct *p)
5374 unsigned long free = 0;
5377 state = p->state ? __ffs(p->state) + 1 : 0;
5378 printk(KERN_INFO "%-13.13s %c", p->comm,
5379 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5380 #if BITS_PER_LONG == 32
5381 if (state == TASK_RUNNING)
5382 printk(KERN_CONT " running ");
5384 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5386 if (state == TASK_RUNNING)
5387 printk(KERN_CONT " running task ");
5389 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5391 #ifdef CONFIG_DEBUG_STACK_USAGE
5392 free = stack_not_used(p);
5394 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5395 task_pid_nr(p), task_pid_nr(p->real_parent),
5396 (unsigned long)task_thread_info(p)->flags);
5398 show_stack(p, NULL);
5401 void show_state_filter(unsigned long state_filter)
5403 struct task_struct *g, *p;
5405 #if BITS_PER_LONG == 32
5407 " task PC stack pid father\n");
5410 " task PC stack pid father\n");
5412 read_lock(&tasklist_lock);
5413 do_each_thread(g, p) {
5415 * reset the NMI-timeout, listing all files on a slow
5416 * console might take alot of time:
5418 touch_nmi_watchdog();
5419 if (!state_filter || (p->state & state_filter))
5421 } while_each_thread(g, p);
5423 touch_all_softlockup_watchdogs();
5425 #ifdef CONFIG_SCHED_DEBUG
5426 sysrq_sched_debug_show();
5428 read_unlock(&tasklist_lock);
5430 * Only show locks if all tasks are dumped:
5433 debug_show_all_locks();
5436 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5438 idle->sched_class = &idle_sched_class;
5442 * init_idle - set up an idle thread for a given CPU
5443 * @idle: task in question
5444 * @cpu: cpu the idle task belongs to
5446 * NOTE: this function does not set the idle thread's NEED_RESCHED
5447 * flag, to make booting more robust.
5449 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5451 struct rq *rq = cpu_rq(cpu);
5452 unsigned long flags;
5454 raw_spin_lock_irqsave(&rq->lock, flags);
5457 idle->state = TASK_RUNNING;
5458 idle->se.exec_start = sched_clock();
5460 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
5462 * We're having a chicken and egg problem, even though we are
5463 * holding rq->lock, the cpu isn't yet set to this cpu so the
5464 * lockdep check in task_group() will fail.
5466 * Similar case to sched_fork(). / Alternatively we could
5467 * use task_rq_lock() here and obtain the other rq->lock.
5472 __set_task_cpu(idle, cpu);
5475 rq->curr = rq->idle = idle;
5476 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5479 raw_spin_unlock_irqrestore(&rq->lock, flags);
5481 /* Set the preempt count _outside_ the spinlocks! */
5482 #if defined(CONFIG_PREEMPT)
5483 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5485 task_thread_info(idle)->preempt_count = 0;
5488 * The idle tasks have their own, simple scheduling class:
5490 idle->sched_class = &idle_sched_class;
5491 ftrace_graph_init_task(idle);
5495 * In a system that switches off the HZ timer nohz_cpu_mask
5496 * indicates which cpus entered this state. This is used
5497 * in the rcu update to wait only for active cpus. For system
5498 * which do not switch off the HZ timer nohz_cpu_mask should
5499 * always be CPU_BITS_NONE.
5501 cpumask_var_t nohz_cpu_mask;
5504 * Increase the granularity value when there are more CPUs,
5505 * because with more CPUs the 'effective latency' as visible
5506 * to users decreases. But the relationship is not linear,
5507 * so pick a second-best guess by going with the log2 of the
5510 * This idea comes from the SD scheduler of Con Kolivas:
5512 static int get_update_sysctl_factor(void)
5514 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5515 unsigned int factor;
5517 switch (sysctl_sched_tunable_scaling) {
5518 case SCHED_TUNABLESCALING_NONE:
5521 case SCHED_TUNABLESCALING_LINEAR:
5524 case SCHED_TUNABLESCALING_LOG:
5526 factor = 1 + ilog2(cpus);
5533 static void update_sysctl(void)
5535 unsigned int factor = get_update_sysctl_factor();
5537 #define SET_SYSCTL(name) \
5538 (sysctl_##name = (factor) * normalized_sysctl_##name)
5539 SET_SYSCTL(sched_min_granularity);
5540 SET_SYSCTL(sched_latency);
5541 SET_SYSCTL(sched_wakeup_granularity);
5542 SET_SYSCTL(sched_shares_ratelimit);
5546 static inline void sched_init_granularity(void)
5553 * This is how migration works:
5555 * 1) we invoke migration_cpu_stop() on the target CPU using
5557 * 2) stopper starts to run (implicitly forcing the migrated thread
5559 * 3) it checks whether the migrated task is still in the wrong runqueue.
5560 * 4) if it's in the wrong runqueue then the migration thread removes
5561 * it and puts it into the right queue.
5562 * 5) stopper completes and stop_one_cpu() returns and the migration
5567 * Change a given task's CPU affinity. Migrate the thread to a
5568 * proper CPU and schedule it away if the CPU it's executing on
5569 * is removed from the allowed bitmask.
5571 * NOTE: the caller must have a valid reference to the task, the
5572 * task must not exit() & deallocate itself prematurely. The
5573 * call is not atomic; no spinlocks may be held.
5575 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5577 unsigned long flags;
5579 unsigned int dest_cpu;
5583 * Serialize against TASK_WAKING so that ttwu() and wunt() can
5584 * drop the rq->lock and still rely on ->cpus_allowed.
5587 while (task_is_waking(p))
5589 rq = task_rq_lock(p, &flags);
5590 if (task_is_waking(p)) {
5591 task_rq_unlock(rq, &flags);
5595 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5600 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5601 !cpumask_equal(&p->cpus_allowed, new_mask))) {
5606 if (p->sched_class->set_cpus_allowed)
5607 p->sched_class->set_cpus_allowed(p, new_mask);
5609 cpumask_copy(&p->cpus_allowed, new_mask);
5610 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5613 /* Can the task run on the task's current CPU? If so, we're done */
5614 if (cpumask_test_cpu(task_cpu(p), new_mask))
5617 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
5618 if (migrate_task(p, dest_cpu)) {
5619 struct migration_arg arg = { p, dest_cpu };
5620 /* Need help from migration thread: drop lock and wait. */
5621 task_rq_unlock(rq, &flags);
5622 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
5623 tlb_migrate_finish(p->mm);
5627 task_rq_unlock(rq, &flags);
5631 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5634 * Move (not current) task off this cpu, onto dest cpu. We're doing
5635 * this because either it can't run here any more (set_cpus_allowed()
5636 * away from this CPU, or CPU going down), or because we're
5637 * attempting to rebalance this task on exec (sched_exec).
5639 * So we race with normal scheduler movements, but that's OK, as long
5640 * as the task is no longer on this CPU.
5642 * Returns non-zero if task was successfully migrated.
5644 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5646 struct rq *rq_dest, *rq_src;
5649 if (unlikely(!cpu_active(dest_cpu)))
5652 rq_src = cpu_rq(src_cpu);
5653 rq_dest = cpu_rq(dest_cpu);
5655 double_rq_lock(rq_src, rq_dest);
5656 /* Already moved. */
5657 if (task_cpu(p) != src_cpu)
5659 /* Affinity changed (again). */
5660 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
5664 * If we're not on a rq, the next wake-up will ensure we're
5668 deactivate_task(rq_src, p, 0);
5669 set_task_cpu(p, dest_cpu);
5670 activate_task(rq_dest, p, 0);
5671 check_preempt_curr(rq_dest, p, 0);
5676 double_rq_unlock(rq_src, rq_dest);
5681 * migration_cpu_stop - this will be executed by a highprio stopper thread
5682 * and performs thread migration by bumping thread off CPU then
5683 * 'pushing' onto another runqueue.
5685 static int migration_cpu_stop(void *data)
5687 struct migration_arg *arg = data;
5690 * The original target cpu might have gone down and we might
5691 * be on another cpu but it doesn't matter.
5693 local_irq_disable();
5694 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
5699 #ifdef CONFIG_HOTPLUG_CPU
5701 * Figure out where task on dead CPU should go, use force if necessary.
5703 void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5705 struct rq *rq = cpu_rq(dead_cpu);
5706 int needs_cpu, uninitialized_var(dest_cpu);
5707 unsigned long flags;
5709 local_irq_save(flags);
5711 raw_spin_lock(&rq->lock);
5712 needs_cpu = (task_cpu(p) == dead_cpu) && (p->state != TASK_WAKING);
5714 dest_cpu = select_fallback_rq(dead_cpu, p);
5715 raw_spin_unlock(&rq->lock);
5717 * It can only fail if we race with set_cpus_allowed(),
5718 * in the racer should migrate the task anyway.
5721 __migrate_task(p, dead_cpu, dest_cpu);
5722 local_irq_restore(flags);
5726 * While a dead CPU has no uninterruptible tasks queued at this point,
5727 * it might still have a nonzero ->nr_uninterruptible counter, because
5728 * for performance reasons the counter is not stricly tracking tasks to
5729 * their home CPUs. So we just add the counter to another CPU's counter,
5730 * to keep the global sum constant after CPU-down:
5732 static void migrate_nr_uninterruptible(struct rq *rq_src)
5734 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5735 unsigned long flags;
5737 local_irq_save(flags);
5738 double_rq_lock(rq_src, rq_dest);
5739 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5740 rq_src->nr_uninterruptible = 0;
5741 double_rq_unlock(rq_src, rq_dest);
5742 local_irq_restore(flags);
5745 /* Run through task list and migrate tasks from the dead cpu. */
5746 static void migrate_live_tasks(int src_cpu)
5748 struct task_struct *p, *t;
5750 read_lock(&tasklist_lock);
5752 do_each_thread(t, p) {
5756 if (task_cpu(p) == src_cpu)
5757 move_task_off_dead_cpu(src_cpu, p);
5758 } while_each_thread(t, p);
5760 read_unlock(&tasklist_lock);
5764 * Schedules idle task to be the next runnable task on current CPU.
5765 * It does so by boosting its priority to highest possible.
5766 * Used by CPU offline code.
5768 void sched_idle_next(void)
5770 int this_cpu = smp_processor_id();
5771 struct rq *rq = cpu_rq(this_cpu);
5772 struct task_struct *p = rq->idle;
5773 unsigned long flags;
5775 /* cpu has to be offline */
5776 BUG_ON(cpu_online(this_cpu));
5779 * Strictly not necessary since rest of the CPUs are stopped by now
5780 * and interrupts disabled on the current cpu.
5782 raw_spin_lock_irqsave(&rq->lock, flags);
5784 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5786 activate_task(rq, p, 0);
5788 raw_spin_unlock_irqrestore(&rq->lock, flags);
5792 * Ensures that the idle task is using init_mm right before its cpu goes
5795 void idle_task_exit(void)
5797 struct mm_struct *mm = current->active_mm;
5799 BUG_ON(cpu_online(smp_processor_id()));
5802 switch_mm(mm, &init_mm, current);
5806 /* called under rq->lock with disabled interrupts */
5807 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5809 struct rq *rq = cpu_rq(dead_cpu);
5811 /* Must be exiting, otherwise would be on tasklist. */
5812 BUG_ON(!p->exit_state);
5814 /* Cannot have done final schedule yet: would have vanished. */
5815 BUG_ON(p->state == TASK_DEAD);
5820 * Drop lock around migration; if someone else moves it,
5821 * that's OK. No task can be added to this CPU, so iteration is
5824 raw_spin_unlock_irq(&rq->lock);
5825 move_task_off_dead_cpu(dead_cpu, p);
5826 raw_spin_lock_irq(&rq->lock);
5831 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5832 static void migrate_dead_tasks(unsigned int dead_cpu)
5834 struct rq *rq = cpu_rq(dead_cpu);
5835 struct task_struct *next;
5838 if (!rq->nr_running)
5840 next = pick_next_task(rq);
5843 next->sched_class->put_prev_task(rq, next);
5844 migrate_dead(dead_cpu, next);
5850 * remove the tasks which were accounted by rq from calc_load_tasks.
5852 static void calc_global_load_remove(struct rq *rq)
5854 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5855 rq->calc_load_active = 0;
5857 #endif /* CONFIG_HOTPLUG_CPU */
5859 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5861 static struct ctl_table sd_ctl_dir[] = {
5863 .procname = "sched_domain",
5869 static struct ctl_table sd_ctl_root[] = {
5871 .procname = "kernel",
5873 .child = sd_ctl_dir,
5878 static struct ctl_table *sd_alloc_ctl_entry(int n)
5880 struct ctl_table *entry =
5881 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5886 static void sd_free_ctl_entry(struct ctl_table **tablep)
5888 struct ctl_table *entry;
5891 * In the intermediate directories, both the child directory and
5892 * procname are dynamically allocated and could fail but the mode
5893 * will always be set. In the lowest directory the names are
5894 * static strings and all have proc handlers.
5896 for (entry = *tablep; entry->mode; entry++) {
5898 sd_free_ctl_entry(&entry->child);
5899 if (entry->proc_handler == NULL)
5900 kfree(entry->procname);
5908 set_table_entry(struct ctl_table *entry,
5909 const char *procname, void *data, int maxlen,
5910 mode_t mode, proc_handler *proc_handler)
5912 entry->procname = procname;
5914 entry->maxlen = maxlen;
5916 entry->proc_handler = proc_handler;
5919 static struct ctl_table *
5920 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5922 struct ctl_table *table = sd_alloc_ctl_entry(13);
5927 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5928 sizeof(long), 0644, proc_doulongvec_minmax);
5929 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5930 sizeof(long), 0644, proc_doulongvec_minmax);
5931 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5932 sizeof(int), 0644, proc_dointvec_minmax);
5933 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5934 sizeof(int), 0644, proc_dointvec_minmax);
5935 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5936 sizeof(int), 0644, proc_dointvec_minmax);
5937 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5938 sizeof(int), 0644, proc_dointvec_minmax);
5939 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5940 sizeof(int), 0644, proc_dointvec_minmax);
5941 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5942 sizeof(int), 0644, proc_dointvec_minmax);
5943 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5944 sizeof(int), 0644, proc_dointvec_minmax);
5945 set_table_entry(&table[9], "cache_nice_tries",
5946 &sd->cache_nice_tries,
5947 sizeof(int), 0644, proc_dointvec_minmax);
5948 set_table_entry(&table[10], "flags", &sd->flags,
5949 sizeof(int), 0644, proc_dointvec_minmax);
5950 set_table_entry(&table[11], "name", sd->name,
5951 CORENAME_MAX_SIZE, 0444, proc_dostring);
5952 /* &table[12] is terminator */
5957 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5959 struct ctl_table *entry, *table;
5960 struct sched_domain *sd;
5961 int domain_num = 0, i;
5964 for_each_domain(cpu, sd)
5966 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5971 for_each_domain(cpu, sd) {
5972 snprintf(buf, 32, "domain%d", i);
5973 entry->procname = kstrdup(buf, GFP_KERNEL);
5975 entry->child = sd_alloc_ctl_domain_table(sd);
5982 static struct ctl_table_header *sd_sysctl_header;
5983 static void register_sched_domain_sysctl(void)
5985 int i, cpu_num = num_possible_cpus();
5986 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5989 WARN_ON(sd_ctl_dir[0].child);
5990 sd_ctl_dir[0].child = entry;
5995 for_each_possible_cpu(i) {
5996 snprintf(buf, 32, "cpu%d", i);
5997 entry->procname = kstrdup(buf, GFP_KERNEL);
5999 entry->child = sd_alloc_ctl_cpu_table(i);
6003 WARN_ON(sd_sysctl_header);
6004 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6007 /* may be called multiple times per register */
6008 static void unregister_sched_domain_sysctl(void)
6010 if (sd_sysctl_header)
6011 unregister_sysctl_table(sd_sysctl_header);
6012 sd_sysctl_header = NULL;
6013 if (sd_ctl_dir[0].child)
6014 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6017 static void register_sched_domain_sysctl(void)
6020 static void unregister_sched_domain_sysctl(void)
6025 static void set_rq_online(struct rq *rq)
6028 const struct sched_class *class;
6030 cpumask_set_cpu(rq->cpu, rq->rd->online);
6033 for_each_class(class) {
6034 if (class->rq_online)
6035 class->rq_online(rq);
6040 static void set_rq_offline(struct rq *rq)
6043 const struct sched_class *class;
6045 for_each_class(class) {
6046 if (class->rq_offline)
6047 class->rq_offline(rq);
6050 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6056 * migration_call - callback that gets triggered when a CPU is added.
6057 * Here we can start up the necessary migration thread for the new CPU.
6059 static int __cpuinit
6060 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6062 int cpu = (long)hcpu;
6063 unsigned long flags;
6064 struct rq *rq = cpu_rq(cpu);
6068 case CPU_UP_PREPARE:
6069 case CPU_UP_PREPARE_FROZEN:
6070 rq->calc_load_update = calc_load_update;
6074 case CPU_ONLINE_FROZEN:
6075 /* Update our root-domain */
6076 raw_spin_lock_irqsave(&rq->lock, flags);
6078 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6082 raw_spin_unlock_irqrestore(&rq->lock, flags);
6085 #ifdef CONFIG_HOTPLUG_CPU
6087 case CPU_DEAD_FROZEN:
6088 migrate_live_tasks(cpu);
6089 /* Idle task back to normal (off runqueue, low prio) */
6090 raw_spin_lock_irq(&rq->lock);
6091 deactivate_task(rq, rq->idle, 0);
6092 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6093 rq->idle->sched_class = &idle_sched_class;
6094 migrate_dead_tasks(cpu);
6095 raw_spin_unlock_irq(&rq->lock);
6096 migrate_nr_uninterruptible(rq);
6097 BUG_ON(rq->nr_running != 0);
6098 calc_global_load_remove(rq);
6102 case CPU_DYING_FROZEN:
6103 /* Update our root-domain */
6104 raw_spin_lock_irqsave(&rq->lock, flags);
6106 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6109 raw_spin_unlock_irqrestore(&rq->lock, flags);
6117 * Register at high priority so that task migration (migrate_all_tasks)
6118 * happens before everything else. This has to be lower priority than
6119 * the notifier in the perf_event subsystem, though.
6121 static struct notifier_block __cpuinitdata migration_notifier = {
6122 .notifier_call = migration_call,
6123 .priority = CPU_PRI_MIGRATION,
6126 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
6127 unsigned long action, void *hcpu)
6129 switch (action & ~CPU_TASKS_FROZEN) {
6131 case CPU_DOWN_FAILED:
6132 set_cpu_active((long)hcpu, true);
6139 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
6140 unsigned long action, void *hcpu)
6142 switch (action & ~CPU_TASKS_FROZEN) {
6143 case CPU_DOWN_PREPARE:
6144 set_cpu_active((long)hcpu, false);
6151 static int __init migration_init(void)
6153 void *cpu = (void *)(long)smp_processor_id();
6156 /* Initialize migration for the boot CPU */
6157 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6158 BUG_ON(err == NOTIFY_BAD);
6159 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6160 register_cpu_notifier(&migration_notifier);
6162 /* Register cpu active notifiers */
6163 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
6164 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
6168 early_initcall(migration_init);
6173 #ifdef CONFIG_SCHED_DEBUG
6175 static __read_mostly int sched_domain_debug_enabled;
6177 static int __init sched_domain_debug_setup(char *str)
6179 sched_domain_debug_enabled = 1;
6183 early_param("sched_debug", sched_domain_debug_setup);
6185 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6186 struct cpumask *groupmask)
6188 struct sched_group *group = sd->groups;
6191 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6192 cpumask_clear(groupmask);
6194 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6196 if (!(sd->flags & SD_LOAD_BALANCE)) {
6197 printk("does not load-balance\n");
6199 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6204 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6206 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6207 printk(KERN_ERR "ERROR: domain->span does not contain "
6210 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6211 printk(KERN_ERR "ERROR: domain->groups does not contain"
6215 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6219 printk(KERN_ERR "ERROR: group is NULL\n");
6223 if (!group->cpu_power) {
6224 printk(KERN_CONT "\n");
6225 printk(KERN_ERR "ERROR: domain->cpu_power not "
6230 if (!cpumask_weight(sched_group_cpus(group))) {
6231 printk(KERN_CONT "\n");
6232 printk(KERN_ERR "ERROR: empty group\n");
6236 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6237 printk(KERN_CONT "\n");
6238 printk(KERN_ERR "ERROR: repeated CPUs\n");
6242 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6244 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6246 printk(KERN_CONT " %s", str);
6247 if (group->cpu_power != SCHED_LOAD_SCALE) {
6248 printk(KERN_CONT " (cpu_power = %d)",
6252 group = group->next;
6253 } while (group != sd->groups);
6254 printk(KERN_CONT "\n");
6256 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6257 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6260 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6261 printk(KERN_ERR "ERROR: parent span is not a superset "
6262 "of domain->span\n");
6266 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6268 cpumask_var_t groupmask;
6271 if (!sched_domain_debug_enabled)
6275 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6279 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6281 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6282 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6287 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6294 free_cpumask_var(groupmask);
6296 #else /* !CONFIG_SCHED_DEBUG */
6297 # define sched_domain_debug(sd, cpu) do { } while (0)
6298 #endif /* CONFIG_SCHED_DEBUG */
6300 static int sd_degenerate(struct sched_domain *sd)
6302 if (cpumask_weight(sched_domain_span(sd)) == 1)
6305 /* Following flags need at least 2 groups */
6306 if (sd->flags & (SD_LOAD_BALANCE |
6307 SD_BALANCE_NEWIDLE |
6311 SD_SHARE_PKG_RESOURCES)) {
6312 if (sd->groups != sd->groups->next)
6316 /* Following flags don't use groups */
6317 if (sd->flags & (SD_WAKE_AFFINE))
6324 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6326 unsigned long cflags = sd->flags, pflags = parent->flags;
6328 if (sd_degenerate(parent))
6331 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6334 /* Flags needing groups don't count if only 1 group in parent */
6335 if (parent->groups == parent->groups->next) {
6336 pflags &= ~(SD_LOAD_BALANCE |
6337 SD_BALANCE_NEWIDLE |
6341 SD_SHARE_PKG_RESOURCES);
6342 if (nr_node_ids == 1)
6343 pflags &= ~SD_SERIALIZE;
6345 if (~cflags & pflags)
6351 static void free_rootdomain(struct root_domain *rd)
6353 synchronize_sched();
6355 cpupri_cleanup(&rd->cpupri);
6357 free_cpumask_var(rd->rto_mask);
6358 free_cpumask_var(rd->online);
6359 free_cpumask_var(rd->span);
6363 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6365 struct root_domain *old_rd = NULL;
6366 unsigned long flags;
6368 raw_spin_lock_irqsave(&rq->lock, flags);
6373 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6376 cpumask_clear_cpu(rq->cpu, old_rd->span);
6379 * If we dont want to free the old_rt yet then
6380 * set old_rd to NULL to skip the freeing later
6383 if (!atomic_dec_and_test(&old_rd->refcount))
6387 atomic_inc(&rd->refcount);
6390 cpumask_set_cpu(rq->cpu, rd->span);
6391 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6394 raw_spin_unlock_irqrestore(&rq->lock, flags);
6397 free_rootdomain(old_rd);
6400 static int init_rootdomain(struct root_domain *rd)
6402 memset(rd, 0, sizeof(*rd));
6404 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
6406 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
6408 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6411 if (cpupri_init(&rd->cpupri) != 0)
6416 free_cpumask_var(rd->rto_mask);
6418 free_cpumask_var(rd->online);
6420 free_cpumask_var(rd->span);
6425 static void init_defrootdomain(void)
6427 init_rootdomain(&def_root_domain);
6429 atomic_set(&def_root_domain.refcount, 1);
6432 static struct root_domain *alloc_rootdomain(void)
6434 struct root_domain *rd;
6436 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6440 if (init_rootdomain(rd) != 0) {
6449 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6450 * hold the hotplug lock.
6453 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6455 struct rq *rq = cpu_rq(cpu);
6456 struct sched_domain *tmp;
6458 for (tmp = sd; tmp; tmp = tmp->parent)
6459 tmp->span_weight = cpumask_weight(sched_domain_span(tmp));
6461 /* Remove the sched domains which do not contribute to scheduling. */
6462 for (tmp = sd; tmp; ) {
6463 struct sched_domain *parent = tmp->parent;
6467 if (sd_parent_degenerate(tmp, parent)) {
6468 tmp->parent = parent->parent;
6470 parent->parent->child = tmp;
6475 if (sd && sd_degenerate(sd)) {
6481 sched_domain_debug(sd, cpu);
6483 rq_attach_root(rq, rd);
6484 rcu_assign_pointer(rq->sd, sd);
6487 /* cpus with isolated domains */
6488 static cpumask_var_t cpu_isolated_map;
6490 /* Setup the mask of cpus configured for isolated domains */
6491 static int __init isolated_cpu_setup(char *str)
6493 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6494 cpulist_parse(str, cpu_isolated_map);
6498 __setup("isolcpus=", isolated_cpu_setup);
6501 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6502 * to a function which identifies what group(along with sched group) a CPU
6503 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6504 * (due to the fact that we keep track of groups covered with a struct cpumask).
6506 * init_sched_build_groups will build a circular linked list of the groups
6507 * covered by the given span, and will set each group's ->cpumask correctly,
6508 * and ->cpu_power to 0.
6511 init_sched_build_groups(const struct cpumask *span,
6512 const struct cpumask *cpu_map,
6513 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
6514 struct sched_group **sg,
6515 struct cpumask *tmpmask),
6516 struct cpumask *covered, struct cpumask *tmpmask)
6518 struct sched_group *first = NULL, *last = NULL;
6521 cpumask_clear(covered);
6523 for_each_cpu(i, span) {
6524 struct sched_group *sg;
6525 int group = group_fn(i, cpu_map, &sg, tmpmask);
6528 if (cpumask_test_cpu(i, covered))
6531 cpumask_clear(sched_group_cpus(sg));
6534 for_each_cpu(j, span) {
6535 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6538 cpumask_set_cpu(j, covered);
6539 cpumask_set_cpu(j, sched_group_cpus(sg));
6550 #define SD_NODES_PER_DOMAIN 16
6555 * find_next_best_node - find the next node to include in a sched_domain
6556 * @node: node whose sched_domain we're building
6557 * @used_nodes: nodes already in the sched_domain
6559 * Find the next node to include in a given scheduling domain. Simply
6560 * finds the closest node not already in the @used_nodes map.
6562 * Should use nodemask_t.
6564 static int find_next_best_node(int node, nodemask_t *used_nodes)
6566 int i, n, val, min_val, best_node = 0;
6570 for (i = 0; i < nr_node_ids; i++) {
6571 /* Start at @node */
6572 n = (node + i) % nr_node_ids;
6574 if (!nr_cpus_node(n))
6577 /* Skip already used nodes */
6578 if (node_isset(n, *used_nodes))
6581 /* Simple min distance search */
6582 val = node_distance(node, n);
6584 if (val < min_val) {
6590 node_set(best_node, *used_nodes);
6595 * sched_domain_node_span - get a cpumask for a node's sched_domain
6596 * @node: node whose cpumask we're constructing
6597 * @span: resulting cpumask
6599 * Given a node, construct a good cpumask for its sched_domain to span. It
6600 * should be one that prevents unnecessary balancing, but also spreads tasks
6603 static void sched_domain_node_span(int node, struct cpumask *span)
6605 nodemask_t used_nodes;
6608 cpumask_clear(span);
6609 nodes_clear(used_nodes);
6611 cpumask_or(span, span, cpumask_of_node(node));
6612 node_set(node, used_nodes);
6614 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6615 int next_node = find_next_best_node(node, &used_nodes);
6617 cpumask_or(span, span, cpumask_of_node(next_node));
6620 #endif /* CONFIG_NUMA */
6622 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6625 * The cpus mask in sched_group and sched_domain hangs off the end.
6627 * ( See the the comments in include/linux/sched.h:struct sched_group
6628 * and struct sched_domain. )
6630 struct static_sched_group {
6631 struct sched_group sg;
6632 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
6635 struct static_sched_domain {
6636 struct sched_domain sd;
6637 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
6643 cpumask_var_t domainspan;
6644 cpumask_var_t covered;
6645 cpumask_var_t notcovered;
6647 cpumask_var_t nodemask;
6648 cpumask_var_t this_sibling_map;
6649 cpumask_var_t this_core_map;
6650 cpumask_var_t send_covered;
6651 cpumask_var_t tmpmask;
6652 struct sched_group **sched_group_nodes;
6653 struct root_domain *rd;
6657 sa_sched_groups = 0,
6662 sa_this_sibling_map,
6664 sa_sched_group_nodes,
6674 * SMT sched-domains:
6676 #ifdef CONFIG_SCHED_SMT
6677 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
6678 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
6681 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
6682 struct sched_group **sg, struct cpumask *unused)
6685 *sg = &per_cpu(sched_groups, cpu).sg;
6688 #endif /* CONFIG_SCHED_SMT */
6691 * multi-core sched-domains:
6693 #ifdef CONFIG_SCHED_MC
6694 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
6695 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
6696 #endif /* CONFIG_SCHED_MC */
6698 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6700 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6701 struct sched_group **sg, struct cpumask *mask)
6705 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6706 group = cpumask_first(mask);
6708 *sg = &per_cpu(sched_group_core, group).sg;
6711 #elif defined(CONFIG_SCHED_MC)
6713 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6714 struct sched_group **sg, struct cpumask *unused)
6717 *sg = &per_cpu(sched_group_core, cpu).sg;
6722 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
6723 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
6726 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
6727 struct sched_group **sg, struct cpumask *mask)
6730 #ifdef CONFIG_SCHED_MC
6731 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6732 group = cpumask_first(mask);
6733 #elif defined(CONFIG_SCHED_SMT)
6734 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6735 group = cpumask_first(mask);
6740 *sg = &per_cpu(sched_group_phys, group).sg;
6746 * The init_sched_build_groups can't handle what we want to do with node
6747 * groups, so roll our own. Now each node has its own list of groups which
6748 * gets dynamically allocated.
6750 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
6751 static struct sched_group ***sched_group_nodes_bycpu;
6753 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
6754 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
6756 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
6757 struct sched_group **sg,
6758 struct cpumask *nodemask)
6762 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
6763 group = cpumask_first(nodemask);
6766 *sg = &per_cpu(sched_group_allnodes, group).sg;
6770 static void init_numa_sched_groups_power(struct sched_group *group_head)
6772 struct sched_group *sg = group_head;
6778 for_each_cpu(j, sched_group_cpus(sg)) {
6779 struct sched_domain *sd;
6781 sd = &per_cpu(phys_domains, j).sd;
6782 if (j != group_first_cpu(sd->groups)) {
6784 * Only add "power" once for each
6790 sg->cpu_power += sd->groups->cpu_power;
6793 } while (sg != group_head);
6796 static int build_numa_sched_groups(struct s_data *d,
6797 const struct cpumask *cpu_map, int num)
6799 struct sched_domain *sd;
6800 struct sched_group *sg, *prev;
6803 cpumask_clear(d->covered);
6804 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
6805 if (cpumask_empty(d->nodemask)) {
6806 d->sched_group_nodes[num] = NULL;
6810 sched_domain_node_span(num, d->domainspan);
6811 cpumask_and(d->domainspan, d->domainspan, cpu_map);
6813 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6816 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
6820 d->sched_group_nodes[num] = sg;
6822 for_each_cpu(j, d->nodemask) {
6823 sd = &per_cpu(node_domains, j).sd;
6828 cpumask_copy(sched_group_cpus(sg), d->nodemask);
6830 cpumask_or(d->covered, d->covered, d->nodemask);
6833 for (j = 0; j < nr_node_ids; j++) {
6834 n = (num + j) % nr_node_ids;
6835 cpumask_complement(d->notcovered, d->covered);
6836 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
6837 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
6838 if (cpumask_empty(d->tmpmask))
6840 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
6841 if (cpumask_empty(d->tmpmask))
6843 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6847 "Can not alloc domain group for node %d\n", j);
6851 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
6852 sg->next = prev->next;
6853 cpumask_or(d->covered, d->covered, d->tmpmask);
6860 #endif /* CONFIG_NUMA */
6863 /* Free memory allocated for various sched_group structures */
6864 static void free_sched_groups(const struct cpumask *cpu_map,
6865 struct cpumask *nodemask)
6869 for_each_cpu(cpu, cpu_map) {
6870 struct sched_group **sched_group_nodes
6871 = sched_group_nodes_bycpu[cpu];
6873 if (!sched_group_nodes)
6876 for (i = 0; i < nr_node_ids; i++) {
6877 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6879 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
6880 if (cpumask_empty(nodemask))
6890 if (oldsg != sched_group_nodes[i])
6893 kfree(sched_group_nodes);
6894 sched_group_nodes_bycpu[cpu] = NULL;
6897 #else /* !CONFIG_NUMA */
6898 static void free_sched_groups(const struct cpumask *cpu_map,
6899 struct cpumask *nodemask)
6902 #endif /* CONFIG_NUMA */
6905 * Initialize sched groups cpu_power.
6907 * cpu_power indicates the capacity of sched group, which is used while
6908 * distributing the load between different sched groups in a sched domain.
6909 * Typically cpu_power for all the groups in a sched domain will be same unless
6910 * there are asymmetries in the topology. If there are asymmetries, group
6911 * having more cpu_power will pickup more load compared to the group having
6914 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6916 struct sched_domain *child;
6917 struct sched_group *group;
6921 WARN_ON(!sd || !sd->groups);
6923 if (cpu != group_first_cpu(sd->groups))
6928 sd->groups->cpu_power = 0;
6931 power = SCHED_LOAD_SCALE;
6932 weight = cpumask_weight(sched_domain_span(sd));
6934 * SMT siblings share the power of a single core.
6935 * Usually multiple threads get a better yield out of
6936 * that one core than a single thread would have,
6937 * reflect that in sd->smt_gain.
6939 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
6940 power *= sd->smt_gain;
6942 power >>= SCHED_LOAD_SHIFT;
6944 sd->groups->cpu_power += power;
6949 * Add cpu_power of each child group to this groups cpu_power.
6951 group = child->groups;
6953 sd->groups->cpu_power += group->cpu_power;
6954 group = group->next;
6955 } while (group != child->groups);
6959 * Initializers for schedule domains
6960 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6963 #ifdef CONFIG_SCHED_DEBUG
6964 # define SD_INIT_NAME(sd, type) sd->name = #type
6966 # define SD_INIT_NAME(sd, type) do { } while (0)
6969 #define SD_INIT(sd, type) sd_init_##type(sd)
6971 #define SD_INIT_FUNC(type) \
6972 static noinline void sd_init_##type(struct sched_domain *sd) \
6974 memset(sd, 0, sizeof(*sd)); \
6975 *sd = SD_##type##_INIT; \
6976 sd->level = SD_LV_##type; \
6977 SD_INIT_NAME(sd, type); \
6982 SD_INIT_FUNC(ALLNODES)
6985 #ifdef CONFIG_SCHED_SMT
6986 SD_INIT_FUNC(SIBLING)
6988 #ifdef CONFIG_SCHED_MC
6992 static int default_relax_domain_level = -1;
6994 static int __init setup_relax_domain_level(char *str)
6998 val = simple_strtoul(str, NULL, 0);
6999 if (val < SD_LV_MAX)
7000 default_relax_domain_level = val;
7004 __setup("relax_domain_level=", setup_relax_domain_level);
7006 static void set_domain_attribute(struct sched_domain *sd,
7007 struct sched_domain_attr *attr)
7011 if (!attr || attr->relax_domain_level < 0) {
7012 if (default_relax_domain_level < 0)
7015 request = default_relax_domain_level;
7017 request = attr->relax_domain_level;
7018 if (request < sd->level) {
7019 /* turn off idle balance on this domain */
7020 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7022 /* turn on idle balance on this domain */
7023 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7027 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
7028 const struct cpumask *cpu_map)
7031 case sa_sched_groups:
7032 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
7033 d->sched_group_nodes = NULL;
7035 free_rootdomain(d->rd); /* fall through */
7037 free_cpumask_var(d->tmpmask); /* fall through */
7038 case sa_send_covered:
7039 free_cpumask_var(d->send_covered); /* fall through */
7040 case sa_this_core_map:
7041 free_cpumask_var(d->this_core_map); /* fall through */
7042 case sa_this_sibling_map:
7043 free_cpumask_var(d->this_sibling_map); /* fall through */
7045 free_cpumask_var(d->nodemask); /* fall through */
7046 case sa_sched_group_nodes:
7048 kfree(d->sched_group_nodes); /* fall through */
7050 free_cpumask_var(d->notcovered); /* fall through */
7052 free_cpumask_var(d->covered); /* fall through */
7054 free_cpumask_var(d->domainspan); /* fall through */
7061 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
7062 const struct cpumask *cpu_map)
7065 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
7067 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
7068 return sa_domainspan;
7069 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
7071 /* Allocate the per-node list of sched groups */
7072 d->sched_group_nodes = kcalloc(nr_node_ids,
7073 sizeof(struct sched_group *), GFP_KERNEL);
7074 if (!d->sched_group_nodes) {
7075 printk(KERN_WARNING "Can not alloc sched group node list\n");
7076 return sa_notcovered;
7078 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
7080 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
7081 return sa_sched_group_nodes;
7082 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
7084 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
7085 return sa_this_sibling_map;
7086 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
7087 return sa_this_core_map;
7088 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
7089 return sa_send_covered;
7090 d->rd = alloc_rootdomain();
7092 printk(KERN_WARNING "Cannot alloc root domain\n");
7095 return sa_rootdomain;
7098 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
7099 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
7101 struct sched_domain *sd = NULL;
7103 struct sched_domain *parent;
7106 if (cpumask_weight(cpu_map) >
7107 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
7108 sd = &per_cpu(allnodes_domains, i).sd;
7109 SD_INIT(sd, ALLNODES);
7110 set_domain_attribute(sd, attr);
7111 cpumask_copy(sched_domain_span(sd), cpu_map);
7112 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
7117 sd = &per_cpu(node_domains, i).sd;
7119 set_domain_attribute(sd, attr);
7120 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
7121 sd->parent = parent;
7124 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
7129 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
7130 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7131 struct sched_domain *parent, int i)
7133 struct sched_domain *sd;
7134 sd = &per_cpu(phys_domains, i).sd;
7136 set_domain_attribute(sd, attr);
7137 cpumask_copy(sched_domain_span(sd), d->nodemask);
7138 sd->parent = parent;
7141 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
7145 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
7146 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7147 struct sched_domain *parent, int i)
7149 struct sched_domain *sd = parent;
7150 #ifdef CONFIG_SCHED_MC
7151 sd = &per_cpu(core_domains, i).sd;
7153 set_domain_attribute(sd, attr);
7154 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
7155 sd->parent = parent;
7157 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
7162 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
7163 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7164 struct sched_domain *parent, int i)
7166 struct sched_domain *sd = parent;
7167 #ifdef CONFIG_SCHED_SMT
7168 sd = &per_cpu(cpu_domains, i).sd;
7169 SD_INIT(sd, SIBLING);
7170 set_domain_attribute(sd, attr);
7171 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
7172 sd->parent = parent;
7174 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
7179 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
7180 const struct cpumask *cpu_map, int cpu)
7183 #ifdef CONFIG_SCHED_SMT
7184 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
7185 cpumask_and(d->this_sibling_map, cpu_map,
7186 topology_thread_cpumask(cpu));
7187 if (cpu == cpumask_first(d->this_sibling_map))
7188 init_sched_build_groups(d->this_sibling_map, cpu_map,
7190 d->send_covered, d->tmpmask);
7193 #ifdef CONFIG_SCHED_MC
7194 case SD_LV_MC: /* set up multi-core groups */
7195 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
7196 if (cpu == cpumask_first(d->this_core_map))
7197 init_sched_build_groups(d->this_core_map, cpu_map,
7199 d->send_covered, d->tmpmask);
7202 case SD_LV_CPU: /* set up physical groups */
7203 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
7204 if (!cpumask_empty(d->nodemask))
7205 init_sched_build_groups(d->nodemask, cpu_map,
7207 d->send_covered, d->tmpmask);
7210 case SD_LV_ALLNODES:
7211 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
7212 d->send_covered, d->tmpmask);
7221 * Build sched domains for a given set of cpus and attach the sched domains
7222 * to the individual cpus
7224 static int __build_sched_domains(const struct cpumask *cpu_map,
7225 struct sched_domain_attr *attr)
7227 enum s_alloc alloc_state = sa_none;
7229 struct sched_domain *sd;
7235 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7236 if (alloc_state != sa_rootdomain)
7238 alloc_state = sa_sched_groups;
7241 * Set up domains for cpus specified by the cpu_map.
7243 for_each_cpu(i, cpu_map) {
7244 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
7247 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
7248 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
7249 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
7250 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
7253 for_each_cpu(i, cpu_map) {
7254 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
7255 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
7258 /* Set up physical groups */
7259 for (i = 0; i < nr_node_ids; i++)
7260 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
7263 /* Set up node groups */
7265 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
7267 for (i = 0; i < nr_node_ids; i++)
7268 if (build_numa_sched_groups(&d, cpu_map, i))
7272 /* Calculate CPU power for physical packages and nodes */
7273 #ifdef CONFIG_SCHED_SMT
7274 for_each_cpu(i, cpu_map) {
7275 sd = &per_cpu(cpu_domains, i).sd;
7276 init_sched_groups_power(i, sd);
7279 #ifdef CONFIG_SCHED_MC
7280 for_each_cpu(i, cpu_map) {
7281 sd = &per_cpu(core_domains, i).sd;
7282 init_sched_groups_power(i, sd);
7286 for_each_cpu(i, cpu_map) {
7287 sd = &per_cpu(phys_domains, i).sd;
7288 init_sched_groups_power(i, sd);
7292 for (i = 0; i < nr_node_ids; i++)
7293 init_numa_sched_groups_power(d.sched_group_nodes[i]);
7295 if (d.sd_allnodes) {
7296 struct sched_group *sg;
7298 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7300 init_numa_sched_groups_power(sg);
7304 /* Attach the domains */
7305 for_each_cpu(i, cpu_map) {
7306 #ifdef CONFIG_SCHED_SMT
7307 sd = &per_cpu(cpu_domains, i).sd;
7308 #elif defined(CONFIG_SCHED_MC)
7309 sd = &per_cpu(core_domains, i).sd;
7311 sd = &per_cpu(phys_domains, i).sd;
7313 cpu_attach_domain(sd, d.rd, i);
7316 d.sched_group_nodes = NULL; /* don't free this we still need it */
7317 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
7321 __free_domain_allocs(&d, alloc_state, cpu_map);
7325 static int build_sched_domains(const struct cpumask *cpu_map)
7327 return __build_sched_domains(cpu_map, NULL);
7330 static cpumask_var_t *doms_cur; /* current sched domains */
7331 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7332 static struct sched_domain_attr *dattr_cur;
7333 /* attribues of custom domains in 'doms_cur' */
7336 * Special case: If a kmalloc of a doms_cur partition (array of
7337 * cpumask) fails, then fallback to a single sched domain,
7338 * as determined by the single cpumask fallback_doms.
7340 static cpumask_var_t fallback_doms;
7343 * arch_update_cpu_topology lets virtualized architectures update the
7344 * cpu core maps. It is supposed to return 1 if the topology changed
7345 * or 0 if it stayed the same.
7347 int __attribute__((weak)) arch_update_cpu_topology(void)
7352 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7355 cpumask_var_t *doms;
7357 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7360 for (i = 0; i < ndoms; i++) {
7361 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7362 free_sched_domains(doms, i);
7369 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7372 for (i = 0; i < ndoms; i++)
7373 free_cpumask_var(doms[i]);
7378 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7379 * For now this just excludes isolated cpus, but could be used to
7380 * exclude other special cases in the future.
7382 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7386 arch_update_cpu_topology();
7388 doms_cur = alloc_sched_domains(ndoms_cur);
7390 doms_cur = &fallback_doms;
7391 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7393 err = build_sched_domains(doms_cur[0]);
7394 register_sched_domain_sysctl();
7399 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7400 struct cpumask *tmpmask)
7402 free_sched_groups(cpu_map, tmpmask);
7406 * Detach sched domains from a group of cpus specified in cpu_map
7407 * These cpus will now be attached to the NULL domain
7409 static void detach_destroy_domains(const struct cpumask *cpu_map)
7411 /* Save because hotplug lock held. */
7412 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7415 for_each_cpu(i, cpu_map)
7416 cpu_attach_domain(NULL, &def_root_domain, i);
7417 synchronize_sched();
7418 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7421 /* handle null as "default" */
7422 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7423 struct sched_domain_attr *new, int idx_new)
7425 struct sched_domain_attr tmp;
7432 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7433 new ? (new + idx_new) : &tmp,
7434 sizeof(struct sched_domain_attr));
7438 * Partition sched domains as specified by the 'ndoms_new'
7439 * cpumasks in the array doms_new[] of cpumasks. This compares
7440 * doms_new[] to the current sched domain partitioning, doms_cur[].
7441 * It destroys each deleted domain and builds each new domain.
7443 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7444 * The masks don't intersect (don't overlap.) We should setup one
7445 * sched domain for each mask. CPUs not in any of the cpumasks will
7446 * not be load balanced. If the same cpumask appears both in the
7447 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7450 * The passed in 'doms_new' should be allocated using
7451 * alloc_sched_domains. This routine takes ownership of it and will
7452 * free_sched_domains it when done with it. If the caller failed the
7453 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7454 * and partition_sched_domains() will fallback to the single partition
7455 * 'fallback_doms', it also forces the domains to be rebuilt.
7457 * If doms_new == NULL it will be replaced with cpu_online_mask.
7458 * ndoms_new == 0 is a special case for destroying existing domains,
7459 * and it will not create the default domain.
7461 * Call with hotplug lock held
7463 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7464 struct sched_domain_attr *dattr_new)
7469 mutex_lock(&sched_domains_mutex);
7471 /* always unregister in case we don't destroy any domains */
7472 unregister_sched_domain_sysctl();
7474 /* Let architecture update cpu core mappings. */
7475 new_topology = arch_update_cpu_topology();
7477 n = doms_new ? ndoms_new : 0;
7479 /* Destroy deleted domains */
7480 for (i = 0; i < ndoms_cur; i++) {
7481 for (j = 0; j < n && !new_topology; j++) {
7482 if (cpumask_equal(doms_cur[i], doms_new[j])
7483 && dattrs_equal(dattr_cur, i, dattr_new, j))
7486 /* no match - a current sched domain not in new doms_new[] */
7487 detach_destroy_domains(doms_cur[i]);
7492 if (doms_new == NULL) {
7494 doms_new = &fallback_doms;
7495 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7496 WARN_ON_ONCE(dattr_new);
7499 /* Build new domains */
7500 for (i = 0; i < ndoms_new; i++) {
7501 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7502 if (cpumask_equal(doms_new[i], doms_cur[j])
7503 && dattrs_equal(dattr_new, i, dattr_cur, j))
7506 /* no match - add a new doms_new */
7507 __build_sched_domains(doms_new[i],
7508 dattr_new ? dattr_new + i : NULL);
7513 /* Remember the new sched domains */
7514 if (doms_cur != &fallback_doms)
7515 free_sched_domains(doms_cur, ndoms_cur);
7516 kfree(dattr_cur); /* kfree(NULL) is safe */
7517 doms_cur = doms_new;
7518 dattr_cur = dattr_new;
7519 ndoms_cur = ndoms_new;
7521 register_sched_domain_sysctl();
7523 mutex_unlock(&sched_domains_mutex);
7526 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7527 static void arch_reinit_sched_domains(void)
7531 /* Destroy domains first to force the rebuild */
7532 partition_sched_domains(0, NULL, NULL);
7534 rebuild_sched_domains();
7538 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7540 unsigned int level = 0;
7542 if (sscanf(buf, "%u", &level) != 1)
7546 * level is always be positive so don't check for
7547 * level < POWERSAVINGS_BALANCE_NONE which is 0
7548 * What happens on 0 or 1 byte write,
7549 * need to check for count as well?
7552 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7556 sched_smt_power_savings = level;
7558 sched_mc_power_savings = level;
7560 arch_reinit_sched_domains();
7565 #ifdef CONFIG_SCHED_MC
7566 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7567 struct sysdev_class_attribute *attr,
7570 return sprintf(page, "%u\n", sched_mc_power_savings);
7572 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7573 struct sysdev_class_attribute *attr,
7574 const char *buf, size_t count)
7576 return sched_power_savings_store(buf, count, 0);
7578 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7579 sched_mc_power_savings_show,
7580 sched_mc_power_savings_store);
7583 #ifdef CONFIG_SCHED_SMT
7584 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7585 struct sysdev_class_attribute *attr,
7588 return sprintf(page, "%u\n", sched_smt_power_savings);
7590 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7591 struct sysdev_class_attribute *attr,
7592 const char *buf, size_t count)
7594 return sched_power_savings_store(buf, count, 1);
7596 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7597 sched_smt_power_savings_show,
7598 sched_smt_power_savings_store);
7601 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7605 #ifdef CONFIG_SCHED_SMT
7607 err = sysfs_create_file(&cls->kset.kobj,
7608 &attr_sched_smt_power_savings.attr);
7610 #ifdef CONFIG_SCHED_MC
7611 if (!err && mc_capable())
7612 err = sysfs_create_file(&cls->kset.kobj,
7613 &attr_sched_mc_power_savings.attr);
7617 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7620 * Update cpusets according to cpu_active mask. If cpusets are
7621 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7622 * around partition_sched_domains().
7624 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7627 switch (action & ~CPU_TASKS_FROZEN) {
7629 case CPU_DOWN_FAILED:
7630 cpuset_update_active_cpus();
7637 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7640 switch (action & ~CPU_TASKS_FROZEN) {
7641 case CPU_DOWN_PREPARE:
7642 cpuset_update_active_cpus();
7649 static int update_runtime(struct notifier_block *nfb,
7650 unsigned long action, void *hcpu)
7652 int cpu = (int)(long)hcpu;
7655 case CPU_DOWN_PREPARE:
7656 case CPU_DOWN_PREPARE_FROZEN:
7657 disable_runtime(cpu_rq(cpu));
7660 case CPU_DOWN_FAILED:
7661 case CPU_DOWN_FAILED_FROZEN:
7663 case CPU_ONLINE_FROZEN:
7664 enable_runtime(cpu_rq(cpu));
7672 void __init sched_init_smp(void)
7674 cpumask_var_t non_isolated_cpus;
7676 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7677 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7679 #if defined(CONFIG_NUMA)
7680 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7682 BUG_ON(sched_group_nodes_bycpu == NULL);
7685 mutex_lock(&sched_domains_mutex);
7686 arch_init_sched_domains(cpu_active_mask);
7687 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7688 if (cpumask_empty(non_isolated_cpus))
7689 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7690 mutex_unlock(&sched_domains_mutex);
7693 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7694 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7696 /* RT runtime code needs to handle some hotplug events */
7697 hotcpu_notifier(update_runtime, 0);
7701 /* Move init over to a non-isolated CPU */
7702 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7704 sched_init_granularity();
7705 free_cpumask_var(non_isolated_cpus);
7707 init_sched_rt_class();
7710 void __init sched_init_smp(void)
7712 sched_init_granularity();
7714 #endif /* CONFIG_SMP */
7716 const_debug unsigned int sysctl_timer_migration = 1;
7718 int in_sched_functions(unsigned long addr)
7720 return in_lock_functions(addr) ||
7721 (addr >= (unsigned long)__sched_text_start
7722 && addr < (unsigned long)__sched_text_end);
7725 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7727 cfs_rq->tasks_timeline = RB_ROOT;
7728 INIT_LIST_HEAD(&cfs_rq->tasks);
7729 #ifdef CONFIG_FAIR_GROUP_SCHED
7732 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7735 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7737 struct rt_prio_array *array;
7740 array = &rt_rq->active;
7741 for (i = 0; i < MAX_RT_PRIO; i++) {
7742 INIT_LIST_HEAD(array->queue + i);
7743 __clear_bit(i, array->bitmap);
7745 /* delimiter for bitsearch: */
7746 __set_bit(MAX_RT_PRIO, array->bitmap);
7748 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7749 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7751 rt_rq->highest_prio.next = MAX_RT_PRIO;
7755 rt_rq->rt_nr_migratory = 0;
7756 rt_rq->overloaded = 0;
7757 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
7761 rt_rq->rt_throttled = 0;
7762 rt_rq->rt_runtime = 0;
7763 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7765 #ifdef CONFIG_RT_GROUP_SCHED
7766 rt_rq->rt_nr_boosted = 0;
7771 #ifdef CONFIG_FAIR_GROUP_SCHED
7772 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7773 struct sched_entity *se, int cpu, int add,
7774 struct sched_entity *parent)
7776 struct rq *rq = cpu_rq(cpu);
7777 tg->cfs_rq[cpu] = cfs_rq;
7778 init_cfs_rq(cfs_rq, rq);
7781 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7784 /* se could be NULL for init_task_group */
7789 se->cfs_rq = &rq->cfs;
7791 se->cfs_rq = parent->my_q;
7794 se->load.weight = tg->shares;
7795 se->load.inv_weight = 0;
7796 se->parent = parent;
7800 #ifdef CONFIG_RT_GROUP_SCHED
7801 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7802 struct sched_rt_entity *rt_se, int cpu, int add,
7803 struct sched_rt_entity *parent)
7805 struct rq *rq = cpu_rq(cpu);
7807 tg->rt_rq[cpu] = rt_rq;
7808 init_rt_rq(rt_rq, rq);
7810 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7812 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7814 tg->rt_se[cpu] = rt_se;
7819 rt_se->rt_rq = &rq->rt;
7821 rt_se->rt_rq = parent->my_q;
7823 rt_se->my_q = rt_rq;
7824 rt_se->parent = parent;
7825 INIT_LIST_HEAD(&rt_se->run_list);
7829 void __init sched_init(void)
7832 unsigned long alloc_size = 0, ptr;
7834 #ifdef CONFIG_FAIR_GROUP_SCHED
7835 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7837 #ifdef CONFIG_RT_GROUP_SCHED
7838 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7840 #ifdef CONFIG_CPUMASK_OFFSTACK
7841 alloc_size += num_possible_cpus() * cpumask_size();
7844 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7846 #ifdef CONFIG_FAIR_GROUP_SCHED
7847 init_task_group.se = (struct sched_entity **)ptr;
7848 ptr += nr_cpu_ids * sizeof(void **);
7850 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7851 ptr += nr_cpu_ids * sizeof(void **);
7853 #endif /* CONFIG_FAIR_GROUP_SCHED */
7854 #ifdef CONFIG_RT_GROUP_SCHED
7855 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7856 ptr += nr_cpu_ids * sizeof(void **);
7858 init_task_group.rt_rq = (struct rt_rq **)ptr;
7859 ptr += nr_cpu_ids * sizeof(void **);
7861 #endif /* CONFIG_RT_GROUP_SCHED */
7862 #ifdef CONFIG_CPUMASK_OFFSTACK
7863 for_each_possible_cpu(i) {
7864 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
7865 ptr += cpumask_size();
7867 #endif /* CONFIG_CPUMASK_OFFSTACK */
7871 init_defrootdomain();
7874 init_rt_bandwidth(&def_rt_bandwidth,
7875 global_rt_period(), global_rt_runtime());
7877 #ifdef CONFIG_RT_GROUP_SCHED
7878 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7879 global_rt_period(), global_rt_runtime());
7880 #endif /* CONFIG_RT_GROUP_SCHED */
7882 #ifdef CONFIG_CGROUP_SCHED
7883 list_add(&init_task_group.list, &task_groups);
7884 INIT_LIST_HEAD(&init_task_group.children);
7886 #endif /* CONFIG_CGROUP_SCHED */
7888 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
7889 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
7890 __alignof__(unsigned long));
7892 for_each_possible_cpu(i) {
7896 raw_spin_lock_init(&rq->lock);
7898 rq->calc_load_active = 0;
7899 rq->calc_load_update = jiffies + LOAD_FREQ;
7900 init_cfs_rq(&rq->cfs, rq);
7901 init_rt_rq(&rq->rt, rq);
7902 #ifdef CONFIG_FAIR_GROUP_SCHED
7903 init_task_group.shares = init_task_group_load;
7904 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7905 #ifdef CONFIG_CGROUP_SCHED
7907 * How much cpu bandwidth does init_task_group get?
7909 * In case of task-groups formed thr' the cgroup filesystem, it
7910 * gets 100% of the cpu resources in the system. This overall
7911 * system cpu resource is divided among the tasks of
7912 * init_task_group and its child task-groups in a fair manner,
7913 * based on each entity's (task or task-group's) weight
7914 * (se->load.weight).
7916 * In other words, if init_task_group has 10 tasks of weight
7917 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7918 * then A0's share of the cpu resource is:
7920 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7922 * We achieve this by letting init_task_group's tasks sit
7923 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7925 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
7927 #endif /* CONFIG_FAIR_GROUP_SCHED */
7929 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7930 #ifdef CONFIG_RT_GROUP_SCHED
7931 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7932 #ifdef CONFIG_CGROUP_SCHED
7933 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
7937 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7938 rq->cpu_load[j] = 0;
7940 rq->last_load_update_tick = jiffies;
7945 rq->cpu_power = SCHED_LOAD_SCALE;
7946 rq->post_schedule = 0;
7947 rq->active_balance = 0;
7948 rq->next_balance = jiffies;
7953 rq->avg_idle = 2*sysctl_sched_migration_cost;
7954 rq_attach_root(rq, &def_root_domain);
7956 rq->nohz_balance_kick = 0;
7957 init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb, i));
7961 atomic_set(&rq->nr_iowait, 0);
7964 set_load_weight(&init_task);
7966 #ifdef CONFIG_PREEMPT_NOTIFIERS
7967 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7971 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7974 #ifdef CONFIG_RT_MUTEXES
7975 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
7979 * The boot idle thread does lazy MMU switching as well:
7981 atomic_inc(&init_mm.mm_count);
7982 enter_lazy_tlb(&init_mm, current);
7985 * Make us the idle thread. Technically, schedule() should not be
7986 * called from this thread, however somewhere below it might be,
7987 * but because we are the idle thread, we just pick up running again
7988 * when this runqueue becomes "idle".
7990 init_idle(current, smp_processor_id());
7992 calc_load_update = jiffies + LOAD_FREQ;
7995 * During early bootup we pretend to be a normal task:
7997 current->sched_class = &fair_sched_class;
7999 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8000 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
8003 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8004 alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT);
8005 atomic_set(&nohz.load_balancer, nr_cpu_ids);
8006 atomic_set(&nohz.first_pick_cpu, nr_cpu_ids);
8007 atomic_set(&nohz.second_pick_cpu, nr_cpu_ids);
8009 /* May be allocated at isolcpus cmdline parse time */
8010 if (cpu_isolated_map == NULL)
8011 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
8016 scheduler_running = 1;
8019 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8020 static inline int preempt_count_equals(int preempt_offset)
8022 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
8024 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
8027 void __might_sleep(const char *file, int line, int preempt_offset)
8030 static unsigned long prev_jiffy; /* ratelimiting */
8032 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
8033 system_state != SYSTEM_RUNNING || oops_in_progress)
8035 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8037 prev_jiffy = jiffies;
8040 "BUG: sleeping function called from invalid context at %s:%d\n",
8043 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8044 in_atomic(), irqs_disabled(),
8045 current->pid, current->comm);
8047 debug_show_held_locks(current);
8048 if (irqs_disabled())
8049 print_irqtrace_events(current);
8053 EXPORT_SYMBOL(__might_sleep);
8056 #ifdef CONFIG_MAGIC_SYSRQ
8057 static void normalize_task(struct rq *rq, struct task_struct *p)
8061 on_rq = p->se.on_rq;
8063 deactivate_task(rq, p, 0);
8064 __setscheduler(rq, p, SCHED_NORMAL, 0);
8066 activate_task(rq, p, 0);
8067 resched_task(rq->curr);
8071 void normalize_rt_tasks(void)
8073 struct task_struct *g, *p;
8074 unsigned long flags;
8077 read_lock_irqsave(&tasklist_lock, flags);
8078 do_each_thread(g, p) {
8080 * Only normalize user tasks:
8085 p->se.exec_start = 0;
8086 #ifdef CONFIG_SCHEDSTATS
8087 p->se.statistics.wait_start = 0;
8088 p->se.statistics.sleep_start = 0;
8089 p->se.statistics.block_start = 0;
8094 * Renice negative nice level userspace
8097 if (TASK_NICE(p) < 0 && p->mm)
8098 set_user_nice(p, 0);
8102 raw_spin_lock(&p->pi_lock);
8103 rq = __task_rq_lock(p);
8105 normalize_task(rq, p);
8107 __task_rq_unlock(rq);
8108 raw_spin_unlock(&p->pi_lock);
8109 } while_each_thread(g, p);
8111 read_unlock_irqrestore(&tasklist_lock, flags);
8114 #endif /* CONFIG_MAGIC_SYSRQ */
8116 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8118 * These functions are only useful for the IA64 MCA handling, or kdb.
8120 * They can only be called when the whole system has been
8121 * stopped - every CPU needs to be quiescent, and no scheduling
8122 * activity can take place. Using them for anything else would
8123 * be a serious bug, and as a result, they aren't even visible
8124 * under any other configuration.
8128 * curr_task - return the current task for a given cpu.
8129 * @cpu: the processor in question.
8131 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8133 struct task_struct *curr_task(int cpu)
8135 return cpu_curr(cpu);
8138 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8142 * set_curr_task - set the current task for a given cpu.
8143 * @cpu: the processor in question.
8144 * @p: the task pointer to set.
8146 * Description: This function must only be used when non-maskable interrupts
8147 * are serviced on a separate stack. It allows the architecture to switch the
8148 * notion of the current task on a cpu in a non-blocking manner. This function
8149 * must be called with all CPU's synchronized, and interrupts disabled, the
8150 * and caller must save the original value of the current task (see
8151 * curr_task() above) and restore that value before reenabling interrupts and
8152 * re-starting the system.
8154 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8156 void set_curr_task(int cpu, struct task_struct *p)
8163 #ifdef CONFIG_FAIR_GROUP_SCHED
8164 static void free_fair_sched_group(struct task_group *tg)
8168 for_each_possible_cpu(i) {
8170 kfree(tg->cfs_rq[i]);
8180 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8182 struct cfs_rq *cfs_rq;
8183 struct sched_entity *se;
8187 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8190 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8194 tg->shares = NICE_0_LOAD;
8196 for_each_possible_cpu(i) {
8199 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8200 GFP_KERNEL, cpu_to_node(i));
8204 se = kzalloc_node(sizeof(struct sched_entity),
8205 GFP_KERNEL, cpu_to_node(i));
8209 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
8220 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8222 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8223 &cpu_rq(cpu)->leaf_cfs_rq_list);
8226 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8228 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8230 #else /* !CONFG_FAIR_GROUP_SCHED */
8231 static inline void free_fair_sched_group(struct task_group *tg)
8236 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8241 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8245 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8248 #endif /* CONFIG_FAIR_GROUP_SCHED */
8250 #ifdef CONFIG_RT_GROUP_SCHED
8251 static void free_rt_sched_group(struct task_group *tg)
8255 destroy_rt_bandwidth(&tg->rt_bandwidth);
8257 for_each_possible_cpu(i) {
8259 kfree(tg->rt_rq[i]);
8261 kfree(tg->rt_se[i]);
8269 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8271 struct rt_rq *rt_rq;
8272 struct sched_rt_entity *rt_se;
8276 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8279 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8283 init_rt_bandwidth(&tg->rt_bandwidth,
8284 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8286 for_each_possible_cpu(i) {
8289 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8290 GFP_KERNEL, cpu_to_node(i));
8294 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8295 GFP_KERNEL, cpu_to_node(i));
8299 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
8310 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8312 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8313 &cpu_rq(cpu)->leaf_rt_rq_list);
8316 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8318 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8320 #else /* !CONFIG_RT_GROUP_SCHED */
8321 static inline void free_rt_sched_group(struct task_group *tg)
8326 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8331 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8335 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8338 #endif /* CONFIG_RT_GROUP_SCHED */
8340 #ifdef CONFIG_CGROUP_SCHED
8341 static void free_sched_group(struct task_group *tg)
8343 free_fair_sched_group(tg);
8344 free_rt_sched_group(tg);
8348 /* allocate runqueue etc for a new task group */
8349 struct task_group *sched_create_group(struct task_group *parent)
8351 struct task_group *tg;
8352 unsigned long flags;
8355 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8357 return ERR_PTR(-ENOMEM);
8359 if (!alloc_fair_sched_group(tg, parent))
8362 if (!alloc_rt_sched_group(tg, parent))
8365 spin_lock_irqsave(&task_group_lock, flags);
8366 for_each_possible_cpu(i) {
8367 register_fair_sched_group(tg, i);
8368 register_rt_sched_group(tg, i);
8370 list_add_rcu(&tg->list, &task_groups);
8372 WARN_ON(!parent); /* root should already exist */
8374 tg->parent = parent;
8375 INIT_LIST_HEAD(&tg->children);
8376 list_add_rcu(&tg->siblings, &parent->children);
8377 spin_unlock_irqrestore(&task_group_lock, flags);
8382 free_sched_group(tg);
8383 return ERR_PTR(-ENOMEM);
8386 /* rcu callback to free various structures associated with a task group */
8387 static void free_sched_group_rcu(struct rcu_head *rhp)
8389 /* now it should be safe to free those cfs_rqs */
8390 free_sched_group(container_of(rhp, struct task_group, rcu));
8393 /* Destroy runqueue etc associated with a task group */
8394 void sched_destroy_group(struct task_group *tg)
8396 unsigned long flags;
8399 spin_lock_irqsave(&task_group_lock, flags);
8400 for_each_possible_cpu(i) {
8401 unregister_fair_sched_group(tg, i);
8402 unregister_rt_sched_group(tg, i);
8404 list_del_rcu(&tg->list);
8405 list_del_rcu(&tg->siblings);
8406 spin_unlock_irqrestore(&task_group_lock, flags);
8408 /* wait for possible concurrent references to cfs_rqs complete */
8409 call_rcu(&tg->rcu, free_sched_group_rcu);
8412 /* change task's runqueue when it moves between groups.
8413 * The caller of this function should have put the task in its new group
8414 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8415 * reflect its new group.
8417 void sched_move_task(struct task_struct *tsk)
8420 unsigned long flags;
8423 rq = task_rq_lock(tsk, &flags);
8425 running = task_current(rq, tsk);
8426 on_rq = tsk->se.on_rq;
8429 dequeue_task(rq, tsk, 0);
8430 if (unlikely(running))
8431 tsk->sched_class->put_prev_task(rq, tsk);
8433 set_task_rq(tsk, task_cpu(tsk));
8435 #ifdef CONFIG_FAIR_GROUP_SCHED
8436 if (tsk->sched_class->moved_group)
8437 tsk->sched_class->moved_group(tsk, on_rq);
8440 if (unlikely(running))
8441 tsk->sched_class->set_curr_task(rq);
8443 enqueue_task(rq, tsk, 0);
8445 task_rq_unlock(rq, &flags);
8447 #endif /* CONFIG_CGROUP_SCHED */
8449 #ifdef CONFIG_FAIR_GROUP_SCHED
8450 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8452 struct cfs_rq *cfs_rq = se->cfs_rq;
8457 dequeue_entity(cfs_rq, se, 0);
8459 se->load.weight = shares;
8460 se->load.inv_weight = 0;
8463 enqueue_entity(cfs_rq, se, 0);
8466 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8468 struct cfs_rq *cfs_rq = se->cfs_rq;
8469 struct rq *rq = cfs_rq->rq;
8470 unsigned long flags;
8472 raw_spin_lock_irqsave(&rq->lock, flags);
8473 __set_se_shares(se, shares);
8474 raw_spin_unlock_irqrestore(&rq->lock, flags);
8477 static DEFINE_MUTEX(shares_mutex);
8479 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8482 unsigned long flags;
8485 * We can't change the weight of the root cgroup.
8490 if (shares < MIN_SHARES)
8491 shares = MIN_SHARES;
8492 else if (shares > MAX_SHARES)
8493 shares = MAX_SHARES;
8495 mutex_lock(&shares_mutex);
8496 if (tg->shares == shares)
8499 spin_lock_irqsave(&task_group_lock, flags);
8500 for_each_possible_cpu(i)
8501 unregister_fair_sched_group(tg, i);
8502 list_del_rcu(&tg->siblings);
8503 spin_unlock_irqrestore(&task_group_lock, flags);
8505 /* wait for any ongoing reference to this group to finish */
8506 synchronize_sched();
8509 * Now we are free to modify the group's share on each cpu
8510 * w/o tripping rebalance_share or load_balance_fair.
8512 tg->shares = shares;
8513 for_each_possible_cpu(i) {
8517 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8518 set_se_shares(tg->se[i], shares);
8522 * Enable load balance activity on this group, by inserting it back on
8523 * each cpu's rq->leaf_cfs_rq_list.
8525 spin_lock_irqsave(&task_group_lock, flags);
8526 for_each_possible_cpu(i)
8527 register_fair_sched_group(tg, i);
8528 list_add_rcu(&tg->siblings, &tg->parent->children);
8529 spin_unlock_irqrestore(&task_group_lock, flags);
8531 mutex_unlock(&shares_mutex);
8535 unsigned long sched_group_shares(struct task_group *tg)
8541 #ifdef CONFIG_RT_GROUP_SCHED
8543 * Ensure that the real time constraints are schedulable.
8545 static DEFINE_MUTEX(rt_constraints_mutex);
8547 static unsigned long to_ratio(u64 period, u64 runtime)
8549 if (runtime == RUNTIME_INF)
8552 return div64_u64(runtime << 20, period);
8555 /* Must be called with tasklist_lock held */
8556 static inline int tg_has_rt_tasks(struct task_group *tg)
8558 struct task_struct *g, *p;
8560 do_each_thread(g, p) {
8561 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8563 } while_each_thread(g, p);
8568 struct rt_schedulable_data {
8569 struct task_group *tg;
8574 static int tg_schedulable(struct task_group *tg, void *data)
8576 struct rt_schedulable_data *d = data;
8577 struct task_group *child;
8578 unsigned long total, sum = 0;
8579 u64 period, runtime;
8581 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8582 runtime = tg->rt_bandwidth.rt_runtime;
8585 period = d->rt_period;
8586 runtime = d->rt_runtime;
8590 * Cannot have more runtime than the period.
8592 if (runtime > period && runtime != RUNTIME_INF)
8596 * Ensure we don't starve existing RT tasks.
8598 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8601 total = to_ratio(period, runtime);
8604 * Nobody can have more than the global setting allows.
8606 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8610 * The sum of our children's runtime should not exceed our own.
8612 list_for_each_entry_rcu(child, &tg->children, siblings) {
8613 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8614 runtime = child->rt_bandwidth.rt_runtime;
8616 if (child == d->tg) {
8617 period = d->rt_period;
8618 runtime = d->rt_runtime;
8621 sum += to_ratio(period, runtime);
8630 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8632 struct rt_schedulable_data data = {
8634 .rt_period = period,
8635 .rt_runtime = runtime,
8638 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8641 static int tg_set_bandwidth(struct task_group *tg,
8642 u64 rt_period, u64 rt_runtime)
8646 mutex_lock(&rt_constraints_mutex);
8647 read_lock(&tasklist_lock);
8648 err = __rt_schedulable(tg, rt_period, rt_runtime);
8652 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8653 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8654 tg->rt_bandwidth.rt_runtime = rt_runtime;
8656 for_each_possible_cpu(i) {
8657 struct rt_rq *rt_rq = tg->rt_rq[i];
8659 raw_spin_lock(&rt_rq->rt_runtime_lock);
8660 rt_rq->rt_runtime = rt_runtime;
8661 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8663 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8665 read_unlock(&tasklist_lock);
8666 mutex_unlock(&rt_constraints_mutex);
8671 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8673 u64 rt_runtime, rt_period;
8675 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8676 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8677 if (rt_runtime_us < 0)
8678 rt_runtime = RUNTIME_INF;
8680 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8683 long sched_group_rt_runtime(struct task_group *tg)
8687 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8690 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8691 do_div(rt_runtime_us, NSEC_PER_USEC);
8692 return rt_runtime_us;
8695 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8697 u64 rt_runtime, rt_period;
8699 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8700 rt_runtime = tg->rt_bandwidth.rt_runtime;
8705 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8708 long sched_group_rt_period(struct task_group *tg)
8712 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8713 do_div(rt_period_us, NSEC_PER_USEC);
8714 return rt_period_us;
8717 static int sched_rt_global_constraints(void)
8719 u64 runtime, period;
8722 if (sysctl_sched_rt_period <= 0)
8725 runtime = global_rt_runtime();
8726 period = global_rt_period();
8729 * Sanity check on the sysctl variables.
8731 if (runtime > period && runtime != RUNTIME_INF)
8734 mutex_lock(&rt_constraints_mutex);
8735 read_lock(&tasklist_lock);
8736 ret = __rt_schedulable(NULL, 0, 0);
8737 read_unlock(&tasklist_lock);
8738 mutex_unlock(&rt_constraints_mutex);
8743 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8745 /* Don't accept realtime tasks when there is no way for them to run */
8746 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8752 #else /* !CONFIG_RT_GROUP_SCHED */
8753 static int sched_rt_global_constraints(void)
8755 unsigned long flags;
8758 if (sysctl_sched_rt_period <= 0)
8762 * There's always some RT tasks in the root group
8763 * -- migration, kstopmachine etc..
8765 if (sysctl_sched_rt_runtime == 0)
8768 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8769 for_each_possible_cpu(i) {
8770 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8772 raw_spin_lock(&rt_rq->rt_runtime_lock);
8773 rt_rq->rt_runtime = global_rt_runtime();
8774 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8776 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8780 #endif /* CONFIG_RT_GROUP_SCHED */
8782 int sched_rt_handler(struct ctl_table *table, int write,
8783 void __user *buffer, size_t *lenp,
8787 int old_period, old_runtime;
8788 static DEFINE_MUTEX(mutex);
8791 old_period = sysctl_sched_rt_period;
8792 old_runtime = sysctl_sched_rt_runtime;
8794 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8796 if (!ret && write) {
8797 ret = sched_rt_global_constraints();
8799 sysctl_sched_rt_period = old_period;
8800 sysctl_sched_rt_runtime = old_runtime;
8802 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8803 def_rt_bandwidth.rt_period =
8804 ns_to_ktime(global_rt_period());
8807 mutex_unlock(&mutex);
8812 #ifdef CONFIG_CGROUP_SCHED
8814 /* return corresponding task_group object of a cgroup */
8815 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8817 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8818 struct task_group, css);
8821 static struct cgroup_subsys_state *
8822 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8824 struct task_group *tg, *parent;
8826 if (!cgrp->parent) {
8827 /* This is early initialization for the top cgroup */
8828 return &init_task_group.css;
8831 parent = cgroup_tg(cgrp->parent);
8832 tg = sched_create_group(parent);
8834 return ERR_PTR(-ENOMEM);
8840 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8842 struct task_group *tg = cgroup_tg(cgrp);
8844 sched_destroy_group(tg);
8848 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8850 #ifdef CONFIG_RT_GROUP_SCHED
8851 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
8854 /* We don't support RT-tasks being in separate groups */
8855 if (tsk->sched_class != &fair_sched_class)
8862 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8863 struct task_struct *tsk, bool threadgroup)
8865 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
8869 struct task_struct *c;
8871 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8872 retval = cpu_cgroup_can_attach_task(cgrp, c);
8884 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8885 struct cgroup *old_cont, struct task_struct *tsk,
8888 sched_move_task(tsk);
8890 struct task_struct *c;
8892 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8899 #ifdef CONFIG_FAIR_GROUP_SCHED
8900 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8903 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8906 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8908 struct task_group *tg = cgroup_tg(cgrp);
8910 return (u64) tg->shares;
8912 #endif /* CONFIG_FAIR_GROUP_SCHED */
8914 #ifdef CONFIG_RT_GROUP_SCHED
8915 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8918 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8921 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8923 return sched_group_rt_runtime(cgroup_tg(cgrp));
8926 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8929 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8932 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8934 return sched_group_rt_period(cgroup_tg(cgrp));
8936 #endif /* CONFIG_RT_GROUP_SCHED */
8938 static struct cftype cpu_files[] = {
8939 #ifdef CONFIG_FAIR_GROUP_SCHED
8942 .read_u64 = cpu_shares_read_u64,
8943 .write_u64 = cpu_shares_write_u64,
8946 #ifdef CONFIG_RT_GROUP_SCHED
8948 .name = "rt_runtime_us",
8949 .read_s64 = cpu_rt_runtime_read,
8950 .write_s64 = cpu_rt_runtime_write,
8953 .name = "rt_period_us",
8954 .read_u64 = cpu_rt_period_read_uint,
8955 .write_u64 = cpu_rt_period_write_uint,
8960 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8962 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8965 struct cgroup_subsys cpu_cgroup_subsys = {
8967 .create = cpu_cgroup_create,
8968 .destroy = cpu_cgroup_destroy,
8969 .can_attach = cpu_cgroup_can_attach,
8970 .attach = cpu_cgroup_attach,
8971 .populate = cpu_cgroup_populate,
8972 .subsys_id = cpu_cgroup_subsys_id,
8976 #endif /* CONFIG_CGROUP_SCHED */
8978 #ifdef CONFIG_CGROUP_CPUACCT
8981 * CPU accounting code for task groups.
8983 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8984 * (balbir@in.ibm.com).
8987 /* track cpu usage of a group of tasks and its child groups */
8989 struct cgroup_subsys_state css;
8990 /* cpuusage holds pointer to a u64-type object on every cpu */
8991 u64 __percpu *cpuusage;
8992 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
8993 struct cpuacct *parent;
8996 struct cgroup_subsys cpuacct_subsys;
8998 /* return cpu accounting group corresponding to this container */
8999 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9001 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9002 struct cpuacct, css);
9005 /* return cpu accounting group to which this task belongs */
9006 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9008 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9009 struct cpuacct, css);
9012 /* create a new cpu accounting group */
9013 static struct cgroup_subsys_state *cpuacct_create(
9014 struct cgroup_subsys *ss, struct cgroup *cgrp)
9016 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9022 ca->cpuusage = alloc_percpu(u64);
9026 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9027 if (percpu_counter_init(&ca->cpustat[i], 0))
9028 goto out_free_counters;
9031 ca->parent = cgroup_ca(cgrp->parent);
9037 percpu_counter_destroy(&ca->cpustat[i]);
9038 free_percpu(ca->cpuusage);
9042 return ERR_PTR(-ENOMEM);
9045 /* destroy an existing cpu accounting group */
9047 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9049 struct cpuacct *ca = cgroup_ca(cgrp);
9052 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9053 percpu_counter_destroy(&ca->cpustat[i]);
9054 free_percpu(ca->cpuusage);
9058 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
9060 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9063 #ifndef CONFIG_64BIT
9065 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9067 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9069 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9077 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9079 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9081 #ifndef CONFIG_64BIT
9083 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9085 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9087 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9093 /* return total cpu usage (in nanoseconds) of a group */
9094 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9096 struct cpuacct *ca = cgroup_ca(cgrp);
9097 u64 totalcpuusage = 0;
9100 for_each_present_cpu(i)
9101 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9103 return totalcpuusage;
9106 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9109 struct cpuacct *ca = cgroup_ca(cgrp);
9118 for_each_present_cpu(i)
9119 cpuacct_cpuusage_write(ca, i, 0);
9125 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9128 struct cpuacct *ca = cgroup_ca(cgroup);
9132 for_each_present_cpu(i) {
9133 percpu = cpuacct_cpuusage_read(ca, i);
9134 seq_printf(m, "%llu ", (unsigned long long) percpu);
9136 seq_printf(m, "\n");
9140 static const char *cpuacct_stat_desc[] = {
9141 [CPUACCT_STAT_USER] = "user",
9142 [CPUACCT_STAT_SYSTEM] = "system",
9145 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
9146 struct cgroup_map_cb *cb)
9148 struct cpuacct *ca = cgroup_ca(cgrp);
9151 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
9152 s64 val = percpu_counter_read(&ca->cpustat[i]);
9153 val = cputime64_to_clock_t(val);
9154 cb->fill(cb, cpuacct_stat_desc[i], val);
9159 static struct cftype files[] = {
9162 .read_u64 = cpuusage_read,
9163 .write_u64 = cpuusage_write,
9166 .name = "usage_percpu",
9167 .read_seq_string = cpuacct_percpu_seq_read,
9171 .read_map = cpuacct_stats_show,
9175 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9177 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9181 * charge this task's execution time to its accounting group.
9183 * called with rq->lock held.
9185 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9190 if (unlikely(!cpuacct_subsys.active))
9193 cpu = task_cpu(tsk);
9199 for (; ca; ca = ca->parent) {
9200 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9201 *cpuusage += cputime;
9208 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9209 * in cputime_t units. As a result, cpuacct_update_stats calls
9210 * percpu_counter_add with values large enough to always overflow the
9211 * per cpu batch limit causing bad SMP scalability.
9213 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9214 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9215 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9218 #define CPUACCT_BATCH \
9219 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9221 #define CPUACCT_BATCH 0
9225 * Charge the system/user time to the task's accounting group.
9227 static void cpuacct_update_stats(struct task_struct *tsk,
9228 enum cpuacct_stat_index idx, cputime_t val)
9231 int batch = CPUACCT_BATCH;
9233 if (unlikely(!cpuacct_subsys.active))
9240 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9246 struct cgroup_subsys cpuacct_subsys = {
9248 .create = cpuacct_create,
9249 .destroy = cpuacct_destroy,
9250 .populate = cpuacct_populate,
9251 .subsys_id = cpuacct_subsys_id,
9253 #endif /* CONFIG_CGROUP_CPUACCT */
9257 void synchronize_sched_expedited(void)
9261 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9263 #else /* #ifndef CONFIG_SMP */
9265 static atomic_t synchronize_sched_expedited_count = ATOMIC_INIT(0);
9267 static int synchronize_sched_expedited_cpu_stop(void *data)
9270 * There must be a full memory barrier on each affected CPU
9271 * between the time that try_stop_cpus() is called and the
9272 * time that it returns.
9274 * In the current initial implementation of cpu_stop, the
9275 * above condition is already met when the control reaches
9276 * this point and the following smp_mb() is not strictly
9277 * necessary. Do smp_mb() anyway for documentation and
9278 * robustness against future implementation changes.
9280 smp_mb(); /* See above comment block. */
9285 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
9286 * approach to force grace period to end quickly. This consumes
9287 * significant time on all CPUs, and is thus not recommended for
9288 * any sort of common-case code.
9290 * Note that it is illegal to call this function while holding any
9291 * lock that is acquired by a CPU-hotplug notifier. Failing to
9292 * observe this restriction will result in deadlock.
9294 void synchronize_sched_expedited(void)
9296 int snap, trycount = 0;
9298 smp_mb(); /* ensure prior mod happens before capturing snap. */
9299 snap = atomic_read(&synchronize_sched_expedited_count) + 1;
9301 while (try_stop_cpus(cpu_online_mask,
9302 synchronize_sched_expedited_cpu_stop,
9305 if (trycount++ < 10)
9306 udelay(trycount * num_online_cpus());
9308 synchronize_sched();
9311 if (atomic_read(&synchronize_sched_expedited_count) - snap > 0) {
9312 smp_mb(); /* ensure test happens before caller kfree */
9317 atomic_inc(&synchronize_sched_expedited_count);
9318 smp_mb__after_atomic_inc(); /* ensure post-GP actions seen after GP. */
9321 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9323 #endif /* #else #ifndef CONFIG_SMP */