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 <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/stop_machine.h>
60 #include <linux/sysctl.h>
61 #include <linux/syscalls.h>
62 #include <linux/times.h>
63 #include <linux/tsacct_kern.h>
64 #include <linux/kprobes.h>
65 #include <linux/delayacct.h>
66 #include <linux/unistd.h>
67 #include <linux/pagemap.h>
68 #include <linux/hrtimer.h>
69 #include <linux/tick.h>
70 #include <linux/debugfs.h>
71 #include <linux/ctype.h>
72 #include <linux/ftrace.h>
73 #include <linux/slab.h>
76 #include <asm/irq_regs.h>
77 #include <asm/mutex.h>
79 #include "sched_cpupri.h"
80 #include "workqueue_sched.h"
81 #include "sched_autogroup.h"
83 #define CREATE_TRACE_POINTS
84 #include <trace/events/sched.h>
87 * Convert user-nice values [ -20 ... 0 ... 19 ]
88 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
91 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
92 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
93 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
96 * 'User priority' is the nice value converted to something we
97 * can work with better when scaling various scheduler parameters,
98 * it's a [ 0 ... 39 ] range.
100 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
101 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
102 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
105 * Helpers for converting nanosecond timing to jiffy resolution
107 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
109 #define NICE_0_LOAD SCHED_LOAD_SCALE
110 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
113 * These are the 'tuning knobs' of the scheduler:
115 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
116 * Timeslices get refilled after they expire.
118 #define DEF_TIMESLICE (100 * HZ / 1000)
121 * single value that denotes runtime == period, ie unlimited time.
123 #define RUNTIME_INF ((u64)~0ULL)
125 static inline int rt_policy(int policy)
127 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
132 static inline int task_has_rt_policy(struct task_struct *p)
134 return rt_policy(p->policy);
138 * This is the priority-queue data structure of the RT scheduling class:
140 struct rt_prio_array {
141 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
142 struct list_head queue[MAX_RT_PRIO];
145 struct rt_bandwidth {
146 /* nests inside the rq lock: */
147 raw_spinlock_t rt_runtime_lock;
150 struct hrtimer rt_period_timer;
153 static struct rt_bandwidth def_rt_bandwidth;
155 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
157 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
159 struct rt_bandwidth *rt_b =
160 container_of(timer, struct rt_bandwidth, rt_period_timer);
166 now = hrtimer_cb_get_time(timer);
167 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
172 idle = do_sched_rt_period_timer(rt_b, overrun);
175 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
179 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
181 rt_b->rt_period = ns_to_ktime(period);
182 rt_b->rt_runtime = runtime;
184 raw_spin_lock_init(&rt_b->rt_runtime_lock);
186 hrtimer_init(&rt_b->rt_period_timer,
187 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
188 rt_b->rt_period_timer.function = sched_rt_period_timer;
191 static inline int rt_bandwidth_enabled(void)
193 return sysctl_sched_rt_runtime >= 0;
196 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
200 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
203 if (hrtimer_active(&rt_b->rt_period_timer))
206 raw_spin_lock(&rt_b->rt_runtime_lock);
211 if (hrtimer_active(&rt_b->rt_period_timer))
214 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
215 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
217 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
218 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
219 delta = ktime_to_ns(ktime_sub(hard, soft));
220 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
221 HRTIMER_MODE_ABS_PINNED, 0);
223 raw_spin_unlock(&rt_b->rt_runtime_lock);
226 #ifdef CONFIG_RT_GROUP_SCHED
227 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
229 hrtimer_cancel(&rt_b->rt_period_timer);
234 * sched_domains_mutex serializes calls to init_sched_domains,
235 * detach_destroy_domains and partition_sched_domains.
237 static DEFINE_MUTEX(sched_domains_mutex);
239 #ifdef CONFIG_CGROUP_SCHED
241 #include <linux/cgroup.h>
245 static LIST_HEAD(task_groups);
247 /* task group related information */
249 struct cgroup_subsys_state css;
251 #ifdef CONFIG_FAIR_GROUP_SCHED
252 /* schedulable entities of this group on each cpu */
253 struct sched_entity **se;
254 /* runqueue "owned" by this group on each cpu */
255 struct cfs_rq **cfs_rq;
256 unsigned long shares;
258 atomic_t load_weight;
261 #ifdef CONFIG_RT_GROUP_SCHED
262 struct sched_rt_entity **rt_se;
263 struct rt_rq **rt_rq;
265 struct rt_bandwidth rt_bandwidth;
269 struct list_head list;
271 struct task_group *parent;
272 struct list_head siblings;
273 struct list_head children;
275 #ifdef CONFIG_SCHED_AUTOGROUP
276 struct autogroup *autogroup;
280 /* task_group_lock serializes the addition/removal of task groups */
281 static DEFINE_SPINLOCK(task_group_lock);
283 #ifdef CONFIG_FAIR_GROUP_SCHED
285 # define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
288 * A weight of 0 or 1 can cause arithmetics problems.
289 * A weight of a cfs_rq is the sum of weights of which entities
290 * are queued on this cfs_rq, so a weight of a entity should not be
291 * too large, so as the shares value of a task group.
292 * (The default weight is 1024 - so there's no practical
293 * limitation from this.)
296 #define MAX_SHARES (1UL << 18)
298 static int root_task_group_load = ROOT_TASK_GROUP_LOAD;
301 /* Default task group.
302 * Every task in system belong to this group at bootup.
304 struct task_group root_task_group;
306 #endif /* CONFIG_CGROUP_SCHED */
308 /* CFS-related fields in a runqueue */
310 struct load_weight load;
311 unsigned long nr_running;
316 u64 min_vruntime_copy;
319 struct rb_root tasks_timeline;
320 struct rb_node *rb_leftmost;
322 struct list_head tasks;
323 struct list_head *balance_iterator;
326 * 'curr' points to currently running entity on this cfs_rq.
327 * It is set to NULL otherwise (i.e when none are currently running).
329 struct sched_entity *curr, *next, *last, *skip;
331 #ifdef CONFIG_SCHED_DEBUG
332 unsigned int nr_spread_over;
335 #ifdef CONFIG_FAIR_GROUP_SCHED
336 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
339 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
340 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
341 * (like users, containers etc.)
343 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
344 * list is used during load balance.
347 struct list_head leaf_cfs_rq_list;
348 struct task_group *tg; /* group that "owns" this runqueue */
352 * the part of load.weight contributed by tasks
354 unsigned long task_weight;
357 * h_load = weight * f(tg)
359 * Where f(tg) is the recursive weight fraction assigned to
362 unsigned long h_load;
365 * Maintaining per-cpu shares distribution for group scheduling
367 * load_stamp is the last time we updated the load average
368 * load_last is the last time we updated the load average and saw load
369 * load_unacc_exec_time is currently unaccounted execution time
373 u64 load_stamp, load_last, load_unacc_exec_time;
375 unsigned long load_contribution;
380 /* Real-Time classes' related field in a runqueue: */
382 struct rt_prio_array active;
383 unsigned long rt_nr_running;
384 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
386 int curr; /* highest queued rt task prio */
388 int next; /* next highest */
393 unsigned long rt_nr_migratory;
394 unsigned long rt_nr_total;
396 struct plist_head pushable_tasks;
401 /* Nests inside the rq lock: */
402 raw_spinlock_t rt_runtime_lock;
404 #ifdef CONFIG_RT_GROUP_SCHED
405 unsigned long rt_nr_boosted;
408 struct list_head leaf_rt_rq_list;
409 struct task_group *tg;
416 * We add the notion of a root-domain which will be used to define per-domain
417 * variables. Each exclusive cpuset essentially defines an island domain by
418 * fully partitioning the member cpus from any other cpuset. Whenever a new
419 * exclusive cpuset is created, we also create and attach a new root-domain
427 cpumask_var_t online;
430 * The "RT overload" flag: it gets set if a CPU has more than
431 * one runnable RT task.
433 cpumask_var_t rto_mask;
435 struct cpupri cpupri;
439 * By default the system creates a single root-domain with all cpus as
440 * members (mimicking the global state we have today).
442 static struct root_domain def_root_domain;
444 #endif /* CONFIG_SMP */
447 * This is the main, per-CPU runqueue data structure.
449 * Locking rule: those places that want to lock multiple runqueues
450 * (such as the load balancing or the thread migration code), lock
451 * acquire operations must be ordered by ascending &runqueue.
458 * nr_running and cpu_load should be in the same cacheline because
459 * remote CPUs use both these fields when doing load calculation.
461 unsigned long nr_running;
462 #define CPU_LOAD_IDX_MAX 5
463 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
464 unsigned long last_load_update_tick;
467 unsigned char nohz_balance_kick;
469 int skip_clock_update;
471 /* capture load from *all* tasks on this cpu: */
472 struct load_weight load;
473 unsigned long nr_load_updates;
479 #ifdef CONFIG_FAIR_GROUP_SCHED
480 /* list of leaf cfs_rq on this cpu: */
481 struct list_head leaf_cfs_rq_list;
483 #ifdef CONFIG_RT_GROUP_SCHED
484 struct list_head leaf_rt_rq_list;
488 * This is part of a global counter where only the total sum
489 * over all CPUs matters. A task can increase this counter on
490 * one CPU and if it got migrated afterwards it may decrease
491 * it on another CPU. Always updated under the runqueue lock:
493 unsigned long nr_uninterruptible;
495 struct task_struct *curr, *idle, *stop;
496 unsigned long next_balance;
497 struct mm_struct *prev_mm;
505 struct root_domain *rd;
506 struct sched_domain *sd;
508 unsigned long cpu_power;
510 unsigned char idle_at_tick;
511 /* For active balancing */
515 struct cpu_stop_work active_balance_work;
516 /* cpu of this runqueue: */
520 unsigned long avg_load_per_task;
528 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
532 /* calc_load related fields */
533 unsigned long calc_load_update;
534 long calc_load_active;
536 #ifdef CONFIG_SCHED_HRTICK
538 int hrtick_csd_pending;
539 struct call_single_data hrtick_csd;
541 struct hrtimer hrtick_timer;
544 #ifdef CONFIG_SCHEDSTATS
546 struct sched_info rq_sched_info;
547 unsigned long long rq_cpu_time;
548 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
550 /* sys_sched_yield() stats */
551 unsigned int yld_count;
553 /* schedule() stats */
554 unsigned int sched_switch;
555 unsigned int sched_count;
556 unsigned int sched_goidle;
558 /* try_to_wake_up() stats */
559 unsigned int ttwu_count;
560 unsigned int ttwu_local;
564 struct task_struct *wake_list;
568 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
571 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags);
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_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(&p->pi_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 task_group *tg;
616 struct cgroup_subsys_state *css;
618 css = task_subsys_state_check(p, cpu_cgroup_subsys_id,
619 lockdep_is_held(&p->pi_lock));
620 tg = container_of(css, struct task_group, css);
622 return autogroup_task_group(p, tg);
625 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
626 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
628 #ifdef CONFIG_FAIR_GROUP_SCHED
629 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
630 p->se.parent = task_group(p)->se[cpu];
633 #ifdef CONFIG_RT_GROUP_SCHED
634 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
635 p->rt.parent = task_group(p)->rt_se[cpu];
639 #else /* CONFIG_CGROUP_SCHED */
641 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
642 static inline struct task_group *task_group(struct task_struct *p)
647 #endif /* CONFIG_CGROUP_SCHED */
649 static void update_rq_clock_task(struct rq *rq, s64 delta);
651 static void update_rq_clock(struct rq *rq)
655 if (rq->skip_clock_update > 0)
658 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
660 update_rq_clock_task(rq, delta);
664 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
666 #ifdef CONFIG_SCHED_DEBUG
667 # define const_debug __read_mostly
669 # define const_debug static const
673 * runqueue_is_locked - Returns true if the current cpu runqueue is locked
674 * @cpu: the processor in question.
676 * This interface allows printk to be called with the runqueue lock
677 * held and know whether or not it is OK to wake up the klogd.
679 int runqueue_is_locked(int cpu)
681 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
685 * Debugging: various feature bits
688 #define SCHED_FEAT(name, enabled) \
689 __SCHED_FEAT_##name ,
692 #include "sched_features.h"
697 #define SCHED_FEAT(name, enabled) \
698 (1UL << __SCHED_FEAT_##name) * enabled |
700 const_debug unsigned int sysctl_sched_features =
701 #include "sched_features.h"
706 #ifdef CONFIG_SCHED_DEBUG
707 #define SCHED_FEAT(name, enabled) \
710 static __read_mostly char *sched_feat_names[] = {
711 #include "sched_features.h"
717 static int sched_feat_show(struct seq_file *m, void *v)
721 for (i = 0; sched_feat_names[i]; i++) {
722 if (!(sysctl_sched_features & (1UL << i)))
724 seq_printf(m, "%s ", sched_feat_names[i]);
732 sched_feat_write(struct file *filp, const char __user *ubuf,
733 size_t cnt, loff_t *ppos)
743 if (copy_from_user(&buf, ubuf, cnt))
749 if (strncmp(cmp, "NO_", 3) == 0) {
754 for (i = 0; sched_feat_names[i]; i++) {
755 if (strcmp(cmp, sched_feat_names[i]) == 0) {
757 sysctl_sched_features &= ~(1UL << i);
759 sysctl_sched_features |= (1UL << i);
764 if (!sched_feat_names[i])
772 static int sched_feat_open(struct inode *inode, struct file *filp)
774 return single_open(filp, sched_feat_show, NULL);
777 static const struct file_operations sched_feat_fops = {
778 .open = sched_feat_open,
779 .write = sched_feat_write,
782 .release = single_release,
785 static __init int sched_init_debug(void)
787 debugfs_create_file("sched_features", 0644, NULL, NULL,
792 late_initcall(sched_init_debug);
796 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
799 * Number of tasks to iterate in a single balance run.
800 * Limited because this is done with IRQs disabled.
802 const_debug unsigned int sysctl_sched_nr_migrate = 32;
805 * period over which we average the RT time consumption, measured
810 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
813 * period over which we measure -rt task cpu usage in us.
816 unsigned int sysctl_sched_rt_period = 1000000;
818 static __read_mostly int scheduler_running;
821 * part of the period that we allow rt tasks to run in us.
824 int sysctl_sched_rt_runtime = 950000;
826 static inline u64 global_rt_period(void)
828 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
831 static inline u64 global_rt_runtime(void)
833 if (sysctl_sched_rt_runtime < 0)
836 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
839 #ifndef prepare_arch_switch
840 # define prepare_arch_switch(next) do { } while (0)
842 #ifndef finish_arch_switch
843 # define finish_arch_switch(prev) do { } while (0)
846 static inline int task_current(struct rq *rq, struct task_struct *p)
848 return rq->curr == p;
851 static inline int task_running(struct rq *rq, struct task_struct *p)
856 return task_current(rq, p);
860 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
861 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
865 * We can optimise this out completely for !SMP, because the
866 * SMP rebalancing from interrupt is the only thing that cares
873 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
877 * After ->on_cpu is cleared, the task can be moved to a different CPU.
878 * We must ensure this doesn't happen until the switch is completely
884 #ifdef CONFIG_DEBUG_SPINLOCK
885 /* this is a valid case when another task releases the spinlock */
886 rq->lock.owner = current;
889 * If we are tracking spinlock dependencies then we have to
890 * fix up the runqueue lock - which gets 'carried over' from
893 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
895 raw_spin_unlock_irq(&rq->lock);
898 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
899 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
903 * We can optimise this out completely for !SMP, because the
904 * SMP rebalancing from interrupt is the only thing that cares
909 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
910 raw_spin_unlock_irq(&rq->lock);
912 raw_spin_unlock(&rq->lock);
916 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
920 * After ->on_cpu is cleared, the task can be moved to a different CPU.
921 * We must ensure this doesn't happen until the switch is completely
927 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
931 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
934 * __task_rq_lock - lock the rq @p resides on.
936 static inline struct rq *__task_rq_lock(struct task_struct *p)
941 lockdep_assert_held(&p->pi_lock);
945 raw_spin_lock(&rq->lock);
946 if (likely(rq == task_rq(p)))
948 raw_spin_unlock(&rq->lock);
953 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
955 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
956 __acquires(p->pi_lock)
962 raw_spin_lock_irqsave(&p->pi_lock, *flags);
964 raw_spin_lock(&rq->lock);
965 if (likely(rq == task_rq(p)))
967 raw_spin_unlock(&rq->lock);
968 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
972 static void __task_rq_unlock(struct rq *rq)
975 raw_spin_unlock(&rq->lock);
979 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
981 __releases(p->pi_lock)
983 raw_spin_unlock(&rq->lock);
984 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
988 * this_rq_lock - lock this runqueue and disable interrupts.
990 static struct rq *this_rq_lock(void)
997 raw_spin_lock(&rq->lock);
1002 #ifdef CONFIG_SCHED_HRTICK
1004 * Use HR-timers to deliver accurate preemption points.
1006 * Its all a bit involved since we cannot program an hrt while holding the
1007 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1010 * When we get rescheduled we reprogram the hrtick_timer outside of the
1016 * - enabled by features
1017 * - hrtimer is actually high res
1019 static inline int hrtick_enabled(struct rq *rq)
1021 if (!sched_feat(HRTICK))
1023 if (!cpu_active(cpu_of(rq)))
1025 return hrtimer_is_hres_active(&rq->hrtick_timer);
1028 static void hrtick_clear(struct rq *rq)
1030 if (hrtimer_active(&rq->hrtick_timer))
1031 hrtimer_cancel(&rq->hrtick_timer);
1035 * High-resolution timer tick.
1036 * Runs from hardirq context with interrupts disabled.
1038 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1040 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1042 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1044 raw_spin_lock(&rq->lock);
1045 update_rq_clock(rq);
1046 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1047 raw_spin_unlock(&rq->lock);
1049 return HRTIMER_NORESTART;
1054 * called from hardirq (IPI) context
1056 static void __hrtick_start(void *arg)
1058 struct rq *rq = arg;
1060 raw_spin_lock(&rq->lock);
1061 hrtimer_restart(&rq->hrtick_timer);
1062 rq->hrtick_csd_pending = 0;
1063 raw_spin_unlock(&rq->lock);
1067 * Called to set the hrtick timer state.
1069 * called with rq->lock held and irqs disabled
1071 static void hrtick_start(struct rq *rq, u64 delay)
1073 struct hrtimer *timer = &rq->hrtick_timer;
1074 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1076 hrtimer_set_expires(timer, time);
1078 if (rq == this_rq()) {
1079 hrtimer_restart(timer);
1080 } else if (!rq->hrtick_csd_pending) {
1081 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1082 rq->hrtick_csd_pending = 1;
1087 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1089 int cpu = (int)(long)hcpu;
1092 case CPU_UP_CANCELED:
1093 case CPU_UP_CANCELED_FROZEN:
1094 case CPU_DOWN_PREPARE:
1095 case CPU_DOWN_PREPARE_FROZEN:
1097 case CPU_DEAD_FROZEN:
1098 hrtick_clear(cpu_rq(cpu));
1105 static __init void init_hrtick(void)
1107 hotcpu_notifier(hotplug_hrtick, 0);
1111 * Called to set the hrtick timer state.
1113 * called with rq->lock held and irqs disabled
1115 static void hrtick_start(struct rq *rq, u64 delay)
1117 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1118 HRTIMER_MODE_REL_PINNED, 0);
1121 static inline void init_hrtick(void)
1124 #endif /* CONFIG_SMP */
1126 static void init_rq_hrtick(struct rq *rq)
1129 rq->hrtick_csd_pending = 0;
1131 rq->hrtick_csd.flags = 0;
1132 rq->hrtick_csd.func = __hrtick_start;
1133 rq->hrtick_csd.info = rq;
1136 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1137 rq->hrtick_timer.function = hrtick;
1139 #else /* CONFIG_SCHED_HRTICK */
1140 static inline void hrtick_clear(struct rq *rq)
1144 static inline void init_rq_hrtick(struct rq *rq)
1148 static inline void init_hrtick(void)
1151 #endif /* CONFIG_SCHED_HRTICK */
1154 * resched_task - mark a task 'to be rescheduled now'.
1156 * On UP this means the setting of the need_resched flag, on SMP it
1157 * might also involve a cross-CPU call to trigger the scheduler on
1162 #ifndef tsk_is_polling
1163 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1166 static void resched_task(struct task_struct *p)
1170 assert_raw_spin_locked(&task_rq(p)->lock);
1172 if (test_tsk_need_resched(p))
1175 set_tsk_need_resched(p);
1178 if (cpu == smp_processor_id())
1181 /* NEED_RESCHED must be visible before we test polling */
1183 if (!tsk_is_polling(p))
1184 smp_send_reschedule(cpu);
1187 static void resched_cpu(int cpu)
1189 struct rq *rq = cpu_rq(cpu);
1190 unsigned long flags;
1192 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1194 resched_task(cpu_curr(cpu));
1195 raw_spin_unlock_irqrestore(&rq->lock, flags);
1200 * In the semi idle case, use the nearest busy cpu for migrating timers
1201 * from an idle cpu. This is good for power-savings.
1203 * We don't do similar optimization for completely idle system, as
1204 * selecting an idle cpu will add more delays to the timers than intended
1205 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1207 int get_nohz_timer_target(void)
1209 int cpu = smp_processor_id();
1211 struct sched_domain *sd;
1214 for_each_domain(cpu, sd) {
1215 for_each_cpu(i, sched_domain_span(sd)) {
1227 * When add_timer_on() enqueues a timer into the timer wheel of an
1228 * idle CPU then this timer might expire before the next timer event
1229 * which is scheduled to wake up that CPU. In case of a completely
1230 * idle system the next event might even be infinite time into the
1231 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1232 * leaves the inner idle loop so the newly added timer is taken into
1233 * account when the CPU goes back to idle and evaluates the timer
1234 * wheel for the next timer event.
1236 void wake_up_idle_cpu(int cpu)
1238 struct rq *rq = cpu_rq(cpu);
1240 if (cpu == smp_processor_id())
1244 * This is safe, as this function is called with the timer
1245 * wheel base lock of (cpu) held. When the CPU is on the way
1246 * to idle and has not yet set rq->curr to idle then it will
1247 * be serialized on the timer wheel base lock and take the new
1248 * timer into account automatically.
1250 if (rq->curr != rq->idle)
1254 * We can set TIF_RESCHED on the idle task of the other CPU
1255 * lockless. The worst case is that the other CPU runs the
1256 * idle task through an additional NOOP schedule()
1258 set_tsk_need_resched(rq->idle);
1260 /* NEED_RESCHED must be visible before we test polling */
1262 if (!tsk_is_polling(rq->idle))
1263 smp_send_reschedule(cpu);
1266 #endif /* CONFIG_NO_HZ */
1268 static u64 sched_avg_period(void)
1270 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1273 static void sched_avg_update(struct rq *rq)
1275 s64 period = sched_avg_period();
1277 while ((s64)(rq->clock - rq->age_stamp) > period) {
1279 * Inline assembly required to prevent the compiler
1280 * optimising this loop into a divmod call.
1281 * See __iter_div_u64_rem() for another example of this.
1283 asm("" : "+rm" (rq->age_stamp));
1284 rq->age_stamp += period;
1289 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1291 rq->rt_avg += rt_delta;
1292 sched_avg_update(rq);
1295 #else /* !CONFIG_SMP */
1296 static void resched_task(struct task_struct *p)
1298 assert_raw_spin_locked(&task_rq(p)->lock);
1299 set_tsk_need_resched(p);
1302 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1306 static void sched_avg_update(struct rq *rq)
1309 #endif /* CONFIG_SMP */
1311 #if BITS_PER_LONG == 32
1312 # define WMULT_CONST (~0UL)
1314 # define WMULT_CONST (1UL << 32)
1317 #define WMULT_SHIFT 32
1320 * Shift right and round:
1322 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1325 * delta *= weight / lw
1327 static unsigned long
1328 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1329 struct load_weight *lw)
1333 tmp = (u64)delta_exec * weight;
1335 if (!lw->inv_weight) {
1336 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1339 lw->inv_weight = WMULT_CONST / lw->weight;
1343 * Check whether we'd overflow the 64-bit multiplication:
1345 if (unlikely(tmp > WMULT_CONST))
1346 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1349 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1351 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1354 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1360 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1366 static inline void update_load_set(struct load_weight *lw, unsigned long w)
1373 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1374 * of tasks with abnormal "nice" values across CPUs the contribution that
1375 * each task makes to its run queue's load is weighted according to its
1376 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1377 * scaled version of the new time slice allocation that they receive on time
1381 #define WEIGHT_IDLEPRIO 3
1382 #define WMULT_IDLEPRIO 1431655765
1385 * Nice levels are multiplicative, with a gentle 10% change for every
1386 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1387 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1388 * that remained on nice 0.
1390 * The "10% effect" is relative and cumulative: from _any_ nice level,
1391 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1392 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1393 * If a task goes up by ~10% and another task goes down by ~10% then
1394 * the relative distance between them is ~25%.)
1396 static const int prio_to_weight[40] = {
1397 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1398 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1399 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1400 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1401 /* 0 */ 1024, 820, 655, 526, 423,
1402 /* 5 */ 335, 272, 215, 172, 137,
1403 /* 10 */ 110, 87, 70, 56, 45,
1404 /* 15 */ 36, 29, 23, 18, 15,
1408 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1410 * In cases where the weight does not change often, we can use the
1411 * precalculated inverse to speed up arithmetics by turning divisions
1412 * into multiplications:
1414 static const u32 prio_to_wmult[40] = {
1415 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1416 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1417 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1418 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1419 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1420 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1421 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1422 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1425 /* Time spent by the tasks of the cpu accounting group executing in ... */
1426 enum cpuacct_stat_index {
1427 CPUACCT_STAT_USER, /* ... user mode */
1428 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1430 CPUACCT_STAT_NSTATS,
1433 #ifdef CONFIG_CGROUP_CPUACCT
1434 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1435 static void cpuacct_update_stats(struct task_struct *tsk,
1436 enum cpuacct_stat_index idx, cputime_t val);
1438 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1439 static inline void cpuacct_update_stats(struct task_struct *tsk,
1440 enum cpuacct_stat_index idx, cputime_t val) {}
1443 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1445 update_load_add(&rq->load, load);
1448 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1450 update_load_sub(&rq->load, load);
1453 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1454 typedef int (*tg_visitor)(struct task_group *, void *);
1457 * Iterate the full tree, calling @down when first entering a node and @up when
1458 * leaving it for the final time.
1460 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1462 struct task_group *parent, *child;
1466 parent = &root_task_group;
1468 ret = (*down)(parent, data);
1471 list_for_each_entry_rcu(child, &parent->children, siblings) {
1478 ret = (*up)(parent, data);
1483 parent = parent->parent;
1492 static int tg_nop(struct task_group *tg, void *data)
1499 /* Used instead of source_load when we know the type == 0 */
1500 static unsigned long weighted_cpuload(const int cpu)
1502 return cpu_rq(cpu)->load.weight;
1506 * Return a low guess at the load of a migration-source cpu weighted
1507 * according to the scheduling class and "nice" value.
1509 * We want to under-estimate the load of migration sources, to
1510 * balance conservatively.
1512 static unsigned long source_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 min(rq->cpu_load[type-1], total);
1524 * Return a high guess at the load of a migration-target cpu weighted
1525 * according to the scheduling class and "nice" value.
1527 static unsigned long target_load(int cpu, int type)
1529 struct rq *rq = cpu_rq(cpu);
1530 unsigned long total = weighted_cpuload(cpu);
1532 if (type == 0 || !sched_feat(LB_BIAS))
1535 return max(rq->cpu_load[type-1], total);
1538 static unsigned long power_of(int cpu)
1540 return cpu_rq(cpu)->cpu_power;
1543 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1545 static unsigned long cpu_avg_load_per_task(int cpu)
1547 struct rq *rq = cpu_rq(cpu);
1548 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1551 rq->avg_load_per_task = rq->load.weight / nr_running;
1553 rq->avg_load_per_task = 0;
1555 return rq->avg_load_per_task;
1558 #ifdef CONFIG_FAIR_GROUP_SCHED
1561 * Compute the cpu's hierarchical load factor for each task group.
1562 * This needs to be done in a top-down fashion because the load of a child
1563 * group is a fraction of its parents load.
1565 static int tg_load_down(struct task_group *tg, void *data)
1568 long cpu = (long)data;
1571 load = cpu_rq(cpu)->load.weight;
1573 load = tg->parent->cfs_rq[cpu]->h_load;
1574 load *= tg->se[cpu]->load.weight;
1575 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1578 tg->cfs_rq[cpu]->h_load = load;
1583 static void update_h_load(long cpu)
1585 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1590 #ifdef CONFIG_PREEMPT
1592 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1595 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1596 * way at the expense of forcing extra atomic operations in all
1597 * invocations. This assures that the double_lock is acquired using the
1598 * same underlying policy as the spinlock_t on this architecture, which
1599 * reduces latency compared to the unfair variant below. However, it
1600 * also adds more overhead and therefore may reduce throughput.
1602 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1603 __releases(this_rq->lock)
1604 __acquires(busiest->lock)
1605 __acquires(this_rq->lock)
1607 raw_spin_unlock(&this_rq->lock);
1608 double_rq_lock(this_rq, busiest);
1615 * Unfair double_lock_balance: Optimizes throughput at the expense of
1616 * latency by eliminating extra atomic operations when the locks are
1617 * already in proper order on entry. This favors lower cpu-ids and will
1618 * grant the double lock to lower cpus over higher ids under contention,
1619 * regardless of entry order into the function.
1621 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1622 __releases(this_rq->lock)
1623 __acquires(busiest->lock)
1624 __acquires(this_rq->lock)
1628 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1629 if (busiest < this_rq) {
1630 raw_spin_unlock(&this_rq->lock);
1631 raw_spin_lock(&busiest->lock);
1632 raw_spin_lock_nested(&this_rq->lock,
1633 SINGLE_DEPTH_NESTING);
1636 raw_spin_lock_nested(&busiest->lock,
1637 SINGLE_DEPTH_NESTING);
1642 #endif /* CONFIG_PREEMPT */
1645 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1647 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1649 if (unlikely(!irqs_disabled())) {
1650 /* printk() doesn't work good under rq->lock */
1651 raw_spin_unlock(&this_rq->lock);
1655 return _double_lock_balance(this_rq, busiest);
1658 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1659 __releases(busiest->lock)
1661 raw_spin_unlock(&busiest->lock);
1662 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1666 * double_rq_lock - safely lock two runqueues
1668 * Note this does not disable interrupts like task_rq_lock,
1669 * you need to do so manually before calling.
1671 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1672 __acquires(rq1->lock)
1673 __acquires(rq2->lock)
1675 BUG_ON(!irqs_disabled());
1677 raw_spin_lock(&rq1->lock);
1678 __acquire(rq2->lock); /* Fake it out ;) */
1681 raw_spin_lock(&rq1->lock);
1682 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1684 raw_spin_lock(&rq2->lock);
1685 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1691 * double_rq_unlock - safely unlock two runqueues
1693 * Note this does not restore interrupts like task_rq_unlock,
1694 * you need to do so manually after calling.
1696 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1697 __releases(rq1->lock)
1698 __releases(rq2->lock)
1700 raw_spin_unlock(&rq1->lock);
1702 raw_spin_unlock(&rq2->lock);
1704 __release(rq2->lock);
1707 #else /* CONFIG_SMP */
1710 * double_rq_lock - safely lock two runqueues
1712 * Note this does not disable interrupts like task_rq_lock,
1713 * you need to do so manually before calling.
1715 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1716 __acquires(rq1->lock)
1717 __acquires(rq2->lock)
1719 BUG_ON(!irqs_disabled());
1721 raw_spin_lock(&rq1->lock);
1722 __acquire(rq2->lock); /* Fake it out ;) */
1726 * double_rq_unlock - safely unlock two runqueues
1728 * Note this does not restore interrupts like task_rq_unlock,
1729 * you need to do so manually after calling.
1731 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1732 __releases(rq1->lock)
1733 __releases(rq2->lock)
1736 raw_spin_unlock(&rq1->lock);
1737 __release(rq2->lock);
1742 static void calc_load_account_idle(struct rq *this_rq);
1743 static void update_sysctl(void);
1744 static int get_update_sysctl_factor(void);
1745 static void update_cpu_load(struct rq *this_rq);
1747 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1749 set_task_rq(p, cpu);
1752 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1753 * successfuly executed on another CPU. We must ensure that updates of
1754 * per-task data have been completed by this moment.
1757 task_thread_info(p)->cpu = cpu;
1761 static const struct sched_class rt_sched_class;
1763 #define sched_class_highest (&stop_sched_class)
1764 #define for_each_class(class) \
1765 for (class = sched_class_highest; class; class = class->next)
1767 #include "sched_stats.h"
1769 static void inc_nr_running(struct rq *rq)
1774 static void dec_nr_running(struct rq *rq)
1779 static void set_load_weight(struct task_struct *p)
1781 int prio = p->static_prio - MAX_RT_PRIO;
1782 struct load_weight *load = &p->se.load;
1785 * SCHED_IDLE tasks get minimal weight:
1787 if (p->policy == SCHED_IDLE) {
1788 load->weight = WEIGHT_IDLEPRIO;
1789 load->inv_weight = WMULT_IDLEPRIO;
1793 load->weight = prio_to_weight[prio];
1794 load->inv_weight = prio_to_wmult[prio];
1797 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1799 update_rq_clock(rq);
1800 sched_info_queued(p);
1801 p->sched_class->enqueue_task(rq, p, flags);
1804 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1806 update_rq_clock(rq);
1807 sched_info_dequeued(p);
1808 p->sched_class->dequeue_task(rq, p, flags);
1812 * activate_task - move a task to the runqueue.
1814 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1816 if (task_contributes_to_load(p))
1817 rq->nr_uninterruptible--;
1819 enqueue_task(rq, p, flags);
1824 * deactivate_task - remove a task from the runqueue.
1826 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1828 if (task_contributes_to_load(p))
1829 rq->nr_uninterruptible++;
1831 dequeue_task(rq, p, flags);
1835 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1838 * There are no locks covering percpu hardirq/softirq time.
1839 * They are only modified in account_system_vtime, on corresponding CPU
1840 * with interrupts disabled. So, writes are safe.
1841 * They are read and saved off onto struct rq in update_rq_clock().
1842 * This may result in other CPU reading this CPU's irq time and can
1843 * race with irq/account_system_vtime on this CPU. We would either get old
1844 * or new value with a side effect of accounting a slice of irq time to wrong
1845 * task when irq is in progress while we read rq->clock. That is a worthy
1846 * compromise in place of having locks on each irq in account_system_time.
1848 static DEFINE_PER_CPU(u64, cpu_hardirq_time);
1849 static DEFINE_PER_CPU(u64, cpu_softirq_time);
1851 static DEFINE_PER_CPU(u64, irq_start_time);
1852 static int sched_clock_irqtime;
1854 void enable_sched_clock_irqtime(void)
1856 sched_clock_irqtime = 1;
1859 void disable_sched_clock_irqtime(void)
1861 sched_clock_irqtime = 0;
1864 #ifndef CONFIG_64BIT
1865 static DEFINE_PER_CPU(seqcount_t, irq_time_seq);
1867 static inline void irq_time_write_begin(void)
1869 __this_cpu_inc(irq_time_seq.sequence);
1873 static inline void irq_time_write_end(void)
1876 __this_cpu_inc(irq_time_seq.sequence);
1879 static inline u64 irq_time_read(int cpu)
1885 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
1886 irq_time = per_cpu(cpu_softirq_time, cpu) +
1887 per_cpu(cpu_hardirq_time, cpu);
1888 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
1892 #else /* CONFIG_64BIT */
1893 static inline void irq_time_write_begin(void)
1897 static inline void irq_time_write_end(void)
1901 static inline u64 irq_time_read(int cpu)
1903 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
1905 #endif /* CONFIG_64BIT */
1908 * Called before incrementing preempt_count on {soft,}irq_enter
1909 * and before decrementing preempt_count on {soft,}irq_exit.
1911 void account_system_vtime(struct task_struct *curr)
1913 unsigned long flags;
1917 if (!sched_clock_irqtime)
1920 local_irq_save(flags);
1922 cpu = smp_processor_id();
1923 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
1924 __this_cpu_add(irq_start_time, delta);
1926 irq_time_write_begin();
1928 * We do not account for softirq time from ksoftirqd here.
1929 * We want to continue accounting softirq time to ksoftirqd thread
1930 * in that case, so as not to confuse scheduler with a special task
1931 * that do not consume any time, but still wants to run.
1933 if (hardirq_count())
1934 __this_cpu_add(cpu_hardirq_time, delta);
1935 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
1936 __this_cpu_add(cpu_softirq_time, delta);
1938 irq_time_write_end();
1939 local_irq_restore(flags);
1941 EXPORT_SYMBOL_GPL(account_system_vtime);
1943 static void update_rq_clock_task(struct rq *rq, s64 delta)
1947 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
1950 * Since irq_time is only updated on {soft,}irq_exit, we might run into
1951 * this case when a previous update_rq_clock() happened inside a
1952 * {soft,}irq region.
1954 * When this happens, we stop ->clock_task and only update the
1955 * prev_irq_time stamp to account for the part that fit, so that a next
1956 * update will consume the rest. This ensures ->clock_task is
1959 * It does however cause some slight miss-attribution of {soft,}irq
1960 * time, a more accurate solution would be to update the irq_time using
1961 * the current rq->clock timestamp, except that would require using
1964 if (irq_delta > delta)
1967 rq->prev_irq_time += irq_delta;
1969 rq->clock_task += delta;
1971 if (irq_delta && sched_feat(NONIRQ_POWER))
1972 sched_rt_avg_update(rq, irq_delta);
1975 static int irqtime_account_hi_update(void)
1977 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
1978 unsigned long flags;
1982 local_irq_save(flags);
1983 latest_ns = this_cpu_read(cpu_hardirq_time);
1984 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->irq))
1986 local_irq_restore(flags);
1990 static int irqtime_account_si_update(void)
1992 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
1993 unsigned long flags;
1997 local_irq_save(flags);
1998 latest_ns = this_cpu_read(cpu_softirq_time);
1999 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->softirq))
2001 local_irq_restore(flags);
2005 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
2007 #define sched_clock_irqtime (0)
2009 static void update_rq_clock_task(struct rq *rq, s64 delta)
2011 rq->clock_task += delta;
2014 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2016 #include "sched_idletask.c"
2017 #include "sched_fair.c"
2018 #include "sched_rt.c"
2019 #include "sched_autogroup.c"
2020 #include "sched_stoptask.c"
2021 #ifdef CONFIG_SCHED_DEBUG
2022 # include "sched_debug.c"
2025 void sched_set_stop_task(int cpu, struct task_struct *stop)
2027 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
2028 struct task_struct *old_stop = cpu_rq(cpu)->stop;
2032 * Make it appear like a SCHED_FIFO task, its something
2033 * userspace knows about and won't get confused about.
2035 * Also, it will make PI more or less work without too
2036 * much confusion -- but then, stop work should not
2037 * rely on PI working anyway.
2039 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
2041 stop->sched_class = &stop_sched_class;
2044 cpu_rq(cpu)->stop = stop;
2048 * Reset it back to a normal scheduling class so that
2049 * it can die in pieces.
2051 old_stop->sched_class = &rt_sched_class;
2056 * __normal_prio - return the priority that is based on the static prio
2058 static inline int __normal_prio(struct task_struct *p)
2060 return p->static_prio;
2064 * Calculate the expected normal priority: i.e. priority
2065 * without taking RT-inheritance into account. Might be
2066 * boosted by interactivity modifiers. Changes upon fork,
2067 * setprio syscalls, and whenever the interactivity
2068 * estimator recalculates.
2070 static inline int normal_prio(struct task_struct *p)
2074 if (task_has_rt_policy(p))
2075 prio = MAX_RT_PRIO-1 - p->rt_priority;
2077 prio = __normal_prio(p);
2082 * Calculate the current priority, i.e. the priority
2083 * taken into account by the scheduler. This value might
2084 * be boosted by RT tasks, or might be boosted by
2085 * interactivity modifiers. Will be RT if the task got
2086 * RT-boosted. If not then it returns p->normal_prio.
2088 static int effective_prio(struct task_struct *p)
2090 p->normal_prio = normal_prio(p);
2092 * If we are RT tasks or we were boosted to RT priority,
2093 * keep the priority unchanged. Otherwise, update priority
2094 * to the normal priority:
2096 if (!rt_prio(p->prio))
2097 return p->normal_prio;
2102 * task_curr - is this task currently executing on a CPU?
2103 * @p: the task in question.
2105 inline int task_curr(const struct task_struct *p)
2107 return cpu_curr(task_cpu(p)) == p;
2110 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2111 const struct sched_class *prev_class,
2114 if (prev_class != p->sched_class) {
2115 if (prev_class->switched_from)
2116 prev_class->switched_from(rq, p);
2117 p->sched_class->switched_to(rq, p);
2118 } else if (oldprio != p->prio)
2119 p->sched_class->prio_changed(rq, p, oldprio);
2122 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2124 const struct sched_class *class;
2126 if (p->sched_class == rq->curr->sched_class) {
2127 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2129 for_each_class(class) {
2130 if (class == rq->curr->sched_class)
2132 if (class == p->sched_class) {
2133 resched_task(rq->curr);
2140 * A queue event has occurred, and we're going to schedule. In
2141 * this case, we can save a useless back to back clock update.
2143 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
2144 rq->skip_clock_update = 1;
2149 * Is this task likely cache-hot:
2152 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2156 if (p->sched_class != &fair_sched_class)
2159 if (unlikely(p->policy == SCHED_IDLE))
2163 * Buddy candidates are cache hot:
2165 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2166 (&p->se == cfs_rq_of(&p->se)->next ||
2167 &p->se == cfs_rq_of(&p->se)->last))
2170 if (sysctl_sched_migration_cost == -1)
2172 if (sysctl_sched_migration_cost == 0)
2175 delta = now - p->se.exec_start;
2177 return delta < (s64)sysctl_sched_migration_cost;
2180 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2182 #ifdef CONFIG_SCHED_DEBUG
2184 * We should never call set_task_cpu() on a blocked task,
2185 * ttwu() will sort out the placement.
2187 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2188 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2190 #ifdef CONFIG_LOCKDEP
2191 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
2192 lockdep_is_held(&task_rq(p)->lock)));
2196 trace_sched_migrate_task(p, new_cpu);
2198 if (task_cpu(p) != new_cpu) {
2199 p->se.nr_migrations++;
2200 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2203 __set_task_cpu(p, new_cpu);
2206 struct migration_arg {
2207 struct task_struct *task;
2211 static int migration_cpu_stop(void *data);
2214 * wait_task_inactive - wait for a thread to unschedule.
2216 * If @match_state is nonzero, it's the @p->state value just checked and
2217 * not expected to change. If it changes, i.e. @p might have woken up,
2218 * then return zero. When we succeed in waiting for @p to be off its CPU,
2219 * we return a positive number (its total switch count). If a second call
2220 * a short while later returns the same number, the caller can be sure that
2221 * @p has remained unscheduled the whole time.
2223 * The caller must ensure that the task *will* unschedule sometime soon,
2224 * else this function might spin for a *long* time. This function can't
2225 * be called with interrupts off, or it may introduce deadlock with
2226 * smp_call_function() if an IPI is sent by the same process we are
2227 * waiting to become inactive.
2229 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2231 unsigned long flags;
2238 * We do the initial early heuristics without holding
2239 * any task-queue locks at all. We'll only try to get
2240 * the runqueue lock when things look like they will
2246 * If the task is actively running on another CPU
2247 * still, just relax and busy-wait without holding
2250 * NOTE! Since we don't hold any locks, it's not
2251 * even sure that "rq" stays as the right runqueue!
2252 * But we don't care, since "task_running()" will
2253 * return false if the runqueue has changed and p
2254 * is actually now running somewhere else!
2256 while (task_running(rq, p)) {
2257 if (match_state && unlikely(p->state != match_state))
2263 * Ok, time to look more closely! We need the rq
2264 * lock now, to be *sure*. If we're wrong, we'll
2265 * just go back and repeat.
2267 rq = task_rq_lock(p, &flags);
2268 trace_sched_wait_task(p);
2269 running = task_running(rq, p);
2272 if (!match_state || p->state == match_state)
2273 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2274 task_rq_unlock(rq, p, &flags);
2277 * If it changed from the expected state, bail out now.
2279 if (unlikely(!ncsw))
2283 * Was it really running after all now that we
2284 * checked with the proper locks actually held?
2286 * Oops. Go back and try again..
2288 if (unlikely(running)) {
2294 * It's not enough that it's not actively running,
2295 * it must be off the runqueue _entirely_, and not
2298 * So if it was still runnable (but just not actively
2299 * running right now), it's preempted, and we should
2300 * yield - it could be a while.
2302 if (unlikely(on_rq)) {
2303 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
2305 set_current_state(TASK_UNINTERRUPTIBLE);
2306 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
2311 * Ahh, all good. It wasn't running, and it wasn't
2312 * runnable, which means that it will never become
2313 * running in the future either. We're all done!
2322 * kick_process - kick a running thread to enter/exit the kernel
2323 * @p: the to-be-kicked thread
2325 * Cause a process which is running on another CPU to enter
2326 * kernel-mode, without any delay. (to get signals handled.)
2328 * NOTE: this function doesn't have to take the runqueue lock,
2329 * because all it wants to ensure is that the remote task enters
2330 * the kernel. If the IPI races and the task has been migrated
2331 * to another CPU then no harm is done and the purpose has been
2334 void kick_process(struct task_struct *p)
2340 if ((cpu != smp_processor_id()) && task_curr(p))
2341 smp_send_reschedule(cpu);
2344 EXPORT_SYMBOL_GPL(kick_process);
2345 #endif /* CONFIG_SMP */
2349 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
2351 static int select_fallback_rq(int cpu, struct task_struct *p)
2354 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2356 /* Look for allowed, online CPU in same node. */
2357 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2358 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2361 /* Any allowed, online CPU? */
2362 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2363 if (dest_cpu < nr_cpu_ids)
2366 /* No more Mr. Nice Guy. */
2367 dest_cpu = cpuset_cpus_allowed_fallback(p);
2369 * Don't tell them about moving exiting tasks or
2370 * kernel threads (both mm NULL), since they never
2373 if (p->mm && printk_ratelimit()) {
2374 printk(KERN_INFO "process %d (%s) no longer affine to cpu%d\n",
2375 task_pid_nr(p), p->comm, cpu);
2382 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
2385 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2387 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2390 * In order not to call set_task_cpu() on a blocking task we need
2391 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2394 * Since this is common to all placement strategies, this lives here.
2396 * [ this allows ->select_task() to simply return task_cpu(p) and
2397 * not worry about this generic constraint ]
2399 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2401 cpu = select_fallback_rq(task_cpu(p), p);
2406 static void update_avg(u64 *avg, u64 sample)
2408 s64 diff = sample - *avg;
2414 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
2416 #ifdef CONFIG_SCHEDSTATS
2417 struct rq *rq = this_rq();
2420 int this_cpu = smp_processor_id();
2422 if (cpu == this_cpu) {
2423 schedstat_inc(rq, ttwu_local);
2424 schedstat_inc(p, se.statistics.nr_wakeups_local);
2426 struct sched_domain *sd;
2428 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2430 for_each_domain(this_cpu, sd) {
2431 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2432 schedstat_inc(sd, ttwu_wake_remote);
2438 #endif /* CONFIG_SMP */
2440 schedstat_inc(rq, ttwu_count);
2441 schedstat_inc(p, se.statistics.nr_wakeups);
2443 if (wake_flags & WF_SYNC)
2444 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2446 if (cpu != task_cpu(p))
2447 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2449 #endif /* CONFIG_SCHEDSTATS */
2452 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
2454 activate_task(rq, p, en_flags);
2457 /* if a worker is waking up, notify workqueue */
2458 if (p->flags & PF_WQ_WORKER)
2459 wq_worker_waking_up(p, cpu_of(rq));
2463 * Mark the task runnable and perform wakeup-preemption.
2466 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
2468 trace_sched_wakeup(p, true);
2469 check_preempt_curr(rq, p, wake_flags);
2471 p->state = TASK_RUNNING;
2473 if (p->sched_class->task_woken)
2474 p->sched_class->task_woken(rq, p);
2476 if (unlikely(rq->idle_stamp)) {
2477 u64 delta = rq->clock - rq->idle_stamp;
2478 u64 max = 2*sysctl_sched_migration_cost;
2483 update_avg(&rq->avg_idle, delta);
2490 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
2493 if (p->sched_contributes_to_load)
2494 rq->nr_uninterruptible--;
2497 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
2498 ttwu_do_wakeup(rq, p, wake_flags);
2502 * Called in case the task @p isn't fully descheduled from its runqueue,
2503 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
2504 * since all we need to do is flip p->state to TASK_RUNNING, since
2505 * the task is still ->on_rq.
2507 static int ttwu_remote(struct task_struct *p, int wake_flags)
2512 rq = __task_rq_lock(p);
2514 ttwu_do_wakeup(rq, p, wake_flags);
2517 __task_rq_unlock(rq);
2523 static void sched_ttwu_pending(void)
2525 struct rq *rq = this_rq();
2526 struct task_struct *list = xchg(&rq->wake_list, NULL);
2531 raw_spin_lock(&rq->lock);
2534 struct task_struct *p = list;
2535 list = list->wake_entry;
2536 ttwu_do_activate(rq, p, 0);
2539 raw_spin_unlock(&rq->lock);
2542 void scheduler_ipi(void)
2544 sched_ttwu_pending();
2547 static void ttwu_queue_remote(struct task_struct *p, int cpu)
2549 struct rq *rq = cpu_rq(cpu);
2550 struct task_struct *next = rq->wake_list;
2553 struct task_struct *old = next;
2555 p->wake_entry = next;
2556 next = cmpxchg(&rq->wake_list, old, p);
2562 smp_send_reschedule(cpu);
2566 static void ttwu_queue(struct task_struct *p, int cpu)
2568 struct rq *rq = cpu_rq(cpu);
2570 #if defined(CONFIG_SMP) && defined(CONFIG_SCHED_TTWU_QUEUE)
2571 if (sched_feat(TTWU_QUEUE) && cpu != smp_processor_id()) {
2572 ttwu_queue_remote(p, cpu);
2577 raw_spin_lock(&rq->lock);
2578 ttwu_do_activate(rq, p, 0);
2579 raw_spin_unlock(&rq->lock);
2583 * try_to_wake_up - wake up a thread
2584 * @p: the thread to be awakened
2585 * @state: the mask of task states that can be woken
2586 * @wake_flags: wake modifier flags (WF_*)
2588 * Put it on the run-queue if it's not already there. The "current"
2589 * thread is always on the run-queue (except when the actual
2590 * re-schedule is in progress), and as such you're allowed to do
2591 * the simpler "current->state = TASK_RUNNING" to mark yourself
2592 * runnable without the overhead of this.
2594 * Returns %true if @p was woken up, %false if it was already running
2595 * or @state didn't match @p's state.
2598 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
2600 unsigned long flags;
2601 int cpu, success = 0;
2604 raw_spin_lock_irqsave(&p->pi_lock, flags);
2605 if (!(p->state & state))
2608 success = 1; /* we're going to change ->state */
2611 if (p->on_rq && ttwu_remote(p, wake_flags))
2616 * If the owning (remote) cpu is still in the middle of schedule() with
2617 * this task as prev, wait until its done referencing the task.
2620 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2622 * If called from interrupt context we could have landed in the
2623 * middle of schedule(), in this case we should take care not
2624 * to spin on ->on_cpu if p is current, since that would
2635 * Pairs with the smp_wmb() in finish_lock_switch().
2639 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2640 p->state = TASK_WAKING;
2642 if (p->sched_class->task_waking)
2643 p->sched_class->task_waking(p);
2645 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2646 if (task_cpu(p) != cpu)
2647 set_task_cpu(p, cpu);
2648 #endif /* CONFIG_SMP */
2652 ttwu_stat(p, cpu, wake_flags);
2654 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2660 * try_to_wake_up_local - try to wake up a local task with rq lock held
2661 * @p: the thread to be awakened
2663 * Put @p on the run-queue if it's not already there. The caller must
2664 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2667 static void try_to_wake_up_local(struct task_struct *p)
2669 struct rq *rq = task_rq(p);
2671 BUG_ON(rq != this_rq());
2672 BUG_ON(p == current);
2673 lockdep_assert_held(&rq->lock);
2675 if (!raw_spin_trylock(&p->pi_lock)) {
2676 raw_spin_unlock(&rq->lock);
2677 raw_spin_lock(&p->pi_lock);
2678 raw_spin_lock(&rq->lock);
2681 if (!(p->state & TASK_NORMAL))
2685 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2687 ttwu_do_wakeup(rq, p, 0);
2688 ttwu_stat(p, smp_processor_id(), 0);
2690 raw_spin_unlock(&p->pi_lock);
2694 * wake_up_process - Wake up a specific process
2695 * @p: The process to be woken up.
2697 * Attempt to wake up the nominated process and move it to the set of runnable
2698 * processes. Returns 1 if the process was woken up, 0 if it was already
2701 * It may be assumed that this function implies a write memory barrier before
2702 * changing the task state if and only if any tasks are woken up.
2704 int wake_up_process(struct task_struct *p)
2706 return try_to_wake_up(p, TASK_ALL, 0);
2708 EXPORT_SYMBOL(wake_up_process);
2710 int wake_up_state(struct task_struct *p, unsigned int state)
2712 return try_to_wake_up(p, state, 0);
2716 * Perform scheduler related setup for a newly forked process p.
2717 * p is forked by current.
2719 * __sched_fork() is basic setup used by init_idle() too:
2721 static void __sched_fork(struct task_struct *p)
2726 p->se.exec_start = 0;
2727 p->se.sum_exec_runtime = 0;
2728 p->se.prev_sum_exec_runtime = 0;
2729 p->se.nr_migrations = 0;
2731 INIT_LIST_HEAD(&p->se.group_node);
2733 #ifdef CONFIG_SCHEDSTATS
2734 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2737 INIT_LIST_HEAD(&p->rt.run_list);
2739 #ifdef CONFIG_PREEMPT_NOTIFIERS
2740 INIT_HLIST_HEAD(&p->preempt_notifiers);
2745 * fork()/clone()-time setup:
2747 void sched_fork(struct task_struct *p)
2749 unsigned long flags;
2750 int cpu = get_cpu();
2754 * We mark the process as running here. This guarantees that
2755 * nobody will actually run it, and a signal or other external
2756 * event cannot wake it up and insert it on the runqueue either.
2758 p->state = TASK_RUNNING;
2761 * Revert to default priority/policy on fork if requested.
2763 if (unlikely(p->sched_reset_on_fork)) {
2764 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2765 p->policy = SCHED_NORMAL;
2766 p->normal_prio = p->static_prio;
2769 if (PRIO_TO_NICE(p->static_prio) < 0) {
2770 p->static_prio = NICE_TO_PRIO(0);
2771 p->normal_prio = p->static_prio;
2776 * We don't need the reset flag anymore after the fork. It has
2777 * fulfilled its duty:
2779 p->sched_reset_on_fork = 0;
2783 * Make sure we do not leak PI boosting priority to the child.
2785 p->prio = current->normal_prio;
2787 if (!rt_prio(p->prio))
2788 p->sched_class = &fair_sched_class;
2790 if (p->sched_class->task_fork)
2791 p->sched_class->task_fork(p);
2794 * The child is not yet in the pid-hash so no cgroup attach races,
2795 * and the cgroup is pinned to this child due to cgroup_fork()
2796 * is ran before sched_fork().
2798 * Silence PROVE_RCU.
2800 raw_spin_lock_irqsave(&p->pi_lock, flags);
2801 set_task_cpu(p, cpu);
2802 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2804 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2805 if (likely(sched_info_on()))
2806 memset(&p->sched_info, 0, sizeof(p->sched_info));
2808 #if defined(CONFIG_SMP)
2811 #ifdef CONFIG_PREEMPT
2812 /* Want to start with kernel preemption disabled. */
2813 task_thread_info(p)->preempt_count = 1;
2816 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2823 * wake_up_new_task - wake up a newly created task for the first time.
2825 * This function will do some initial scheduler statistics housekeeping
2826 * that must be done for every newly created context, then puts the task
2827 * on the runqueue and wakes it.
2829 void wake_up_new_task(struct task_struct *p)
2831 unsigned long flags;
2834 raw_spin_lock_irqsave(&p->pi_lock, flags);
2837 * Fork balancing, do it here and not earlier because:
2838 * - cpus_allowed can change in the fork path
2839 * - any previously selected cpu might disappear through hotplug
2841 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
2844 rq = __task_rq_lock(p);
2845 activate_task(rq, p, 0);
2847 trace_sched_wakeup_new(p, true);
2848 check_preempt_curr(rq, p, WF_FORK);
2850 if (p->sched_class->task_woken)
2851 p->sched_class->task_woken(rq, p);
2853 task_rq_unlock(rq, p, &flags);
2856 #ifdef CONFIG_PREEMPT_NOTIFIERS
2859 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2860 * @notifier: notifier struct to register
2862 void preempt_notifier_register(struct preempt_notifier *notifier)
2864 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2866 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2869 * preempt_notifier_unregister - no longer interested in preemption notifications
2870 * @notifier: notifier struct to unregister
2872 * This is safe to call from within a preemption notifier.
2874 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2876 hlist_del(¬ifier->link);
2878 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2880 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2882 struct preempt_notifier *notifier;
2883 struct hlist_node *node;
2885 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2886 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2890 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2891 struct task_struct *next)
2893 struct preempt_notifier *notifier;
2894 struct hlist_node *node;
2896 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2897 notifier->ops->sched_out(notifier, next);
2900 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2902 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2907 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2908 struct task_struct *next)
2912 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2915 * prepare_task_switch - prepare to switch tasks
2916 * @rq: the runqueue preparing to switch
2917 * @prev: the current task that is being switched out
2918 * @next: the task we are going to switch to.
2920 * This is called with the rq lock held and interrupts off. It must
2921 * be paired with a subsequent finish_task_switch after the context
2924 * prepare_task_switch sets up locking and calls architecture specific
2928 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2929 struct task_struct *next)
2931 sched_info_switch(prev, next);
2932 perf_event_task_sched_out(prev, next);
2933 fire_sched_out_preempt_notifiers(prev, next);
2934 prepare_lock_switch(rq, next);
2935 prepare_arch_switch(next);
2936 trace_sched_switch(prev, next);
2940 * finish_task_switch - clean up after a task-switch
2941 * @rq: runqueue associated with task-switch
2942 * @prev: the thread we just switched away from.
2944 * finish_task_switch must be called after the context switch, paired
2945 * with a prepare_task_switch call before the context switch.
2946 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2947 * and do any other architecture-specific cleanup actions.
2949 * Note that we may have delayed dropping an mm in context_switch(). If
2950 * so, we finish that here outside of the runqueue lock. (Doing it
2951 * with the lock held can cause deadlocks; see schedule() for
2954 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2955 __releases(rq->lock)
2957 struct mm_struct *mm = rq->prev_mm;
2963 * A task struct has one reference for the use as "current".
2964 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2965 * schedule one last time. The schedule call will never return, and
2966 * the scheduled task must drop that reference.
2967 * The test for TASK_DEAD must occur while the runqueue locks are
2968 * still held, otherwise prev could be scheduled on another cpu, die
2969 * there before we look at prev->state, and then the reference would
2971 * Manfred Spraul <manfred@colorfullife.com>
2973 prev_state = prev->state;
2974 finish_arch_switch(prev);
2975 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2976 local_irq_disable();
2977 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2978 perf_event_task_sched_in(current);
2979 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2981 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2982 finish_lock_switch(rq, prev);
2984 fire_sched_in_preempt_notifiers(current);
2987 if (unlikely(prev_state == TASK_DEAD)) {
2989 * Remove function-return probe instances associated with this
2990 * task and put them back on the free list.
2992 kprobe_flush_task(prev);
2993 put_task_struct(prev);
2999 /* assumes rq->lock is held */
3000 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
3002 if (prev->sched_class->pre_schedule)
3003 prev->sched_class->pre_schedule(rq, prev);
3006 /* rq->lock is NOT held, but preemption is disabled */
3007 static inline void post_schedule(struct rq *rq)
3009 if (rq->post_schedule) {
3010 unsigned long flags;
3012 raw_spin_lock_irqsave(&rq->lock, flags);
3013 if (rq->curr->sched_class->post_schedule)
3014 rq->curr->sched_class->post_schedule(rq);
3015 raw_spin_unlock_irqrestore(&rq->lock, flags);
3017 rq->post_schedule = 0;
3023 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
3027 static inline void post_schedule(struct rq *rq)
3034 * schedule_tail - first thing a freshly forked thread must call.
3035 * @prev: the thread we just switched away from.
3037 asmlinkage void schedule_tail(struct task_struct *prev)
3038 __releases(rq->lock)
3040 struct rq *rq = this_rq();
3042 finish_task_switch(rq, prev);
3045 * FIXME: do we need to worry about rq being invalidated by the
3050 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
3051 /* In this case, finish_task_switch does not reenable preemption */
3054 if (current->set_child_tid)
3055 put_user(task_pid_vnr(current), current->set_child_tid);
3059 * context_switch - switch to the new MM and the new
3060 * thread's register state.
3063 context_switch(struct rq *rq, struct task_struct *prev,
3064 struct task_struct *next)
3066 struct mm_struct *mm, *oldmm;
3068 prepare_task_switch(rq, prev, next);
3071 oldmm = prev->active_mm;
3073 * For paravirt, this is coupled with an exit in switch_to to
3074 * combine the page table reload and the switch backend into
3077 arch_start_context_switch(prev);
3080 next->active_mm = oldmm;
3081 atomic_inc(&oldmm->mm_count);
3082 enter_lazy_tlb(oldmm, next);
3084 switch_mm(oldmm, mm, next);
3087 prev->active_mm = NULL;
3088 rq->prev_mm = oldmm;
3091 * Since the runqueue lock will be released by the next
3092 * task (which is an invalid locking op but in the case
3093 * of the scheduler it's an obvious special-case), so we
3094 * do an early lockdep release here:
3096 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
3097 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3100 /* Here we just switch the register state and the stack. */
3101 switch_to(prev, next, prev);
3105 * this_rq must be evaluated again because prev may have moved
3106 * CPUs since it called schedule(), thus the 'rq' on its stack
3107 * frame will be invalid.
3109 finish_task_switch(this_rq(), prev);
3113 * nr_running, nr_uninterruptible and nr_context_switches:
3115 * externally visible scheduler statistics: current number of runnable
3116 * threads, current number of uninterruptible-sleeping threads, total
3117 * number of context switches performed since bootup.
3119 unsigned long nr_running(void)
3121 unsigned long i, sum = 0;
3123 for_each_online_cpu(i)
3124 sum += cpu_rq(i)->nr_running;
3129 unsigned long nr_uninterruptible(void)
3131 unsigned long i, sum = 0;
3133 for_each_possible_cpu(i)
3134 sum += cpu_rq(i)->nr_uninterruptible;
3137 * Since we read the counters lockless, it might be slightly
3138 * inaccurate. Do not allow it to go below zero though:
3140 if (unlikely((long)sum < 0))
3146 unsigned long long nr_context_switches(void)
3149 unsigned long long sum = 0;
3151 for_each_possible_cpu(i)
3152 sum += cpu_rq(i)->nr_switches;
3157 unsigned long nr_iowait(void)
3159 unsigned long i, sum = 0;
3161 for_each_possible_cpu(i)
3162 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3167 unsigned long nr_iowait_cpu(int cpu)
3169 struct rq *this = cpu_rq(cpu);
3170 return atomic_read(&this->nr_iowait);
3173 unsigned long this_cpu_load(void)
3175 struct rq *this = this_rq();
3176 return this->cpu_load[0];
3180 /* Variables and functions for calc_load */
3181 static atomic_long_t calc_load_tasks;
3182 static unsigned long calc_load_update;
3183 unsigned long avenrun[3];
3184 EXPORT_SYMBOL(avenrun);
3186 static long calc_load_fold_active(struct rq *this_rq)
3188 long nr_active, delta = 0;
3190 nr_active = this_rq->nr_running;
3191 nr_active += (long) this_rq->nr_uninterruptible;
3193 if (nr_active != this_rq->calc_load_active) {
3194 delta = nr_active - this_rq->calc_load_active;
3195 this_rq->calc_load_active = nr_active;
3201 static unsigned long
3202 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3205 load += active * (FIXED_1 - exp);
3206 load += 1UL << (FSHIFT - 1);
3207 return load >> FSHIFT;
3212 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
3214 * When making the ILB scale, we should try to pull this in as well.
3216 static atomic_long_t calc_load_tasks_idle;
3218 static void calc_load_account_idle(struct rq *this_rq)
3222 delta = calc_load_fold_active(this_rq);
3224 atomic_long_add(delta, &calc_load_tasks_idle);
3227 static long calc_load_fold_idle(void)
3232 * Its got a race, we don't care...
3234 if (atomic_long_read(&calc_load_tasks_idle))
3235 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
3241 * fixed_power_int - compute: x^n, in O(log n) time
3243 * @x: base of the power
3244 * @frac_bits: fractional bits of @x
3245 * @n: power to raise @x to.
3247 * By exploiting the relation between the definition of the natural power
3248 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
3249 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
3250 * (where: n_i \elem {0, 1}, the binary vector representing n),
3251 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
3252 * of course trivially computable in O(log_2 n), the length of our binary
3255 static unsigned long
3256 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
3258 unsigned long result = 1UL << frac_bits;
3263 result += 1UL << (frac_bits - 1);
3264 result >>= frac_bits;
3270 x += 1UL << (frac_bits - 1);
3278 * a1 = a0 * e + a * (1 - e)
3280 * a2 = a1 * e + a * (1 - e)
3281 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
3282 * = a0 * e^2 + a * (1 - e) * (1 + e)
3284 * a3 = a2 * e + a * (1 - e)
3285 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
3286 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
3290 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
3291 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
3292 * = a0 * e^n + a * (1 - e^n)
3294 * [1] application of the geometric series:
3297 * S_n := \Sum x^i = -------------
3300 static unsigned long
3301 calc_load_n(unsigned long load, unsigned long exp,
3302 unsigned long active, unsigned int n)
3305 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
3309 * NO_HZ can leave us missing all per-cpu ticks calling
3310 * calc_load_account_active(), but since an idle CPU folds its delta into
3311 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
3312 * in the pending idle delta if our idle period crossed a load cycle boundary.
3314 * Once we've updated the global active value, we need to apply the exponential
3315 * weights adjusted to the number of cycles missed.
3317 static void calc_global_nohz(unsigned long ticks)
3319 long delta, active, n;
3321 if (time_before(jiffies, calc_load_update))
3325 * If we crossed a calc_load_update boundary, make sure to fold
3326 * any pending idle changes, the respective CPUs might have
3327 * missed the tick driven calc_load_account_active() update
3330 delta = calc_load_fold_idle();
3332 atomic_long_add(delta, &calc_load_tasks);
3335 * If we were idle for multiple load cycles, apply them.
3337 if (ticks >= LOAD_FREQ) {
3338 n = ticks / LOAD_FREQ;
3340 active = atomic_long_read(&calc_load_tasks);
3341 active = active > 0 ? active * FIXED_1 : 0;
3343 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
3344 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
3345 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
3347 calc_load_update += n * LOAD_FREQ;
3351 * Its possible the remainder of the above division also crosses
3352 * a LOAD_FREQ period, the regular check in calc_global_load()
3353 * which comes after this will take care of that.
3355 * Consider us being 11 ticks before a cycle completion, and us
3356 * sleeping for 4*LOAD_FREQ + 22 ticks, then the above code will
3357 * age us 4 cycles, and the test in calc_global_load() will
3358 * pick up the final one.
3362 static void calc_load_account_idle(struct rq *this_rq)
3366 static inline long calc_load_fold_idle(void)
3371 static void calc_global_nohz(unsigned long ticks)
3377 * get_avenrun - get the load average array
3378 * @loads: pointer to dest load array
3379 * @offset: offset to add
3380 * @shift: shift count to shift the result left
3382 * These values are estimates at best, so no need for locking.
3384 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3386 loads[0] = (avenrun[0] + offset) << shift;
3387 loads[1] = (avenrun[1] + offset) << shift;
3388 loads[2] = (avenrun[2] + offset) << shift;
3392 * calc_load - update the avenrun load estimates 10 ticks after the
3393 * CPUs have updated calc_load_tasks.
3395 void calc_global_load(unsigned long ticks)
3399 calc_global_nohz(ticks);
3401 if (time_before(jiffies, calc_load_update + 10))
3404 active = atomic_long_read(&calc_load_tasks);
3405 active = active > 0 ? active * FIXED_1 : 0;
3407 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3408 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3409 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3411 calc_load_update += LOAD_FREQ;
3415 * Called from update_cpu_load() to periodically update this CPU's
3418 static void calc_load_account_active(struct rq *this_rq)
3422 if (time_before(jiffies, this_rq->calc_load_update))
3425 delta = calc_load_fold_active(this_rq);
3426 delta += calc_load_fold_idle();
3428 atomic_long_add(delta, &calc_load_tasks);
3430 this_rq->calc_load_update += LOAD_FREQ;
3434 * The exact cpuload at various idx values, calculated at every tick would be
3435 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3437 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3438 * on nth tick when cpu may be busy, then we have:
3439 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3440 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3442 * decay_load_missed() below does efficient calculation of
3443 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3444 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3446 * The calculation is approximated on a 128 point scale.
3447 * degrade_zero_ticks is the number of ticks after which load at any
3448 * particular idx is approximated to be zero.
3449 * degrade_factor is a precomputed table, a row for each load idx.
3450 * Each column corresponds to degradation factor for a power of two ticks,
3451 * based on 128 point scale.
3453 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3454 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3456 * With this power of 2 load factors, we can degrade the load n times
3457 * by looking at 1 bits in n and doing as many mult/shift instead of
3458 * n mult/shifts needed by the exact degradation.
3460 #define DEGRADE_SHIFT 7
3461 static const unsigned char
3462 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3463 static const unsigned char
3464 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3465 {0, 0, 0, 0, 0, 0, 0, 0},
3466 {64, 32, 8, 0, 0, 0, 0, 0},
3467 {96, 72, 40, 12, 1, 0, 0},
3468 {112, 98, 75, 43, 15, 1, 0},
3469 {120, 112, 98, 76, 45, 16, 2} };
3472 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3473 * would be when CPU is idle and so we just decay the old load without
3474 * adding any new load.
3476 static unsigned long
3477 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3481 if (!missed_updates)
3484 if (missed_updates >= degrade_zero_ticks[idx])
3488 return load >> missed_updates;
3490 while (missed_updates) {
3491 if (missed_updates % 2)
3492 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3494 missed_updates >>= 1;
3501 * Update rq->cpu_load[] statistics. This function is usually called every
3502 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3503 * every tick. We fix it up based on jiffies.
3505 static void update_cpu_load(struct rq *this_rq)
3507 unsigned long this_load = this_rq->load.weight;
3508 unsigned long curr_jiffies = jiffies;
3509 unsigned long pending_updates;
3512 this_rq->nr_load_updates++;
3514 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3515 if (curr_jiffies == this_rq->last_load_update_tick)
3518 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3519 this_rq->last_load_update_tick = curr_jiffies;
3521 /* Update our load: */
3522 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3523 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3524 unsigned long old_load, new_load;
3526 /* scale is effectively 1 << i now, and >> i divides by scale */
3528 old_load = this_rq->cpu_load[i];
3529 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3530 new_load = this_load;
3532 * Round up the averaging division if load is increasing. This
3533 * prevents us from getting stuck on 9 if the load is 10, for
3536 if (new_load > old_load)
3537 new_load += scale - 1;
3539 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3542 sched_avg_update(this_rq);
3545 static void update_cpu_load_active(struct rq *this_rq)
3547 update_cpu_load(this_rq);
3549 calc_load_account_active(this_rq);
3555 * sched_exec - execve() is a valuable balancing opportunity, because at
3556 * this point the task has the smallest effective memory and cache footprint.
3558 void sched_exec(void)
3560 struct task_struct *p = current;
3561 unsigned long flags;
3564 raw_spin_lock_irqsave(&p->pi_lock, flags);
3565 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
3566 if (dest_cpu == smp_processor_id())
3569 if (likely(cpu_active(dest_cpu))) {
3570 struct migration_arg arg = { p, dest_cpu };
3572 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3573 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
3577 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3582 DEFINE_PER_CPU(struct kernel_stat, kstat);
3584 EXPORT_PER_CPU_SYMBOL(kstat);
3587 * Return any ns on the sched_clock that have not yet been accounted in
3588 * @p in case that task is currently running.
3590 * Called with task_rq_lock() held on @rq.
3592 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3596 if (task_current(rq, p)) {
3597 update_rq_clock(rq);
3598 ns = rq->clock_task - p->se.exec_start;
3606 unsigned long long task_delta_exec(struct task_struct *p)
3608 unsigned long flags;
3612 rq = task_rq_lock(p, &flags);
3613 ns = do_task_delta_exec(p, rq);
3614 task_rq_unlock(rq, p, &flags);
3620 * Return accounted runtime for the task.
3621 * In case the task is currently running, return the runtime plus current's
3622 * pending runtime that have not been accounted yet.
3624 unsigned long long task_sched_runtime(struct task_struct *p)
3626 unsigned long flags;
3630 rq = task_rq_lock(p, &flags);
3631 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3632 task_rq_unlock(rq, p, &flags);
3638 * Return sum_exec_runtime for the thread group.
3639 * In case the task is currently running, return the sum plus current's
3640 * pending runtime that have not been accounted yet.
3642 * Note that the thread group might have other running tasks as well,
3643 * so the return value not includes other pending runtime that other
3644 * running tasks might have.
3646 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3648 struct task_cputime totals;
3649 unsigned long flags;
3653 rq = task_rq_lock(p, &flags);
3654 thread_group_cputime(p, &totals);
3655 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3656 task_rq_unlock(rq, p, &flags);
3662 * Account user cpu time to a process.
3663 * @p: the process that the cpu time gets accounted to
3664 * @cputime: the cpu time spent in user space since the last update
3665 * @cputime_scaled: cputime scaled by cpu frequency
3667 void account_user_time(struct task_struct *p, cputime_t cputime,
3668 cputime_t cputime_scaled)
3670 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3673 /* Add user time to process. */
3674 p->utime = cputime_add(p->utime, cputime);
3675 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3676 account_group_user_time(p, cputime);
3678 /* Add user time to cpustat. */
3679 tmp = cputime_to_cputime64(cputime);
3680 if (TASK_NICE(p) > 0)
3681 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3683 cpustat->user = cputime64_add(cpustat->user, tmp);
3685 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3686 /* Account for user time used */
3687 acct_update_integrals(p);
3691 * Account guest cpu time to a process.
3692 * @p: the process that the cpu time gets accounted to
3693 * @cputime: the cpu time spent in virtual machine since the last update
3694 * @cputime_scaled: cputime scaled by cpu frequency
3696 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3697 cputime_t cputime_scaled)
3700 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3702 tmp = cputime_to_cputime64(cputime);
3704 /* Add guest time to process. */
3705 p->utime = cputime_add(p->utime, cputime);
3706 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3707 account_group_user_time(p, cputime);
3708 p->gtime = cputime_add(p->gtime, cputime);
3710 /* Add guest time to cpustat. */
3711 if (TASK_NICE(p) > 0) {
3712 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3713 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3715 cpustat->user = cputime64_add(cpustat->user, tmp);
3716 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3721 * Account system cpu time to a process and desired cpustat field
3722 * @p: the process that the cpu time gets accounted to
3723 * @cputime: the cpu time spent in kernel space since the last update
3724 * @cputime_scaled: cputime scaled by cpu frequency
3725 * @target_cputime64: pointer to cpustat field that has to be updated
3728 void __account_system_time(struct task_struct *p, cputime_t cputime,
3729 cputime_t cputime_scaled, cputime64_t *target_cputime64)
3731 cputime64_t tmp = cputime_to_cputime64(cputime);
3733 /* Add system time to process. */
3734 p->stime = cputime_add(p->stime, cputime);
3735 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3736 account_group_system_time(p, cputime);
3738 /* Add system time to cpustat. */
3739 *target_cputime64 = cputime64_add(*target_cputime64, tmp);
3740 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3742 /* Account for system time used */
3743 acct_update_integrals(p);
3747 * Account system cpu time to a process.
3748 * @p: the process that the cpu time gets accounted to
3749 * @hardirq_offset: the offset to subtract from hardirq_count()
3750 * @cputime: the cpu time spent in kernel space since the last update
3751 * @cputime_scaled: cputime scaled by cpu frequency
3753 void account_system_time(struct task_struct *p, int hardirq_offset,
3754 cputime_t cputime, cputime_t cputime_scaled)
3756 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3757 cputime64_t *target_cputime64;
3759 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3760 account_guest_time(p, cputime, cputime_scaled);
3764 if (hardirq_count() - hardirq_offset)
3765 target_cputime64 = &cpustat->irq;
3766 else if (in_serving_softirq())
3767 target_cputime64 = &cpustat->softirq;
3769 target_cputime64 = &cpustat->system;
3771 __account_system_time(p, cputime, cputime_scaled, target_cputime64);
3775 * Account for involuntary wait time.
3776 * @cputime: the cpu time spent in involuntary wait
3778 void account_steal_time(cputime_t cputime)
3780 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3781 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3783 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3787 * Account for idle time.
3788 * @cputime: the cpu time spent in idle wait
3790 void account_idle_time(cputime_t cputime)
3792 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3793 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3794 struct rq *rq = this_rq();
3796 if (atomic_read(&rq->nr_iowait) > 0)
3797 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3799 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3802 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3804 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
3806 * Account a tick to a process and cpustat
3807 * @p: the process that the cpu time gets accounted to
3808 * @user_tick: is the tick from userspace
3809 * @rq: the pointer to rq
3811 * Tick demultiplexing follows the order
3812 * - pending hardirq update
3813 * - pending softirq update
3817 * - check for guest_time
3818 * - else account as system_time
3820 * Check for hardirq is done both for system and user time as there is
3821 * no timer going off while we are on hardirq and hence we may never get an
3822 * opportunity to update it solely in system time.
3823 * p->stime and friends are only updated on system time and not on irq
3824 * softirq as those do not count in task exec_runtime any more.
3826 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
3829 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3830 cputime64_t tmp = cputime_to_cputime64(cputime_one_jiffy);
3831 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3833 if (irqtime_account_hi_update()) {
3834 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3835 } else if (irqtime_account_si_update()) {
3836 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3837 } else if (this_cpu_ksoftirqd() == p) {
3839 * ksoftirqd time do not get accounted in cpu_softirq_time.
3840 * So, we have to handle it separately here.
3841 * Also, p->stime needs to be updated for ksoftirqd.
3843 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
3845 } else if (user_tick) {
3846 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3847 } else if (p == rq->idle) {
3848 account_idle_time(cputime_one_jiffy);
3849 } else if (p->flags & PF_VCPU) { /* System time or guest time */
3850 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled);
3852 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
3857 static void irqtime_account_idle_ticks(int ticks)
3860 struct rq *rq = this_rq();
3862 for (i = 0; i < ticks; i++)
3863 irqtime_account_process_tick(current, 0, rq);
3865 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
3866 static void irqtime_account_idle_ticks(int ticks) {}
3867 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
3869 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
3872 * Account a single tick of cpu time.
3873 * @p: the process that the cpu time gets accounted to
3874 * @user_tick: indicates if the tick is a user or a system tick
3876 void account_process_tick(struct task_struct *p, int user_tick)
3878 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3879 struct rq *rq = this_rq();
3881 if (sched_clock_irqtime) {
3882 irqtime_account_process_tick(p, user_tick, rq);
3887 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3888 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3889 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3892 account_idle_time(cputime_one_jiffy);
3896 * Account multiple ticks of steal time.
3897 * @p: the process from which the cpu time has been stolen
3898 * @ticks: number of stolen ticks
3900 void account_steal_ticks(unsigned long ticks)
3902 account_steal_time(jiffies_to_cputime(ticks));
3906 * Account multiple ticks of idle time.
3907 * @ticks: number of stolen ticks
3909 void account_idle_ticks(unsigned long ticks)
3912 if (sched_clock_irqtime) {
3913 irqtime_account_idle_ticks(ticks);
3917 account_idle_time(jiffies_to_cputime(ticks));
3923 * Use precise platform statistics if available:
3925 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3926 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3932 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3934 struct task_cputime cputime;
3936 thread_group_cputime(p, &cputime);
3938 *ut = cputime.utime;
3939 *st = cputime.stime;
3943 #ifndef nsecs_to_cputime
3944 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3947 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3949 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
3952 * Use CFS's precise accounting:
3954 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3960 do_div(temp, total);
3961 utime = (cputime_t)temp;
3966 * Compare with previous values, to keep monotonicity:
3968 p->prev_utime = max(p->prev_utime, utime);
3969 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
3971 *ut = p->prev_utime;
3972 *st = p->prev_stime;
3976 * Must be called with siglock held.
3978 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3980 struct signal_struct *sig = p->signal;
3981 struct task_cputime cputime;
3982 cputime_t rtime, utime, total;
3984 thread_group_cputime(p, &cputime);
3986 total = cputime_add(cputime.utime, cputime.stime);
3987 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3992 temp *= cputime.utime;
3993 do_div(temp, total);
3994 utime = (cputime_t)temp;
3998 sig->prev_utime = max(sig->prev_utime, utime);
3999 sig->prev_stime = max(sig->prev_stime,
4000 cputime_sub(rtime, sig->prev_utime));
4002 *ut = sig->prev_utime;
4003 *st = sig->prev_stime;
4008 * This function gets called by the timer code, with HZ frequency.
4009 * We call it with interrupts disabled.
4011 void scheduler_tick(void)
4013 int cpu = smp_processor_id();
4014 struct rq *rq = cpu_rq(cpu);
4015 struct task_struct *curr = rq->curr;
4019 raw_spin_lock(&rq->lock);
4020 update_rq_clock(rq);
4021 update_cpu_load_active(rq);
4022 curr->sched_class->task_tick(rq, curr, 0);
4023 raw_spin_unlock(&rq->lock);
4025 perf_event_task_tick();
4028 rq->idle_at_tick = idle_cpu(cpu);
4029 trigger_load_balance(rq, cpu);
4033 notrace unsigned long get_parent_ip(unsigned long addr)
4035 if (in_lock_functions(addr)) {
4036 addr = CALLER_ADDR2;
4037 if (in_lock_functions(addr))
4038 addr = CALLER_ADDR3;
4043 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4044 defined(CONFIG_PREEMPT_TRACER))
4046 void __kprobes add_preempt_count(int val)
4048 #ifdef CONFIG_DEBUG_PREEMPT
4052 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4055 preempt_count() += val;
4056 #ifdef CONFIG_DEBUG_PREEMPT
4058 * Spinlock count overflowing soon?
4060 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4063 if (preempt_count() == val)
4064 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4066 EXPORT_SYMBOL(add_preempt_count);
4068 void __kprobes sub_preempt_count(int val)
4070 #ifdef CONFIG_DEBUG_PREEMPT
4074 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4077 * Is the spinlock portion underflowing?
4079 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4080 !(preempt_count() & PREEMPT_MASK)))
4084 if (preempt_count() == val)
4085 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4086 preempt_count() -= val;
4088 EXPORT_SYMBOL(sub_preempt_count);
4093 * Print scheduling while atomic bug:
4095 static noinline void __schedule_bug(struct task_struct *prev)
4097 struct pt_regs *regs = get_irq_regs();
4099 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4100 prev->comm, prev->pid, preempt_count());
4102 debug_show_held_locks(prev);
4104 if (irqs_disabled())
4105 print_irqtrace_events(prev);
4114 * Various schedule()-time debugging checks and statistics:
4116 static inline void schedule_debug(struct task_struct *prev)
4119 * Test if we are atomic. Since do_exit() needs to call into
4120 * schedule() atomically, we ignore that path for now.
4121 * Otherwise, whine if we are scheduling when we should not be.
4123 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4124 __schedule_bug(prev);
4126 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4128 schedstat_inc(this_rq(), sched_count);
4131 static void put_prev_task(struct rq *rq, struct task_struct *prev)
4133 if (prev->on_rq || rq->skip_clock_update < 0)
4134 update_rq_clock(rq);
4135 prev->sched_class->put_prev_task(rq, prev);
4139 * Pick up the highest-prio task:
4141 static inline struct task_struct *
4142 pick_next_task(struct rq *rq)
4144 const struct sched_class *class;
4145 struct task_struct *p;
4148 * Optimization: we know that if all tasks are in
4149 * the fair class we can call that function directly:
4151 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4152 p = fair_sched_class.pick_next_task(rq);
4157 for_each_class(class) {
4158 p = class->pick_next_task(rq);
4163 BUG(); /* the idle class will always have a runnable task */
4167 * schedule() is the main scheduler function.
4169 asmlinkage void __sched schedule(void)
4171 struct task_struct *prev, *next;
4172 unsigned long *switch_count;
4178 cpu = smp_processor_id();
4180 rcu_note_context_switch(cpu);
4183 schedule_debug(prev);
4185 if (sched_feat(HRTICK))
4188 raw_spin_lock_irq(&rq->lock);
4190 switch_count = &prev->nivcsw;
4191 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4192 if (unlikely(signal_pending_state(prev->state, prev))) {
4193 prev->state = TASK_RUNNING;
4195 deactivate_task(rq, prev, DEQUEUE_SLEEP);
4199 * If a worker went to sleep, notify and ask workqueue
4200 * whether it wants to wake up a task to maintain
4203 if (prev->flags & PF_WQ_WORKER) {
4204 struct task_struct *to_wakeup;
4206 to_wakeup = wq_worker_sleeping(prev, cpu);
4208 try_to_wake_up_local(to_wakeup);
4212 * If we are going to sleep and we have plugged IO
4213 * queued, make sure to submit it to avoid deadlocks.
4215 if (blk_needs_flush_plug(prev)) {
4216 raw_spin_unlock(&rq->lock);
4217 blk_schedule_flush_plug(prev);
4218 raw_spin_lock(&rq->lock);
4221 switch_count = &prev->nvcsw;
4224 pre_schedule(rq, prev);
4226 if (unlikely(!rq->nr_running))
4227 idle_balance(cpu, rq);
4229 put_prev_task(rq, prev);
4230 next = pick_next_task(rq);
4231 clear_tsk_need_resched(prev);
4232 rq->skip_clock_update = 0;
4234 if (likely(prev != next)) {
4239 context_switch(rq, prev, next); /* unlocks the rq */
4241 * The context switch have flipped the stack from under us
4242 * and restored the local variables which were saved when
4243 * this task called schedule() in the past. prev == current
4244 * is still correct, but it can be moved to another cpu/rq.
4246 cpu = smp_processor_id();
4249 raw_spin_unlock_irq(&rq->lock);
4253 preempt_enable_no_resched();
4257 EXPORT_SYMBOL(schedule);
4259 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
4261 static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
4266 if (lock->owner != owner)
4270 * Ensure we emit the owner->on_cpu, dereference _after_ checking
4271 * lock->owner still matches owner, if that fails, owner might
4272 * point to free()d memory, if it still matches, the rcu_read_lock()
4273 * ensures the memory stays valid.
4277 ret = owner->on_cpu;
4285 * Look out! "owner" is an entirely speculative pointer
4286 * access and not reliable.
4288 int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
4290 if (!sched_feat(OWNER_SPIN))
4293 while (owner_running(lock, owner)) {
4297 arch_mutex_cpu_relax();
4301 * If the owner changed to another task there is likely
4302 * heavy contention, stop spinning.
4311 #ifdef CONFIG_PREEMPT
4313 * this is the entry point to schedule() from in-kernel preemption
4314 * off of preempt_enable. Kernel preemptions off return from interrupt
4315 * occur there and call schedule directly.
4317 asmlinkage void __sched notrace preempt_schedule(void)
4319 struct thread_info *ti = current_thread_info();
4322 * If there is a non-zero preempt_count or interrupts are disabled,
4323 * we do not want to preempt the current task. Just return..
4325 if (likely(ti->preempt_count || irqs_disabled()))
4329 add_preempt_count_notrace(PREEMPT_ACTIVE);
4331 sub_preempt_count_notrace(PREEMPT_ACTIVE);
4334 * Check again in case we missed a preemption opportunity
4335 * between schedule and now.
4338 } while (need_resched());
4340 EXPORT_SYMBOL(preempt_schedule);
4343 * this is the entry point to schedule() from kernel preemption
4344 * off of irq context.
4345 * Note, that this is called and return with irqs disabled. This will
4346 * protect us against recursive calling from irq.
4348 asmlinkage void __sched preempt_schedule_irq(void)
4350 struct thread_info *ti = current_thread_info();
4352 /* Catch callers which need to be fixed */
4353 BUG_ON(ti->preempt_count || !irqs_disabled());
4356 add_preempt_count(PREEMPT_ACTIVE);
4359 local_irq_disable();
4360 sub_preempt_count(PREEMPT_ACTIVE);
4363 * Check again in case we missed a preemption opportunity
4364 * between schedule and now.
4367 } while (need_resched());
4370 #endif /* CONFIG_PREEMPT */
4372 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
4375 return try_to_wake_up(curr->private, mode, wake_flags);
4377 EXPORT_SYMBOL(default_wake_function);
4380 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4381 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4382 * number) then we wake all the non-exclusive tasks and one exclusive task.
4384 * There are circumstances in which we can try to wake a task which has already
4385 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4386 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4388 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4389 int nr_exclusive, int wake_flags, void *key)
4391 wait_queue_t *curr, *next;
4393 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4394 unsigned flags = curr->flags;
4396 if (curr->func(curr, mode, wake_flags, key) &&
4397 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4403 * __wake_up - wake up threads blocked on a waitqueue.
4405 * @mode: which threads
4406 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4407 * @key: is directly passed to the wakeup function
4409 * It may be assumed that this function implies a write memory barrier before
4410 * changing the task state if and only if any tasks are woken up.
4412 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4413 int nr_exclusive, void *key)
4415 unsigned long flags;
4417 spin_lock_irqsave(&q->lock, flags);
4418 __wake_up_common(q, mode, nr_exclusive, 0, key);
4419 spin_unlock_irqrestore(&q->lock, flags);
4421 EXPORT_SYMBOL(__wake_up);
4424 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4426 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4428 __wake_up_common(q, mode, 1, 0, NULL);
4430 EXPORT_SYMBOL_GPL(__wake_up_locked);
4432 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
4434 __wake_up_common(q, mode, 1, 0, key);
4436 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
4439 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4441 * @mode: which threads
4442 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4443 * @key: opaque value to be passed to wakeup targets
4445 * The sync wakeup differs that the waker knows that it will schedule
4446 * away soon, so while the target thread will be woken up, it will not
4447 * be migrated to another CPU - ie. the two threads are 'synchronized'
4448 * with each other. This can prevent needless bouncing between CPUs.
4450 * On UP it can prevent extra preemption.
4452 * It may be assumed that this function implies a write memory barrier before
4453 * changing the task state if and only if any tasks are woken up.
4455 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4456 int nr_exclusive, void *key)
4458 unsigned long flags;
4459 int wake_flags = WF_SYNC;
4464 if (unlikely(!nr_exclusive))
4467 spin_lock_irqsave(&q->lock, flags);
4468 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4469 spin_unlock_irqrestore(&q->lock, flags);
4471 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4474 * __wake_up_sync - see __wake_up_sync_key()
4476 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4478 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4480 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4483 * complete: - signals a single thread waiting on this completion
4484 * @x: holds the state of this particular completion
4486 * This will wake up a single thread waiting on this completion. Threads will be
4487 * awakened in the same order in which they were queued.
4489 * See also complete_all(), wait_for_completion() and related routines.
4491 * It may be assumed that this function implies a write memory barrier before
4492 * changing the task state if and only if any tasks are woken up.
4494 void complete(struct completion *x)
4496 unsigned long flags;
4498 spin_lock_irqsave(&x->wait.lock, flags);
4500 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4501 spin_unlock_irqrestore(&x->wait.lock, flags);
4503 EXPORT_SYMBOL(complete);
4506 * complete_all: - signals all threads waiting on this completion
4507 * @x: holds the state of this particular completion
4509 * This will wake up all threads waiting on this particular completion event.
4511 * It may be assumed that this function implies a write memory barrier before
4512 * changing the task state if and only if any tasks are woken up.
4514 void complete_all(struct completion *x)
4516 unsigned long flags;
4518 spin_lock_irqsave(&x->wait.lock, flags);
4519 x->done += UINT_MAX/2;
4520 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4521 spin_unlock_irqrestore(&x->wait.lock, flags);
4523 EXPORT_SYMBOL(complete_all);
4525 static inline long __sched
4526 do_wait_for_common(struct completion *x, long timeout, int state)
4529 DECLARE_WAITQUEUE(wait, current);
4531 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4533 if (signal_pending_state(state, current)) {
4534 timeout = -ERESTARTSYS;
4537 __set_current_state(state);
4538 spin_unlock_irq(&x->wait.lock);
4539 timeout = schedule_timeout(timeout);
4540 spin_lock_irq(&x->wait.lock);
4541 } while (!x->done && timeout);
4542 __remove_wait_queue(&x->wait, &wait);
4547 return timeout ?: 1;
4551 wait_for_common(struct completion *x, long timeout, int state)
4555 spin_lock_irq(&x->wait.lock);
4556 timeout = do_wait_for_common(x, timeout, state);
4557 spin_unlock_irq(&x->wait.lock);
4562 * wait_for_completion: - waits for completion of a task
4563 * @x: holds the state of this particular completion
4565 * This waits to be signaled for completion of a specific task. It is NOT
4566 * interruptible and there is no timeout.
4568 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4569 * and interrupt capability. Also see complete().
4571 void __sched wait_for_completion(struct completion *x)
4573 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4575 EXPORT_SYMBOL(wait_for_completion);
4578 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4579 * @x: holds the state of this particular completion
4580 * @timeout: timeout value in jiffies
4582 * This waits for either a completion of a specific task to be signaled or for a
4583 * specified timeout to expire. The timeout is in jiffies. It is not
4586 unsigned long __sched
4587 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4589 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4591 EXPORT_SYMBOL(wait_for_completion_timeout);
4594 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4595 * @x: holds the state of this particular completion
4597 * This waits for completion of a specific task to be signaled. It is
4600 int __sched wait_for_completion_interruptible(struct completion *x)
4602 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4603 if (t == -ERESTARTSYS)
4607 EXPORT_SYMBOL(wait_for_completion_interruptible);
4610 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4611 * @x: holds the state of this particular completion
4612 * @timeout: timeout value in jiffies
4614 * This waits for either a completion of a specific task to be signaled or for a
4615 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4618 wait_for_completion_interruptible_timeout(struct completion *x,
4619 unsigned long timeout)
4621 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4623 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4626 * wait_for_completion_killable: - waits for completion of a task (killable)
4627 * @x: holds the state of this particular completion
4629 * This waits to be signaled for completion of a specific task. It can be
4630 * interrupted by a kill signal.
4632 int __sched wait_for_completion_killable(struct completion *x)
4634 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4635 if (t == -ERESTARTSYS)
4639 EXPORT_SYMBOL(wait_for_completion_killable);
4642 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4643 * @x: holds the state of this particular completion
4644 * @timeout: timeout value in jiffies
4646 * This waits for either a completion of a specific task to be
4647 * signaled or for a specified timeout to expire. It can be
4648 * interrupted by a kill signal. The timeout is in jiffies.
4651 wait_for_completion_killable_timeout(struct completion *x,
4652 unsigned long timeout)
4654 return wait_for_common(x, timeout, TASK_KILLABLE);
4656 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
4659 * try_wait_for_completion - try to decrement a completion without blocking
4660 * @x: completion structure
4662 * Returns: 0 if a decrement cannot be done without blocking
4663 * 1 if a decrement succeeded.
4665 * If a completion is being used as a counting completion,
4666 * attempt to decrement the counter without blocking. This
4667 * enables us to avoid waiting if the resource the completion
4668 * is protecting is not available.
4670 bool try_wait_for_completion(struct completion *x)
4672 unsigned long flags;
4675 spin_lock_irqsave(&x->wait.lock, flags);
4680 spin_unlock_irqrestore(&x->wait.lock, flags);
4683 EXPORT_SYMBOL(try_wait_for_completion);
4686 * completion_done - Test to see if a completion has any waiters
4687 * @x: completion structure
4689 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4690 * 1 if there are no waiters.
4693 bool completion_done(struct completion *x)
4695 unsigned long flags;
4698 spin_lock_irqsave(&x->wait.lock, flags);
4701 spin_unlock_irqrestore(&x->wait.lock, flags);
4704 EXPORT_SYMBOL(completion_done);
4707 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4709 unsigned long flags;
4712 init_waitqueue_entry(&wait, current);
4714 __set_current_state(state);
4716 spin_lock_irqsave(&q->lock, flags);
4717 __add_wait_queue(q, &wait);
4718 spin_unlock(&q->lock);
4719 timeout = schedule_timeout(timeout);
4720 spin_lock_irq(&q->lock);
4721 __remove_wait_queue(q, &wait);
4722 spin_unlock_irqrestore(&q->lock, flags);
4727 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4729 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4731 EXPORT_SYMBOL(interruptible_sleep_on);
4734 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4736 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4738 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4740 void __sched sleep_on(wait_queue_head_t *q)
4742 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4744 EXPORT_SYMBOL(sleep_on);
4746 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4748 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4750 EXPORT_SYMBOL(sleep_on_timeout);
4752 #ifdef CONFIG_RT_MUTEXES
4755 * rt_mutex_setprio - set the current priority of a task
4757 * @prio: prio value (kernel-internal form)
4759 * This function changes the 'effective' priority of a task. It does
4760 * not touch ->normal_prio like __setscheduler().
4762 * Used by the rt_mutex code to implement priority inheritance logic.
4764 void rt_mutex_setprio(struct task_struct *p, int prio)
4766 int oldprio, on_rq, running;
4768 const struct sched_class *prev_class;
4770 BUG_ON(prio < 0 || prio > MAX_PRIO);
4772 rq = __task_rq_lock(p);
4774 trace_sched_pi_setprio(p, prio);
4776 prev_class = p->sched_class;
4778 running = task_current(rq, p);
4780 dequeue_task(rq, p, 0);
4782 p->sched_class->put_prev_task(rq, p);
4785 p->sched_class = &rt_sched_class;
4787 p->sched_class = &fair_sched_class;
4792 p->sched_class->set_curr_task(rq);
4794 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4796 check_class_changed(rq, p, prev_class, oldprio);
4797 __task_rq_unlock(rq);
4802 void set_user_nice(struct task_struct *p, long nice)
4804 int old_prio, delta, on_rq;
4805 unsigned long flags;
4808 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4811 * We have to be careful, if called from sys_setpriority(),
4812 * the task might be in the middle of scheduling on another CPU.
4814 rq = task_rq_lock(p, &flags);
4816 * The RT priorities are set via sched_setscheduler(), but we still
4817 * allow the 'normal' nice value to be set - but as expected
4818 * it wont have any effect on scheduling until the task is
4819 * SCHED_FIFO/SCHED_RR:
4821 if (task_has_rt_policy(p)) {
4822 p->static_prio = NICE_TO_PRIO(nice);
4827 dequeue_task(rq, p, 0);
4829 p->static_prio = NICE_TO_PRIO(nice);
4832 p->prio = effective_prio(p);
4833 delta = p->prio - old_prio;
4836 enqueue_task(rq, p, 0);
4838 * If the task increased its priority or is running and
4839 * lowered its priority, then reschedule its CPU:
4841 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4842 resched_task(rq->curr);
4845 task_rq_unlock(rq, p, &flags);
4847 EXPORT_SYMBOL(set_user_nice);
4850 * can_nice - check if a task can reduce its nice value
4854 int can_nice(const struct task_struct *p, const int nice)
4856 /* convert nice value [19,-20] to rlimit style value [1,40] */
4857 int nice_rlim = 20 - nice;
4859 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4860 capable(CAP_SYS_NICE));
4863 #ifdef __ARCH_WANT_SYS_NICE
4866 * sys_nice - change the priority of the current process.
4867 * @increment: priority increment
4869 * sys_setpriority is a more generic, but much slower function that
4870 * does similar things.
4872 SYSCALL_DEFINE1(nice, int, increment)
4877 * Setpriority might change our priority at the same moment.
4878 * We don't have to worry. Conceptually one call occurs first
4879 * and we have a single winner.
4881 if (increment < -40)
4886 nice = TASK_NICE(current) + increment;
4892 if (increment < 0 && !can_nice(current, nice))
4895 retval = security_task_setnice(current, nice);
4899 set_user_nice(current, nice);
4906 * task_prio - return the priority value of a given task.
4907 * @p: the task in question.
4909 * This is the priority value as seen by users in /proc.
4910 * RT tasks are offset by -200. Normal tasks are centered
4911 * around 0, value goes from -16 to +15.
4913 int task_prio(const struct task_struct *p)
4915 return p->prio - MAX_RT_PRIO;
4919 * task_nice - return the nice value of a given task.
4920 * @p: the task in question.
4922 int task_nice(const struct task_struct *p)
4924 return TASK_NICE(p);
4926 EXPORT_SYMBOL(task_nice);
4929 * idle_cpu - is a given cpu idle currently?
4930 * @cpu: the processor in question.
4932 int idle_cpu(int cpu)
4934 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4938 * idle_task - return the idle task for a given cpu.
4939 * @cpu: the processor in question.
4941 struct task_struct *idle_task(int cpu)
4943 return cpu_rq(cpu)->idle;
4947 * find_process_by_pid - find a process with a matching PID value.
4948 * @pid: the pid in question.
4950 static struct task_struct *find_process_by_pid(pid_t pid)
4952 return pid ? find_task_by_vpid(pid) : current;
4955 /* Actually do priority change: must hold rq lock. */
4957 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4960 p->rt_priority = prio;
4961 p->normal_prio = normal_prio(p);
4962 /* we are holding p->pi_lock already */
4963 p->prio = rt_mutex_getprio(p);
4964 if (rt_prio(p->prio))
4965 p->sched_class = &rt_sched_class;
4967 p->sched_class = &fair_sched_class;
4972 * check the target process has a UID that matches the current process's
4974 static bool check_same_owner(struct task_struct *p)
4976 const struct cred *cred = current_cred(), *pcred;
4980 pcred = __task_cred(p);
4981 if (cred->user->user_ns == pcred->user->user_ns)
4982 match = (cred->euid == pcred->euid ||
4983 cred->euid == pcred->uid);
4990 static int __sched_setscheduler(struct task_struct *p, int policy,
4991 const struct sched_param *param, bool user)
4993 int retval, oldprio, oldpolicy = -1, on_rq, running;
4994 unsigned long flags;
4995 const struct sched_class *prev_class;
4999 /* may grab non-irq protected spin_locks */
5000 BUG_ON(in_interrupt());
5002 /* double check policy once rq lock held */
5004 reset_on_fork = p->sched_reset_on_fork;
5005 policy = oldpolicy = p->policy;
5007 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
5008 policy &= ~SCHED_RESET_ON_FORK;
5010 if (policy != SCHED_FIFO && policy != SCHED_RR &&
5011 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5012 policy != SCHED_IDLE)
5017 * Valid priorities for SCHED_FIFO and SCHED_RR are
5018 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5019 * SCHED_BATCH and SCHED_IDLE is 0.
5021 if (param->sched_priority < 0 ||
5022 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5023 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5025 if (rt_policy(policy) != (param->sched_priority != 0))
5029 * Allow unprivileged RT tasks to decrease priority:
5031 if (user && !capable(CAP_SYS_NICE)) {
5032 if (rt_policy(policy)) {
5033 unsigned long rlim_rtprio =
5034 task_rlimit(p, RLIMIT_RTPRIO);
5036 /* can't set/change the rt policy */
5037 if (policy != p->policy && !rlim_rtprio)
5040 /* can't increase priority */
5041 if (param->sched_priority > p->rt_priority &&
5042 param->sched_priority > rlim_rtprio)
5047 * Treat SCHED_IDLE as nice 20. Only allow a switch to
5048 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
5050 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
5051 if (!can_nice(p, TASK_NICE(p)))
5055 /* can't change other user's priorities */
5056 if (!check_same_owner(p))
5059 /* Normal users shall not reset the sched_reset_on_fork flag */
5060 if (p->sched_reset_on_fork && !reset_on_fork)
5065 retval = security_task_setscheduler(p);
5071 * make sure no PI-waiters arrive (or leave) while we are
5072 * changing the priority of the task:
5074 * To be able to change p->policy safely, the appropriate
5075 * runqueue lock must be held.
5077 rq = task_rq_lock(p, &flags);
5080 * Changing the policy of the stop threads its a very bad idea
5082 if (p == rq->stop) {
5083 task_rq_unlock(rq, p, &flags);
5088 * If not changing anything there's no need to proceed further:
5090 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
5091 param->sched_priority == p->rt_priority))) {
5093 __task_rq_unlock(rq);
5094 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5098 #ifdef CONFIG_RT_GROUP_SCHED
5101 * Do not allow realtime tasks into groups that have no runtime
5104 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5105 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
5106 !task_group_is_autogroup(task_group(p))) {
5107 task_rq_unlock(rq, p, &flags);
5113 /* recheck policy now with rq lock held */
5114 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5115 policy = oldpolicy = -1;
5116 task_rq_unlock(rq, p, &flags);
5120 running = task_current(rq, p);
5122 deactivate_task(rq, p, 0);
5124 p->sched_class->put_prev_task(rq, p);
5126 p->sched_reset_on_fork = reset_on_fork;
5129 prev_class = p->sched_class;
5130 __setscheduler(rq, p, policy, param->sched_priority);
5133 p->sched_class->set_curr_task(rq);
5135 activate_task(rq, p, 0);
5137 check_class_changed(rq, p, prev_class, oldprio);
5138 task_rq_unlock(rq, p, &flags);
5140 rt_mutex_adjust_pi(p);
5146 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5147 * @p: the task in question.
5148 * @policy: new policy.
5149 * @param: structure containing the new RT priority.
5151 * NOTE that the task may be already dead.
5153 int sched_setscheduler(struct task_struct *p, int policy,
5154 const struct sched_param *param)
5156 return __sched_setscheduler(p, policy, param, true);
5158 EXPORT_SYMBOL_GPL(sched_setscheduler);
5161 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5162 * @p: the task in question.
5163 * @policy: new policy.
5164 * @param: structure containing the new RT priority.
5166 * Just like sched_setscheduler, only don't bother checking if the
5167 * current context has permission. For example, this is needed in
5168 * stop_machine(): we create temporary high priority worker threads,
5169 * but our caller might not have that capability.
5171 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5172 const struct sched_param *param)
5174 return __sched_setscheduler(p, policy, param, false);
5178 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5180 struct sched_param lparam;
5181 struct task_struct *p;
5184 if (!param || pid < 0)
5186 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5191 p = find_process_by_pid(pid);
5193 retval = sched_setscheduler(p, policy, &lparam);
5200 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5201 * @pid: the pid in question.
5202 * @policy: new policy.
5203 * @param: structure containing the new RT priority.
5205 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
5206 struct sched_param __user *, param)
5208 /* negative values for policy are not valid */
5212 return do_sched_setscheduler(pid, policy, param);
5216 * sys_sched_setparam - set/change the RT priority of a thread
5217 * @pid: the pid in question.
5218 * @param: structure containing the new RT priority.
5220 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5222 return do_sched_setscheduler(pid, -1, param);
5226 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5227 * @pid: the pid in question.
5229 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5231 struct task_struct *p;
5239 p = find_process_by_pid(pid);
5241 retval = security_task_getscheduler(p);
5244 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
5251 * sys_sched_getparam - get the RT priority of a thread
5252 * @pid: the pid in question.
5253 * @param: structure containing the RT priority.
5255 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5257 struct sched_param lp;
5258 struct task_struct *p;
5261 if (!param || pid < 0)
5265 p = find_process_by_pid(pid);
5270 retval = security_task_getscheduler(p);
5274 lp.sched_priority = p->rt_priority;
5278 * This one might sleep, we cannot do it with a spinlock held ...
5280 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5289 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5291 cpumask_var_t cpus_allowed, new_mask;
5292 struct task_struct *p;
5298 p = find_process_by_pid(pid);
5305 /* Prevent p going away */
5309 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5313 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5315 goto out_free_cpus_allowed;
5318 if (!check_same_owner(p) && !task_ns_capable(p, CAP_SYS_NICE))
5321 retval = security_task_setscheduler(p);
5325 cpuset_cpus_allowed(p, cpus_allowed);
5326 cpumask_and(new_mask, in_mask, cpus_allowed);
5328 retval = set_cpus_allowed_ptr(p, new_mask);
5331 cpuset_cpus_allowed(p, cpus_allowed);
5332 if (!cpumask_subset(new_mask, cpus_allowed)) {
5334 * We must have raced with a concurrent cpuset
5335 * update. Just reset the cpus_allowed to the
5336 * cpuset's cpus_allowed
5338 cpumask_copy(new_mask, cpus_allowed);
5343 free_cpumask_var(new_mask);
5344 out_free_cpus_allowed:
5345 free_cpumask_var(cpus_allowed);
5352 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5353 struct cpumask *new_mask)
5355 if (len < cpumask_size())
5356 cpumask_clear(new_mask);
5357 else if (len > cpumask_size())
5358 len = cpumask_size();
5360 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5364 * sys_sched_setaffinity - set the cpu affinity of a process
5365 * @pid: pid of the process
5366 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5367 * @user_mask_ptr: user-space pointer to the new cpu mask
5369 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5370 unsigned long __user *, user_mask_ptr)
5372 cpumask_var_t new_mask;
5375 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5378 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5380 retval = sched_setaffinity(pid, new_mask);
5381 free_cpumask_var(new_mask);
5385 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5387 struct task_struct *p;
5388 unsigned long flags;
5395 p = find_process_by_pid(pid);
5399 retval = security_task_getscheduler(p);
5403 raw_spin_lock_irqsave(&p->pi_lock, flags);
5404 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5405 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5415 * sys_sched_getaffinity - get the cpu affinity of a process
5416 * @pid: pid of the process
5417 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5418 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5420 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5421 unsigned long __user *, user_mask_ptr)
5426 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5428 if (len & (sizeof(unsigned long)-1))
5431 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5434 ret = sched_getaffinity(pid, mask);
5436 size_t retlen = min_t(size_t, len, cpumask_size());
5438 if (copy_to_user(user_mask_ptr, mask, retlen))
5443 free_cpumask_var(mask);
5449 * sys_sched_yield - yield the current processor to other threads.
5451 * This function yields the current CPU to other tasks. If there are no
5452 * other threads running on this CPU then this function will return.
5454 SYSCALL_DEFINE0(sched_yield)
5456 struct rq *rq = this_rq_lock();
5458 schedstat_inc(rq, yld_count);
5459 current->sched_class->yield_task(rq);
5462 * Since we are going to call schedule() anyway, there's
5463 * no need to preempt or enable interrupts:
5465 __release(rq->lock);
5466 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5467 do_raw_spin_unlock(&rq->lock);
5468 preempt_enable_no_resched();
5475 static inline int should_resched(void)
5477 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5480 static void __cond_resched(void)
5482 add_preempt_count(PREEMPT_ACTIVE);
5484 sub_preempt_count(PREEMPT_ACTIVE);
5487 int __sched _cond_resched(void)
5489 if (should_resched()) {
5495 EXPORT_SYMBOL(_cond_resched);
5498 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5499 * call schedule, and on return reacquire the lock.
5501 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5502 * operations here to prevent schedule() from being called twice (once via
5503 * spin_unlock(), once by hand).
5505 int __cond_resched_lock(spinlock_t *lock)
5507 int resched = should_resched();
5510 lockdep_assert_held(lock);
5512 if (spin_needbreak(lock) || resched) {
5523 EXPORT_SYMBOL(__cond_resched_lock);
5525 int __sched __cond_resched_softirq(void)
5527 BUG_ON(!in_softirq());
5529 if (should_resched()) {
5537 EXPORT_SYMBOL(__cond_resched_softirq);
5540 * yield - yield the current processor to other threads.
5542 * This is a shortcut for kernel-space yielding - it marks the
5543 * thread runnable and calls sys_sched_yield().
5545 void __sched yield(void)
5547 set_current_state(TASK_RUNNING);
5550 EXPORT_SYMBOL(yield);
5553 * yield_to - yield the current processor to another thread in
5554 * your thread group, or accelerate that thread toward the
5555 * processor it's on.
5557 * @preempt: whether task preemption is allowed or not
5559 * It's the caller's job to ensure that the target task struct
5560 * can't go away on us before we can do any checks.
5562 * Returns true if we indeed boosted the target task.
5564 bool __sched yield_to(struct task_struct *p, bool preempt)
5566 struct task_struct *curr = current;
5567 struct rq *rq, *p_rq;
5568 unsigned long flags;
5571 local_irq_save(flags);
5576 double_rq_lock(rq, p_rq);
5577 while (task_rq(p) != p_rq) {
5578 double_rq_unlock(rq, p_rq);
5582 if (!curr->sched_class->yield_to_task)
5585 if (curr->sched_class != p->sched_class)
5588 if (task_running(p_rq, p) || p->state)
5591 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5593 schedstat_inc(rq, yld_count);
5595 * Make p's CPU reschedule; pick_next_entity takes care of
5598 if (preempt && rq != p_rq)
5599 resched_task(p_rq->curr);
5603 double_rq_unlock(rq, p_rq);
5604 local_irq_restore(flags);
5611 EXPORT_SYMBOL_GPL(yield_to);
5614 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5615 * that process accounting knows that this is a task in IO wait state.
5617 void __sched io_schedule(void)
5619 struct rq *rq = raw_rq();
5621 delayacct_blkio_start();
5622 atomic_inc(&rq->nr_iowait);
5623 blk_flush_plug(current);
5624 current->in_iowait = 1;
5626 current->in_iowait = 0;
5627 atomic_dec(&rq->nr_iowait);
5628 delayacct_blkio_end();
5630 EXPORT_SYMBOL(io_schedule);
5632 long __sched io_schedule_timeout(long timeout)
5634 struct rq *rq = raw_rq();
5637 delayacct_blkio_start();
5638 atomic_inc(&rq->nr_iowait);
5639 blk_flush_plug(current);
5640 current->in_iowait = 1;
5641 ret = schedule_timeout(timeout);
5642 current->in_iowait = 0;
5643 atomic_dec(&rq->nr_iowait);
5644 delayacct_blkio_end();
5649 * sys_sched_get_priority_max - return maximum RT priority.
5650 * @policy: scheduling class.
5652 * this syscall returns the maximum rt_priority that can be used
5653 * by a given scheduling class.
5655 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5662 ret = MAX_USER_RT_PRIO-1;
5674 * sys_sched_get_priority_min - return minimum RT priority.
5675 * @policy: scheduling class.
5677 * this syscall returns the minimum rt_priority that can be used
5678 * by a given scheduling class.
5680 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5698 * sys_sched_rr_get_interval - return the default timeslice of a process.
5699 * @pid: pid of the process.
5700 * @interval: userspace pointer to the timeslice value.
5702 * this syscall writes the default timeslice value of a given process
5703 * into the user-space timespec buffer. A value of '0' means infinity.
5705 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5706 struct timespec __user *, interval)
5708 struct task_struct *p;
5709 unsigned int time_slice;
5710 unsigned long flags;
5720 p = find_process_by_pid(pid);
5724 retval = security_task_getscheduler(p);
5728 rq = task_rq_lock(p, &flags);
5729 time_slice = p->sched_class->get_rr_interval(rq, p);
5730 task_rq_unlock(rq, p, &flags);
5733 jiffies_to_timespec(time_slice, &t);
5734 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5742 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5744 void sched_show_task(struct task_struct *p)
5746 unsigned long free = 0;
5749 state = p->state ? __ffs(p->state) + 1 : 0;
5750 printk(KERN_INFO "%-15.15s %c", p->comm,
5751 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5752 #if BITS_PER_LONG == 32
5753 if (state == TASK_RUNNING)
5754 printk(KERN_CONT " running ");
5756 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5758 if (state == TASK_RUNNING)
5759 printk(KERN_CONT " running task ");
5761 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5763 #ifdef CONFIG_DEBUG_STACK_USAGE
5764 free = stack_not_used(p);
5766 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5767 task_pid_nr(p), task_pid_nr(p->real_parent),
5768 (unsigned long)task_thread_info(p)->flags);
5770 show_stack(p, NULL);
5773 void show_state_filter(unsigned long state_filter)
5775 struct task_struct *g, *p;
5777 #if BITS_PER_LONG == 32
5779 " task PC stack pid father\n");
5782 " task PC stack pid father\n");
5784 read_lock(&tasklist_lock);
5785 do_each_thread(g, p) {
5787 * reset the NMI-timeout, listing all files on a slow
5788 * console might take a lot of time:
5790 touch_nmi_watchdog();
5791 if (!state_filter || (p->state & state_filter))
5793 } while_each_thread(g, p);
5795 touch_all_softlockup_watchdogs();
5797 #ifdef CONFIG_SCHED_DEBUG
5798 sysrq_sched_debug_show();
5800 read_unlock(&tasklist_lock);
5802 * Only show locks if all tasks are dumped:
5805 debug_show_all_locks();
5808 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5810 idle->sched_class = &idle_sched_class;
5814 * init_idle - set up an idle thread for a given CPU
5815 * @idle: task in question
5816 * @cpu: cpu the idle task belongs to
5818 * NOTE: this function does not set the idle thread's NEED_RESCHED
5819 * flag, to make booting more robust.
5821 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5823 struct rq *rq = cpu_rq(cpu);
5824 unsigned long flags;
5826 raw_spin_lock_irqsave(&rq->lock, flags);
5829 idle->state = TASK_RUNNING;
5830 idle->se.exec_start = sched_clock();
5832 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
5834 * We're having a chicken and egg problem, even though we are
5835 * holding rq->lock, the cpu isn't yet set to this cpu so the
5836 * lockdep check in task_group() will fail.
5838 * Similar case to sched_fork(). / Alternatively we could
5839 * use task_rq_lock() here and obtain the other rq->lock.
5844 __set_task_cpu(idle, cpu);
5847 rq->curr = rq->idle = idle;
5848 #if defined(CONFIG_SMP)
5851 raw_spin_unlock_irqrestore(&rq->lock, flags);
5853 /* Set the preempt count _outside_ the spinlocks! */
5854 task_thread_info(idle)->preempt_count = 0;
5857 * The idle tasks have their own, simple scheduling class:
5859 idle->sched_class = &idle_sched_class;
5860 ftrace_graph_init_idle_task(idle, cpu);
5864 * In a system that switches off the HZ timer nohz_cpu_mask
5865 * indicates which cpus entered this state. This is used
5866 * in the rcu update to wait only for active cpus. For system
5867 * which do not switch off the HZ timer nohz_cpu_mask should
5868 * always be CPU_BITS_NONE.
5870 cpumask_var_t nohz_cpu_mask;
5873 * Increase the granularity value when there are more CPUs,
5874 * because with more CPUs the 'effective latency' as visible
5875 * to users decreases. But the relationship is not linear,
5876 * so pick a second-best guess by going with the log2 of the
5879 * This idea comes from the SD scheduler of Con Kolivas:
5881 static int get_update_sysctl_factor(void)
5883 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5884 unsigned int factor;
5886 switch (sysctl_sched_tunable_scaling) {
5887 case SCHED_TUNABLESCALING_NONE:
5890 case SCHED_TUNABLESCALING_LINEAR:
5893 case SCHED_TUNABLESCALING_LOG:
5895 factor = 1 + ilog2(cpus);
5902 static void update_sysctl(void)
5904 unsigned int factor = get_update_sysctl_factor();
5906 #define SET_SYSCTL(name) \
5907 (sysctl_##name = (factor) * normalized_sysctl_##name)
5908 SET_SYSCTL(sched_min_granularity);
5909 SET_SYSCTL(sched_latency);
5910 SET_SYSCTL(sched_wakeup_granularity);
5914 static inline void sched_init_granularity(void)
5921 * This is how migration works:
5923 * 1) we invoke migration_cpu_stop() on the target CPU using
5925 * 2) stopper starts to run (implicitly forcing the migrated thread
5927 * 3) it checks whether the migrated task is still in the wrong runqueue.
5928 * 4) if it's in the wrong runqueue then the migration thread removes
5929 * it and puts it into the right queue.
5930 * 5) stopper completes and stop_one_cpu() returns and the migration
5935 * Change a given task's CPU affinity. Migrate the thread to a
5936 * proper CPU and schedule it away if the CPU it's executing on
5937 * is removed from the allowed bitmask.
5939 * NOTE: the caller must have a valid reference to the task, the
5940 * task must not exit() & deallocate itself prematurely. The
5941 * call is not atomic; no spinlocks may be held.
5943 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5945 unsigned long flags;
5947 unsigned int dest_cpu;
5950 rq = task_rq_lock(p, &flags);
5952 if (cpumask_equal(&p->cpus_allowed, new_mask))
5955 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5960 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
5965 if (p->sched_class->set_cpus_allowed)
5966 p->sched_class->set_cpus_allowed(p, new_mask);
5968 cpumask_copy(&p->cpus_allowed, new_mask);
5969 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5972 /* Can the task run on the task's current CPU? If so, we're done */
5973 if (cpumask_test_cpu(task_cpu(p), new_mask))
5976 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
5978 struct migration_arg arg = { p, dest_cpu };
5979 /* Need help from migration thread: drop lock and wait. */
5980 task_rq_unlock(rq, p, &flags);
5981 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
5982 tlb_migrate_finish(p->mm);
5986 task_rq_unlock(rq, p, &flags);
5990 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5993 * Move (not current) task off this cpu, onto dest cpu. We're doing
5994 * this because either it can't run here any more (set_cpus_allowed()
5995 * away from this CPU, or CPU going down), or because we're
5996 * attempting to rebalance this task on exec (sched_exec).
5998 * So we race with normal scheduler movements, but that's OK, as long
5999 * as the task is no longer on this CPU.
6001 * Returns non-zero if task was successfully migrated.
6003 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6005 struct rq *rq_dest, *rq_src;
6008 if (unlikely(!cpu_active(dest_cpu)))
6011 rq_src = cpu_rq(src_cpu);
6012 rq_dest = cpu_rq(dest_cpu);
6014 raw_spin_lock(&p->pi_lock);
6015 double_rq_lock(rq_src, rq_dest);
6016 /* Already moved. */
6017 if (task_cpu(p) != src_cpu)
6019 /* Affinity changed (again). */
6020 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6024 * If we're not on a rq, the next wake-up will ensure we're
6028 deactivate_task(rq_src, p, 0);
6029 set_task_cpu(p, dest_cpu);
6030 activate_task(rq_dest, p, 0);
6031 check_preempt_curr(rq_dest, p, 0);
6036 double_rq_unlock(rq_src, rq_dest);
6037 raw_spin_unlock(&p->pi_lock);
6042 * migration_cpu_stop - this will be executed by a highprio stopper thread
6043 * and performs thread migration by bumping thread off CPU then
6044 * 'pushing' onto another runqueue.
6046 static int migration_cpu_stop(void *data)
6048 struct migration_arg *arg = data;
6051 * The original target cpu might have gone down and we might
6052 * be on another cpu but it doesn't matter.
6054 local_irq_disable();
6055 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
6060 #ifdef CONFIG_HOTPLUG_CPU
6063 * Ensures that the idle task is using init_mm right before its cpu goes
6066 void idle_task_exit(void)
6068 struct mm_struct *mm = current->active_mm;
6070 BUG_ON(cpu_online(smp_processor_id()));
6073 switch_mm(mm, &init_mm, current);
6078 * While a dead CPU has no uninterruptible tasks queued at this point,
6079 * it might still have a nonzero ->nr_uninterruptible counter, because
6080 * for performance reasons the counter is not stricly tracking tasks to
6081 * their home CPUs. So we just add the counter to another CPU's counter,
6082 * to keep the global sum constant after CPU-down:
6084 static void migrate_nr_uninterruptible(struct rq *rq_src)
6086 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
6088 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6089 rq_src->nr_uninterruptible = 0;
6093 * remove the tasks which were accounted by rq from calc_load_tasks.
6095 static void calc_global_load_remove(struct rq *rq)
6097 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
6098 rq->calc_load_active = 0;
6102 * Migrate all tasks from the rq, sleeping tasks will be migrated by
6103 * try_to_wake_up()->select_task_rq().
6105 * Called with rq->lock held even though we'er in stop_machine() and
6106 * there's no concurrency possible, we hold the required locks anyway
6107 * because of lock validation efforts.
6109 static void migrate_tasks(unsigned int dead_cpu)
6111 struct rq *rq = cpu_rq(dead_cpu);
6112 struct task_struct *next, *stop = rq->stop;
6116 * Fudge the rq selection such that the below task selection loop
6117 * doesn't get stuck on the currently eligible stop task.
6119 * We're currently inside stop_machine() and the rq is either stuck
6120 * in the stop_machine_cpu_stop() loop, or we're executing this code,
6121 * either way we should never end up calling schedule() until we're
6128 * There's this thread running, bail when that's the only
6131 if (rq->nr_running == 1)
6134 next = pick_next_task(rq);
6136 next->sched_class->put_prev_task(rq, next);
6138 /* Find suitable destination for @next, with force if needed. */
6139 dest_cpu = select_fallback_rq(dead_cpu, next);
6140 raw_spin_unlock(&rq->lock);
6142 __migrate_task(next, dead_cpu, dest_cpu);
6144 raw_spin_lock(&rq->lock);
6150 #endif /* CONFIG_HOTPLUG_CPU */
6152 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6154 static struct ctl_table sd_ctl_dir[] = {
6156 .procname = "sched_domain",
6162 static struct ctl_table sd_ctl_root[] = {
6164 .procname = "kernel",
6166 .child = sd_ctl_dir,
6171 static struct ctl_table *sd_alloc_ctl_entry(int n)
6173 struct ctl_table *entry =
6174 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6179 static void sd_free_ctl_entry(struct ctl_table **tablep)
6181 struct ctl_table *entry;
6184 * In the intermediate directories, both the child directory and
6185 * procname are dynamically allocated and could fail but the mode
6186 * will always be set. In the lowest directory the names are
6187 * static strings and all have proc handlers.
6189 for (entry = *tablep; entry->mode; entry++) {
6191 sd_free_ctl_entry(&entry->child);
6192 if (entry->proc_handler == NULL)
6193 kfree(entry->procname);
6201 set_table_entry(struct ctl_table *entry,
6202 const char *procname, void *data, int maxlen,
6203 mode_t mode, proc_handler *proc_handler)
6205 entry->procname = procname;
6207 entry->maxlen = maxlen;
6209 entry->proc_handler = proc_handler;
6212 static struct ctl_table *
6213 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6215 struct ctl_table *table = sd_alloc_ctl_entry(13);
6220 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6221 sizeof(long), 0644, proc_doulongvec_minmax);
6222 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6223 sizeof(long), 0644, proc_doulongvec_minmax);
6224 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6225 sizeof(int), 0644, proc_dointvec_minmax);
6226 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6227 sizeof(int), 0644, proc_dointvec_minmax);
6228 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6229 sizeof(int), 0644, proc_dointvec_minmax);
6230 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6231 sizeof(int), 0644, proc_dointvec_minmax);
6232 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6233 sizeof(int), 0644, proc_dointvec_minmax);
6234 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6235 sizeof(int), 0644, proc_dointvec_minmax);
6236 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6237 sizeof(int), 0644, proc_dointvec_minmax);
6238 set_table_entry(&table[9], "cache_nice_tries",
6239 &sd->cache_nice_tries,
6240 sizeof(int), 0644, proc_dointvec_minmax);
6241 set_table_entry(&table[10], "flags", &sd->flags,
6242 sizeof(int), 0644, proc_dointvec_minmax);
6243 set_table_entry(&table[11], "name", sd->name,
6244 CORENAME_MAX_SIZE, 0444, proc_dostring);
6245 /* &table[12] is terminator */
6250 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6252 struct ctl_table *entry, *table;
6253 struct sched_domain *sd;
6254 int domain_num = 0, i;
6257 for_each_domain(cpu, sd)
6259 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6264 for_each_domain(cpu, sd) {
6265 snprintf(buf, 32, "domain%d", i);
6266 entry->procname = kstrdup(buf, GFP_KERNEL);
6268 entry->child = sd_alloc_ctl_domain_table(sd);
6275 static struct ctl_table_header *sd_sysctl_header;
6276 static void register_sched_domain_sysctl(void)
6278 int i, cpu_num = num_possible_cpus();
6279 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6282 WARN_ON(sd_ctl_dir[0].child);
6283 sd_ctl_dir[0].child = entry;
6288 for_each_possible_cpu(i) {
6289 snprintf(buf, 32, "cpu%d", i);
6290 entry->procname = kstrdup(buf, GFP_KERNEL);
6292 entry->child = sd_alloc_ctl_cpu_table(i);
6296 WARN_ON(sd_sysctl_header);
6297 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6300 /* may be called multiple times per register */
6301 static void unregister_sched_domain_sysctl(void)
6303 if (sd_sysctl_header)
6304 unregister_sysctl_table(sd_sysctl_header);
6305 sd_sysctl_header = NULL;
6306 if (sd_ctl_dir[0].child)
6307 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6310 static void register_sched_domain_sysctl(void)
6313 static void unregister_sched_domain_sysctl(void)
6318 static void set_rq_online(struct rq *rq)
6321 const struct sched_class *class;
6323 cpumask_set_cpu(rq->cpu, rq->rd->online);
6326 for_each_class(class) {
6327 if (class->rq_online)
6328 class->rq_online(rq);
6333 static void set_rq_offline(struct rq *rq)
6336 const struct sched_class *class;
6338 for_each_class(class) {
6339 if (class->rq_offline)
6340 class->rq_offline(rq);
6343 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6349 * migration_call - callback that gets triggered when a CPU is added.
6350 * Here we can start up the necessary migration thread for the new CPU.
6352 static int __cpuinit
6353 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6355 int cpu = (long)hcpu;
6356 unsigned long flags;
6357 struct rq *rq = cpu_rq(cpu);
6359 switch (action & ~CPU_TASKS_FROZEN) {
6361 case CPU_UP_PREPARE:
6362 rq->calc_load_update = calc_load_update;
6366 /* Update our root-domain */
6367 raw_spin_lock_irqsave(&rq->lock, flags);
6369 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6373 raw_spin_unlock_irqrestore(&rq->lock, flags);
6376 #ifdef CONFIG_HOTPLUG_CPU
6378 sched_ttwu_pending();
6379 /* Update our root-domain */
6380 raw_spin_lock_irqsave(&rq->lock, flags);
6382 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6386 BUG_ON(rq->nr_running != 1); /* the migration thread */
6387 raw_spin_unlock_irqrestore(&rq->lock, flags);
6389 migrate_nr_uninterruptible(rq);
6390 calc_global_load_remove(rq);
6395 update_max_interval();
6401 * Register at high priority so that task migration (migrate_all_tasks)
6402 * happens before everything else. This has to be lower priority than
6403 * the notifier in the perf_event subsystem, though.
6405 static struct notifier_block __cpuinitdata migration_notifier = {
6406 .notifier_call = migration_call,
6407 .priority = CPU_PRI_MIGRATION,
6410 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
6411 unsigned long action, void *hcpu)
6413 switch (action & ~CPU_TASKS_FROZEN) {
6415 case CPU_DOWN_FAILED:
6416 set_cpu_active((long)hcpu, true);
6423 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
6424 unsigned long action, void *hcpu)
6426 switch (action & ~CPU_TASKS_FROZEN) {
6427 case CPU_DOWN_PREPARE:
6428 set_cpu_active((long)hcpu, false);
6435 static int __init migration_init(void)
6437 void *cpu = (void *)(long)smp_processor_id();
6440 /* Initialize migration for the boot CPU */
6441 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6442 BUG_ON(err == NOTIFY_BAD);
6443 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6444 register_cpu_notifier(&migration_notifier);
6446 /* Register cpu active notifiers */
6447 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
6448 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
6452 early_initcall(migration_init);
6457 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
6459 #ifdef CONFIG_SCHED_DEBUG
6461 static __read_mostly int sched_domain_debug_enabled;
6463 static int __init sched_domain_debug_setup(char *str)
6465 sched_domain_debug_enabled = 1;
6469 early_param("sched_debug", sched_domain_debug_setup);
6471 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6472 struct cpumask *groupmask)
6474 struct sched_group *group = sd->groups;
6477 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6478 cpumask_clear(groupmask);
6480 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6482 if (!(sd->flags & SD_LOAD_BALANCE)) {
6483 printk("does not load-balance\n");
6485 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6490 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6492 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6493 printk(KERN_ERR "ERROR: domain->span does not contain "
6496 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6497 printk(KERN_ERR "ERROR: domain->groups does not contain"
6501 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6505 printk(KERN_ERR "ERROR: group is NULL\n");
6509 if (!group->cpu_power) {
6510 printk(KERN_CONT "\n");
6511 printk(KERN_ERR "ERROR: domain->cpu_power not "
6516 if (!cpumask_weight(sched_group_cpus(group))) {
6517 printk(KERN_CONT "\n");
6518 printk(KERN_ERR "ERROR: empty group\n");
6522 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6523 printk(KERN_CONT "\n");
6524 printk(KERN_ERR "ERROR: repeated CPUs\n");
6528 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6530 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6532 printk(KERN_CONT " %s", str);
6533 if (group->cpu_power != SCHED_POWER_SCALE) {
6534 printk(KERN_CONT " (cpu_power = %d)",
6538 group = group->next;
6539 } while (group != sd->groups);
6540 printk(KERN_CONT "\n");
6542 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6543 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6546 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6547 printk(KERN_ERR "ERROR: parent span is not a superset "
6548 "of domain->span\n");
6552 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6556 if (!sched_domain_debug_enabled)
6560 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6564 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6567 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
6575 #else /* !CONFIG_SCHED_DEBUG */
6576 # define sched_domain_debug(sd, cpu) do { } while (0)
6577 #endif /* CONFIG_SCHED_DEBUG */
6579 static int sd_degenerate(struct sched_domain *sd)
6581 if (cpumask_weight(sched_domain_span(sd)) == 1)
6584 /* Following flags need at least 2 groups */
6585 if (sd->flags & (SD_LOAD_BALANCE |
6586 SD_BALANCE_NEWIDLE |
6590 SD_SHARE_PKG_RESOURCES)) {
6591 if (sd->groups != sd->groups->next)
6595 /* Following flags don't use groups */
6596 if (sd->flags & (SD_WAKE_AFFINE))
6603 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6605 unsigned long cflags = sd->flags, pflags = parent->flags;
6607 if (sd_degenerate(parent))
6610 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6613 /* Flags needing groups don't count if only 1 group in parent */
6614 if (parent->groups == parent->groups->next) {
6615 pflags &= ~(SD_LOAD_BALANCE |
6616 SD_BALANCE_NEWIDLE |
6620 SD_SHARE_PKG_RESOURCES);
6621 if (nr_node_ids == 1)
6622 pflags &= ~SD_SERIALIZE;
6624 if (~cflags & pflags)
6630 static void free_rootdomain(struct rcu_head *rcu)
6632 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
6634 cpupri_cleanup(&rd->cpupri);
6635 free_cpumask_var(rd->rto_mask);
6636 free_cpumask_var(rd->online);
6637 free_cpumask_var(rd->span);
6641 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6643 struct root_domain *old_rd = NULL;
6644 unsigned long flags;
6646 raw_spin_lock_irqsave(&rq->lock, flags);
6651 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6654 cpumask_clear_cpu(rq->cpu, old_rd->span);
6657 * If we dont want to free the old_rt yet then
6658 * set old_rd to NULL to skip the freeing later
6661 if (!atomic_dec_and_test(&old_rd->refcount))
6665 atomic_inc(&rd->refcount);
6668 cpumask_set_cpu(rq->cpu, rd->span);
6669 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6672 raw_spin_unlock_irqrestore(&rq->lock, flags);
6675 call_rcu_sched(&old_rd->rcu, free_rootdomain);
6678 static int init_rootdomain(struct root_domain *rd)
6680 memset(rd, 0, sizeof(*rd));
6682 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
6684 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
6686 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6689 if (cpupri_init(&rd->cpupri) != 0)
6694 free_cpumask_var(rd->rto_mask);
6696 free_cpumask_var(rd->online);
6698 free_cpumask_var(rd->span);
6703 static void init_defrootdomain(void)
6705 init_rootdomain(&def_root_domain);
6707 atomic_set(&def_root_domain.refcount, 1);
6710 static struct root_domain *alloc_rootdomain(void)
6712 struct root_domain *rd;
6714 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6718 if (init_rootdomain(rd) != 0) {
6726 static void free_sched_domain(struct rcu_head *rcu)
6728 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
6729 if (atomic_dec_and_test(&sd->groups->ref))
6734 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
6736 call_rcu(&sd->rcu, free_sched_domain);
6739 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
6741 for (; sd; sd = sd->parent)
6742 destroy_sched_domain(sd, cpu);
6746 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6747 * hold the hotplug lock.
6750 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6752 struct rq *rq = cpu_rq(cpu);
6753 struct sched_domain *tmp;
6755 /* Remove the sched domains which do not contribute to scheduling. */
6756 for (tmp = sd; tmp; ) {
6757 struct sched_domain *parent = tmp->parent;
6761 if (sd_parent_degenerate(tmp, parent)) {
6762 tmp->parent = parent->parent;
6764 parent->parent->child = tmp;
6765 destroy_sched_domain(parent, cpu);
6770 if (sd && sd_degenerate(sd)) {
6773 destroy_sched_domain(tmp, cpu);
6778 sched_domain_debug(sd, cpu);
6780 rq_attach_root(rq, rd);
6782 rcu_assign_pointer(rq->sd, sd);
6783 destroy_sched_domains(tmp, cpu);
6786 /* cpus with isolated domains */
6787 static cpumask_var_t cpu_isolated_map;
6789 /* Setup the mask of cpus configured for isolated domains */
6790 static int __init isolated_cpu_setup(char *str)
6792 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6793 cpulist_parse(str, cpu_isolated_map);
6797 __setup("isolcpus=", isolated_cpu_setup);
6799 #define SD_NODES_PER_DOMAIN 16
6804 * find_next_best_node - find the next node to include in a sched_domain
6805 * @node: node whose sched_domain we're building
6806 * @used_nodes: nodes already in the sched_domain
6808 * Find the next node to include in a given scheduling domain. Simply
6809 * finds the closest node not already in the @used_nodes map.
6811 * Should use nodemask_t.
6813 static int find_next_best_node(int node, nodemask_t *used_nodes)
6815 int i, n, val, min_val, best_node = -1;
6819 for (i = 0; i < nr_node_ids; i++) {
6820 /* Start at @node */
6821 n = (node + i) % nr_node_ids;
6823 if (!nr_cpus_node(n))
6826 /* Skip already used nodes */
6827 if (node_isset(n, *used_nodes))
6830 /* Simple min distance search */
6831 val = node_distance(node, n);
6833 if (val < min_val) {
6839 if (best_node != -1)
6840 node_set(best_node, *used_nodes);
6845 * sched_domain_node_span - get a cpumask for a node's sched_domain
6846 * @node: node whose cpumask we're constructing
6847 * @span: resulting cpumask
6849 * Given a node, construct a good cpumask for its sched_domain to span. It
6850 * should be one that prevents unnecessary balancing, but also spreads tasks
6853 static void sched_domain_node_span(int node, struct cpumask *span)
6855 nodemask_t used_nodes;
6858 cpumask_clear(span);
6859 nodes_clear(used_nodes);
6861 cpumask_or(span, span, cpumask_of_node(node));
6862 node_set(node, used_nodes);
6864 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6865 int next_node = find_next_best_node(node, &used_nodes);
6868 cpumask_or(span, span, cpumask_of_node(next_node));
6872 static const struct cpumask *cpu_node_mask(int cpu)
6874 lockdep_assert_held(&sched_domains_mutex);
6876 sched_domain_node_span(cpu_to_node(cpu), sched_domains_tmpmask);
6878 return sched_domains_tmpmask;
6881 static const struct cpumask *cpu_allnodes_mask(int cpu)
6883 return cpu_possible_mask;
6885 #endif /* CONFIG_NUMA */
6887 static const struct cpumask *cpu_cpu_mask(int cpu)
6889 return cpumask_of_node(cpu_to_node(cpu));
6892 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6895 struct sched_domain **__percpu sd;
6896 struct sched_group **__percpu sg;
6900 struct sched_domain ** __percpu sd;
6901 struct root_domain *rd;
6911 struct sched_domain_topology_level;
6913 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
6914 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
6916 struct sched_domain_topology_level {
6917 sched_domain_init_f init;
6918 sched_domain_mask_f mask;
6919 struct sd_data data;
6923 * Assumes the sched_domain tree is fully constructed
6925 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6927 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6928 struct sched_domain *child = sd->child;
6931 cpu = cpumask_first(sched_domain_span(child));
6934 *sg = *per_cpu_ptr(sdd->sg, cpu);
6940 * build_sched_groups takes the cpumask we wish to span, and a pointer
6941 * to a function which identifies what group(along with sched group) a CPU
6942 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6943 * (due to the fact that we keep track of groups covered with a struct cpumask).
6945 * build_sched_groups will build a circular linked list of the groups
6946 * covered by the given span, and will set each group's ->cpumask correctly,
6947 * and ->cpu_power to 0.
6950 build_sched_groups(struct sched_domain *sd)
6952 struct sched_group *first = NULL, *last = NULL;
6953 struct sd_data *sdd = sd->private;
6954 const struct cpumask *span = sched_domain_span(sd);
6955 struct cpumask *covered;
6958 lockdep_assert_held(&sched_domains_mutex);
6959 covered = sched_domains_tmpmask;
6961 cpumask_clear(covered);
6963 for_each_cpu(i, span) {
6964 struct sched_group *sg;
6965 int group = get_group(i, sdd, &sg);
6968 if (cpumask_test_cpu(i, covered))
6971 cpumask_clear(sched_group_cpus(sg));
6974 for_each_cpu(j, span) {
6975 if (get_group(j, sdd, NULL) != group)
6978 cpumask_set_cpu(j, covered);
6979 cpumask_set_cpu(j, sched_group_cpus(sg));
6992 * Initialize sched groups cpu_power.
6994 * cpu_power indicates the capacity of sched group, which is used while
6995 * distributing the load between different sched groups in a sched domain.
6996 * Typically cpu_power for all the groups in a sched domain will be same unless
6997 * there are asymmetries in the topology. If there are asymmetries, group
6998 * having more cpu_power will pickup more load compared to the group having
7001 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7003 WARN_ON(!sd || !sd->groups);
7005 if (cpu != group_first_cpu(sd->groups))
7008 sd->groups->group_weight = cpumask_weight(sched_group_cpus(sd->groups));
7010 update_group_power(sd, cpu);
7014 * Initializers for schedule domains
7015 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7018 #ifdef CONFIG_SCHED_DEBUG
7019 # define SD_INIT_NAME(sd, type) sd->name = #type
7021 # define SD_INIT_NAME(sd, type) do { } while (0)
7024 #define SD_INIT_FUNC(type) \
7025 static noinline struct sched_domain * \
7026 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
7028 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
7029 *sd = SD_##type##_INIT; \
7030 SD_INIT_NAME(sd, type); \
7031 sd->private = &tl->data; \
7037 SD_INIT_FUNC(ALLNODES)
7040 #ifdef CONFIG_SCHED_SMT
7041 SD_INIT_FUNC(SIBLING)
7043 #ifdef CONFIG_SCHED_MC
7046 #ifdef CONFIG_SCHED_BOOK
7050 static int default_relax_domain_level = -1;
7051 int sched_domain_level_max;
7053 static int __init setup_relax_domain_level(char *str)
7057 val = simple_strtoul(str, NULL, 0);
7058 if (val < sched_domain_level_max)
7059 default_relax_domain_level = val;
7063 __setup("relax_domain_level=", setup_relax_domain_level);
7065 static void set_domain_attribute(struct sched_domain *sd,
7066 struct sched_domain_attr *attr)
7070 if (!attr || attr->relax_domain_level < 0) {
7071 if (default_relax_domain_level < 0)
7074 request = default_relax_domain_level;
7076 request = attr->relax_domain_level;
7077 if (request < sd->level) {
7078 /* turn off idle balance on this domain */
7079 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7081 /* turn on idle balance on this domain */
7082 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7086 static void __sdt_free(const struct cpumask *cpu_map);
7087 static int __sdt_alloc(const struct cpumask *cpu_map);
7089 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
7090 const struct cpumask *cpu_map)
7094 if (!atomic_read(&d->rd->refcount))
7095 free_rootdomain(&d->rd->rcu); /* fall through */
7097 free_percpu(d->sd); /* fall through */
7099 __sdt_free(cpu_map); /* fall through */
7105 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
7106 const struct cpumask *cpu_map)
7108 memset(d, 0, sizeof(*d));
7110 if (__sdt_alloc(cpu_map))
7111 return sa_sd_storage;
7112 d->sd = alloc_percpu(struct sched_domain *);
7114 return sa_sd_storage;
7115 d->rd = alloc_rootdomain();
7118 return sa_rootdomain;
7122 * NULL the sd_data elements we've used to build the sched_domain and
7123 * sched_group structure so that the subsequent __free_domain_allocs()
7124 * will not free the data we're using.
7126 static void claim_allocations(int cpu, struct sched_domain *sd)
7128 struct sd_data *sdd = sd->private;
7129 struct sched_group *sg = sd->groups;
7131 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
7132 *per_cpu_ptr(sdd->sd, cpu) = NULL;
7134 if (cpu == cpumask_first(sched_group_cpus(sg))) {
7135 WARN_ON_ONCE(*per_cpu_ptr(sdd->sg, cpu) != sg);
7136 *per_cpu_ptr(sdd->sg, cpu) = NULL;
7140 #ifdef CONFIG_SCHED_SMT
7141 static const struct cpumask *cpu_smt_mask(int cpu)
7143 return topology_thread_cpumask(cpu);
7148 * Topology list, bottom-up.
7150 static struct sched_domain_topology_level default_topology[] = {
7151 #ifdef CONFIG_SCHED_SMT
7152 { sd_init_SIBLING, cpu_smt_mask, },
7154 #ifdef CONFIG_SCHED_MC
7155 { sd_init_MC, cpu_coregroup_mask, },
7157 #ifdef CONFIG_SCHED_BOOK
7158 { sd_init_BOOK, cpu_book_mask, },
7160 { sd_init_CPU, cpu_cpu_mask, },
7162 { sd_init_NODE, cpu_node_mask, },
7163 { sd_init_ALLNODES, cpu_allnodes_mask, },
7168 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
7170 static int __sdt_alloc(const struct cpumask *cpu_map)
7172 struct sched_domain_topology_level *tl;
7175 for (tl = sched_domain_topology; tl->init; tl++) {
7176 struct sd_data *sdd = &tl->data;
7178 sdd->sd = alloc_percpu(struct sched_domain *);
7182 sdd->sg = alloc_percpu(struct sched_group *);
7186 for_each_cpu(j, cpu_map) {
7187 struct sched_domain *sd;
7188 struct sched_group *sg;
7190 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
7191 GFP_KERNEL, cpu_to_node(j));
7195 *per_cpu_ptr(sdd->sd, j) = sd;
7197 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7198 GFP_KERNEL, cpu_to_node(j));
7202 *per_cpu_ptr(sdd->sg, j) = sg;
7209 static void __sdt_free(const struct cpumask *cpu_map)
7211 struct sched_domain_topology_level *tl;
7214 for (tl = sched_domain_topology; tl->init; tl++) {
7215 struct sd_data *sdd = &tl->data;
7217 for_each_cpu(j, cpu_map) {
7218 kfree(*per_cpu_ptr(sdd->sd, j));
7219 kfree(*per_cpu_ptr(sdd->sg, j));
7221 free_percpu(sdd->sd);
7222 free_percpu(sdd->sg);
7226 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
7227 struct s_data *d, const struct cpumask *cpu_map,
7228 struct sched_domain_attr *attr, struct sched_domain *child,
7231 struct sched_domain *sd = tl->init(tl, cpu);
7235 set_domain_attribute(sd, attr);
7236 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
7238 sd->level = child->level + 1;
7239 sched_domain_level_max = max(sched_domain_level_max, sd->level);
7248 * Build sched domains for a given set of cpus and attach the sched domains
7249 * to the individual cpus
7251 static int build_sched_domains(const struct cpumask *cpu_map,
7252 struct sched_domain_attr *attr)
7254 enum s_alloc alloc_state = sa_none;
7255 struct sched_domain *sd;
7257 int i, ret = -ENOMEM;
7259 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7260 if (alloc_state != sa_rootdomain)
7263 /* Set up domains for cpus specified by the cpu_map. */
7264 for_each_cpu(i, cpu_map) {
7265 struct sched_domain_topology_level *tl;
7268 for (tl = sched_domain_topology; tl->init; tl++)
7269 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
7274 *per_cpu_ptr(d.sd, i) = sd;
7277 /* Build the groups for the domains */
7278 for_each_cpu(i, cpu_map) {
7279 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7280 sd->span_weight = cpumask_weight(sched_domain_span(sd));
7281 get_group(i, sd->private, &sd->groups);
7282 atomic_inc(&sd->groups->ref);
7284 if (i != cpumask_first(sched_domain_span(sd)))
7287 build_sched_groups(sd);
7291 /* Calculate CPU power for physical packages and nodes */
7292 for (i = nr_cpumask_bits-1; i >= 0; i--) {
7293 if (!cpumask_test_cpu(i, cpu_map))
7296 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7297 claim_allocations(i, sd);
7298 init_sched_groups_power(i, sd);
7302 /* Attach the domains */
7304 for_each_cpu(i, cpu_map) {
7305 sd = *per_cpu_ptr(d.sd, i);
7306 cpu_attach_domain(sd, d.rd, i);
7312 __free_domain_allocs(&d, alloc_state, cpu_map);
7316 static cpumask_var_t *doms_cur; /* current sched domains */
7317 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7318 static struct sched_domain_attr *dattr_cur;
7319 /* attribues of custom domains in 'doms_cur' */
7322 * Special case: If a kmalloc of a doms_cur partition (array of
7323 * cpumask) fails, then fallback to a single sched domain,
7324 * as determined by the single cpumask fallback_doms.
7326 static cpumask_var_t fallback_doms;
7329 * arch_update_cpu_topology lets virtualized architectures update the
7330 * cpu core maps. It is supposed to return 1 if the topology changed
7331 * or 0 if it stayed the same.
7333 int __attribute__((weak)) arch_update_cpu_topology(void)
7338 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7341 cpumask_var_t *doms;
7343 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7346 for (i = 0; i < ndoms; i++) {
7347 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7348 free_sched_domains(doms, i);
7355 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7358 for (i = 0; i < ndoms; i++)
7359 free_cpumask_var(doms[i]);
7364 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7365 * For now this just excludes isolated cpus, but could be used to
7366 * exclude other special cases in the future.
7368 static int init_sched_domains(const struct cpumask *cpu_map)
7372 arch_update_cpu_topology();
7374 doms_cur = alloc_sched_domains(ndoms_cur);
7376 doms_cur = &fallback_doms;
7377 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7379 err = build_sched_domains(doms_cur[0], NULL);
7380 register_sched_domain_sysctl();
7386 * Detach sched domains from a group of cpus specified in cpu_map
7387 * These cpus will now be attached to the NULL domain
7389 static void detach_destroy_domains(const struct cpumask *cpu_map)
7394 for_each_cpu(i, cpu_map)
7395 cpu_attach_domain(NULL, &def_root_domain, i);
7399 /* handle null as "default" */
7400 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7401 struct sched_domain_attr *new, int idx_new)
7403 struct sched_domain_attr tmp;
7410 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7411 new ? (new + idx_new) : &tmp,
7412 sizeof(struct sched_domain_attr));
7416 * Partition sched domains as specified by the 'ndoms_new'
7417 * cpumasks in the array doms_new[] of cpumasks. This compares
7418 * doms_new[] to the current sched domain partitioning, doms_cur[].
7419 * It destroys each deleted domain and builds each new domain.
7421 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7422 * The masks don't intersect (don't overlap.) We should setup one
7423 * sched domain for each mask. CPUs not in any of the cpumasks will
7424 * not be load balanced. If the same cpumask appears both in the
7425 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7428 * The passed in 'doms_new' should be allocated using
7429 * alloc_sched_domains. This routine takes ownership of it and will
7430 * free_sched_domains it when done with it. If the caller failed the
7431 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7432 * and partition_sched_domains() will fallback to the single partition
7433 * 'fallback_doms', it also forces the domains to be rebuilt.
7435 * If doms_new == NULL it will be replaced with cpu_online_mask.
7436 * ndoms_new == 0 is a special case for destroying existing domains,
7437 * and it will not create the default domain.
7439 * Call with hotplug lock held
7441 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7442 struct sched_domain_attr *dattr_new)
7447 mutex_lock(&sched_domains_mutex);
7449 /* always unregister in case we don't destroy any domains */
7450 unregister_sched_domain_sysctl();
7452 /* Let architecture update cpu core mappings. */
7453 new_topology = arch_update_cpu_topology();
7455 n = doms_new ? ndoms_new : 0;
7457 /* Destroy deleted domains */
7458 for (i = 0; i < ndoms_cur; i++) {
7459 for (j = 0; j < n && !new_topology; j++) {
7460 if (cpumask_equal(doms_cur[i], doms_new[j])
7461 && dattrs_equal(dattr_cur, i, dattr_new, j))
7464 /* no match - a current sched domain not in new doms_new[] */
7465 detach_destroy_domains(doms_cur[i]);
7470 if (doms_new == NULL) {
7472 doms_new = &fallback_doms;
7473 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7474 WARN_ON_ONCE(dattr_new);
7477 /* Build new domains */
7478 for (i = 0; i < ndoms_new; i++) {
7479 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7480 if (cpumask_equal(doms_new[i], doms_cur[j])
7481 && dattrs_equal(dattr_new, i, dattr_cur, j))
7484 /* no match - add a new doms_new */
7485 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7490 /* Remember the new sched domains */
7491 if (doms_cur != &fallback_doms)
7492 free_sched_domains(doms_cur, ndoms_cur);
7493 kfree(dattr_cur); /* kfree(NULL) is safe */
7494 doms_cur = doms_new;
7495 dattr_cur = dattr_new;
7496 ndoms_cur = ndoms_new;
7498 register_sched_domain_sysctl();
7500 mutex_unlock(&sched_domains_mutex);
7503 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7504 static void reinit_sched_domains(void)
7508 /* Destroy domains first to force the rebuild */
7509 partition_sched_domains(0, NULL, NULL);
7511 rebuild_sched_domains();
7515 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7517 unsigned int level = 0;
7519 if (sscanf(buf, "%u", &level) != 1)
7523 * level is always be positive so don't check for
7524 * level < POWERSAVINGS_BALANCE_NONE which is 0
7525 * What happens on 0 or 1 byte write,
7526 * need to check for count as well?
7529 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7533 sched_smt_power_savings = level;
7535 sched_mc_power_savings = level;
7537 reinit_sched_domains();
7542 #ifdef CONFIG_SCHED_MC
7543 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7544 struct sysdev_class_attribute *attr,
7547 return sprintf(page, "%u\n", sched_mc_power_savings);
7549 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7550 struct sysdev_class_attribute *attr,
7551 const char *buf, size_t count)
7553 return sched_power_savings_store(buf, count, 0);
7555 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7556 sched_mc_power_savings_show,
7557 sched_mc_power_savings_store);
7560 #ifdef CONFIG_SCHED_SMT
7561 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7562 struct sysdev_class_attribute *attr,
7565 return sprintf(page, "%u\n", sched_smt_power_savings);
7567 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7568 struct sysdev_class_attribute *attr,
7569 const char *buf, size_t count)
7571 return sched_power_savings_store(buf, count, 1);
7573 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7574 sched_smt_power_savings_show,
7575 sched_smt_power_savings_store);
7578 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7582 #ifdef CONFIG_SCHED_SMT
7584 err = sysfs_create_file(&cls->kset.kobj,
7585 &attr_sched_smt_power_savings.attr);
7587 #ifdef CONFIG_SCHED_MC
7588 if (!err && mc_capable())
7589 err = sysfs_create_file(&cls->kset.kobj,
7590 &attr_sched_mc_power_savings.attr);
7594 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7597 * Update cpusets according to cpu_active mask. If cpusets are
7598 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7599 * around partition_sched_domains().
7601 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7604 switch (action & ~CPU_TASKS_FROZEN) {
7606 case CPU_DOWN_FAILED:
7607 cpuset_update_active_cpus();
7614 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7617 switch (action & ~CPU_TASKS_FROZEN) {
7618 case CPU_DOWN_PREPARE:
7619 cpuset_update_active_cpus();
7626 static int update_runtime(struct notifier_block *nfb,
7627 unsigned long action, void *hcpu)
7629 int cpu = (int)(long)hcpu;
7632 case CPU_DOWN_PREPARE:
7633 case CPU_DOWN_PREPARE_FROZEN:
7634 disable_runtime(cpu_rq(cpu));
7637 case CPU_DOWN_FAILED:
7638 case CPU_DOWN_FAILED_FROZEN:
7640 case CPU_ONLINE_FROZEN:
7641 enable_runtime(cpu_rq(cpu));
7649 void __init sched_init_smp(void)
7651 cpumask_var_t non_isolated_cpus;
7653 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7654 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7657 mutex_lock(&sched_domains_mutex);
7658 init_sched_domains(cpu_active_mask);
7659 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7660 if (cpumask_empty(non_isolated_cpus))
7661 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7662 mutex_unlock(&sched_domains_mutex);
7665 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7666 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7668 /* RT runtime code needs to handle some hotplug events */
7669 hotcpu_notifier(update_runtime, 0);
7673 /* Move init over to a non-isolated CPU */
7674 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7676 sched_init_granularity();
7677 free_cpumask_var(non_isolated_cpus);
7679 init_sched_rt_class();
7682 void __init sched_init_smp(void)
7684 sched_init_granularity();
7686 #endif /* CONFIG_SMP */
7688 const_debug unsigned int sysctl_timer_migration = 1;
7690 int in_sched_functions(unsigned long addr)
7692 return in_lock_functions(addr) ||
7693 (addr >= (unsigned long)__sched_text_start
7694 && addr < (unsigned long)__sched_text_end);
7697 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7699 cfs_rq->tasks_timeline = RB_ROOT;
7700 INIT_LIST_HEAD(&cfs_rq->tasks);
7701 #ifdef CONFIG_FAIR_GROUP_SCHED
7703 /* allow initial update_cfs_load() to truncate */
7705 cfs_rq->load_stamp = 1;
7708 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7711 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7713 struct rt_prio_array *array;
7716 array = &rt_rq->active;
7717 for (i = 0; i < MAX_RT_PRIO; i++) {
7718 INIT_LIST_HEAD(array->queue + i);
7719 __clear_bit(i, array->bitmap);
7721 /* delimiter for bitsearch: */
7722 __set_bit(MAX_RT_PRIO, array->bitmap);
7724 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7725 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7727 rt_rq->highest_prio.next = MAX_RT_PRIO;
7731 rt_rq->rt_nr_migratory = 0;
7732 rt_rq->overloaded = 0;
7733 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
7737 rt_rq->rt_throttled = 0;
7738 rt_rq->rt_runtime = 0;
7739 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7741 #ifdef CONFIG_RT_GROUP_SCHED
7742 rt_rq->rt_nr_boosted = 0;
7747 #ifdef CONFIG_FAIR_GROUP_SCHED
7748 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7749 struct sched_entity *se, int cpu,
7750 struct sched_entity *parent)
7752 struct rq *rq = cpu_rq(cpu);
7753 tg->cfs_rq[cpu] = cfs_rq;
7754 init_cfs_rq(cfs_rq, rq);
7758 /* se could be NULL for root_task_group */
7763 se->cfs_rq = &rq->cfs;
7765 se->cfs_rq = parent->my_q;
7768 update_load_set(&se->load, 0);
7769 se->parent = parent;
7773 #ifdef CONFIG_RT_GROUP_SCHED
7774 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7775 struct sched_rt_entity *rt_se, int cpu,
7776 struct sched_rt_entity *parent)
7778 struct rq *rq = cpu_rq(cpu);
7780 tg->rt_rq[cpu] = rt_rq;
7781 init_rt_rq(rt_rq, rq);
7783 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7785 tg->rt_se[cpu] = rt_se;
7790 rt_se->rt_rq = &rq->rt;
7792 rt_se->rt_rq = parent->my_q;
7794 rt_se->my_q = rt_rq;
7795 rt_se->parent = parent;
7796 INIT_LIST_HEAD(&rt_se->run_list);
7800 void __init sched_init(void)
7803 unsigned long alloc_size = 0, ptr;
7805 #ifdef CONFIG_FAIR_GROUP_SCHED
7806 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7808 #ifdef CONFIG_RT_GROUP_SCHED
7809 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7811 #ifdef CONFIG_CPUMASK_OFFSTACK
7812 alloc_size += num_possible_cpus() * cpumask_size();
7815 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7817 #ifdef CONFIG_FAIR_GROUP_SCHED
7818 root_task_group.se = (struct sched_entity **)ptr;
7819 ptr += nr_cpu_ids * sizeof(void **);
7821 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7822 ptr += nr_cpu_ids * sizeof(void **);
7824 #endif /* CONFIG_FAIR_GROUP_SCHED */
7825 #ifdef CONFIG_RT_GROUP_SCHED
7826 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7827 ptr += nr_cpu_ids * sizeof(void **);
7829 root_task_group.rt_rq = (struct rt_rq **)ptr;
7830 ptr += nr_cpu_ids * sizeof(void **);
7832 #endif /* CONFIG_RT_GROUP_SCHED */
7833 #ifdef CONFIG_CPUMASK_OFFSTACK
7834 for_each_possible_cpu(i) {
7835 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
7836 ptr += cpumask_size();
7838 #endif /* CONFIG_CPUMASK_OFFSTACK */
7842 init_defrootdomain();
7845 init_rt_bandwidth(&def_rt_bandwidth,
7846 global_rt_period(), global_rt_runtime());
7848 #ifdef CONFIG_RT_GROUP_SCHED
7849 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7850 global_rt_period(), global_rt_runtime());
7851 #endif /* CONFIG_RT_GROUP_SCHED */
7853 #ifdef CONFIG_CGROUP_SCHED
7854 list_add(&root_task_group.list, &task_groups);
7855 INIT_LIST_HEAD(&root_task_group.children);
7856 autogroup_init(&init_task);
7857 #endif /* CONFIG_CGROUP_SCHED */
7859 for_each_possible_cpu(i) {
7863 raw_spin_lock_init(&rq->lock);
7865 rq->calc_load_active = 0;
7866 rq->calc_load_update = jiffies + LOAD_FREQ;
7867 init_cfs_rq(&rq->cfs, rq);
7868 init_rt_rq(&rq->rt, rq);
7869 #ifdef CONFIG_FAIR_GROUP_SCHED
7870 root_task_group.shares = root_task_group_load;
7871 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7873 * How much cpu bandwidth does root_task_group get?
7875 * In case of task-groups formed thr' the cgroup filesystem, it
7876 * gets 100% of the cpu resources in the system. This overall
7877 * system cpu resource is divided among the tasks of
7878 * root_task_group and its child task-groups in a fair manner,
7879 * based on each entity's (task or task-group's) weight
7880 * (se->load.weight).
7882 * In other words, if root_task_group has 10 tasks of weight
7883 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7884 * then A0's share of the cpu resource is:
7886 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7888 * We achieve this by letting root_task_group's tasks sit
7889 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7891 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7892 #endif /* CONFIG_FAIR_GROUP_SCHED */
7894 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7895 #ifdef CONFIG_RT_GROUP_SCHED
7896 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7897 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7900 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7901 rq->cpu_load[j] = 0;
7903 rq->last_load_update_tick = jiffies;
7908 rq->cpu_power = SCHED_POWER_SCALE;
7909 rq->post_schedule = 0;
7910 rq->active_balance = 0;
7911 rq->next_balance = jiffies;
7916 rq->avg_idle = 2*sysctl_sched_migration_cost;
7917 rq_attach_root(rq, &def_root_domain);
7919 rq->nohz_balance_kick = 0;
7920 init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb, i));
7924 atomic_set(&rq->nr_iowait, 0);
7927 set_load_weight(&init_task);
7929 #ifdef CONFIG_PREEMPT_NOTIFIERS
7930 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7934 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7937 #ifdef CONFIG_RT_MUTEXES
7938 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
7942 * The boot idle thread does lazy MMU switching as well:
7944 atomic_inc(&init_mm.mm_count);
7945 enter_lazy_tlb(&init_mm, current);
7948 * Make us the idle thread. Technically, schedule() should not be
7949 * called from this thread, however somewhere below it might be,
7950 * but because we are the idle thread, we just pick up running again
7951 * when this runqueue becomes "idle".
7953 init_idle(current, smp_processor_id());
7955 calc_load_update = jiffies + LOAD_FREQ;
7958 * During early bootup we pretend to be a normal task:
7960 current->sched_class = &fair_sched_class;
7962 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
7963 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
7965 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7967 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
7968 alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT);
7969 atomic_set(&nohz.load_balancer, nr_cpu_ids);
7970 atomic_set(&nohz.first_pick_cpu, nr_cpu_ids);
7971 atomic_set(&nohz.second_pick_cpu, nr_cpu_ids);
7973 /* May be allocated at isolcpus cmdline parse time */
7974 if (cpu_isolated_map == NULL)
7975 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7978 scheduler_running = 1;
7981 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7982 static inline int preempt_count_equals(int preempt_offset)
7984 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7986 return (nested == preempt_offset);
7989 void __might_sleep(const char *file, int line, int preempt_offset)
7992 static unsigned long prev_jiffy; /* ratelimiting */
7994 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7995 system_state != SYSTEM_RUNNING || oops_in_progress)
7997 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7999 prev_jiffy = jiffies;
8002 "BUG: sleeping function called from invalid context at %s:%d\n",
8005 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8006 in_atomic(), irqs_disabled(),
8007 current->pid, current->comm);
8009 debug_show_held_locks(current);
8010 if (irqs_disabled())
8011 print_irqtrace_events(current);
8015 EXPORT_SYMBOL(__might_sleep);
8018 #ifdef CONFIG_MAGIC_SYSRQ
8019 static void normalize_task(struct rq *rq, struct task_struct *p)
8021 const struct sched_class *prev_class = p->sched_class;
8022 int old_prio = p->prio;
8027 deactivate_task(rq, p, 0);
8028 __setscheduler(rq, p, SCHED_NORMAL, 0);
8030 activate_task(rq, p, 0);
8031 resched_task(rq->curr);
8034 check_class_changed(rq, p, prev_class, old_prio);
8037 void normalize_rt_tasks(void)
8039 struct task_struct *g, *p;
8040 unsigned long flags;
8043 read_lock_irqsave(&tasklist_lock, flags);
8044 do_each_thread(g, p) {
8046 * Only normalize user tasks:
8051 p->se.exec_start = 0;
8052 #ifdef CONFIG_SCHEDSTATS
8053 p->se.statistics.wait_start = 0;
8054 p->se.statistics.sleep_start = 0;
8055 p->se.statistics.block_start = 0;
8060 * Renice negative nice level userspace
8063 if (TASK_NICE(p) < 0 && p->mm)
8064 set_user_nice(p, 0);
8068 raw_spin_lock(&p->pi_lock);
8069 rq = __task_rq_lock(p);
8071 normalize_task(rq, p);
8073 __task_rq_unlock(rq);
8074 raw_spin_unlock(&p->pi_lock);
8075 } while_each_thread(g, p);
8077 read_unlock_irqrestore(&tasklist_lock, flags);
8080 #endif /* CONFIG_MAGIC_SYSRQ */
8082 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8084 * These functions are only useful for the IA64 MCA handling, or kdb.
8086 * They can only be called when the whole system has been
8087 * stopped - every CPU needs to be quiescent, and no scheduling
8088 * activity can take place. Using them for anything else would
8089 * be a serious bug, and as a result, they aren't even visible
8090 * under any other configuration.
8094 * curr_task - return the current task for a given cpu.
8095 * @cpu: the processor in question.
8097 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8099 struct task_struct *curr_task(int cpu)
8101 return cpu_curr(cpu);
8104 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8108 * set_curr_task - set the current task for a given cpu.
8109 * @cpu: the processor in question.
8110 * @p: the task pointer to set.
8112 * Description: This function must only be used when non-maskable interrupts
8113 * are serviced on a separate stack. It allows the architecture to switch the
8114 * notion of the current task on a cpu in a non-blocking manner. This function
8115 * must be called with all CPU's synchronized, and interrupts disabled, the
8116 * and caller must save the original value of the current task (see
8117 * curr_task() above) and restore that value before reenabling interrupts and
8118 * re-starting the system.
8120 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8122 void set_curr_task(int cpu, struct task_struct *p)
8129 #ifdef CONFIG_FAIR_GROUP_SCHED
8130 static void free_fair_sched_group(struct task_group *tg)
8134 for_each_possible_cpu(i) {
8136 kfree(tg->cfs_rq[i]);
8146 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8148 struct cfs_rq *cfs_rq;
8149 struct sched_entity *se;
8152 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8155 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8159 tg->shares = NICE_0_LOAD;
8161 for_each_possible_cpu(i) {
8162 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8163 GFP_KERNEL, cpu_to_node(i));
8167 se = kzalloc_node(sizeof(struct sched_entity),
8168 GFP_KERNEL, cpu_to_node(i));
8172 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8183 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8185 struct rq *rq = cpu_rq(cpu);
8186 unsigned long flags;
8189 * Only empty task groups can be destroyed; so we can speculatively
8190 * check on_list without danger of it being re-added.
8192 if (!tg->cfs_rq[cpu]->on_list)
8195 raw_spin_lock_irqsave(&rq->lock, flags);
8196 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8197 raw_spin_unlock_irqrestore(&rq->lock, flags);
8199 #else /* !CONFG_FAIR_GROUP_SCHED */
8200 static inline void free_fair_sched_group(struct task_group *tg)
8205 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8210 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8213 #endif /* CONFIG_FAIR_GROUP_SCHED */
8215 #ifdef CONFIG_RT_GROUP_SCHED
8216 static void free_rt_sched_group(struct task_group *tg)
8220 destroy_rt_bandwidth(&tg->rt_bandwidth);
8222 for_each_possible_cpu(i) {
8224 kfree(tg->rt_rq[i]);
8226 kfree(tg->rt_se[i]);
8234 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8236 struct rt_rq *rt_rq;
8237 struct sched_rt_entity *rt_se;
8240 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8243 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8247 init_rt_bandwidth(&tg->rt_bandwidth,
8248 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8250 for_each_possible_cpu(i) {
8251 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8252 GFP_KERNEL, cpu_to_node(i));
8256 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8257 GFP_KERNEL, cpu_to_node(i));
8261 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
8271 #else /* !CONFIG_RT_GROUP_SCHED */
8272 static inline void free_rt_sched_group(struct task_group *tg)
8277 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8281 #endif /* CONFIG_RT_GROUP_SCHED */
8283 #ifdef CONFIG_CGROUP_SCHED
8284 static void free_sched_group(struct task_group *tg)
8286 free_fair_sched_group(tg);
8287 free_rt_sched_group(tg);
8292 /* allocate runqueue etc for a new task group */
8293 struct task_group *sched_create_group(struct task_group *parent)
8295 struct task_group *tg;
8296 unsigned long flags;
8298 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8300 return ERR_PTR(-ENOMEM);
8302 if (!alloc_fair_sched_group(tg, parent))
8305 if (!alloc_rt_sched_group(tg, parent))
8308 spin_lock_irqsave(&task_group_lock, flags);
8309 list_add_rcu(&tg->list, &task_groups);
8311 WARN_ON(!parent); /* root should already exist */
8313 tg->parent = parent;
8314 INIT_LIST_HEAD(&tg->children);
8315 list_add_rcu(&tg->siblings, &parent->children);
8316 spin_unlock_irqrestore(&task_group_lock, flags);
8321 free_sched_group(tg);
8322 return ERR_PTR(-ENOMEM);
8325 /* rcu callback to free various structures associated with a task group */
8326 static void free_sched_group_rcu(struct rcu_head *rhp)
8328 /* now it should be safe to free those cfs_rqs */
8329 free_sched_group(container_of(rhp, struct task_group, rcu));
8332 /* Destroy runqueue etc associated with a task group */
8333 void sched_destroy_group(struct task_group *tg)
8335 unsigned long flags;
8338 /* end participation in shares distribution */
8339 for_each_possible_cpu(i)
8340 unregister_fair_sched_group(tg, i);
8342 spin_lock_irqsave(&task_group_lock, flags);
8343 list_del_rcu(&tg->list);
8344 list_del_rcu(&tg->siblings);
8345 spin_unlock_irqrestore(&task_group_lock, flags);
8347 /* wait for possible concurrent references to cfs_rqs complete */
8348 call_rcu(&tg->rcu, free_sched_group_rcu);
8351 /* change task's runqueue when it moves between groups.
8352 * The caller of this function should have put the task in its new group
8353 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8354 * reflect its new group.
8356 void sched_move_task(struct task_struct *tsk)
8359 unsigned long flags;
8362 rq = task_rq_lock(tsk, &flags);
8364 running = task_current(rq, tsk);
8368 dequeue_task(rq, tsk, 0);
8369 if (unlikely(running))
8370 tsk->sched_class->put_prev_task(rq, tsk);
8372 #ifdef CONFIG_FAIR_GROUP_SCHED
8373 if (tsk->sched_class->task_move_group)
8374 tsk->sched_class->task_move_group(tsk, on_rq);
8377 set_task_rq(tsk, task_cpu(tsk));
8379 if (unlikely(running))
8380 tsk->sched_class->set_curr_task(rq);
8382 enqueue_task(rq, tsk, 0);
8384 task_rq_unlock(rq, tsk, &flags);
8386 #endif /* CONFIG_CGROUP_SCHED */
8388 #ifdef CONFIG_FAIR_GROUP_SCHED
8389 static DEFINE_MUTEX(shares_mutex);
8391 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8394 unsigned long flags;
8397 * We can't change the weight of the root cgroup.
8402 if (shares < MIN_SHARES)
8403 shares = MIN_SHARES;
8404 else if (shares > MAX_SHARES)
8405 shares = MAX_SHARES;
8407 mutex_lock(&shares_mutex);
8408 if (tg->shares == shares)
8411 tg->shares = shares;
8412 for_each_possible_cpu(i) {
8413 struct rq *rq = cpu_rq(i);
8414 struct sched_entity *se;
8417 /* Propagate contribution to hierarchy */
8418 raw_spin_lock_irqsave(&rq->lock, flags);
8419 for_each_sched_entity(se)
8420 update_cfs_shares(group_cfs_rq(se));
8421 raw_spin_unlock_irqrestore(&rq->lock, flags);
8425 mutex_unlock(&shares_mutex);
8429 unsigned long sched_group_shares(struct task_group *tg)
8435 #ifdef CONFIG_RT_GROUP_SCHED
8437 * Ensure that the real time constraints are schedulable.
8439 static DEFINE_MUTEX(rt_constraints_mutex);
8441 static unsigned long to_ratio(u64 period, u64 runtime)
8443 if (runtime == RUNTIME_INF)
8446 return div64_u64(runtime << 20, period);
8449 /* Must be called with tasklist_lock held */
8450 static inline int tg_has_rt_tasks(struct task_group *tg)
8452 struct task_struct *g, *p;
8454 do_each_thread(g, p) {
8455 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8457 } while_each_thread(g, p);
8462 struct rt_schedulable_data {
8463 struct task_group *tg;
8468 static int tg_schedulable(struct task_group *tg, void *data)
8470 struct rt_schedulable_data *d = data;
8471 struct task_group *child;
8472 unsigned long total, sum = 0;
8473 u64 period, runtime;
8475 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8476 runtime = tg->rt_bandwidth.rt_runtime;
8479 period = d->rt_period;
8480 runtime = d->rt_runtime;
8484 * Cannot have more runtime than the period.
8486 if (runtime > period && runtime != RUNTIME_INF)
8490 * Ensure we don't starve existing RT tasks.
8492 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8495 total = to_ratio(period, runtime);
8498 * Nobody can have more than the global setting allows.
8500 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8504 * The sum of our children's runtime should not exceed our own.
8506 list_for_each_entry_rcu(child, &tg->children, siblings) {
8507 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8508 runtime = child->rt_bandwidth.rt_runtime;
8510 if (child == d->tg) {
8511 period = d->rt_period;
8512 runtime = d->rt_runtime;
8515 sum += to_ratio(period, runtime);
8524 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8526 struct rt_schedulable_data data = {
8528 .rt_period = period,
8529 .rt_runtime = runtime,
8532 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8535 static int tg_set_bandwidth(struct task_group *tg,
8536 u64 rt_period, u64 rt_runtime)
8540 mutex_lock(&rt_constraints_mutex);
8541 read_lock(&tasklist_lock);
8542 err = __rt_schedulable(tg, rt_period, rt_runtime);
8546 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8547 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8548 tg->rt_bandwidth.rt_runtime = rt_runtime;
8550 for_each_possible_cpu(i) {
8551 struct rt_rq *rt_rq = tg->rt_rq[i];
8553 raw_spin_lock(&rt_rq->rt_runtime_lock);
8554 rt_rq->rt_runtime = rt_runtime;
8555 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8557 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8559 read_unlock(&tasklist_lock);
8560 mutex_unlock(&rt_constraints_mutex);
8565 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8567 u64 rt_runtime, rt_period;
8569 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8570 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8571 if (rt_runtime_us < 0)
8572 rt_runtime = RUNTIME_INF;
8574 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8577 long sched_group_rt_runtime(struct task_group *tg)
8581 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8584 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8585 do_div(rt_runtime_us, NSEC_PER_USEC);
8586 return rt_runtime_us;
8589 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8591 u64 rt_runtime, rt_period;
8593 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8594 rt_runtime = tg->rt_bandwidth.rt_runtime;
8599 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8602 long sched_group_rt_period(struct task_group *tg)
8606 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8607 do_div(rt_period_us, NSEC_PER_USEC);
8608 return rt_period_us;
8611 static int sched_rt_global_constraints(void)
8613 u64 runtime, period;
8616 if (sysctl_sched_rt_period <= 0)
8619 runtime = global_rt_runtime();
8620 period = global_rt_period();
8623 * Sanity check on the sysctl variables.
8625 if (runtime > period && runtime != RUNTIME_INF)
8628 mutex_lock(&rt_constraints_mutex);
8629 read_lock(&tasklist_lock);
8630 ret = __rt_schedulable(NULL, 0, 0);
8631 read_unlock(&tasklist_lock);
8632 mutex_unlock(&rt_constraints_mutex);
8637 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8639 /* Don't accept realtime tasks when there is no way for them to run */
8640 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8646 #else /* !CONFIG_RT_GROUP_SCHED */
8647 static int sched_rt_global_constraints(void)
8649 unsigned long flags;
8652 if (sysctl_sched_rt_period <= 0)
8656 * There's always some RT tasks in the root group
8657 * -- migration, kstopmachine etc..
8659 if (sysctl_sched_rt_runtime == 0)
8662 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8663 for_each_possible_cpu(i) {
8664 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8666 raw_spin_lock(&rt_rq->rt_runtime_lock);
8667 rt_rq->rt_runtime = global_rt_runtime();
8668 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8670 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8674 #endif /* CONFIG_RT_GROUP_SCHED */
8676 int sched_rt_handler(struct ctl_table *table, int write,
8677 void __user *buffer, size_t *lenp,
8681 int old_period, old_runtime;
8682 static DEFINE_MUTEX(mutex);
8685 old_period = sysctl_sched_rt_period;
8686 old_runtime = sysctl_sched_rt_runtime;
8688 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8690 if (!ret && write) {
8691 ret = sched_rt_global_constraints();
8693 sysctl_sched_rt_period = old_period;
8694 sysctl_sched_rt_runtime = old_runtime;
8696 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8697 def_rt_bandwidth.rt_period =
8698 ns_to_ktime(global_rt_period());
8701 mutex_unlock(&mutex);
8706 #ifdef CONFIG_CGROUP_SCHED
8708 /* return corresponding task_group object of a cgroup */
8709 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8711 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8712 struct task_group, css);
8715 static struct cgroup_subsys_state *
8716 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8718 struct task_group *tg, *parent;
8720 if (!cgrp->parent) {
8721 /* This is early initialization for the top cgroup */
8722 return &root_task_group.css;
8725 parent = cgroup_tg(cgrp->parent);
8726 tg = sched_create_group(parent);
8728 return ERR_PTR(-ENOMEM);
8734 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8736 struct task_group *tg = cgroup_tg(cgrp);
8738 sched_destroy_group(tg);
8742 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8744 #ifdef CONFIG_RT_GROUP_SCHED
8745 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
8748 /* We don't support RT-tasks being in separate groups */
8749 if (tsk->sched_class != &fair_sched_class)
8756 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8757 struct task_struct *tsk, bool threadgroup)
8759 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
8763 struct task_struct *c;
8765 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8766 retval = cpu_cgroup_can_attach_task(cgrp, c);
8778 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8779 struct cgroup *old_cont, struct task_struct *tsk,
8782 sched_move_task(tsk);
8784 struct task_struct *c;
8786 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8794 cpu_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
8795 struct cgroup *old_cgrp, struct task_struct *task)
8798 * cgroup_exit() is called in the copy_process() failure path.
8799 * Ignore this case since the task hasn't ran yet, this avoids
8800 * trying to poke a half freed task state from generic code.
8802 if (!(task->flags & PF_EXITING))
8805 sched_move_task(task);
8808 #ifdef CONFIG_FAIR_GROUP_SCHED
8809 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8812 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8815 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8817 struct task_group *tg = cgroup_tg(cgrp);
8819 return (u64) tg->shares;
8821 #endif /* CONFIG_FAIR_GROUP_SCHED */
8823 #ifdef CONFIG_RT_GROUP_SCHED
8824 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8827 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8830 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8832 return sched_group_rt_runtime(cgroup_tg(cgrp));
8835 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8838 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8841 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8843 return sched_group_rt_period(cgroup_tg(cgrp));
8845 #endif /* CONFIG_RT_GROUP_SCHED */
8847 static struct cftype cpu_files[] = {
8848 #ifdef CONFIG_FAIR_GROUP_SCHED
8851 .read_u64 = cpu_shares_read_u64,
8852 .write_u64 = cpu_shares_write_u64,
8855 #ifdef CONFIG_RT_GROUP_SCHED
8857 .name = "rt_runtime_us",
8858 .read_s64 = cpu_rt_runtime_read,
8859 .write_s64 = cpu_rt_runtime_write,
8862 .name = "rt_period_us",
8863 .read_u64 = cpu_rt_period_read_uint,
8864 .write_u64 = cpu_rt_period_write_uint,
8869 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8871 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8874 struct cgroup_subsys cpu_cgroup_subsys = {
8876 .create = cpu_cgroup_create,
8877 .destroy = cpu_cgroup_destroy,
8878 .can_attach = cpu_cgroup_can_attach,
8879 .attach = cpu_cgroup_attach,
8880 .exit = cpu_cgroup_exit,
8881 .populate = cpu_cgroup_populate,
8882 .subsys_id = cpu_cgroup_subsys_id,
8886 #endif /* CONFIG_CGROUP_SCHED */
8888 #ifdef CONFIG_CGROUP_CPUACCT
8891 * CPU accounting code for task groups.
8893 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8894 * (balbir@in.ibm.com).
8897 /* track cpu usage of a group of tasks and its child groups */
8899 struct cgroup_subsys_state css;
8900 /* cpuusage holds pointer to a u64-type object on every cpu */
8901 u64 __percpu *cpuusage;
8902 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
8903 struct cpuacct *parent;
8906 struct cgroup_subsys cpuacct_subsys;
8908 /* return cpu accounting group corresponding to this container */
8909 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8911 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8912 struct cpuacct, css);
8915 /* return cpu accounting group to which this task belongs */
8916 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8918 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8919 struct cpuacct, css);
8922 /* create a new cpu accounting group */
8923 static struct cgroup_subsys_state *cpuacct_create(
8924 struct cgroup_subsys *ss, struct cgroup *cgrp)
8926 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8932 ca->cpuusage = alloc_percpu(u64);
8936 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8937 if (percpu_counter_init(&ca->cpustat[i], 0))
8938 goto out_free_counters;
8941 ca->parent = cgroup_ca(cgrp->parent);
8947 percpu_counter_destroy(&ca->cpustat[i]);
8948 free_percpu(ca->cpuusage);
8952 return ERR_PTR(-ENOMEM);
8955 /* destroy an existing cpu accounting group */
8957 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8959 struct cpuacct *ca = cgroup_ca(cgrp);
8962 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8963 percpu_counter_destroy(&ca->cpustat[i]);
8964 free_percpu(ca->cpuusage);
8968 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
8970 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8973 #ifndef CONFIG_64BIT
8975 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8977 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8979 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8987 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
8989 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8991 #ifndef CONFIG_64BIT
8993 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8995 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8997 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9003 /* return total cpu usage (in nanoseconds) of a group */
9004 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9006 struct cpuacct *ca = cgroup_ca(cgrp);
9007 u64 totalcpuusage = 0;
9010 for_each_present_cpu(i)
9011 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9013 return totalcpuusage;
9016 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9019 struct cpuacct *ca = cgroup_ca(cgrp);
9028 for_each_present_cpu(i)
9029 cpuacct_cpuusage_write(ca, i, 0);
9035 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9038 struct cpuacct *ca = cgroup_ca(cgroup);
9042 for_each_present_cpu(i) {
9043 percpu = cpuacct_cpuusage_read(ca, i);
9044 seq_printf(m, "%llu ", (unsigned long long) percpu);
9046 seq_printf(m, "\n");
9050 static const char *cpuacct_stat_desc[] = {
9051 [CPUACCT_STAT_USER] = "user",
9052 [CPUACCT_STAT_SYSTEM] = "system",
9055 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
9056 struct cgroup_map_cb *cb)
9058 struct cpuacct *ca = cgroup_ca(cgrp);
9061 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
9062 s64 val = percpu_counter_read(&ca->cpustat[i]);
9063 val = cputime64_to_clock_t(val);
9064 cb->fill(cb, cpuacct_stat_desc[i], val);
9069 static struct cftype files[] = {
9072 .read_u64 = cpuusage_read,
9073 .write_u64 = cpuusage_write,
9076 .name = "usage_percpu",
9077 .read_seq_string = cpuacct_percpu_seq_read,
9081 .read_map = cpuacct_stats_show,
9085 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9087 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9091 * charge this task's execution time to its accounting group.
9093 * called with rq->lock held.
9095 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9100 if (unlikely(!cpuacct_subsys.active))
9103 cpu = task_cpu(tsk);
9109 for (; ca; ca = ca->parent) {
9110 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9111 *cpuusage += cputime;
9118 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9119 * in cputime_t units. As a result, cpuacct_update_stats calls
9120 * percpu_counter_add with values large enough to always overflow the
9121 * per cpu batch limit causing bad SMP scalability.
9123 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9124 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9125 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9128 #define CPUACCT_BATCH \
9129 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9131 #define CPUACCT_BATCH 0
9135 * Charge the system/user time to the task's accounting group.
9137 static void cpuacct_update_stats(struct task_struct *tsk,
9138 enum cpuacct_stat_index idx, cputime_t val)
9141 int batch = CPUACCT_BATCH;
9143 if (unlikely(!cpuacct_subsys.active))
9150 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9156 struct cgroup_subsys cpuacct_subsys = {
9158 .create = cpuacct_create,
9159 .destroy = cpuacct_destroy,
9160 .populate = cpuacct_populate,
9161 .subsys_id = cpuacct_subsys_id,
9163 #endif /* CONFIG_CGROUP_CPUACCT */