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 + SCHED_LOAD_RESOLUTION))
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)
1334 * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched
1335 * entities since MIN_SHARES = 2. Treat weight as 1 if less than
1336 * 2^SCHED_LOAD_RESOLUTION.
1338 if (likely(weight > (1UL << SCHED_LOAD_RESOLUTION)))
1339 tmp = (u64)delta_exec * scale_load_down(weight);
1341 tmp = (u64)delta_exec;
1343 if (!lw->inv_weight) {
1344 unsigned long w = scale_load_down(lw->weight);
1346 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
1348 else if (unlikely(!w))
1349 lw->inv_weight = WMULT_CONST;
1351 lw->inv_weight = WMULT_CONST / w;
1355 * Check whether we'd overflow the 64-bit multiplication:
1357 if (unlikely(tmp > WMULT_CONST))
1358 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1361 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1363 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1366 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1372 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1378 static inline void update_load_set(struct load_weight *lw, unsigned long w)
1385 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1386 * of tasks with abnormal "nice" values across CPUs the contribution that
1387 * each task makes to its run queue's load is weighted according to its
1388 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1389 * scaled version of the new time slice allocation that they receive on time
1393 #define WEIGHT_IDLEPRIO 3
1394 #define WMULT_IDLEPRIO 1431655765
1397 * Nice levels are multiplicative, with a gentle 10% change for every
1398 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1399 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1400 * that remained on nice 0.
1402 * The "10% effect" is relative and cumulative: from _any_ nice level,
1403 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1404 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1405 * If a task goes up by ~10% and another task goes down by ~10% then
1406 * the relative distance between them is ~25%.)
1408 static const int prio_to_weight[40] = {
1409 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1410 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1411 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1412 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1413 /* 0 */ 1024, 820, 655, 526, 423,
1414 /* 5 */ 335, 272, 215, 172, 137,
1415 /* 10 */ 110, 87, 70, 56, 45,
1416 /* 15 */ 36, 29, 23, 18, 15,
1420 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1422 * In cases where the weight does not change often, we can use the
1423 * precalculated inverse to speed up arithmetics by turning divisions
1424 * into multiplications:
1426 static const u32 prio_to_wmult[40] = {
1427 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1428 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1429 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1430 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1431 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1432 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1433 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1434 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1437 /* Time spent by the tasks of the cpu accounting group executing in ... */
1438 enum cpuacct_stat_index {
1439 CPUACCT_STAT_USER, /* ... user mode */
1440 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1442 CPUACCT_STAT_NSTATS,
1445 #ifdef CONFIG_CGROUP_CPUACCT
1446 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1447 static void cpuacct_update_stats(struct task_struct *tsk,
1448 enum cpuacct_stat_index idx, cputime_t val);
1450 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1451 static inline void cpuacct_update_stats(struct task_struct *tsk,
1452 enum cpuacct_stat_index idx, cputime_t val) {}
1455 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1457 update_load_add(&rq->load, load);
1460 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1462 update_load_sub(&rq->load, load);
1465 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1466 typedef int (*tg_visitor)(struct task_group *, void *);
1469 * Iterate the full tree, calling @down when first entering a node and @up when
1470 * leaving it for the final time.
1472 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1474 struct task_group *parent, *child;
1478 parent = &root_task_group;
1480 ret = (*down)(parent, data);
1483 list_for_each_entry_rcu(child, &parent->children, siblings) {
1490 ret = (*up)(parent, data);
1495 parent = parent->parent;
1504 static int tg_nop(struct task_group *tg, void *data)
1511 /* Used instead of source_load when we know the type == 0 */
1512 static unsigned long weighted_cpuload(const int cpu)
1514 return cpu_rq(cpu)->load.weight;
1518 * Return a low guess at the load of a migration-source cpu weighted
1519 * according to the scheduling class and "nice" value.
1521 * We want to under-estimate the load of migration sources, to
1522 * balance conservatively.
1524 static unsigned long source_load(int cpu, int type)
1526 struct rq *rq = cpu_rq(cpu);
1527 unsigned long total = weighted_cpuload(cpu);
1529 if (type == 0 || !sched_feat(LB_BIAS))
1532 return min(rq->cpu_load[type-1], total);
1536 * Return a high guess at the load of a migration-target cpu weighted
1537 * according to the scheduling class and "nice" value.
1539 static unsigned long target_load(int cpu, int type)
1541 struct rq *rq = cpu_rq(cpu);
1542 unsigned long total = weighted_cpuload(cpu);
1544 if (type == 0 || !sched_feat(LB_BIAS))
1547 return max(rq->cpu_load[type-1], total);
1550 static unsigned long power_of(int cpu)
1552 return cpu_rq(cpu)->cpu_power;
1555 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1557 static unsigned long cpu_avg_load_per_task(int cpu)
1559 struct rq *rq = cpu_rq(cpu);
1560 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1563 rq->avg_load_per_task = rq->load.weight / nr_running;
1565 rq->avg_load_per_task = 0;
1567 return rq->avg_load_per_task;
1570 #ifdef CONFIG_FAIR_GROUP_SCHED
1573 * Compute the cpu's hierarchical load factor for each task group.
1574 * This needs to be done in a top-down fashion because the load of a child
1575 * group is a fraction of its parents load.
1577 static int tg_load_down(struct task_group *tg, void *data)
1580 long cpu = (long)data;
1583 load = cpu_rq(cpu)->load.weight;
1585 load = tg->parent->cfs_rq[cpu]->h_load;
1586 load *= tg->se[cpu]->load.weight;
1587 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1590 tg->cfs_rq[cpu]->h_load = load;
1595 static void update_h_load(long cpu)
1597 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1602 #ifdef CONFIG_PREEMPT
1604 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1607 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1608 * way at the expense of forcing extra atomic operations in all
1609 * invocations. This assures that the double_lock is acquired using the
1610 * same underlying policy as the spinlock_t on this architecture, which
1611 * reduces latency compared to the unfair variant below. However, it
1612 * also adds more overhead and therefore may reduce throughput.
1614 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1615 __releases(this_rq->lock)
1616 __acquires(busiest->lock)
1617 __acquires(this_rq->lock)
1619 raw_spin_unlock(&this_rq->lock);
1620 double_rq_lock(this_rq, busiest);
1627 * Unfair double_lock_balance: Optimizes throughput at the expense of
1628 * latency by eliminating extra atomic operations when the locks are
1629 * already in proper order on entry. This favors lower cpu-ids and will
1630 * grant the double lock to lower cpus over higher ids under contention,
1631 * regardless of entry order into the function.
1633 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1634 __releases(this_rq->lock)
1635 __acquires(busiest->lock)
1636 __acquires(this_rq->lock)
1640 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1641 if (busiest < this_rq) {
1642 raw_spin_unlock(&this_rq->lock);
1643 raw_spin_lock(&busiest->lock);
1644 raw_spin_lock_nested(&this_rq->lock,
1645 SINGLE_DEPTH_NESTING);
1648 raw_spin_lock_nested(&busiest->lock,
1649 SINGLE_DEPTH_NESTING);
1654 #endif /* CONFIG_PREEMPT */
1657 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1659 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1661 if (unlikely(!irqs_disabled())) {
1662 /* printk() doesn't work good under rq->lock */
1663 raw_spin_unlock(&this_rq->lock);
1667 return _double_lock_balance(this_rq, busiest);
1670 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1671 __releases(busiest->lock)
1673 raw_spin_unlock(&busiest->lock);
1674 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1678 * double_rq_lock - safely lock two runqueues
1680 * Note this does not disable interrupts like task_rq_lock,
1681 * you need to do so manually before calling.
1683 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1684 __acquires(rq1->lock)
1685 __acquires(rq2->lock)
1687 BUG_ON(!irqs_disabled());
1689 raw_spin_lock(&rq1->lock);
1690 __acquire(rq2->lock); /* Fake it out ;) */
1693 raw_spin_lock(&rq1->lock);
1694 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1696 raw_spin_lock(&rq2->lock);
1697 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1703 * double_rq_unlock - safely unlock two runqueues
1705 * Note this does not restore interrupts like task_rq_unlock,
1706 * you need to do so manually after calling.
1708 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1709 __releases(rq1->lock)
1710 __releases(rq2->lock)
1712 raw_spin_unlock(&rq1->lock);
1714 raw_spin_unlock(&rq2->lock);
1716 __release(rq2->lock);
1719 #else /* CONFIG_SMP */
1722 * double_rq_lock - safely lock two runqueues
1724 * Note this does not disable interrupts like task_rq_lock,
1725 * you need to do so manually before calling.
1727 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1728 __acquires(rq1->lock)
1729 __acquires(rq2->lock)
1731 BUG_ON(!irqs_disabled());
1733 raw_spin_lock(&rq1->lock);
1734 __acquire(rq2->lock); /* Fake it out ;) */
1738 * double_rq_unlock - safely unlock two runqueues
1740 * Note this does not restore interrupts like task_rq_unlock,
1741 * you need to do so manually after calling.
1743 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1744 __releases(rq1->lock)
1745 __releases(rq2->lock)
1748 raw_spin_unlock(&rq1->lock);
1749 __release(rq2->lock);
1754 static void calc_load_account_idle(struct rq *this_rq);
1755 static void update_sysctl(void);
1756 static int get_update_sysctl_factor(void);
1757 static void update_cpu_load(struct rq *this_rq);
1759 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1761 set_task_rq(p, cpu);
1764 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1765 * successfuly executed on another CPU. We must ensure that updates of
1766 * per-task data have been completed by this moment.
1769 task_thread_info(p)->cpu = cpu;
1773 static const struct sched_class rt_sched_class;
1775 #define sched_class_highest (&stop_sched_class)
1776 #define for_each_class(class) \
1777 for (class = sched_class_highest; class; class = class->next)
1779 #include "sched_stats.h"
1781 static void inc_nr_running(struct rq *rq)
1786 static void dec_nr_running(struct rq *rq)
1791 static void set_load_weight(struct task_struct *p)
1793 int prio = p->static_prio - MAX_RT_PRIO;
1794 struct load_weight *load = &p->se.load;
1797 * SCHED_IDLE tasks get minimal weight:
1799 if (p->policy == SCHED_IDLE) {
1800 load->weight = scale_load(WEIGHT_IDLEPRIO);
1801 load->inv_weight = WMULT_IDLEPRIO;
1805 load->weight = scale_load(prio_to_weight[prio]);
1806 load->inv_weight = prio_to_wmult[prio];
1809 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1811 update_rq_clock(rq);
1812 sched_info_queued(p);
1813 p->sched_class->enqueue_task(rq, p, flags);
1816 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1818 update_rq_clock(rq);
1819 sched_info_dequeued(p);
1820 p->sched_class->dequeue_task(rq, p, flags);
1824 * activate_task - move a task to the runqueue.
1826 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1828 if (task_contributes_to_load(p))
1829 rq->nr_uninterruptible--;
1831 enqueue_task(rq, p, flags);
1836 * deactivate_task - remove a task from the runqueue.
1838 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1840 if (task_contributes_to_load(p))
1841 rq->nr_uninterruptible++;
1843 dequeue_task(rq, p, flags);
1847 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1850 * There are no locks covering percpu hardirq/softirq time.
1851 * They are only modified in account_system_vtime, on corresponding CPU
1852 * with interrupts disabled. So, writes are safe.
1853 * They are read and saved off onto struct rq in update_rq_clock().
1854 * This may result in other CPU reading this CPU's irq time and can
1855 * race with irq/account_system_vtime on this CPU. We would either get old
1856 * or new value with a side effect of accounting a slice of irq time to wrong
1857 * task when irq is in progress while we read rq->clock. That is a worthy
1858 * compromise in place of having locks on each irq in account_system_time.
1860 static DEFINE_PER_CPU(u64, cpu_hardirq_time);
1861 static DEFINE_PER_CPU(u64, cpu_softirq_time);
1863 static DEFINE_PER_CPU(u64, irq_start_time);
1864 static int sched_clock_irqtime;
1866 void enable_sched_clock_irqtime(void)
1868 sched_clock_irqtime = 1;
1871 void disable_sched_clock_irqtime(void)
1873 sched_clock_irqtime = 0;
1876 #ifndef CONFIG_64BIT
1877 static DEFINE_PER_CPU(seqcount_t, irq_time_seq);
1879 static inline void irq_time_write_begin(void)
1881 __this_cpu_inc(irq_time_seq.sequence);
1885 static inline void irq_time_write_end(void)
1888 __this_cpu_inc(irq_time_seq.sequence);
1891 static inline u64 irq_time_read(int cpu)
1897 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
1898 irq_time = per_cpu(cpu_softirq_time, cpu) +
1899 per_cpu(cpu_hardirq_time, cpu);
1900 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
1904 #else /* CONFIG_64BIT */
1905 static inline void irq_time_write_begin(void)
1909 static inline void irq_time_write_end(void)
1913 static inline u64 irq_time_read(int cpu)
1915 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
1917 #endif /* CONFIG_64BIT */
1920 * Called before incrementing preempt_count on {soft,}irq_enter
1921 * and before decrementing preempt_count on {soft,}irq_exit.
1923 void account_system_vtime(struct task_struct *curr)
1925 unsigned long flags;
1929 if (!sched_clock_irqtime)
1932 local_irq_save(flags);
1934 cpu = smp_processor_id();
1935 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
1936 __this_cpu_add(irq_start_time, delta);
1938 irq_time_write_begin();
1940 * We do not account for softirq time from ksoftirqd here.
1941 * We want to continue accounting softirq time to ksoftirqd thread
1942 * in that case, so as not to confuse scheduler with a special task
1943 * that do not consume any time, but still wants to run.
1945 if (hardirq_count())
1946 __this_cpu_add(cpu_hardirq_time, delta);
1947 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
1948 __this_cpu_add(cpu_softirq_time, delta);
1950 irq_time_write_end();
1951 local_irq_restore(flags);
1953 EXPORT_SYMBOL_GPL(account_system_vtime);
1955 static void update_rq_clock_task(struct rq *rq, s64 delta)
1959 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
1962 * Since irq_time is only updated on {soft,}irq_exit, we might run into
1963 * this case when a previous update_rq_clock() happened inside a
1964 * {soft,}irq region.
1966 * When this happens, we stop ->clock_task and only update the
1967 * prev_irq_time stamp to account for the part that fit, so that a next
1968 * update will consume the rest. This ensures ->clock_task is
1971 * It does however cause some slight miss-attribution of {soft,}irq
1972 * time, a more accurate solution would be to update the irq_time using
1973 * the current rq->clock timestamp, except that would require using
1976 if (irq_delta > delta)
1979 rq->prev_irq_time += irq_delta;
1981 rq->clock_task += delta;
1983 if (irq_delta && sched_feat(NONIRQ_POWER))
1984 sched_rt_avg_update(rq, irq_delta);
1987 static int irqtime_account_hi_update(void)
1989 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
1990 unsigned long flags;
1994 local_irq_save(flags);
1995 latest_ns = this_cpu_read(cpu_hardirq_time);
1996 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->irq))
1998 local_irq_restore(flags);
2002 static int irqtime_account_si_update(void)
2004 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2005 unsigned long flags;
2009 local_irq_save(flags);
2010 latest_ns = this_cpu_read(cpu_softirq_time);
2011 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->softirq))
2013 local_irq_restore(flags);
2017 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
2019 #define sched_clock_irqtime (0)
2021 static void update_rq_clock_task(struct rq *rq, s64 delta)
2023 rq->clock_task += delta;
2026 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2028 #include "sched_idletask.c"
2029 #include "sched_fair.c"
2030 #include "sched_rt.c"
2031 #include "sched_autogroup.c"
2032 #include "sched_stoptask.c"
2033 #ifdef CONFIG_SCHED_DEBUG
2034 # include "sched_debug.c"
2037 void sched_set_stop_task(int cpu, struct task_struct *stop)
2039 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
2040 struct task_struct *old_stop = cpu_rq(cpu)->stop;
2044 * Make it appear like a SCHED_FIFO task, its something
2045 * userspace knows about and won't get confused about.
2047 * Also, it will make PI more or less work without too
2048 * much confusion -- but then, stop work should not
2049 * rely on PI working anyway.
2051 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
2053 stop->sched_class = &stop_sched_class;
2056 cpu_rq(cpu)->stop = stop;
2060 * Reset it back to a normal scheduling class so that
2061 * it can die in pieces.
2063 old_stop->sched_class = &rt_sched_class;
2068 * __normal_prio - return the priority that is based on the static prio
2070 static inline int __normal_prio(struct task_struct *p)
2072 return p->static_prio;
2076 * Calculate the expected normal priority: i.e. priority
2077 * without taking RT-inheritance into account. Might be
2078 * boosted by interactivity modifiers. Changes upon fork,
2079 * setprio syscalls, and whenever the interactivity
2080 * estimator recalculates.
2082 static inline int normal_prio(struct task_struct *p)
2086 if (task_has_rt_policy(p))
2087 prio = MAX_RT_PRIO-1 - p->rt_priority;
2089 prio = __normal_prio(p);
2094 * Calculate the current priority, i.e. the priority
2095 * taken into account by the scheduler. This value might
2096 * be boosted by RT tasks, or might be boosted by
2097 * interactivity modifiers. Will be RT if the task got
2098 * RT-boosted. If not then it returns p->normal_prio.
2100 static int effective_prio(struct task_struct *p)
2102 p->normal_prio = normal_prio(p);
2104 * If we are RT tasks or we were boosted to RT priority,
2105 * keep the priority unchanged. Otherwise, update priority
2106 * to the normal priority:
2108 if (!rt_prio(p->prio))
2109 return p->normal_prio;
2114 * task_curr - is this task currently executing on a CPU?
2115 * @p: the task in question.
2117 inline int task_curr(const struct task_struct *p)
2119 return cpu_curr(task_cpu(p)) == p;
2122 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2123 const struct sched_class *prev_class,
2126 if (prev_class != p->sched_class) {
2127 if (prev_class->switched_from)
2128 prev_class->switched_from(rq, p);
2129 p->sched_class->switched_to(rq, p);
2130 } else if (oldprio != p->prio)
2131 p->sched_class->prio_changed(rq, p, oldprio);
2134 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2136 const struct sched_class *class;
2138 if (p->sched_class == rq->curr->sched_class) {
2139 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2141 for_each_class(class) {
2142 if (class == rq->curr->sched_class)
2144 if (class == p->sched_class) {
2145 resched_task(rq->curr);
2152 * A queue event has occurred, and we're going to schedule. In
2153 * this case, we can save a useless back to back clock update.
2155 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
2156 rq->skip_clock_update = 1;
2161 * Is this task likely cache-hot:
2164 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2168 if (p->sched_class != &fair_sched_class)
2171 if (unlikely(p->policy == SCHED_IDLE))
2175 * Buddy candidates are cache hot:
2177 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2178 (&p->se == cfs_rq_of(&p->se)->next ||
2179 &p->se == cfs_rq_of(&p->se)->last))
2182 if (sysctl_sched_migration_cost == -1)
2184 if (sysctl_sched_migration_cost == 0)
2187 delta = now - p->se.exec_start;
2189 return delta < (s64)sysctl_sched_migration_cost;
2192 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2194 #ifdef CONFIG_SCHED_DEBUG
2196 * We should never call set_task_cpu() on a blocked task,
2197 * ttwu() will sort out the placement.
2199 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2200 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2202 #ifdef CONFIG_LOCKDEP
2203 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
2204 lockdep_is_held(&task_rq(p)->lock)));
2208 trace_sched_migrate_task(p, new_cpu);
2210 if (task_cpu(p) != new_cpu) {
2211 p->se.nr_migrations++;
2212 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2215 __set_task_cpu(p, new_cpu);
2218 struct migration_arg {
2219 struct task_struct *task;
2223 static int migration_cpu_stop(void *data);
2226 * wait_task_inactive - wait for a thread to unschedule.
2228 * If @match_state is nonzero, it's the @p->state value just checked and
2229 * not expected to change. If it changes, i.e. @p might have woken up,
2230 * then return zero. When we succeed in waiting for @p to be off its CPU,
2231 * we return a positive number (its total switch count). If a second call
2232 * a short while later returns the same number, the caller can be sure that
2233 * @p has remained unscheduled the whole time.
2235 * The caller must ensure that the task *will* unschedule sometime soon,
2236 * else this function might spin for a *long* time. This function can't
2237 * be called with interrupts off, or it may introduce deadlock with
2238 * smp_call_function() if an IPI is sent by the same process we are
2239 * waiting to become inactive.
2241 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2243 unsigned long flags;
2250 * We do the initial early heuristics without holding
2251 * any task-queue locks at all. We'll only try to get
2252 * the runqueue lock when things look like they will
2258 * If the task is actively running on another CPU
2259 * still, just relax and busy-wait without holding
2262 * NOTE! Since we don't hold any locks, it's not
2263 * even sure that "rq" stays as the right runqueue!
2264 * But we don't care, since "task_running()" will
2265 * return false if the runqueue has changed and p
2266 * is actually now running somewhere else!
2268 while (task_running(rq, p)) {
2269 if (match_state && unlikely(p->state != match_state))
2275 * Ok, time to look more closely! We need the rq
2276 * lock now, to be *sure*. If we're wrong, we'll
2277 * just go back and repeat.
2279 rq = task_rq_lock(p, &flags);
2280 trace_sched_wait_task(p);
2281 running = task_running(rq, p);
2284 if (!match_state || p->state == match_state)
2285 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2286 task_rq_unlock(rq, p, &flags);
2289 * If it changed from the expected state, bail out now.
2291 if (unlikely(!ncsw))
2295 * Was it really running after all now that we
2296 * checked with the proper locks actually held?
2298 * Oops. Go back and try again..
2300 if (unlikely(running)) {
2306 * It's not enough that it's not actively running,
2307 * it must be off the runqueue _entirely_, and not
2310 * So if it was still runnable (but just not actively
2311 * running right now), it's preempted, and we should
2312 * yield - it could be a while.
2314 if (unlikely(on_rq)) {
2315 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
2317 set_current_state(TASK_UNINTERRUPTIBLE);
2318 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
2323 * Ahh, all good. It wasn't running, and it wasn't
2324 * runnable, which means that it will never become
2325 * running in the future either. We're all done!
2334 * kick_process - kick a running thread to enter/exit the kernel
2335 * @p: the to-be-kicked thread
2337 * Cause a process which is running on another CPU to enter
2338 * kernel-mode, without any delay. (to get signals handled.)
2340 * NOTE: this function doesn't have to take the runqueue lock,
2341 * because all it wants to ensure is that the remote task enters
2342 * the kernel. If the IPI races and the task has been migrated
2343 * to another CPU then no harm is done and the purpose has been
2346 void kick_process(struct task_struct *p)
2352 if ((cpu != smp_processor_id()) && task_curr(p))
2353 smp_send_reschedule(cpu);
2356 EXPORT_SYMBOL_GPL(kick_process);
2357 #endif /* CONFIG_SMP */
2361 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
2363 static int select_fallback_rq(int cpu, struct task_struct *p)
2366 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2368 /* Look for allowed, online CPU in same node. */
2369 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2370 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2373 /* Any allowed, online CPU? */
2374 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2375 if (dest_cpu < nr_cpu_ids)
2378 /* No more Mr. Nice Guy. */
2379 dest_cpu = cpuset_cpus_allowed_fallback(p);
2381 * Don't tell them about moving exiting tasks or
2382 * kernel threads (both mm NULL), since they never
2385 if (p->mm && printk_ratelimit()) {
2386 printk(KERN_INFO "process %d (%s) no longer affine to cpu%d\n",
2387 task_pid_nr(p), p->comm, cpu);
2394 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
2397 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2399 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2402 * In order not to call set_task_cpu() on a blocking task we need
2403 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2406 * Since this is common to all placement strategies, this lives here.
2408 * [ this allows ->select_task() to simply return task_cpu(p) and
2409 * not worry about this generic constraint ]
2411 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2413 cpu = select_fallback_rq(task_cpu(p), p);
2418 static void update_avg(u64 *avg, u64 sample)
2420 s64 diff = sample - *avg;
2426 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
2428 #ifdef CONFIG_SCHEDSTATS
2429 struct rq *rq = this_rq();
2432 int this_cpu = smp_processor_id();
2434 if (cpu == this_cpu) {
2435 schedstat_inc(rq, ttwu_local);
2436 schedstat_inc(p, se.statistics.nr_wakeups_local);
2438 struct sched_domain *sd;
2440 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2442 for_each_domain(this_cpu, sd) {
2443 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2444 schedstat_inc(sd, ttwu_wake_remote);
2450 #endif /* CONFIG_SMP */
2452 schedstat_inc(rq, ttwu_count);
2453 schedstat_inc(p, se.statistics.nr_wakeups);
2455 if (wake_flags & WF_SYNC)
2456 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2458 if (cpu != task_cpu(p))
2459 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2461 #endif /* CONFIG_SCHEDSTATS */
2464 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
2466 activate_task(rq, p, en_flags);
2469 /* if a worker is waking up, notify workqueue */
2470 if (p->flags & PF_WQ_WORKER)
2471 wq_worker_waking_up(p, cpu_of(rq));
2475 * Mark the task runnable and perform wakeup-preemption.
2478 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
2480 trace_sched_wakeup(p, true);
2481 check_preempt_curr(rq, p, wake_flags);
2483 p->state = TASK_RUNNING;
2485 if (p->sched_class->task_woken)
2486 p->sched_class->task_woken(rq, p);
2488 if (unlikely(rq->idle_stamp)) {
2489 u64 delta = rq->clock - rq->idle_stamp;
2490 u64 max = 2*sysctl_sched_migration_cost;
2495 update_avg(&rq->avg_idle, delta);
2502 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
2505 if (p->sched_contributes_to_load)
2506 rq->nr_uninterruptible--;
2509 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
2510 ttwu_do_wakeup(rq, p, wake_flags);
2514 * Called in case the task @p isn't fully descheduled from its runqueue,
2515 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
2516 * since all we need to do is flip p->state to TASK_RUNNING, since
2517 * the task is still ->on_rq.
2519 static int ttwu_remote(struct task_struct *p, int wake_flags)
2524 rq = __task_rq_lock(p);
2526 ttwu_do_wakeup(rq, p, wake_flags);
2529 __task_rq_unlock(rq);
2535 static void sched_ttwu_pending(void)
2537 struct rq *rq = this_rq();
2538 struct task_struct *list = xchg(&rq->wake_list, NULL);
2543 raw_spin_lock(&rq->lock);
2546 struct task_struct *p = list;
2547 list = list->wake_entry;
2548 ttwu_do_activate(rq, p, 0);
2551 raw_spin_unlock(&rq->lock);
2554 void scheduler_ipi(void)
2556 sched_ttwu_pending();
2559 static void ttwu_queue_remote(struct task_struct *p, int cpu)
2561 struct rq *rq = cpu_rq(cpu);
2562 struct task_struct *next = rq->wake_list;
2565 struct task_struct *old = next;
2567 p->wake_entry = next;
2568 next = cmpxchg(&rq->wake_list, old, p);
2574 smp_send_reschedule(cpu);
2577 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2578 static int ttwu_activate_remote(struct task_struct *p, int wake_flags)
2583 rq = __task_rq_lock(p);
2585 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2586 ttwu_do_wakeup(rq, p, wake_flags);
2589 __task_rq_unlock(rq);
2594 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2595 #endif /* CONFIG_SMP */
2597 static void ttwu_queue(struct task_struct *p, int cpu)
2599 struct rq *rq = cpu_rq(cpu);
2601 #if defined(CONFIG_SMP)
2602 if (sched_feat(TTWU_QUEUE) && cpu != smp_processor_id()) {
2603 sched_clock_cpu(cpu); /* sync clocks x-cpu */
2604 ttwu_queue_remote(p, cpu);
2609 raw_spin_lock(&rq->lock);
2610 ttwu_do_activate(rq, p, 0);
2611 raw_spin_unlock(&rq->lock);
2615 * try_to_wake_up - wake up a thread
2616 * @p: the thread to be awakened
2617 * @state: the mask of task states that can be woken
2618 * @wake_flags: wake modifier flags (WF_*)
2620 * Put it on the run-queue if it's not already there. The "current"
2621 * thread is always on the run-queue (except when the actual
2622 * re-schedule is in progress), and as such you're allowed to do
2623 * the simpler "current->state = TASK_RUNNING" to mark yourself
2624 * runnable without the overhead of this.
2626 * Returns %true if @p was woken up, %false if it was already running
2627 * or @state didn't match @p's state.
2630 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
2632 unsigned long flags;
2633 int cpu, success = 0;
2636 raw_spin_lock_irqsave(&p->pi_lock, flags);
2637 if (!(p->state & state))
2640 success = 1; /* we're going to change ->state */
2643 if (p->on_rq && ttwu_remote(p, wake_flags))
2648 * If the owning (remote) cpu is still in the middle of schedule() with
2649 * this task as prev, wait until its done referencing the task.
2652 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2654 * In case the architecture enables interrupts in
2655 * context_switch(), we cannot busy wait, since that
2656 * would lead to deadlocks when an interrupt hits and
2657 * tries to wake up @prev. So bail and do a complete
2660 if (ttwu_activate_remote(p, wake_flags))
2667 * Pairs with the smp_wmb() in finish_lock_switch().
2671 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2672 p->state = TASK_WAKING;
2674 if (p->sched_class->task_waking)
2675 p->sched_class->task_waking(p);
2677 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2678 if (task_cpu(p) != cpu)
2679 set_task_cpu(p, cpu);
2680 #endif /* CONFIG_SMP */
2684 ttwu_stat(p, cpu, wake_flags);
2686 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2692 * try_to_wake_up_local - try to wake up a local task with rq lock held
2693 * @p: the thread to be awakened
2695 * Put @p on the run-queue if it's not already there. The caller must
2696 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2699 static void try_to_wake_up_local(struct task_struct *p)
2701 struct rq *rq = task_rq(p);
2703 BUG_ON(rq != this_rq());
2704 BUG_ON(p == current);
2705 lockdep_assert_held(&rq->lock);
2707 if (!raw_spin_trylock(&p->pi_lock)) {
2708 raw_spin_unlock(&rq->lock);
2709 raw_spin_lock(&p->pi_lock);
2710 raw_spin_lock(&rq->lock);
2713 if (!(p->state & TASK_NORMAL))
2717 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2719 ttwu_do_wakeup(rq, p, 0);
2720 ttwu_stat(p, smp_processor_id(), 0);
2722 raw_spin_unlock(&p->pi_lock);
2726 * wake_up_process - Wake up a specific process
2727 * @p: The process to be woken up.
2729 * Attempt to wake up the nominated process and move it to the set of runnable
2730 * processes. Returns 1 if the process was woken up, 0 if it was already
2733 * It may be assumed that this function implies a write memory barrier before
2734 * changing the task state if and only if any tasks are woken up.
2736 int wake_up_process(struct task_struct *p)
2738 return try_to_wake_up(p, TASK_ALL, 0);
2740 EXPORT_SYMBOL(wake_up_process);
2742 int wake_up_state(struct task_struct *p, unsigned int state)
2744 return try_to_wake_up(p, state, 0);
2748 * Perform scheduler related setup for a newly forked process p.
2749 * p is forked by current.
2751 * __sched_fork() is basic setup used by init_idle() too:
2753 static void __sched_fork(struct task_struct *p)
2758 p->se.exec_start = 0;
2759 p->se.sum_exec_runtime = 0;
2760 p->se.prev_sum_exec_runtime = 0;
2761 p->se.nr_migrations = 0;
2763 INIT_LIST_HEAD(&p->se.group_node);
2765 #ifdef CONFIG_SCHEDSTATS
2766 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2769 INIT_LIST_HEAD(&p->rt.run_list);
2771 #ifdef CONFIG_PREEMPT_NOTIFIERS
2772 INIT_HLIST_HEAD(&p->preempt_notifiers);
2777 * fork()/clone()-time setup:
2779 void sched_fork(struct task_struct *p)
2781 unsigned long flags;
2782 int cpu = get_cpu();
2786 * We mark the process as running here. This guarantees that
2787 * nobody will actually run it, and a signal or other external
2788 * event cannot wake it up and insert it on the runqueue either.
2790 p->state = TASK_RUNNING;
2793 * Revert to default priority/policy on fork if requested.
2795 if (unlikely(p->sched_reset_on_fork)) {
2796 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2797 p->policy = SCHED_NORMAL;
2798 p->normal_prio = p->static_prio;
2801 if (PRIO_TO_NICE(p->static_prio) < 0) {
2802 p->static_prio = NICE_TO_PRIO(0);
2803 p->normal_prio = p->static_prio;
2808 * We don't need the reset flag anymore after the fork. It has
2809 * fulfilled its duty:
2811 p->sched_reset_on_fork = 0;
2815 * Make sure we do not leak PI boosting priority to the child.
2817 p->prio = current->normal_prio;
2819 if (!rt_prio(p->prio))
2820 p->sched_class = &fair_sched_class;
2822 if (p->sched_class->task_fork)
2823 p->sched_class->task_fork(p);
2826 * The child is not yet in the pid-hash so no cgroup attach races,
2827 * and the cgroup is pinned to this child due to cgroup_fork()
2828 * is ran before sched_fork().
2830 * Silence PROVE_RCU.
2832 raw_spin_lock_irqsave(&p->pi_lock, flags);
2833 set_task_cpu(p, cpu);
2834 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2836 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2837 if (likely(sched_info_on()))
2838 memset(&p->sched_info, 0, sizeof(p->sched_info));
2840 #if defined(CONFIG_SMP)
2843 #ifdef CONFIG_PREEMPT
2844 /* Want to start with kernel preemption disabled. */
2845 task_thread_info(p)->preempt_count = 1;
2848 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2855 * wake_up_new_task - wake up a newly created task for the first time.
2857 * This function will do some initial scheduler statistics housekeeping
2858 * that must be done for every newly created context, then puts the task
2859 * on the runqueue and wakes it.
2861 void wake_up_new_task(struct task_struct *p)
2863 unsigned long flags;
2866 raw_spin_lock_irqsave(&p->pi_lock, flags);
2869 * Fork balancing, do it here and not earlier because:
2870 * - cpus_allowed can change in the fork path
2871 * - any previously selected cpu might disappear through hotplug
2873 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
2876 rq = __task_rq_lock(p);
2877 activate_task(rq, p, 0);
2879 trace_sched_wakeup_new(p, true);
2880 check_preempt_curr(rq, p, WF_FORK);
2882 if (p->sched_class->task_woken)
2883 p->sched_class->task_woken(rq, p);
2885 task_rq_unlock(rq, p, &flags);
2888 #ifdef CONFIG_PREEMPT_NOTIFIERS
2891 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2892 * @notifier: notifier struct to register
2894 void preempt_notifier_register(struct preempt_notifier *notifier)
2896 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2898 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2901 * preempt_notifier_unregister - no longer interested in preemption notifications
2902 * @notifier: notifier struct to unregister
2904 * This is safe to call from within a preemption notifier.
2906 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2908 hlist_del(¬ifier->link);
2910 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2912 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2914 struct preempt_notifier *notifier;
2915 struct hlist_node *node;
2917 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2918 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2922 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2923 struct task_struct *next)
2925 struct preempt_notifier *notifier;
2926 struct hlist_node *node;
2928 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2929 notifier->ops->sched_out(notifier, next);
2932 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2934 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2939 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2940 struct task_struct *next)
2944 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2947 * prepare_task_switch - prepare to switch tasks
2948 * @rq: the runqueue preparing to switch
2949 * @prev: the current task that is being switched out
2950 * @next: the task we are going to switch to.
2952 * This is called with the rq lock held and interrupts off. It must
2953 * be paired with a subsequent finish_task_switch after the context
2956 * prepare_task_switch sets up locking and calls architecture specific
2960 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2961 struct task_struct *next)
2963 sched_info_switch(prev, next);
2964 perf_event_task_sched_out(prev, next);
2965 fire_sched_out_preempt_notifiers(prev, next);
2966 prepare_lock_switch(rq, next);
2967 prepare_arch_switch(next);
2968 trace_sched_switch(prev, next);
2972 * finish_task_switch - clean up after a task-switch
2973 * @rq: runqueue associated with task-switch
2974 * @prev: the thread we just switched away from.
2976 * finish_task_switch must be called after the context switch, paired
2977 * with a prepare_task_switch call before the context switch.
2978 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2979 * and do any other architecture-specific cleanup actions.
2981 * Note that we may have delayed dropping an mm in context_switch(). If
2982 * so, we finish that here outside of the runqueue lock. (Doing it
2983 * with the lock held can cause deadlocks; see schedule() for
2986 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2987 __releases(rq->lock)
2989 struct mm_struct *mm = rq->prev_mm;
2995 * A task struct has one reference for the use as "current".
2996 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2997 * schedule one last time. The schedule call will never return, and
2998 * the scheduled task must drop that reference.
2999 * The test for TASK_DEAD must occur while the runqueue locks are
3000 * still held, otherwise prev could be scheduled on another cpu, die
3001 * there before we look at prev->state, and then the reference would
3003 * Manfred Spraul <manfred@colorfullife.com>
3005 prev_state = prev->state;
3006 finish_arch_switch(prev);
3007 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3008 local_irq_disable();
3009 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3010 perf_event_task_sched_in(current);
3011 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3013 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3014 finish_lock_switch(rq, prev);
3016 fire_sched_in_preempt_notifiers(current);
3019 if (unlikely(prev_state == TASK_DEAD)) {
3021 * Remove function-return probe instances associated with this
3022 * task and put them back on the free list.
3024 kprobe_flush_task(prev);
3025 put_task_struct(prev);
3031 /* assumes rq->lock is held */
3032 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
3034 if (prev->sched_class->pre_schedule)
3035 prev->sched_class->pre_schedule(rq, prev);
3038 /* rq->lock is NOT held, but preemption is disabled */
3039 static inline void post_schedule(struct rq *rq)
3041 if (rq->post_schedule) {
3042 unsigned long flags;
3044 raw_spin_lock_irqsave(&rq->lock, flags);
3045 if (rq->curr->sched_class->post_schedule)
3046 rq->curr->sched_class->post_schedule(rq);
3047 raw_spin_unlock_irqrestore(&rq->lock, flags);
3049 rq->post_schedule = 0;
3055 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
3059 static inline void post_schedule(struct rq *rq)
3066 * schedule_tail - first thing a freshly forked thread must call.
3067 * @prev: the thread we just switched away from.
3069 asmlinkage void schedule_tail(struct task_struct *prev)
3070 __releases(rq->lock)
3072 struct rq *rq = this_rq();
3074 finish_task_switch(rq, prev);
3077 * FIXME: do we need to worry about rq being invalidated by the
3082 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
3083 /* In this case, finish_task_switch does not reenable preemption */
3086 if (current->set_child_tid)
3087 put_user(task_pid_vnr(current), current->set_child_tid);
3091 * context_switch - switch to the new MM and the new
3092 * thread's register state.
3095 context_switch(struct rq *rq, struct task_struct *prev,
3096 struct task_struct *next)
3098 struct mm_struct *mm, *oldmm;
3100 prepare_task_switch(rq, prev, next);
3103 oldmm = prev->active_mm;
3105 * For paravirt, this is coupled with an exit in switch_to to
3106 * combine the page table reload and the switch backend into
3109 arch_start_context_switch(prev);
3112 next->active_mm = oldmm;
3113 atomic_inc(&oldmm->mm_count);
3114 enter_lazy_tlb(oldmm, next);
3116 switch_mm(oldmm, mm, next);
3119 prev->active_mm = NULL;
3120 rq->prev_mm = oldmm;
3123 * Since the runqueue lock will be released by the next
3124 * task (which is an invalid locking op but in the case
3125 * of the scheduler it's an obvious special-case), so we
3126 * do an early lockdep release here:
3128 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
3129 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3132 /* Here we just switch the register state and the stack. */
3133 switch_to(prev, next, prev);
3137 * this_rq must be evaluated again because prev may have moved
3138 * CPUs since it called schedule(), thus the 'rq' on its stack
3139 * frame will be invalid.
3141 finish_task_switch(this_rq(), prev);
3145 * nr_running, nr_uninterruptible and nr_context_switches:
3147 * externally visible scheduler statistics: current number of runnable
3148 * threads, current number of uninterruptible-sleeping threads, total
3149 * number of context switches performed since bootup.
3151 unsigned long nr_running(void)
3153 unsigned long i, sum = 0;
3155 for_each_online_cpu(i)
3156 sum += cpu_rq(i)->nr_running;
3161 unsigned long nr_uninterruptible(void)
3163 unsigned long i, sum = 0;
3165 for_each_possible_cpu(i)
3166 sum += cpu_rq(i)->nr_uninterruptible;
3169 * Since we read the counters lockless, it might be slightly
3170 * inaccurate. Do not allow it to go below zero though:
3172 if (unlikely((long)sum < 0))
3178 unsigned long long nr_context_switches(void)
3181 unsigned long long sum = 0;
3183 for_each_possible_cpu(i)
3184 sum += cpu_rq(i)->nr_switches;
3189 unsigned long nr_iowait(void)
3191 unsigned long i, sum = 0;
3193 for_each_possible_cpu(i)
3194 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3199 unsigned long nr_iowait_cpu(int cpu)
3201 struct rq *this = cpu_rq(cpu);
3202 return atomic_read(&this->nr_iowait);
3205 unsigned long this_cpu_load(void)
3207 struct rq *this = this_rq();
3208 return this->cpu_load[0];
3212 /* Variables and functions for calc_load */
3213 static atomic_long_t calc_load_tasks;
3214 static unsigned long calc_load_update;
3215 unsigned long avenrun[3];
3216 EXPORT_SYMBOL(avenrun);
3218 static long calc_load_fold_active(struct rq *this_rq)
3220 long nr_active, delta = 0;
3222 nr_active = this_rq->nr_running;
3223 nr_active += (long) this_rq->nr_uninterruptible;
3225 if (nr_active != this_rq->calc_load_active) {
3226 delta = nr_active - this_rq->calc_load_active;
3227 this_rq->calc_load_active = nr_active;
3233 static unsigned long
3234 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3237 load += active * (FIXED_1 - exp);
3238 load += 1UL << (FSHIFT - 1);
3239 return load >> FSHIFT;
3244 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
3246 * When making the ILB scale, we should try to pull this in as well.
3248 static atomic_long_t calc_load_tasks_idle;
3250 static void calc_load_account_idle(struct rq *this_rq)
3254 delta = calc_load_fold_active(this_rq);
3256 atomic_long_add(delta, &calc_load_tasks_idle);
3259 static long calc_load_fold_idle(void)
3264 * Its got a race, we don't care...
3266 if (atomic_long_read(&calc_load_tasks_idle))
3267 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
3273 * fixed_power_int - compute: x^n, in O(log n) time
3275 * @x: base of the power
3276 * @frac_bits: fractional bits of @x
3277 * @n: power to raise @x to.
3279 * By exploiting the relation between the definition of the natural power
3280 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
3281 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
3282 * (where: n_i \elem {0, 1}, the binary vector representing n),
3283 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
3284 * of course trivially computable in O(log_2 n), the length of our binary
3287 static unsigned long
3288 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
3290 unsigned long result = 1UL << frac_bits;
3295 result += 1UL << (frac_bits - 1);
3296 result >>= frac_bits;
3302 x += 1UL << (frac_bits - 1);
3310 * a1 = a0 * e + a * (1 - e)
3312 * a2 = a1 * e + a * (1 - e)
3313 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
3314 * = a0 * e^2 + a * (1 - e) * (1 + e)
3316 * a3 = a2 * e + a * (1 - e)
3317 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
3318 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
3322 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
3323 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
3324 * = a0 * e^n + a * (1 - e^n)
3326 * [1] application of the geometric series:
3329 * S_n := \Sum x^i = -------------
3332 static unsigned long
3333 calc_load_n(unsigned long load, unsigned long exp,
3334 unsigned long active, unsigned int n)
3337 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
3341 * NO_HZ can leave us missing all per-cpu ticks calling
3342 * calc_load_account_active(), but since an idle CPU folds its delta into
3343 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
3344 * in the pending idle delta if our idle period crossed a load cycle boundary.
3346 * Once we've updated the global active value, we need to apply the exponential
3347 * weights adjusted to the number of cycles missed.
3349 static void calc_global_nohz(unsigned long ticks)
3351 long delta, active, n;
3353 if (time_before(jiffies, calc_load_update))
3357 * If we crossed a calc_load_update boundary, make sure to fold
3358 * any pending idle changes, the respective CPUs might have
3359 * missed the tick driven calc_load_account_active() update
3362 delta = calc_load_fold_idle();
3364 atomic_long_add(delta, &calc_load_tasks);
3367 * If we were idle for multiple load cycles, apply them.
3369 if (ticks >= LOAD_FREQ) {
3370 n = ticks / LOAD_FREQ;
3372 active = atomic_long_read(&calc_load_tasks);
3373 active = active > 0 ? active * FIXED_1 : 0;
3375 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
3376 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
3377 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
3379 calc_load_update += n * LOAD_FREQ;
3383 * Its possible the remainder of the above division also crosses
3384 * a LOAD_FREQ period, the regular check in calc_global_load()
3385 * which comes after this will take care of that.
3387 * Consider us being 11 ticks before a cycle completion, and us
3388 * sleeping for 4*LOAD_FREQ + 22 ticks, then the above code will
3389 * age us 4 cycles, and the test in calc_global_load() will
3390 * pick up the final one.
3394 static void calc_load_account_idle(struct rq *this_rq)
3398 static inline long calc_load_fold_idle(void)
3403 static void calc_global_nohz(unsigned long ticks)
3409 * get_avenrun - get the load average array
3410 * @loads: pointer to dest load array
3411 * @offset: offset to add
3412 * @shift: shift count to shift the result left
3414 * These values are estimates at best, so no need for locking.
3416 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3418 loads[0] = (avenrun[0] + offset) << shift;
3419 loads[1] = (avenrun[1] + offset) << shift;
3420 loads[2] = (avenrun[2] + offset) << shift;
3424 * calc_load - update the avenrun load estimates 10 ticks after the
3425 * CPUs have updated calc_load_tasks.
3427 void calc_global_load(unsigned long ticks)
3431 calc_global_nohz(ticks);
3433 if (time_before(jiffies, calc_load_update + 10))
3436 active = atomic_long_read(&calc_load_tasks);
3437 active = active > 0 ? active * FIXED_1 : 0;
3439 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3440 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3441 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3443 calc_load_update += LOAD_FREQ;
3447 * Called from update_cpu_load() to periodically update this CPU's
3450 static void calc_load_account_active(struct rq *this_rq)
3454 if (time_before(jiffies, this_rq->calc_load_update))
3457 delta = calc_load_fold_active(this_rq);
3458 delta += calc_load_fold_idle();
3460 atomic_long_add(delta, &calc_load_tasks);
3462 this_rq->calc_load_update += LOAD_FREQ;
3466 * The exact cpuload at various idx values, calculated at every tick would be
3467 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3469 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3470 * on nth tick when cpu may be busy, then we have:
3471 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3472 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3474 * decay_load_missed() below does efficient calculation of
3475 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3476 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3478 * The calculation is approximated on a 128 point scale.
3479 * degrade_zero_ticks is the number of ticks after which load at any
3480 * particular idx is approximated to be zero.
3481 * degrade_factor is a precomputed table, a row for each load idx.
3482 * Each column corresponds to degradation factor for a power of two ticks,
3483 * based on 128 point scale.
3485 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3486 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3488 * With this power of 2 load factors, we can degrade the load n times
3489 * by looking at 1 bits in n and doing as many mult/shift instead of
3490 * n mult/shifts needed by the exact degradation.
3492 #define DEGRADE_SHIFT 7
3493 static const unsigned char
3494 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3495 static const unsigned char
3496 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3497 {0, 0, 0, 0, 0, 0, 0, 0},
3498 {64, 32, 8, 0, 0, 0, 0, 0},
3499 {96, 72, 40, 12, 1, 0, 0},
3500 {112, 98, 75, 43, 15, 1, 0},
3501 {120, 112, 98, 76, 45, 16, 2} };
3504 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3505 * would be when CPU is idle and so we just decay the old load without
3506 * adding any new load.
3508 static unsigned long
3509 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3513 if (!missed_updates)
3516 if (missed_updates >= degrade_zero_ticks[idx])
3520 return load >> missed_updates;
3522 while (missed_updates) {
3523 if (missed_updates % 2)
3524 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3526 missed_updates >>= 1;
3533 * Update rq->cpu_load[] statistics. This function is usually called every
3534 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3535 * every tick. We fix it up based on jiffies.
3537 static void update_cpu_load(struct rq *this_rq)
3539 unsigned long this_load = this_rq->load.weight;
3540 unsigned long curr_jiffies = jiffies;
3541 unsigned long pending_updates;
3544 this_rq->nr_load_updates++;
3546 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3547 if (curr_jiffies == this_rq->last_load_update_tick)
3550 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3551 this_rq->last_load_update_tick = curr_jiffies;
3553 /* Update our load: */
3554 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3555 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3556 unsigned long old_load, new_load;
3558 /* scale is effectively 1 << i now, and >> i divides by scale */
3560 old_load = this_rq->cpu_load[i];
3561 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3562 new_load = this_load;
3564 * Round up the averaging division if load is increasing. This
3565 * prevents us from getting stuck on 9 if the load is 10, for
3568 if (new_load > old_load)
3569 new_load += scale - 1;
3571 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3574 sched_avg_update(this_rq);
3577 static void update_cpu_load_active(struct rq *this_rq)
3579 update_cpu_load(this_rq);
3581 calc_load_account_active(this_rq);
3587 * sched_exec - execve() is a valuable balancing opportunity, because at
3588 * this point the task has the smallest effective memory and cache footprint.
3590 void sched_exec(void)
3592 struct task_struct *p = current;
3593 unsigned long flags;
3596 raw_spin_lock_irqsave(&p->pi_lock, flags);
3597 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
3598 if (dest_cpu == smp_processor_id())
3601 if (likely(cpu_active(dest_cpu))) {
3602 struct migration_arg arg = { p, dest_cpu };
3604 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3605 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
3609 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3614 DEFINE_PER_CPU(struct kernel_stat, kstat);
3616 EXPORT_PER_CPU_SYMBOL(kstat);
3619 * Return any ns on the sched_clock that have not yet been accounted in
3620 * @p in case that task is currently running.
3622 * Called with task_rq_lock() held on @rq.
3624 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3628 if (task_current(rq, p)) {
3629 update_rq_clock(rq);
3630 ns = rq->clock_task - p->se.exec_start;
3638 unsigned long long task_delta_exec(struct task_struct *p)
3640 unsigned long flags;
3644 rq = task_rq_lock(p, &flags);
3645 ns = do_task_delta_exec(p, rq);
3646 task_rq_unlock(rq, p, &flags);
3652 * Return accounted runtime for the task.
3653 * In case the task is currently running, return the runtime plus current's
3654 * pending runtime that have not been accounted yet.
3656 unsigned long long task_sched_runtime(struct task_struct *p)
3658 unsigned long flags;
3662 rq = task_rq_lock(p, &flags);
3663 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3664 task_rq_unlock(rq, p, &flags);
3670 * Return sum_exec_runtime for the thread group.
3671 * In case the task is currently running, return the sum plus current's
3672 * pending runtime that have not been accounted yet.
3674 * Note that the thread group might have other running tasks as well,
3675 * so the return value not includes other pending runtime that other
3676 * running tasks might have.
3678 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3680 struct task_cputime totals;
3681 unsigned long flags;
3685 rq = task_rq_lock(p, &flags);
3686 thread_group_cputime(p, &totals);
3687 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3688 task_rq_unlock(rq, p, &flags);
3694 * Account user cpu time to a process.
3695 * @p: the process that the cpu time gets accounted to
3696 * @cputime: the cpu time spent in user space since the last update
3697 * @cputime_scaled: cputime scaled by cpu frequency
3699 void account_user_time(struct task_struct *p, cputime_t cputime,
3700 cputime_t cputime_scaled)
3702 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3705 /* Add user time to process. */
3706 p->utime = cputime_add(p->utime, cputime);
3707 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3708 account_group_user_time(p, cputime);
3710 /* Add user time to cpustat. */
3711 tmp = cputime_to_cputime64(cputime);
3712 if (TASK_NICE(p) > 0)
3713 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3715 cpustat->user = cputime64_add(cpustat->user, tmp);
3717 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3718 /* Account for user time used */
3719 acct_update_integrals(p);
3723 * Account guest cpu time to a process.
3724 * @p: the process that the cpu time gets accounted to
3725 * @cputime: the cpu time spent in virtual machine since the last update
3726 * @cputime_scaled: cputime scaled by cpu frequency
3728 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3729 cputime_t cputime_scaled)
3732 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3734 tmp = cputime_to_cputime64(cputime);
3736 /* Add guest time to process. */
3737 p->utime = cputime_add(p->utime, cputime);
3738 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3739 account_group_user_time(p, cputime);
3740 p->gtime = cputime_add(p->gtime, cputime);
3742 /* Add guest time to cpustat. */
3743 if (TASK_NICE(p) > 0) {
3744 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3745 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3747 cpustat->user = cputime64_add(cpustat->user, tmp);
3748 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3753 * Account system cpu time to a process and desired cpustat field
3754 * @p: the process that the cpu time gets accounted to
3755 * @cputime: the cpu time spent in kernel space since the last update
3756 * @cputime_scaled: cputime scaled by cpu frequency
3757 * @target_cputime64: pointer to cpustat field that has to be updated
3760 void __account_system_time(struct task_struct *p, cputime_t cputime,
3761 cputime_t cputime_scaled, cputime64_t *target_cputime64)
3763 cputime64_t tmp = cputime_to_cputime64(cputime);
3765 /* Add system time to process. */
3766 p->stime = cputime_add(p->stime, cputime);
3767 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3768 account_group_system_time(p, cputime);
3770 /* Add system time to cpustat. */
3771 *target_cputime64 = cputime64_add(*target_cputime64, tmp);
3772 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3774 /* Account for system time used */
3775 acct_update_integrals(p);
3779 * Account system cpu time to a process.
3780 * @p: the process that the cpu time gets accounted to
3781 * @hardirq_offset: the offset to subtract from hardirq_count()
3782 * @cputime: the cpu time spent in kernel space since the last update
3783 * @cputime_scaled: cputime scaled by cpu frequency
3785 void account_system_time(struct task_struct *p, int hardirq_offset,
3786 cputime_t cputime, cputime_t cputime_scaled)
3788 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3789 cputime64_t *target_cputime64;
3791 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3792 account_guest_time(p, cputime, cputime_scaled);
3796 if (hardirq_count() - hardirq_offset)
3797 target_cputime64 = &cpustat->irq;
3798 else if (in_serving_softirq())
3799 target_cputime64 = &cpustat->softirq;
3801 target_cputime64 = &cpustat->system;
3803 __account_system_time(p, cputime, cputime_scaled, target_cputime64);
3807 * Account for involuntary wait time.
3808 * @cputime: the cpu time spent in involuntary wait
3810 void account_steal_time(cputime_t cputime)
3812 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3813 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3815 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3819 * Account for idle time.
3820 * @cputime: the cpu time spent in idle wait
3822 void account_idle_time(cputime_t cputime)
3824 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3825 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3826 struct rq *rq = this_rq();
3828 if (atomic_read(&rq->nr_iowait) > 0)
3829 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3831 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3834 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3836 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
3838 * Account a tick to a process and cpustat
3839 * @p: the process that the cpu time gets accounted to
3840 * @user_tick: is the tick from userspace
3841 * @rq: the pointer to rq
3843 * Tick demultiplexing follows the order
3844 * - pending hardirq update
3845 * - pending softirq update
3849 * - check for guest_time
3850 * - else account as system_time
3852 * Check for hardirq is done both for system and user time as there is
3853 * no timer going off while we are on hardirq and hence we may never get an
3854 * opportunity to update it solely in system time.
3855 * p->stime and friends are only updated on system time and not on irq
3856 * softirq as those do not count in task exec_runtime any more.
3858 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
3861 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3862 cputime64_t tmp = cputime_to_cputime64(cputime_one_jiffy);
3863 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3865 if (irqtime_account_hi_update()) {
3866 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3867 } else if (irqtime_account_si_update()) {
3868 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3869 } else if (this_cpu_ksoftirqd() == p) {
3871 * ksoftirqd time do not get accounted in cpu_softirq_time.
3872 * So, we have to handle it separately here.
3873 * Also, p->stime needs to be updated for ksoftirqd.
3875 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
3877 } else if (user_tick) {
3878 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3879 } else if (p == rq->idle) {
3880 account_idle_time(cputime_one_jiffy);
3881 } else if (p->flags & PF_VCPU) { /* System time or guest time */
3882 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled);
3884 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
3889 static void irqtime_account_idle_ticks(int ticks)
3892 struct rq *rq = this_rq();
3894 for (i = 0; i < ticks; i++)
3895 irqtime_account_process_tick(current, 0, rq);
3897 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
3898 static void irqtime_account_idle_ticks(int ticks) {}
3899 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
3901 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
3904 * Account a single tick of cpu time.
3905 * @p: the process that the cpu time gets accounted to
3906 * @user_tick: indicates if the tick is a user or a system tick
3908 void account_process_tick(struct task_struct *p, int user_tick)
3910 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3911 struct rq *rq = this_rq();
3913 if (sched_clock_irqtime) {
3914 irqtime_account_process_tick(p, user_tick, rq);
3919 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3920 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3921 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3924 account_idle_time(cputime_one_jiffy);
3928 * Account multiple ticks of steal time.
3929 * @p: the process from which the cpu time has been stolen
3930 * @ticks: number of stolen ticks
3932 void account_steal_ticks(unsigned long ticks)
3934 account_steal_time(jiffies_to_cputime(ticks));
3938 * Account multiple ticks of idle time.
3939 * @ticks: number of stolen ticks
3941 void account_idle_ticks(unsigned long ticks)
3944 if (sched_clock_irqtime) {
3945 irqtime_account_idle_ticks(ticks);
3949 account_idle_time(jiffies_to_cputime(ticks));
3955 * Use precise platform statistics if available:
3957 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3958 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3964 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3966 struct task_cputime cputime;
3968 thread_group_cputime(p, &cputime);
3970 *ut = cputime.utime;
3971 *st = cputime.stime;
3975 #ifndef nsecs_to_cputime
3976 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3979 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3981 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
3984 * Use CFS's precise accounting:
3986 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3992 do_div(temp, total);
3993 utime = (cputime_t)temp;
3998 * Compare with previous values, to keep monotonicity:
4000 p->prev_utime = max(p->prev_utime, utime);
4001 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
4003 *ut = p->prev_utime;
4004 *st = p->prev_stime;
4008 * Must be called with siglock held.
4010 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4012 struct signal_struct *sig = p->signal;
4013 struct task_cputime cputime;
4014 cputime_t rtime, utime, total;
4016 thread_group_cputime(p, &cputime);
4018 total = cputime_add(cputime.utime, cputime.stime);
4019 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
4024 temp *= cputime.utime;
4025 do_div(temp, total);
4026 utime = (cputime_t)temp;
4030 sig->prev_utime = max(sig->prev_utime, utime);
4031 sig->prev_stime = max(sig->prev_stime,
4032 cputime_sub(rtime, sig->prev_utime));
4034 *ut = sig->prev_utime;
4035 *st = sig->prev_stime;
4040 * This function gets called by the timer code, with HZ frequency.
4041 * We call it with interrupts disabled.
4043 void scheduler_tick(void)
4045 int cpu = smp_processor_id();
4046 struct rq *rq = cpu_rq(cpu);
4047 struct task_struct *curr = rq->curr;
4051 raw_spin_lock(&rq->lock);
4052 update_rq_clock(rq);
4053 update_cpu_load_active(rq);
4054 curr->sched_class->task_tick(rq, curr, 0);
4055 raw_spin_unlock(&rq->lock);
4057 perf_event_task_tick();
4060 rq->idle_at_tick = idle_cpu(cpu);
4061 trigger_load_balance(rq, cpu);
4065 notrace unsigned long get_parent_ip(unsigned long addr)
4067 if (in_lock_functions(addr)) {
4068 addr = CALLER_ADDR2;
4069 if (in_lock_functions(addr))
4070 addr = CALLER_ADDR3;
4075 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4076 defined(CONFIG_PREEMPT_TRACER))
4078 void __kprobes add_preempt_count(int val)
4080 #ifdef CONFIG_DEBUG_PREEMPT
4084 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4087 preempt_count() += val;
4088 #ifdef CONFIG_DEBUG_PREEMPT
4090 * Spinlock count overflowing soon?
4092 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4095 if (preempt_count() == val)
4096 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4098 EXPORT_SYMBOL(add_preempt_count);
4100 void __kprobes sub_preempt_count(int val)
4102 #ifdef CONFIG_DEBUG_PREEMPT
4106 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4109 * Is the spinlock portion underflowing?
4111 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4112 !(preempt_count() & PREEMPT_MASK)))
4116 if (preempt_count() == val)
4117 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4118 preempt_count() -= val;
4120 EXPORT_SYMBOL(sub_preempt_count);
4125 * Print scheduling while atomic bug:
4127 static noinline void __schedule_bug(struct task_struct *prev)
4129 struct pt_regs *regs = get_irq_regs();
4131 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4132 prev->comm, prev->pid, preempt_count());
4134 debug_show_held_locks(prev);
4136 if (irqs_disabled())
4137 print_irqtrace_events(prev);
4146 * Various schedule()-time debugging checks and statistics:
4148 static inline void schedule_debug(struct task_struct *prev)
4151 * Test if we are atomic. Since do_exit() needs to call into
4152 * schedule() atomically, we ignore that path for now.
4153 * Otherwise, whine if we are scheduling when we should not be.
4155 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4156 __schedule_bug(prev);
4158 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4160 schedstat_inc(this_rq(), sched_count);
4163 static void put_prev_task(struct rq *rq, struct task_struct *prev)
4165 if (prev->on_rq || rq->skip_clock_update < 0)
4166 update_rq_clock(rq);
4167 prev->sched_class->put_prev_task(rq, prev);
4171 * Pick up the highest-prio task:
4173 static inline struct task_struct *
4174 pick_next_task(struct rq *rq)
4176 const struct sched_class *class;
4177 struct task_struct *p;
4180 * Optimization: we know that if all tasks are in
4181 * the fair class we can call that function directly:
4183 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4184 p = fair_sched_class.pick_next_task(rq);
4189 for_each_class(class) {
4190 p = class->pick_next_task(rq);
4195 BUG(); /* the idle class will always have a runnable task */
4199 * schedule() is the main scheduler function.
4201 asmlinkage void __sched schedule(void)
4203 struct task_struct *prev, *next;
4204 unsigned long *switch_count;
4210 cpu = smp_processor_id();
4212 rcu_note_context_switch(cpu);
4215 schedule_debug(prev);
4217 if (sched_feat(HRTICK))
4220 raw_spin_lock_irq(&rq->lock);
4222 switch_count = &prev->nivcsw;
4223 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4224 if (unlikely(signal_pending_state(prev->state, prev))) {
4225 prev->state = TASK_RUNNING;
4227 deactivate_task(rq, prev, DEQUEUE_SLEEP);
4231 * If a worker went to sleep, notify and ask workqueue
4232 * whether it wants to wake up a task to maintain
4235 if (prev->flags & PF_WQ_WORKER) {
4236 struct task_struct *to_wakeup;
4238 to_wakeup = wq_worker_sleeping(prev, cpu);
4240 try_to_wake_up_local(to_wakeup);
4244 * If we are going to sleep and we have plugged IO
4245 * queued, make sure to submit it to avoid deadlocks.
4247 if (blk_needs_flush_plug(prev)) {
4248 raw_spin_unlock(&rq->lock);
4249 blk_schedule_flush_plug(prev);
4250 raw_spin_lock(&rq->lock);
4253 switch_count = &prev->nvcsw;
4256 pre_schedule(rq, prev);
4258 if (unlikely(!rq->nr_running))
4259 idle_balance(cpu, rq);
4261 put_prev_task(rq, prev);
4262 next = pick_next_task(rq);
4263 clear_tsk_need_resched(prev);
4264 rq->skip_clock_update = 0;
4266 if (likely(prev != next)) {
4271 context_switch(rq, prev, next); /* unlocks the rq */
4273 * The context switch have flipped the stack from under us
4274 * and restored the local variables which were saved when
4275 * this task called schedule() in the past. prev == current
4276 * is still correct, but it can be moved to another cpu/rq.
4278 cpu = smp_processor_id();
4281 raw_spin_unlock_irq(&rq->lock);
4285 preempt_enable_no_resched();
4289 EXPORT_SYMBOL(schedule);
4291 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
4293 static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
4298 if (lock->owner != owner)
4302 * Ensure we emit the owner->on_cpu, dereference _after_ checking
4303 * lock->owner still matches owner, if that fails, owner might
4304 * point to free()d memory, if it still matches, the rcu_read_lock()
4305 * ensures the memory stays valid.
4309 ret = owner->on_cpu;
4317 * Look out! "owner" is an entirely speculative pointer
4318 * access and not reliable.
4320 int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
4322 if (!sched_feat(OWNER_SPIN))
4325 while (owner_running(lock, owner)) {
4329 arch_mutex_cpu_relax();
4333 * If the owner changed to another task there is likely
4334 * heavy contention, stop spinning.
4343 #ifdef CONFIG_PREEMPT
4345 * this is the entry point to schedule() from in-kernel preemption
4346 * off of preempt_enable. Kernel preemptions off return from interrupt
4347 * occur there and call schedule directly.
4349 asmlinkage void __sched notrace preempt_schedule(void)
4351 struct thread_info *ti = current_thread_info();
4354 * If there is a non-zero preempt_count or interrupts are disabled,
4355 * we do not want to preempt the current task. Just return..
4357 if (likely(ti->preempt_count || irqs_disabled()))
4361 add_preempt_count_notrace(PREEMPT_ACTIVE);
4363 sub_preempt_count_notrace(PREEMPT_ACTIVE);
4366 * Check again in case we missed a preemption opportunity
4367 * between schedule and now.
4370 } while (need_resched());
4372 EXPORT_SYMBOL(preempt_schedule);
4375 * this is the entry point to schedule() from kernel preemption
4376 * off of irq context.
4377 * Note, that this is called and return with irqs disabled. This will
4378 * protect us against recursive calling from irq.
4380 asmlinkage void __sched preempt_schedule_irq(void)
4382 struct thread_info *ti = current_thread_info();
4384 /* Catch callers which need to be fixed */
4385 BUG_ON(ti->preempt_count || !irqs_disabled());
4388 add_preempt_count(PREEMPT_ACTIVE);
4391 local_irq_disable();
4392 sub_preempt_count(PREEMPT_ACTIVE);
4395 * Check again in case we missed a preemption opportunity
4396 * between schedule and now.
4399 } while (need_resched());
4402 #endif /* CONFIG_PREEMPT */
4404 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
4407 return try_to_wake_up(curr->private, mode, wake_flags);
4409 EXPORT_SYMBOL(default_wake_function);
4412 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4413 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4414 * number) then we wake all the non-exclusive tasks and one exclusive task.
4416 * There are circumstances in which we can try to wake a task which has already
4417 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4418 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4420 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4421 int nr_exclusive, int wake_flags, void *key)
4423 wait_queue_t *curr, *next;
4425 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4426 unsigned flags = curr->flags;
4428 if (curr->func(curr, mode, wake_flags, key) &&
4429 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4435 * __wake_up - wake up threads blocked on a waitqueue.
4437 * @mode: which threads
4438 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4439 * @key: is directly passed to the wakeup function
4441 * It may be assumed that this function implies a write memory barrier before
4442 * changing the task state if and only if any tasks are woken up.
4444 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4445 int nr_exclusive, void *key)
4447 unsigned long flags;
4449 spin_lock_irqsave(&q->lock, flags);
4450 __wake_up_common(q, mode, nr_exclusive, 0, key);
4451 spin_unlock_irqrestore(&q->lock, flags);
4453 EXPORT_SYMBOL(__wake_up);
4456 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4458 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4460 __wake_up_common(q, mode, 1, 0, NULL);
4462 EXPORT_SYMBOL_GPL(__wake_up_locked);
4464 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
4466 __wake_up_common(q, mode, 1, 0, key);
4468 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
4471 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4473 * @mode: which threads
4474 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4475 * @key: opaque value to be passed to wakeup targets
4477 * The sync wakeup differs that the waker knows that it will schedule
4478 * away soon, so while the target thread will be woken up, it will not
4479 * be migrated to another CPU - ie. the two threads are 'synchronized'
4480 * with each other. This can prevent needless bouncing between CPUs.
4482 * On UP it can prevent extra preemption.
4484 * It may be assumed that this function implies a write memory barrier before
4485 * changing the task state if and only if any tasks are woken up.
4487 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4488 int nr_exclusive, void *key)
4490 unsigned long flags;
4491 int wake_flags = WF_SYNC;
4496 if (unlikely(!nr_exclusive))
4499 spin_lock_irqsave(&q->lock, flags);
4500 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4501 spin_unlock_irqrestore(&q->lock, flags);
4503 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4506 * __wake_up_sync - see __wake_up_sync_key()
4508 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4510 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4512 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4515 * complete: - signals a single thread waiting on this completion
4516 * @x: holds the state of this particular completion
4518 * This will wake up a single thread waiting on this completion. Threads will be
4519 * awakened in the same order in which they were queued.
4521 * See also complete_all(), wait_for_completion() and related routines.
4523 * It may be assumed that this function implies a write memory barrier before
4524 * changing the task state if and only if any tasks are woken up.
4526 void complete(struct completion *x)
4528 unsigned long flags;
4530 spin_lock_irqsave(&x->wait.lock, flags);
4532 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4533 spin_unlock_irqrestore(&x->wait.lock, flags);
4535 EXPORT_SYMBOL(complete);
4538 * complete_all: - signals all threads waiting on this completion
4539 * @x: holds the state of this particular completion
4541 * This will wake up all threads waiting on this particular completion event.
4543 * It may be assumed that this function implies a write memory barrier before
4544 * changing the task state if and only if any tasks are woken up.
4546 void complete_all(struct completion *x)
4548 unsigned long flags;
4550 spin_lock_irqsave(&x->wait.lock, flags);
4551 x->done += UINT_MAX/2;
4552 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4553 spin_unlock_irqrestore(&x->wait.lock, flags);
4555 EXPORT_SYMBOL(complete_all);
4557 static inline long __sched
4558 do_wait_for_common(struct completion *x, long timeout, int state)
4561 DECLARE_WAITQUEUE(wait, current);
4563 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4565 if (signal_pending_state(state, current)) {
4566 timeout = -ERESTARTSYS;
4569 __set_current_state(state);
4570 spin_unlock_irq(&x->wait.lock);
4571 timeout = schedule_timeout(timeout);
4572 spin_lock_irq(&x->wait.lock);
4573 } while (!x->done && timeout);
4574 __remove_wait_queue(&x->wait, &wait);
4579 return timeout ?: 1;
4583 wait_for_common(struct completion *x, long timeout, int state)
4587 spin_lock_irq(&x->wait.lock);
4588 timeout = do_wait_for_common(x, timeout, state);
4589 spin_unlock_irq(&x->wait.lock);
4594 * wait_for_completion: - waits for completion of a task
4595 * @x: holds the state of this particular completion
4597 * This waits to be signaled for completion of a specific task. It is NOT
4598 * interruptible and there is no timeout.
4600 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4601 * and interrupt capability. Also see complete().
4603 void __sched wait_for_completion(struct completion *x)
4605 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4607 EXPORT_SYMBOL(wait_for_completion);
4610 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
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. The timeout is in jiffies. It is not
4618 unsigned long __sched
4619 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4621 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4623 EXPORT_SYMBOL(wait_for_completion_timeout);
4626 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4627 * @x: holds the state of this particular completion
4629 * This waits for completion of a specific task to be signaled. It is
4632 int __sched wait_for_completion_interruptible(struct completion *x)
4634 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4635 if (t == -ERESTARTSYS)
4639 EXPORT_SYMBOL(wait_for_completion_interruptible);
4642 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
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 signaled or for a
4647 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4650 wait_for_completion_interruptible_timeout(struct completion *x,
4651 unsigned long timeout)
4653 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4655 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4658 * wait_for_completion_killable: - waits for completion of a task (killable)
4659 * @x: holds the state of this particular completion
4661 * This waits to be signaled for completion of a specific task. It can be
4662 * interrupted by a kill signal.
4664 int __sched wait_for_completion_killable(struct completion *x)
4666 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4667 if (t == -ERESTARTSYS)
4671 EXPORT_SYMBOL(wait_for_completion_killable);
4674 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4675 * @x: holds the state of this particular completion
4676 * @timeout: timeout value in jiffies
4678 * This waits for either a completion of a specific task to be
4679 * signaled or for a specified timeout to expire. It can be
4680 * interrupted by a kill signal. The timeout is in jiffies.
4683 wait_for_completion_killable_timeout(struct completion *x,
4684 unsigned long timeout)
4686 return wait_for_common(x, timeout, TASK_KILLABLE);
4688 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
4691 * try_wait_for_completion - try to decrement a completion without blocking
4692 * @x: completion structure
4694 * Returns: 0 if a decrement cannot be done without blocking
4695 * 1 if a decrement succeeded.
4697 * If a completion is being used as a counting completion,
4698 * attempt to decrement the counter without blocking. This
4699 * enables us to avoid waiting if the resource the completion
4700 * is protecting is not available.
4702 bool try_wait_for_completion(struct completion *x)
4704 unsigned long flags;
4707 spin_lock_irqsave(&x->wait.lock, flags);
4712 spin_unlock_irqrestore(&x->wait.lock, flags);
4715 EXPORT_SYMBOL(try_wait_for_completion);
4718 * completion_done - Test to see if a completion has any waiters
4719 * @x: completion structure
4721 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4722 * 1 if there are no waiters.
4725 bool completion_done(struct completion *x)
4727 unsigned long flags;
4730 spin_lock_irqsave(&x->wait.lock, flags);
4733 spin_unlock_irqrestore(&x->wait.lock, flags);
4736 EXPORT_SYMBOL(completion_done);
4739 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4741 unsigned long flags;
4744 init_waitqueue_entry(&wait, current);
4746 __set_current_state(state);
4748 spin_lock_irqsave(&q->lock, flags);
4749 __add_wait_queue(q, &wait);
4750 spin_unlock(&q->lock);
4751 timeout = schedule_timeout(timeout);
4752 spin_lock_irq(&q->lock);
4753 __remove_wait_queue(q, &wait);
4754 spin_unlock_irqrestore(&q->lock, flags);
4759 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4761 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4763 EXPORT_SYMBOL(interruptible_sleep_on);
4766 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4768 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4770 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4772 void __sched sleep_on(wait_queue_head_t *q)
4774 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4776 EXPORT_SYMBOL(sleep_on);
4778 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4780 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4782 EXPORT_SYMBOL(sleep_on_timeout);
4784 #ifdef CONFIG_RT_MUTEXES
4787 * rt_mutex_setprio - set the current priority of a task
4789 * @prio: prio value (kernel-internal form)
4791 * This function changes the 'effective' priority of a task. It does
4792 * not touch ->normal_prio like __setscheduler().
4794 * Used by the rt_mutex code to implement priority inheritance logic.
4796 void rt_mutex_setprio(struct task_struct *p, int prio)
4798 int oldprio, on_rq, running;
4800 const struct sched_class *prev_class;
4802 BUG_ON(prio < 0 || prio > MAX_PRIO);
4804 rq = __task_rq_lock(p);
4806 trace_sched_pi_setprio(p, prio);
4808 prev_class = p->sched_class;
4810 running = task_current(rq, p);
4812 dequeue_task(rq, p, 0);
4814 p->sched_class->put_prev_task(rq, p);
4817 p->sched_class = &rt_sched_class;
4819 p->sched_class = &fair_sched_class;
4824 p->sched_class->set_curr_task(rq);
4826 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4828 check_class_changed(rq, p, prev_class, oldprio);
4829 __task_rq_unlock(rq);
4834 void set_user_nice(struct task_struct *p, long nice)
4836 int old_prio, delta, on_rq;
4837 unsigned long flags;
4840 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4843 * We have to be careful, if called from sys_setpriority(),
4844 * the task might be in the middle of scheduling on another CPU.
4846 rq = task_rq_lock(p, &flags);
4848 * The RT priorities are set via sched_setscheduler(), but we still
4849 * allow the 'normal' nice value to be set - but as expected
4850 * it wont have any effect on scheduling until the task is
4851 * SCHED_FIFO/SCHED_RR:
4853 if (task_has_rt_policy(p)) {
4854 p->static_prio = NICE_TO_PRIO(nice);
4859 dequeue_task(rq, p, 0);
4861 p->static_prio = NICE_TO_PRIO(nice);
4864 p->prio = effective_prio(p);
4865 delta = p->prio - old_prio;
4868 enqueue_task(rq, p, 0);
4870 * If the task increased its priority or is running and
4871 * lowered its priority, then reschedule its CPU:
4873 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4874 resched_task(rq->curr);
4877 task_rq_unlock(rq, p, &flags);
4879 EXPORT_SYMBOL(set_user_nice);
4882 * can_nice - check if a task can reduce its nice value
4886 int can_nice(const struct task_struct *p, const int nice)
4888 /* convert nice value [19,-20] to rlimit style value [1,40] */
4889 int nice_rlim = 20 - nice;
4891 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4892 capable(CAP_SYS_NICE));
4895 #ifdef __ARCH_WANT_SYS_NICE
4898 * sys_nice - change the priority of the current process.
4899 * @increment: priority increment
4901 * sys_setpriority is a more generic, but much slower function that
4902 * does similar things.
4904 SYSCALL_DEFINE1(nice, int, increment)
4909 * Setpriority might change our priority at the same moment.
4910 * We don't have to worry. Conceptually one call occurs first
4911 * and we have a single winner.
4913 if (increment < -40)
4918 nice = TASK_NICE(current) + increment;
4924 if (increment < 0 && !can_nice(current, nice))
4927 retval = security_task_setnice(current, nice);
4931 set_user_nice(current, nice);
4938 * task_prio - return the priority value of a given task.
4939 * @p: the task in question.
4941 * This is the priority value as seen by users in /proc.
4942 * RT tasks are offset by -200. Normal tasks are centered
4943 * around 0, value goes from -16 to +15.
4945 int task_prio(const struct task_struct *p)
4947 return p->prio - MAX_RT_PRIO;
4951 * task_nice - return the nice value of a given task.
4952 * @p: the task in question.
4954 int task_nice(const struct task_struct *p)
4956 return TASK_NICE(p);
4958 EXPORT_SYMBOL(task_nice);
4961 * idle_cpu - is a given cpu idle currently?
4962 * @cpu: the processor in question.
4964 int idle_cpu(int cpu)
4966 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4970 * idle_task - return the idle task for a given cpu.
4971 * @cpu: the processor in question.
4973 struct task_struct *idle_task(int cpu)
4975 return cpu_rq(cpu)->idle;
4979 * find_process_by_pid - find a process with a matching PID value.
4980 * @pid: the pid in question.
4982 static struct task_struct *find_process_by_pid(pid_t pid)
4984 return pid ? find_task_by_vpid(pid) : current;
4987 /* Actually do priority change: must hold rq lock. */
4989 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4992 p->rt_priority = prio;
4993 p->normal_prio = normal_prio(p);
4994 /* we are holding p->pi_lock already */
4995 p->prio = rt_mutex_getprio(p);
4996 if (rt_prio(p->prio))
4997 p->sched_class = &rt_sched_class;
4999 p->sched_class = &fair_sched_class;
5004 * check the target process has a UID that matches the current process's
5006 static bool check_same_owner(struct task_struct *p)
5008 const struct cred *cred = current_cred(), *pcred;
5012 pcred = __task_cred(p);
5013 if (cred->user->user_ns == pcred->user->user_ns)
5014 match = (cred->euid == pcred->euid ||
5015 cred->euid == pcred->uid);
5022 static int __sched_setscheduler(struct task_struct *p, int policy,
5023 const struct sched_param *param, bool user)
5025 int retval, oldprio, oldpolicy = -1, on_rq, running;
5026 unsigned long flags;
5027 const struct sched_class *prev_class;
5031 /* may grab non-irq protected spin_locks */
5032 BUG_ON(in_interrupt());
5034 /* double check policy once rq lock held */
5036 reset_on_fork = p->sched_reset_on_fork;
5037 policy = oldpolicy = p->policy;
5039 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
5040 policy &= ~SCHED_RESET_ON_FORK;
5042 if (policy != SCHED_FIFO && policy != SCHED_RR &&
5043 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5044 policy != SCHED_IDLE)
5049 * Valid priorities for SCHED_FIFO and SCHED_RR are
5050 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5051 * SCHED_BATCH and SCHED_IDLE is 0.
5053 if (param->sched_priority < 0 ||
5054 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5055 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5057 if (rt_policy(policy) != (param->sched_priority != 0))
5061 * Allow unprivileged RT tasks to decrease priority:
5063 if (user && !capable(CAP_SYS_NICE)) {
5064 if (rt_policy(policy)) {
5065 unsigned long rlim_rtprio =
5066 task_rlimit(p, RLIMIT_RTPRIO);
5068 /* can't set/change the rt policy */
5069 if (policy != p->policy && !rlim_rtprio)
5072 /* can't increase priority */
5073 if (param->sched_priority > p->rt_priority &&
5074 param->sched_priority > rlim_rtprio)
5079 * Treat SCHED_IDLE as nice 20. Only allow a switch to
5080 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
5082 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
5083 if (!can_nice(p, TASK_NICE(p)))
5087 /* can't change other user's priorities */
5088 if (!check_same_owner(p))
5091 /* Normal users shall not reset the sched_reset_on_fork flag */
5092 if (p->sched_reset_on_fork && !reset_on_fork)
5097 retval = security_task_setscheduler(p);
5103 * make sure no PI-waiters arrive (or leave) while we are
5104 * changing the priority of the task:
5106 * To be able to change p->policy safely, the appropriate
5107 * runqueue lock must be held.
5109 rq = task_rq_lock(p, &flags);
5112 * Changing the policy of the stop threads its a very bad idea
5114 if (p == rq->stop) {
5115 task_rq_unlock(rq, p, &flags);
5120 * If not changing anything there's no need to proceed further:
5122 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
5123 param->sched_priority == p->rt_priority))) {
5125 __task_rq_unlock(rq);
5126 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5130 #ifdef CONFIG_RT_GROUP_SCHED
5133 * Do not allow realtime tasks into groups that have no runtime
5136 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5137 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
5138 !task_group_is_autogroup(task_group(p))) {
5139 task_rq_unlock(rq, p, &flags);
5145 /* recheck policy now with rq lock held */
5146 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5147 policy = oldpolicy = -1;
5148 task_rq_unlock(rq, p, &flags);
5152 running = task_current(rq, p);
5154 deactivate_task(rq, p, 0);
5156 p->sched_class->put_prev_task(rq, p);
5158 p->sched_reset_on_fork = reset_on_fork;
5161 prev_class = p->sched_class;
5162 __setscheduler(rq, p, policy, param->sched_priority);
5165 p->sched_class->set_curr_task(rq);
5167 activate_task(rq, p, 0);
5169 check_class_changed(rq, p, prev_class, oldprio);
5170 task_rq_unlock(rq, p, &flags);
5172 rt_mutex_adjust_pi(p);
5178 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5179 * @p: the task in question.
5180 * @policy: new policy.
5181 * @param: structure containing the new RT priority.
5183 * NOTE that the task may be already dead.
5185 int sched_setscheduler(struct task_struct *p, int policy,
5186 const struct sched_param *param)
5188 return __sched_setscheduler(p, policy, param, true);
5190 EXPORT_SYMBOL_GPL(sched_setscheduler);
5193 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5194 * @p: the task in question.
5195 * @policy: new policy.
5196 * @param: structure containing the new RT priority.
5198 * Just like sched_setscheduler, only don't bother checking if the
5199 * current context has permission. For example, this is needed in
5200 * stop_machine(): we create temporary high priority worker threads,
5201 * but our caller might not have that capability.
5203 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5204 const struct sched_param *param)
5206 return __sched_setscheduler(p, policy, param, false);
5210 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5212 struct sched_param lparam;
5213 struct task_struct *p;
5216 if (!param || pid < 0)
5218 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5223 p = find_process_by_pid(pid);
5225 retval = sched_setscheduler(p, policy, &lparam);
5232 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5233 * @pid: the pid in question.
5234 * @policy: new policy.
5235 * @param: structure containing the new RT priority.
5237 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
5238 struct sched_param __user *, param)
5240 /* negative values for policy are not valid */
5244 return do_sched_setscheduler(pid, policy, param);
5248 * sys_sched_setparam - set/change the RT priority of a thread
5249 * @pid: the pid in question.
5250 * @param: structure containing the new RT priority.
5252 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5254 return do_sched_setscheduler(pid, -1, param);
5258 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5259 * @pid: the pid in question.
5261 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5263 struct task_struct *p;
5271 p = find_process_by_pid(pid);
5273 retval = security_task_getscheduler(p);
5276 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
5283 * sys_sched_getparam - get the RT priority of a thread
5284 * @pid: the pid in question.
5285 * @param: structure containing the RT priority.
5287 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5289 struct sched_param lp;
5290 struct task_struct *p;
5293 if (!param || pid < 0)
5297 p = find_process_by_pid(pid);
5302 retval = security_task_getscheduler(p);
5306 lp.sched_priority = p->rt_priority;
5310 * This one might sleep, we cannot do it with a spinlock held ...
5312 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5321 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5323 cpumask_var_t cpus_allowed, new_mask;
5324 struct task_struct *p;
5330 p = find_process_by_pid(pid);
5337 /* Prevent p going away */
5341 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5345 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5347 goto out_free_cpus_allowed;
5350 if (!check_same_owner(p) && !task_ns_capable(p, CAP_SYS_NICE))
5353 retval = security_task_setscheduler(p);
5357 cpuset_cpus_allowed(p, cpus_allowed);
5358 cpumask_and(new_mask, in_mask, cpus_allowed);
5360 retval = set_cpus_allowed_ptr(p, new_mask);
5363 cpuset_cpus_allowed(p, cpus_allowed);
5364 if (!cpumask_subset(new_mask, cpus_allowed)) {
5366 * We must have raced with a concurrent cpuset
5367 * update. Just reset the cpus_allowed to the
5368 * cpuset's cpus_allowed
5370 cpumask_copy(new_mask, cpus_allowed);
5375 free_cpumask_var(new_mask);
5376 out_free_cpus_allowed:
5377 free_cpumask_var(cpus_allowed);
5384 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5385 struct cpumask *new_mask)
5387 if (len < cpumask_size())
5388 cpumask_clear(new_mask);
5389 else if (len > cpumask_size())
5390 len = cpumask_size();
5392 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5396 * sys_sched_setaffinity - set the cpu affinity of a process
5397 * @pid: pid of the process
5398 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5399 * @user_mask_ptr: user-space pointer to the new cpu mask
5401 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5402 unsigned long __user *, user_mask_ptr)
5404 cpumask_var_t new_mask;
5407 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5410 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5412 retval = sched_setaffinity(pid, new_mask);
5413 free_cpumask_var(new_mask);
5417 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5419 struct task_struct *p;
5420 unsigned long flags;
5427 p = find_process_by_pid(pid);
5431 retval = security_task_getscheduler(p);
5435 raw_spin_lock_irqsave(&p->pi_lock, flags);
5436 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5437 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5447 * sys_sched_getaffinity - get the cpu affinity of a process
5448 * @pid: pid of the process
5449 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5450 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5452 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5453 unsigned long __user *, user_mask_ptr)
5458 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5460 if (len & (sizeof(unsigned long)-1))
5463 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5466 ret = sched_getaffinity(pid, mask);
5468 size_t retlen = min_t(size_t, len, cpumask_size());
5470 if (copy_to_user(user_mask_ptr, mask, retlen))
5475 free_cpumask_var(mask);
5481 * sys_sched_yield - yield the current processor to other threads.
5483 * This function yields the current CPU to other tasks. If there are no
5484 * other threads running on this CPU then this function will return.
5486 SYSCALL_DEFINE0(sched_yield)
5488 struct rq *rq = this_rq_lock();
5490 schedstat_inc(rq, yld_count);
5491 current->sched_class->yield_task(rq);
5494 * Since we are going to call schedule() anyway, there's
5495 * no need to preempt or enable interrupts:
5497 __release(rq->lock);
5498 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5499 do_raw_spin_unlock(&rq->lock);
5500 preempt_enable_no_resched();
5507 static inline int should_resched(void)
5509 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5512 static void __cond_resched(void)
5514 add_preempt_count(PREEMPT_ACTIVE);
5516 sub_preempt_count(PREEMPT_ACTIVE);
5519 int __sched _cond_resched(void)
5521 if (should_resched()) {
5527 EXPORT_SYMBOL(_cond_resched);
5530 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5531 * call schedule, and on return reacquire the lock.
5533 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5534 * operations here to prevent schedule() from being called twice (once via
5535 * spin_unlock(), once by hand).
5537 int __cond_resched_lock(spinlock_t *lock)
5539 int resched = should_resched();
5542 lockdep_assert_held(lock);
5544 if (spin_needbreak(lock) || resched) {
5555 EXPORT_SYMBOL(__cond_resched_lock);
5557 int __sched __cond_resched_softirq(void)
5559 BUG_ON(!in_softirq());
5561 if (should_resched()) {
5569 EXPORT_SYMBOL(__cond_resched_softirq);
5572 * yield - yield the current processor to other threads.
5574 * This is a shortcut for kernel-space yielding - it marks the
5575 * thread runnable and calls sys_sched_yield().
5577 void __sched yield(void)
5579 set_current_state(TASK_RUNNING);
5582 EXPORT_SYMBOL(yield);
5585 * yield_to - yield the current processor to another thread in
5586 * your thread group, or accelerate that thread toward the
5587 * processor it's on.
5589 * @preempt: whether task preemption is allowed or not
5591 * It's the caller's job to ensure that the target task struct
5592 * can't go away on us before we can do any checks.
5594 * Returns true if we indeed boosted the target task.
5596 bool __sched yield_to(struct task_struct *p, bool preempt)
5598 struct task_struct *curr = current;
5599 struct rq *rq, *p_rq;
5600 unsigned long flags;
5603 local_irq_save(flags);
5608 double_rq_lock(rq, p_rq);
5609 while (task_rq(p) != p_rq) {
5610 double_rq_unlock(rq, p_rq);
5614 if (!curr->sched_class->yield_to_task)
5617 if (curr->sched_class != p->sched_class)
5620 if (task_running(p_rq, p) || p->state)
5623 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5625 schedstat_inc(rq, yld_count);
5627 * Make p's CPU reschedule; pick_next_entity takes care of
5630 if (preempt && rq != p_rq)
5631 resched_task(p_rq->curr);
5635 double_rq_unlock(rq, p_rq);
5636 local_irq_restore(flags);
5643 EXPORT_SYMBOL_GPL(yield_to);
5646 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5647 * that process accounting knows that this is a task in IO wait state.
5649 void __sched io_schedule(void)
5651 struct rq *rq = raw_rq();
5653 delayacct_blkio_start();
5654 atomic_inc(&rq->nr_iowait);
5655 blk_flush_plug(current);
5656 current->in_iowait = 1;
5658 current->in_iowait = 0;
5659 atomic_dec(&rq->nr_iowait);
5660 delayacct_blkio_end();
5662 EXPORT_SYMBOL(io_schedule);
5664 long __sched io_schedule_timeout(long timeout)
5666 struct rq *rq = raw_rq();
5669 delayacct_blkio_start();
5670 atomic_inc(&rq->nr_iowait);
5671 blk_flush_plug(current);
5672 current->in_iowait = 1;
5673 ret = schedule_timeout(timeout);
5674 current->in_iowait = 0;
5675 atomic_dec(&rq->nr_iowait);
5676 delayacct_blkio_end();
5681 * sys_sched_get_priority_max - return maximum RT priority.
5682 * @policy: scheduling class.
5684 * this syscall returns the maximum rt_priority that can be used
5685 * by a given scheduling class.
5687 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5694 ret = MAX_USER_RT_PRIO-1;
5706 * sys_sched_get_priority_min - return minimum RT priority.
5707 * @policy: scheduling class.
5709 * this syscall returns the minimum rt_priority that can be used
5710 * by a given scheduling class.
5712 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5730 * sys_sched_rr_get_interval - return the default timeslice of a process.
5731 * @pid: pid of the process.
5732 * @interval: userspace pointer to the timeslice value.
5734 * this syscall writes the default timeslice value of a given process
5735 * into the user-space timespec buffer. A value of '0' means infinity.
5737 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5738 struct timespec __user *, interval)
5740 struct task_struct *p;
5741 unsigned int time_slice;
5742 unsigned long flags;
5752 p = find_process_by_pid(pid);
5756 retval = security_task_getscheduler(p);
5760 rq = task_rq_lock(p, &flags);
5761 time_slice = p->sched_class->get_rr_interval(rq, p);
5762 task_rq_unlock(rq, p, &flags);
5765 jiffies_to_timespec(time_slice, &t);
5766 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5774 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5776 void sched_show_task(struct task_struct *p)
5778 unsigned long free = 0;
5781 state = p->state ? __ffs(p->state) + 1 : 0;
5782 printk(KERN_INFO "%-15.15s %c", p->comm,
5783 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5784 #if BITS_PER_LONG == 32
5785 if (state == TASK_RUNNING)
5786 printk(KERN_CONT " running ");
5788 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5790 if (state == TASK_RUNNING)
5791 printk(KERN_CONT " running task ");
5793 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5795 #ifdef CONFIG_DEBUG_STACK_USAGE
5796 free = stack_not_used(p);
5798 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5799 task_pid_nr(p), task_pid_nr(p->real_parent),
5800 (unsigned long)task_thread_info(p)->flags);
5802 show_stack(p, NULL);
5805 void show_state_filter(unsigned long state_filter)
5807 struct task_struct *g, *p;
5809 #if BITS_PER_LONG == 32
5811 " task PC stack pid father\n");
5814 " task PC stack pid father\n");
5816 read_lock(&tasklist_lock);
5817 do_each_thread(g, p) {
5819 * reset the NMI-timeout, listing all files on a slow
5820 * console might take a lot of time:
5822 touch_nmi_watchdog();
5823 if (!state_filter || (p->state & state_filter))
5825 } while_each_thread(g, p);
5827 touch_all_softlockup_watchdogs();
5829 #ifdef CONFIG_SCHED_DEBUG
5830 sysrq_sched_debug_show();
5832 read_unlock(&tasklist_lock);
5834 * Only show locks if all tasks are dumped:
5837 debug_show_all_locks();
5840 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5842 idle->sched_class = &idle_sched_class;
5846 * init_idle - set up an idle thread for a given CPU
5847 * @idle: task in question
5848 * @cpu: cpu the idle task belongs to
5850 * NOTE: this function does not set the idle thread's NEED_RESCHED
5851 * flag, to make booting more robust.
5853 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5855 struct rq *rq = cpu_rq(cpu);
5856 unsigned long flags;
5858 raw_spin_lock_irqsave(&rq->lock, flags);
5861 idle->state = TASK_RUNNING;
5862 idle->se.exec_start = sched_clock();
5864 do_set_cpus_allowed(idle, cpumask_of(cpu));
5866 * We're having a chicken and egg problem, even though we are
5867 * holding rq->lock, the cpu isn't yet set to this cpu so the
5868 * lockdep check in task_group() will fail.
5870 * Similar case to sched_fork(). / Alternatively we could
5871 * use task_rq_lock() here and obtain the other rq->lock.
5876 __set_task_cpu(idle, cpu);
5879 rq->curr = rq->idle = idle;
5880 #if defined(CONFIG_SMP)
5883 raw_spin_unlock_irqrestore(&rq->lock, flags);
5885 /* Set the preempt count _outside_ the spinlocks! */
5886 task_thread_info(idle)->preempt_count = 0;
5889 * The idle tasks have their own, simple scheduling class:
5891 idle->sched_class = &idle_sched_class;
5892 ftrace_graph_init_idle_task(idle, cpu);
5896 * In a system that switches off the HZ timer nohz_cpu_mask
5897 * indicates which cpus entered this state. This is used
5898 * in the rcu update to wait only for active cpus. For system
5899 * which do not switch off the HZ timer nohz_cpu_mask should
5900 * always be CPU_BITS_NONE.
5902 cpumask_var_t nohz_cpu_mask;
5905 * Increase the granularity value when there are more CPUs,
5906 * because with more CPUs the 'effective latency' as visible
5907 * to users decreases. But the relationship is not linear,
5908 * so pick a second-best guess by going with the log2 of the
5911 * This idea comes from the SD scheduler of Con Kolivas:
5913 static int get_update_sysctl_factor(void)
5915 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5916 unsigned int factor;
5918 switch (sysctl_sched_tunable_scaling) {
5919 case SCHED_TUNABLESCALING_NONE:
5922 case SCHED_TUNABLESCALING_LINEAR:
5925 case SCHED_TUNABLESCALING_LOG:
5927 factor = 1 + ilog2(cpus);
5934 static void update_sysctl(void)
5936 unsigned int factor = get_update_sysctl_factor();
5938 #define SET_SYSCTL(name) \
5939 (sysctl_##name = (factor) * normalized_sysctl_##name)
5940 SET_SYSCTL(sched_min_granularity);
5941 SET_SYSCTL(sched_latency);
5942 SET_SYSCTL(sched_wakeup_granularity);
5946 static inline void sched_init_granularity(void)
5952 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
5954 if (p->sched_class && p->sched_class->set_cpus_allowed)
5955 p->sched_class->set_cpus_allowed(p, new_mask);
5957 cpumask_copy(&p->cpus_allowed, new_mask);
5958 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5963 * This is how migration works:
5965 * 1) we invoke migration_cpu_stop() on the target CPU using
5967 * 2) stopper starts to run (implicitly forcing the migrated thread
5969 * 3) it checks whether the migrated task is still in the wrong runqueue.
5970 * 4) if it's in the wrong runqueue then the migration thread removes
5971 * it and puts it into the right queue.
5972 * 5) stopper completes and stop_one_cpu() returns and the migration
5977 * Change a given task's CPU affinity. Migrate the thread to a
5978 * proper CPU and schedule it away if the CPU it's executing on
5979 * is removed from the allowed bitmask.
5981 * NOTE: the caller must have a valid reference to the task, the
5982 * task must not exit() & deallocate itself prematurely. The
5983 * call is not atomic; no spinlocks may be held.
5985 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5987 unsigned long flags;
5989 unsigned int dest_cpu;
5992 rq = task_rq_lock(p, &flags);
5994 if (cpumask_equal(&p->cpus_allowed, new_mask))
5997 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
6002 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
6007 do_set_cpus_allowed(p, new_mask);
6009 /* Can the task run on the task's current CPU? If so, we're done */
6010 if (cpumask_test_cpu(task_cpu(p), new_mask))
6013 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
6015 struct migration_arg arg = { p, dest_cpu };
6016 /* Need help from migration thread: drop lock and wait. */
6017 task_rq_unlock(rq, p, &flags);
6018 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
6019 tlb_migrate_finish(p->mm);
6023 task_rq_unlock(rq, p, &flags);
6027 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
6030 * Move (not current) task off this cpu, onto dest cpu. We're doing
6031 * this because either it can't run here any more (set_cpus_allowed()
6032 * away from this CPU, or CPU going down), or because we're
6033 * attempting to rebalance this task on exec (sched_exec).
6035 * So we race with normal scheduler movements, but that's OK, as long
6036 * as the task is no longer on this CPU.
6038 * Returns non-zero if task was successfully migrated.
6040 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6042 struct rq *rq_dest, *rq_src;
6045 if (unlikely(!cpu_active(dest_cpu)))
6048 rq_src = cpu_rq(src_cpu);
6049 rq_dest = cpu_rq(dest_cpu);
6051 raw_spin_lock(&p->pi_lock);
6052 double_rq_lock(rq_src, rq_dest);
6053 /* Already moved. */
6054 if (task_cpu(p) != src_cpu)
6056 /* Affinity changed (again). */
6057 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6061 * If we're not on a rq, the next wake-up will ensure we're
6065 deactivate_task(rq_src, p, 0);
6066 set_task_cpu(p, dest_cpu);
6067 activate_task(rq_dest, p, 0);
6068 check_preempt_curr(rq_dest, p, 0);
6073 double_rq_unlock(rq_src, rq_dest);
6074 raw_spin_unlock(&p->pi_lock);
6079 * migration_cpu_stop - this will be executed by a highprio stopper thread
6080 * and performs thread migration by bumping thread off CPU then
6081 * 'pushing' onto another runqueue.
6083 static int migration_cpu_stop(void *data)
6085 struct migration_arg *arg = data;
6088 * The original target cpu might have gone down and we might
6089 * be on another cpu but it doesn't matter.
6091 local_irq_disable();
6092 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
6097 #ifdef CONFIG_HOTPLUG_CPU
6100 * Ensures that the idle task is using init_mm right before its cpu goes
6103 void idle_task_exit(void)
6105 struct mm_struct *mm = current->active_mm;
6107 BUG_ON(cpu_online(smp_processor_id()));
6110 switch_mm(mm, &init_mm, current);
6115 * While a dead CPU has no uninterruptible tasks queued at this point,
6116 * it might still have a nonzero ->nr_uninterruptible counter, because
6117 * for performance reasons the counter is not stricly tracking tasks to
6118 * their home CPUs. So we just add the counter to another CPU's counter,
6119 * to keep the global sum constant after CPU-down:
6121 static void migrate_nr_uninterruptible(struct rq *rq_src)
6123 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
6125 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6126 rq_src->nr_uninterruptible = 0;
6130 * remove the tasks which were accounted by rq from calc_load_tasks.
6132 static void calc_global_load_remove(struct rq *rq)
6134 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
6135 rq->calc_load_active = 0;
6139 * Migrate all tasks from the rq, sleeping tasks will be migrated by
6140 * try_to_wake_up()->select_task_rq().
6142 * Called with rq->lock held even though we'er in stop_machine() and
6143 * there's no concurrency possible, we hold the required locks anyway
6144 * because of lock validation efforts.
6146 static void migrate_tasks(unsigned int dead_cpu)
6148 struct rq *rq = cpu_rq(dead_cpu);
6149 struct task_struct *next, *stop = rq->stop;
6153 * Fudge the rq selection such that the below task selection loop
6154 * doesn't get stuck on the currently eligible stop task.
6156 * We're currently inside stop_machine() and the rq is either stuck
6157 * in the stop_machine_cpu_stop() loop, or we're executing this code,
6158 * either way we should never end up calling schedule() until we're
6165 * There's this thread running, bail when that's the only
6168 if (rq->nr_running == 1)
6171 next = pick_next_task(rq);
6173 next->sched_class->put_prev_task(rq, next);
6175 /* Find suitable destination for @next, with force if needed. */
6176 dest_cpu = select_fallback_rq(dead_cpu, next);
6177 raw_spin_unlock(&rq->lock);
6179 __migrate_task(next, dead_cpu, dest_cpu);
6181 raw_spin_lock(&rq->lock);
6187 #endif /* CONFIG_HOTPLUG_CPU */
6189 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6191 static struct ctl_table sd_ctl_dir[] = {
6193 .procname = "sched_domain",
6199 static struct ctl_table sd_ctl_root[] = {
6201 .procname = "kernel",
6203 .child = sd_ctl_dir,
6208 static struct ctl_table *sd_alloc_ctl_entry(int n)
6210 struct ctl_table *entry =
6211 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6216 static void sd_free_ctl_entry(struct ctl_table **tablep)
6218 struct ctl_table *entry;
6221 * In the intermediate directories, both the child directory and
6222 * procname are dynamically allocated and could fail but the mode
6223 * will always be set. In the lowest directory the names are
6224 * static strings and all have proc handlers.
6226 for (entry = *tablep; entry->mode; entry++) {
6228 sd_free_ctl_entry(&entry->child);
6229 if (entry->proc_handler == NULL)
6230 kfree(entry->procname);
6238 set_table_entry(struct ctl_table *entry,
6239 const char *procname, void *data, int maxlen,
6240 mode_t mode, proc_handler *proc_handler)
6242 entry->procname = procname;
6244 entry->maxlen = maxlen;
6246 entry->proc_handler = proc_handler;
6249 static struct ctl_table *
6250 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6252 struct ctl_table *table = sd_alloc_ctl_entry(13);
6257 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6258 sizeof(long), 0644, proc_doulongvec_minmax);
6259 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6260 sizeof(long), 0644, proc_doulongvec_minmax);
6261 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6262 sizeof(int), 0644, proc_dointvec_minmax);
6263 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6264 sizeof(int), 0644, proc_dointvec_minmax);
6265 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6266 sizeof(int), 0644, proc_dointvec_minmax);
6267 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6268 sizeof(int), 0644, proc_dointvec_minmax);
6269 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6270 sizeof(int), 0644, proc_dointvec_minmax);
6271 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6272 sizeof(int), 0644, proc_dointvec_minmax);
6273 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6274 sizeof(int), 0644, proc_dointvec_minmax);
6275 set_table_entry(&table[9], "cache_nice_tries",
6276 &sd->cache_nice_tries,
6277 sizeof(int), 0644, proc_dointvec_minmax);
6278 set_table_entry(&table[10], "flags", &sd->flags,
6279 sizeof(int), 0644, proc_dointvec_minmax);
6280 set_table_entry(&table[11], "name", sd->name,
6281 CORENAME_MAX_SIZE, 0444, proc_dostring);
6282 /* &table[12] is terminator */
6287 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6289 struct ctl_table *entry, *table;
6290 struct sched_domain *sd;
6291 int domain_num = 0, i;
6294 for_each_domain(cpu, sd)
6296 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6301 for_each_domain(cpu, sd) {
6302 snprintf(buf, 32, "domain%d", i);
6303 entry->procname = kstrdup(buf, GFP_KERNEL);
6305 entry->child = sd_alloc_ctl_domain_table(sd);
6312 static struct ctl_table_header *sd_sysctl_header;
6313 static void register_sched_domain_sysctl(void)
6315 int i, cpu_num = num_possible_cpus();
6316 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6319 WARN_ON(sd_ctl_dir[0].child);
6320 sd_ctl_dir[0].child = entry;
6325 for_each_possible_cpu(i) {
6326 snprintf(buf, 32, "cpu%d", i);
6327 entry->procname = kstrdup(buf, GFP_KERNEL);
6329 entry->child = sd_alloc_ctl_cpu_table(i);
6333 WARN_ON(sd_sysctl_header);
6334 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6337 /* may be called multiple times per register */
6338 static void unregister_sched_domain_sysctl(void)
6340 if (sd_sysctl_header)
6341 unregister_sysctl_table(sd_sysctl_header);
6342 sd_sysctl_header = NULL;
6343 if (sd_ctl_dir[0].child)
6344 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6347 static void register_sched_domain_sysctl(void)
6350 static void unregister_sched_domain_sysctl(void)
6355 static void set_rq_online(struct rq *rq)
6358 const struct sched_class *class;
6360 cpumask_set_cpu(rq->cpu, rq->rd->online);
6363 for_each_class(class) {
6364 if (class->rq_online)
6365 class->rq_online(rq);
6370 static void set_rq_offline(struct rq *rq)
6373 const struct sched_class *class;
6375 for_each_class(class) {
6376 if (class->rq_offline)
6377 class->rq_offline(rq);
6380 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6386 * migration_call - callback that gets triggered when a CPU is added.
6387 * Here we can start up the necessary migration thread for the new CPU.
6389 static int __cpuinit
6390 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6392 int cpu = (long)hcpu;
6393 unsigned long flags;
6394 struct rq *rq = cpu_rq(cpu);
6396 switch (action & ~CPU_TASKS_FROZEN) {
6398 case CPU_UP_PREPARE:
6399 rq->calc_load_update = calc_load_update;
6403 /* Update our root-domain */
6404 raw_spin_lock_irqsave(&rq->lock, flags);
6406 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6410 raw_spin_unlock_irqrestore(&rq->lock, flags);
6413 #ifdef CONFIG_HOTPLUG_CPU
6415 sched_ttwu_pending();
6416 /* Update our root-domain */
6417 raw_spin_lock_irqsave(&rq->lock, flags);
6419 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6423 BUG_ON(rq->nr_running != 1); /* the migration thread */
6424 raw_spin_unlock_irqrestore(&rq->lock, flags);
6426 migrate_nr_uninterruptible(rq);
6427 calc_global_load_remove(rq);
6432 update_max_interval();
6438 * Register at high priority so that task migration (migrate_all_tasks)
6439 * happens before everything else. This has to be lower priority than
6440 * the notifier in the perf_event subsystem, though.
6442 static struct notifier_block __cpuinitdata migration_notifier = {
6443 .notifier_call = migration_call,
6444 .priority = CPU_PRI_MIGRATION,
6447 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
6448 unsigned long action, void *hcpu)
6450 switch (action & ~CPU_TASKS_FROZEN) {
6452 case CPU_DOWN_FAILED:
6453 set_cpu_active((long)hcpu, true);
6460 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
6461 unsigned long action, void *hcpu)
6463 switch (action & ~CPU_TASKS_FROZEN) {
6464 case CPU_DOWN_PREPARE:
6465 set_cpu_active((long)hcpu, false);
6472 static int __init migration_init(void)
6474 void *cpu = (void *)(long)smp_processor_id();
6477 /* Initialize migration for the boot CPU */
6478 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6479 BUG_ON(err == NOTIFY_BAD);
6480 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6481 register_cpu_notifier(&migration_notifier);
6483 /* Register cpu active notifiers */
6484 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
6485 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
6489 early_initcall(migration_init);
6494 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
6496 #ifdef CONFIG_SCHED_DEBUG
6498 static __read_mostly int sched_domain_debug_enabled;
6500 static int __init sched_domain_debug_setup(char *str)
6502 sched_domain_debug_enabled = 1;
6506 early_param("sched_debug", sched_domain_debug_setup);
6508 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6509 struct cpumask *groupmask)
6511 struct sched_group *group = sd->groups;
6514 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6515 cpumask_clear(groupmask);
6517 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6519 if (!(sd->flags & SD_LOAD_BALANCE)) {
6520 printk("does not load-balance\n");
6522 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6527 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6529 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6530 printk(KERN_ERR "ERROR: domain->span does not contain "
6533 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6534 printk(KERN_ERR "ERROR: domain->groups does not contain"
6538 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6542 printk(KERN_ERR "ERROR: group is NULL\n");
6546 if (!group->cpu_power) {
6547 printk(KERN_CONT "\n");
6548 printk(KERN_ERR "ERROR: domain->cpu_power not "
6553 if (!cpumask_weight(sched_group_cpus(group))) {
6554 printk(KERN_CONT "\n");
6555 printk(KERN_ERR "ERROR: empty group\n");
6559 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6560 printk(KERN_CONT "\n");
6561 printk(KERN_ERR "ERROR: repeated CPUs\n");
6565 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6567 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6569 printk(KERN_CONT " %s", str);
6570 if (group->cpu_power != SCHED_POWER_SCALE) {
6571 printk(KERN_CONT " (cpu_power = %d)",
6575 group = group->next;
6576 } while (group != sd->groups);
6577 printk(KERN_CONT "\n");
6579 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6580 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6583 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6584 printk(KERN_ERR "ERROR: parent span is not a superset "
6585 "of domain->span\n");
6589 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6593 if (!sched_domain_debug_enabled)
6597 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6601 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6604 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
6612 #else /* !CONFIG_SCHED_DEBUG */
6613 # define sched_domain_debug(sd, cpu) do { } while (0)
6614 #endif /* CONFIG_SCHED_DEBUG */
6616 static int sd_degenerate(struct sched_domain *sd)
6618 if (cpumask_weight(sched_domain_span(sd)) == 1)
6621 /* Following flags need at least 2 groups */
6622 if (sd->flags & (SD_LOAD_BALANCE |
6623 SD_BALANCE_NEWIDLE |
6627 SD_SHARE_PKG_RESOURCES)) {
6628 if (sd->groups != sd->groups->next)
6632 /* Following flags don't use groups */
6633 if (sd->flags & (SD_WAKE_AFFINE))
6640 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6642 unsigned long cflags = sd->flags, pflags = parent->flags;
6644 if (sd_degenerate(parent))
6647 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6650 /* Flags needing groups don't count if only 1 group in parent */
6651 if (parent->groups == parent->groups->next) {
6652 pflags &= ~(SD_LOAD_BALANCE |
6653 SD_BALANCE_NEWIDLE |
6657 SD_SHARE_PKG_RESOURCES);
6658 if (nr_node_ids == 1)
6659 pflags &= ~SD_SERIALIZE;
6661 if (~cflags & pflags)
6667 static void free_rootdomain(struct rcu_head *rcu)
6669 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
6671 cpupri_cleanup(&rd->cpupri);
6672 free_cpumask_var(rd->rto_mask);
6673 free_cpumask_var(rd->online);
6674 free_cpumask_var(rd->span);
6678 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6680 struct root_domain *old_rd = NULL;
6681 unsigned long flags;
6683 raw_spin_lock_irqsave(&rq->lock, flags);
6688 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6691 cpumask_clear_cpu(rq->cpu, old_rd->span);
6694 * If we dont want to free the old_rt yet then
6695 * set old_rd to NULL to skip the freeing later
6698 if (!atomic_dec_and_test(&old_rd->refcount))
6702 atomic_inc(&rd->refcount);
6705 cpumask_set_cpu(rq->cpu, rd->span);
6706 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6709 raw_spin_unlock_irqrestore(&rq->lock, flags);
6712 call_rcu_sched(&old_rd->rcu, free_rootdomain);
6715 static int init_rootdomain(struct root_domain *rd)
6717 memset(rd, 0, sizeof(*rd));
6719 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
6721 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
6723 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6726 if (cpupri_init(&rd->cpupri) != 0)
6731 free_cpumask_var(rd->rto_mask);
6733 free_cpumask_var(rd->online);
6735 free_cpumask_var(rd->span);
6740 static void init_defrootdomain(void)
6742 init_rootdomain(&def_root_domain);
6744 atomic_set(&def_root_domain.refcount, 1);
6747 static struct root_domain *alloc_rootdomain(void)
6749 struct root_domain *rd;
6751 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6755 if (init_rootdomain(rd) != 0) {
6763 static void free_sched_domain(struct rcu_head *rcu)
6765 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
6766 if (atomic_dec_and_test(&sd->groups->ref))
6771 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
6773 call_rcu(&sd->rcu, free_sched_domain);
6776 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
6778 for (; sd; sd = sd->parent)
6779 destroy_sched_domain(sd, cpu);
6783 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6784 * hold the hotplug lock.
6787 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6789 struct rq *rq = cpu_rq(cpu);
6790 struct sched_domain *tmp;
6792 /* Remove the sched domains which do not contribute to scheduling. */
6793 for (tmp = sd; tmp; ) {
6794 struct sched_domain *parent = tmp->parent;
6798 if (sd_parent_degenerate(tmp, parent)) {
6799 tmp->parent = parent->parent;
6801 parent->parent->child = tmp;
6802 destroy_sched_domain(parent, cpu);
6807 if (sd && sd_degenerate(sd)) {
6810 destroy_sched_domain(tmp, cpu);
6815 sched_domain_debug(sd, cpu);
6817 rq_attach_root(rq, rd);
6819 rcu_assign_pointer(rq->sd, sd);
6820 destroy_sched_domains(tmp, cpu);
6823 /* cpus with isolated domains */
6824 static cpumask_var_t cpu_isolated_map;
6826 /* Setup the mask of cpus configured for isolated domains */
6827 static int __init isolated_cpu_setup(char *str)
6829 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6830 cpulist_parse(str, cpu_isolated_map);
6834 __setup("isolcpus=", isolated_cpu_setup);
6836 #define SD_NODES_PER_DOMAIN 16
6841 * find_next_best_node - find the next node to include in a sched_domain
6842 * @node: node whose sched_domain we're building
6843 * @used_nodes: nodes already in the sched_domain
6845 * Find the next node to include in a given scheduling domain. Simply
6846 * finds the closest node not already in the @used_nodes map.
6848 * Should use nodemask_t.
6850 static int find_next_best_node(int node, nodemask_t *used_nodes)
6852 int i, n, val, min_val, best_node = -1;
6856 for (i = 0; i < nr_node_ids; i++) {
6857 /* Start at @node */
6858 n = (node + i) % nr_node_ids;
6860 if (!nr_cpus_node(n))
6863 /* Skip already used nodes */
6864 if (node_isset(n, *used_nodes))
6867 /* Simple min distance search */
6868 val = node_distance(node, n);
6870 if (val < min_val) {
6876 if (best_node != -1)
6877 node_set(best_node, *used_nodes);
6882 * sched_domain_node_span - get a cpumask for a node's sched_domain
6883 * @node: node whose cpumask we're constructing
6884 * @span: resulting cpumask
6886 * Given a node, construct a good cpumask for its sched_domain to span. It
6887 * should be one that prevents unnecessary balancing, but also spreads tasks
6890 static void sched_domain_node_span(int node, struct cpumask *span)
6892 nodemask_t used_nodes;
6895 cpumask_clear(span);
6896 nodes_clear(used_nodes);
6898 cpumask_or(span, span, cpumask_of_node(node));
6899 node_set(node, used_nodes);
6901 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6902 int next_node = find_next_best_node(node, &used_nodes);
6905 cpumask_or(span, span, cpumask_of_node(next_node));
6909 static const struct cpumask *cpu_node_mask(int cpu)
6911 lockdep_assert_held(&sched_domains_mutex);
6913 sched_domain_node_span(cpu_to_node(cpu), sched_domains_tmpmask);
6915 return sched_domains_tmpmask;
6918 static const struct cpumask *cpu_allnodes_mask(int cpu)
6920 return cpu_possible_mask;
6922 #endif /* CONFIG_NUMA */
6924 static const struct cpumask *cpu_cpu_mask(int cpu)
6926 return cpumask_of_node(cpu_to_node(cpu));
6929 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6932 struct sched_domain **__percpu sd;
6933 struct sched_group **__percpu sg;
6937 struct sched_domain ** __percpu sd;
6938 struct root_domain *rd;
6948 struct sched_domain_topology_level;
6950 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
6951 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
6953 struct sched_domain_topology_level {
6954 sched_domain_init_f init;
6955 sched_domain_mask_f mask;
6956 struct sd_data data;
6960 * Assumes the sched_domain tree is fully constructed
6962 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6964 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6965 struct sched_domain *child = sd->child;
6968 cpu = cpumask_first(sched_domain_span(child));
6971 *sg = *per_cpu_ptr(sdd->sg, cpu);
6977 * build_sched_groups takes the cpumask we wish to span, and a pointer
6978 * to a function which identifies what group(along with sched group) a CPU
6979 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6980 * (due to the fact that we keep track of groups covered with a struct cpumask).
6982 * build_sched_groups will build a circular linked list of the groups
6983 * covered by the given span, and will set each group's ->cpumask correctly,
6984 * and ->cpu_power to 0.
6987 build_sched_groups(struct sched_domain *sd)
6989 struct sched_group *first = NULL, *last = NULL;
6990 struct sd_data *sdd = sd->private;
6991 const struct cpumask *span = sched_domain_span(sd);
6992 struct cpumask *covered;
6995 lockdep_assert_held(&sched_domains_mutex);
6996 covered = sched_domains_tmpmask;
6998 cpumask_clear(covered);
7000 for_each_cpu(i, span) {
7001 struct sched_group *sg;
7002 int group = get_group(i, sdd, &sg);
7005 if (cpumask_test_cpu(i, covered))
7008 cpumask_clear(sched_group_cpus(sg));
7011 for_each_cpu(j, span) {
7012 if (get_group(j, sdd, NULL) != group)
7015 cpumask_set_cpu(j, covered);
7016 cpumask_set_cpu(j, sched_group_cpus(sg));
7029 * Initialize sched groups cpu_power.
7031 * cpu_power indicates the capacity of sched group, which is used while
7032 * distributing the load between different sched groups in a sched domain.
7033 * Typically cpu_power for all the groups in a sched domain will be same unless
7034 * there are asymmetries in the topology. If there are asymmetries, group
7035 * having more cpu_power will pickup more load compared to the group having
7038 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7040 WARN_ON(!sd || !sd->groups);
7042 if (cpu != group_first_cpu(sd->groups))
7045 sd->groups->group_weight = cpumask_weight(sched_group_cpus(sd->groups));
7047 update_group_power(sd, cpu);
7051 * Initializers for schedule domains
7052 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7055 #ifdef CONFIG_SCHED_DEBUG
7056 # define SD_INIT_NAME(sd, type) sd->name = #type
7058 # define SD_INIT_NAME(sd, type) do { } while (0)
7061 #define SD_INIT_FUNC(type) \
7062 static noinline struct sched_domain * \
7063 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
7065 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
7066 *sd = SD_##type##_INIT; \
7067 SD_INIT_NAME(sd, type); \
7068 sd->private = &tl->data; \
7074 SD_INIT_FUNC(ALLNODES)
7077 #ifdef CONFIG_SCHED_SMT
7078 SD_INIT_FUNC(SIBLING)
7080 #ifdef CONFIG_SCHED_MC
7083 #ifdef CONFIG_SCHED_BOOK
7087 static int default_relax_domain_level = -1;
7088 int sched_domain_level_max;
7090 static int __init setup_relax_domain_level(char *str)
7094 val = simple_strtoul(str, NULL, 0);
7095 if (val < sched_domain_level_max)
7096 default_relax_domain_level = val;
7100 __setup("relax_domain_level=", setup_relax_domain_level);
7102 static void set_domain_attribute(struct sched_domain *sd,
7103 struct sched_domain_attr *attr)
7107 if (!attr || attr->relax_domain_level < 0) {
7108 if (default_relax_domain_level < 0)
7111 request = default_relax_domain_level;
7113 request = attr->relax_domain_level;
7114 if (request < sd->level) {
7115 /* turn off idle balance on this domain */
7116 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7118 /* turn on idle balance on this domain */
7119 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7123 static void __sdt_free(const struct cpumask *cpu_map);
7124 static int __sdt_alloc(const struct cpumask *cpu_map);
7126 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
7127 const struct cpumask *cpu_map)
7131 if (!atomic_read(&d->rd->refcount))
7132 free_rootdomain(&d->rd->rcu); /* fall through */
7134 free_percpu(d->sd); /* fall through */
7136 __sdt_free(cpu_map); /* fall through */
7142 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
7143 const struct cpumask *cpu_map)
7145 memset(d, 0, sizeof(*d));
7147 if (__sdt_alloc(cpu_map))
7148 return sa_sd_storage;
7149 d->sd = alloc_percpu(struct sched_domain *);
7151 return sa_sd_storage;
7152 d->rd = alloc_rootdomain();
7155 return sa_rootdomain;
7159 * NULL the sd_data elements we've used to build the sched_domain and
7160 * sched_group structure so that the subsequent __free_domain_allocs()
7161 * will not free the data we're using.
7163 static void claim_allocations(int cpu, struct sched_domain *sd)
7165 struct sd_data *sdd = sd->private;
7166 struct sched_group *sg = sd->groups;
7168 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
7169 *per_cpu_ptr(sdd->sd, cpu) = NULL;
7171 if (cpu == cpumask_first(sched_group_cpus(sg))) {
7172 WARN_ON_ONCE(*per_cpu_ptr(sdd->sg, cpu) != sg);
7173 *per_cpu_ptr(sdd->sg, cpu) = NULL;
7177 #ifdef CONFIG_SCHED_SMT
7178 static const struct cpumask *cpu_smt_mask(int cpu)
7180 return topology_thread_cpumask(cpu);
7185 * Topology list, bottom-up.
7187 static struct sched_domain_topology_level default_topology[] = {
7188 #ifdef CONFIG_SCHED_SMT
7189 { sd_init_SIBLING, cpu_smt_mask, },
7191 #ifdef CONFIG_SCHED_MC
7192 { sd_init_MC, cpu_coregroup_mask, },
7194 #ifdef CONFIG_SCHED_BOOK
7195 { sd_init_BOOK, cpu_book_mask, },
7197 { sd_init_CPU, cpu_cpu_mask, },
7199 { sd_init_NODE, cpu_node_mask, },
7200 { sd_init_ALLNODES, cpu_allnodes_mask, },
7205 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
7207 static int __sdt_alloc(const struct cpumask *cpu_map)
7209 struct sched_domain_topology_level *tl;
7212 for (tl = sched_domain_topology; tl->init; tl++) {
7213 struct sd_data *sdd = &tl->data;
7215 sdd->sd = alloc_percpu(struct sched_domain *);
7219 sdd->sg = alloc_percpu(struct sched_group *);
7223 for_each_cpu(j, cpu_map) {
7224 struct sched_domain *sd;
7225 struct sched_group *sg;
7227 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
7228 GFP_KERNEL, cpu_to_node(j));
7232 *per_cpu_ptr(sdd->sd, j) = sd;
7234 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7235 GFP_KERNEL, cpu_to_node(j));
7239 *per_cpu_ptr(sdd->sg, j) = sg;
7246 static void __sdt_free(const struct cpumask *cpu_map)
7248 struct sched_domain_topology_level *tl;
7251 for (tl = sched_domain_topology; tl->init; tl++) {
7252 struct sd_data *sdd = &tl->data;
7254 for_each_cpu(j, cpu_map) {
7255 kfree(*per_cpu_ptr(sdd->sd, j));
7256 kfree(*per_cpu_ptr(sdd->sg, j));
7258 free_percpu(sdd->sd);
7259 free_percpu(sdd->sg);
7263 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
7264 struct s_data *d, const struct cpumask *cpu_map,
7265 struct sched_domain_attr *attr, struct sched_domain *child,
7268 struct sched_domain *sd = tl->init(tl, cpu);
7272 set_domain_attribute(sd, attr);
7273 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
7275 sd->level = child->level + 1;
7276 sched_domain_level_max = max(sched_domain_level_max, sd->level);
7285 * Build sched domains for a given set of cpus and attach the sched domains
7286 * to the individual cpus
7288 static int build_sched_domains(const struct cpumask *cpu_map,
7289 struct sched_domain_attr *attr)
7291 enum s_alloc alloc_state = sa_none;
7292 struct sched_domain *sd;
7294 int i, ret = -ENOMEM;
7296 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7297 if (alloc_state != sa_rootdomain)
7300 /* Set up domains for cpus specified by the cpu_map. */
7301 for_each_cpu(i, cpu_map) {
7302 struct sched_domain_topology_level *tl;
7305 for (tl = sched_domain_topology; tl->init; tl++)
7306 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
7311 *per_cpu_ptr(d.sd, i) = sd;
7314 /* Build the groups for the domains */
7315 for_each_cpu(i, cpu_map) {
7316 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7317 sd->span_weight = cpumask_weight(sched_domain_span(sd));
7318 get_group(i, sd->private, &sd->groups);
7319 atomic_inc(&sd->groups->ref);
7321 if (i != cpumask_first(sched_domain_span(sd)))
7324 build_sched_groups(sd);
7328 /* Calculate CPU power for physical packages and nodes */
7329 for (i = nr_cpumask_bits-1; i >= 0; i--) {
7330 if (!cpumask_test_cpu(i, cpu_map))
7333 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7334 claim_allocations(i, sd);
7335 init_sched_groups_power(i, sd);
7339 /* Attach the domains */
7341 for_each_cpu(i, cpu_map) {
7342 sd = *per_cpu_ptr(d.sd, i);
7343 cpu_attach_domain(sd, d.rd, i);
7349 __free_domain_allocs(&d, alloc_state, cpu_map);
7353 static cpumask_var_t *doms_cur; /* current sched domains */
7354 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7355 static struct sched_domain_attr *dattr_cur;
7356 /* attribues of custom domains in 'doms_cur' */
7359 * Special case: If a kmalloc of a doms_cur partition (array of
7360 * cpumask) fails, then fallback to a single sched domain,
7361 * as determined by the single cpumask fallback_doms.
7363 static cpumask_var_t fallback_doms;
7366 * arch_update_cpu_topology lets virtualized architectures update the
7367 * cpu core maps. It is supposed to return 1 if the topology changed
7368 * or 0 if it stayed the same.
7370 int __attribute__((weak)) arch_update_cpu_topology(void)
7375 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7378 cpumask_var_t *doms;
7380 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7383 for (i = 0; i < ndoms; i++) {
7384 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7385 free_sched_domains(doms, i);
7392 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7395 for (i = 0; i < ndoms; i++)
7396 free_cpumask_var(doms[i]);
7401 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7402 * For now this just excludes isolated cpus, but could be used to
7403 * exclude other special cases in the future.
7405 static int init_sched_domains(const struct cpumask *cpu_map)
7409 arch_update_cpu_topology();
7411 doms_cur = alloc_sched_domains(ndoms_cur);
7413 doms_cur = &fallback_doms;
7414 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7416 err = build_sched_domains(doms_cur[0], NULL);
7417 register_sched_domain_sysctl();
7423 * Detach sched domains from a group of cpus specified in cpu_map
7424 * These cpus will now be attached to the NULL domain
7426 static void detach_destroy_domains(const struct cpumask *cpu_map)
7431 for_each_cpu(i, cpu_map)
7432 cpu_attach_domain(NULL, &def_root_domain, i);
7436 /* handle null as "default" */
7437 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7438 struct sched_domain_attr *new, int idx_new)
7440 struct sched_domain_attr tmp;
7447 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7448 new ? (new + idx_new) : &tmp,
7449 sizeof(struct sched_domain_attr));
7453 * Partition sched domains as specified by the 'ndoms_new'
7454 * cpumasks in the array doms_new[] of cpumasks. This compares
7455 * doms_new[] to the current sched domain partitioning, doms_cur[].
7456 * It destroys each deleted domain and builds each new domain.
7458 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7459 * The masks don't intersect (don't overlap.) We should setup one
7460 * sched domain for each mask. CPUs not in any of the cpumasks will
7461 * not be load balanced. If the same cpumask appears both in the
7462 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7465 * The passed in 'doms_new' should be allocated using
7466 * alloc_sched_domains. This routine takes ownership of it and will
7467 * free_sched_domains it when done with it. If the caller failed the
7468 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7469 * and partition_sched_domains() will fallback to the single partition
7470 * 'fallback_doms', it also forces the domains to be rebuilt.
7472 * If doms_new == NULL it will be replaced with cpu_online_mask.
7473 * ndoms_new == 0 is a special case for destroying existing domains,
7474 * and it will not create the default domain.
7476 * Call with hotplug lock held
7478 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7479 struct sched_domain_attr *dattr_new)
7484 mutex_lock(&sched_domains_mutex);
7486 /* always unregister in case we don't destroy any domains */
7487 unregister_sched_domain_sysctl();
7489 /* Let architecture update cpu core mappings. */
7490 new_topology = arch_update_cpu_topology();
7492 n = doms_new ? ndoms_new : 0;
7494 /* Destroy deleted domains */
7495 for (i = 0; i < ndoms_cur; i++) {
7496 for (j = 0; j < n && !new_topology; j++) {
7497 if (cpumask_equal(doms_cur[i], doms_new[j])
7498 && dattrs_equal(dattr_cur, i, dattr_new, j))
7501 /* no match - a current sched domain not in new doms_new[] */
7502 detach_destroy_domains(doms_cur[i]);
7507 if (doms_new == NULL) {
7509 doms_new = &fallback_doms;
7510 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7511 WARN_ON_ONCE(dattr_new);
7514 /* Build new domains */
7515 for (i = 0; i < ndoms_new; i++) {
7516 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7517 if (cpumask_equal(doms_new[i], doms_cur[j])
7518 && dattrs_equal(dattr_new, i, dattr_cur, j))
7521 /* no match - add a new doms_new */
7522 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7527 /* Remember the new sched domains */
7528 if (doms_cur != &fallback_doms)
7529 free_sched_domains(doms_cur, ndoms_cur);
7530 kfree(dattr_cur); /* kfree(NULL) is safe */
7531 doms_cur = doms_new;
7532 dattr_cur = dattr_new;
7533 ndoms_cur = ndoms_new;
7535 register_sched_domain_sysctl();
7537 mutex_unlock(&sched_domains_mutex);
7540 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7541 static void reinit_sched_domains(void)
7545 /* Destroy domains first to force the rebuild */
7546 partition_sched_domains(0, NULL, NULL);
7548 rebuild_sched_domains();
7552 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7554 unsigned int level = 0;
7556 if (sscanf(buf, "%u", &level) != 1)
7560 * level is always be positive so don't check for
7561 * level < POWERSAVINGS_BALANCE_NONE which is 0
7562 * What happens on 0 or 1 byte write,
7563 * need to check for count as well?
7566 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7570 sched_smt_power_savings = level;
7572 sched_mc_power_savings = level;
7574 reinit_sched_domains();
7579 #ifdef CONFIG_SCHED_MC
7580 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7581 struct sysdev_class_attribute *attr,
7584 return sprintf(page, "%u\n", sched_mc_power_savings);
7586 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7587 struct sysdev_class_attribute *attr,
7588 const char *buf, size_t count)
7590 return sched_power_savings_store(buf, count, 0);
7592 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7593 sched_mc_power_savings_show,
7594 sched_mc_power_savings_store);
7597 #ifdef CONFIG_SCHED_SMT
7598 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7599 struct sysdev_class_attribute *attr,
7602 return sprintf(page, "%u\n", sched_smt_power_savings);
7604 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7605 struct sysdev_class_attribute *attr,
7606 const char *buf, size_t count)
7608 return sched_power_savings_store(buf, count, 1);
7610 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7611 sched_smt_power_savings_show,
7612 sched_smt_power_savings_store);
7615 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7619 #ifdef CONFIG_SCHED_SMT
7621 err = sysfs_create_file(&cls->kset.kobj,
7622 &attr_sched_smt_power_savings.attr);
7624 #ifdef CONFIG_SCHED_MC
7625 if (!err && mc_capable())
7626 err = sysfs_create_file(&cls->kset.kobj,
7627 &attr_sched_mc_power_savings.attr);
7631 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7634 * Update cpusets according to cpu_active mask. If cpusets are
7635 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7636 * around partition_sched_domains().
7638 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7641 switch (action & ~CPU_TASKS_FROZEN) {
7643 case CPU_DOWN_FAILED:
7644 cpuset_update_active_cpus();
7651 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7654 switch (action & ~CPU_TASKS_FROZEN) {
7655 case CPU_DOWN_PREPARE:
7656 cpuset_update_active_cpus();
7663 static int update_runtime(struct notifier_block *nfb,
7664 unsigned long action, void *hcpu)
7666 int cpu = (int)(long)hcpu;
7669 case CPU_DOWN_PREPARE:
7670 case CPU_DOWN_PREPARE_FROZEN:
7671 disable_runtime(cpu_rq(cpu));
7674 case CPU_DOWN_FAILED:
7675 case CPU_DOWN_FAILED_FROZEN:
7677 case CPU_ONLINE_FROZEN:
7678 enable_runtime(cpu_rq(cpu));
7686 void __init sched_init_smp(void)
7688 cpumask_var_t non_isolated_cpus;
7690 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7691 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7694 mutex_lock(&sched_domains_mutex);
7695 init_sched_domains(cpu_active_mask);
7696 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7697 if (cpumask_empty(non_isolated_cpus))
7698 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7699 mutex_unlock(&sched_domains_mutex);
7702 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7703 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7705 /* RT runtime code needs to handle some hotplug events */
7706 hotcpu_notifier(update_runtime, 0);
7710 /* Move init over to a non-isolated CPU */
7711 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7713 sched_init_granularity();
7714 free_cpumask_var(non_isolated_cpus);
7716 init_sched_rt_class();
7719 void __init sched_init_smp(void)
7721 sched_init_granularity();
7723 #endif /* CONFIG_SMP */
7725 const_debug unsigned int sysctl_timer_migration = 1;
7727 int in_sched_functions(unsigned long addr)
7729 return in_lock_functions(addr) ||
7730 (addr >= (unsigned long)__sched_text_start
7731 && addr < (unsigned long)__sched_text_end);
7734 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7736 cfs_rq->tasks_timeline = RB_ROOT;
7737 INIT_LIST_HEAD(&cfs_rq->tasks);
7738 #ifdef CONFIG_FAIR_GROUP_SCHED
7740 /* allow initial update_cfs_load() to truncate */
7742 cfs_rq->load_stamp = 1;
7745 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7748 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7750 struct rt_prio_array *array;
7753 array = &rt_rq->active;
7754 for (i = 0; i < MAX_RT_PRIO; i++) {
7755 INIT_LIST_HEAD(array->queue + i);
7756 __clear_bit(i, array->bitmap);
7758 /* delimiter for bitsearch: */
7759 __set_bit(MAX_RT_PRIO, array->bitmap);
7761 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7762 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7764 rt_rq->highest_prio.next = MAX_RT_PRIO;
7768 rt_rq->rt_nr_migratory = 0;
7769 rt_rq->overloaded = 0;
7770 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
7774 rt_rq->rt_throttled = 0;
7775 rt_rq->rt_runtime = 0;
7776 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7778 #ifdef CONFIG_RT_GROUP_SCHED
7779 rt_rq->rt_nr_boosted = 0;
7784 #ifdef CONFIG_FAIR_GROUP_SCHED
7785 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7786 struct sched_entity *se, int cpu,
7787 struct sched_entity *parent)
7789 struct rq *rq = cpu_rq(cpu);
7790 tg->cfs_rq[cpu] = cfs_rq;
7791 init_cfs_rq(cfs_rq, rq);
7795 /* se could be NULL for root_task_group */
7800 se->cfs_rq = &rq->cfs;
7802 se->cfs_rq = parent->my_q;
7805 update_load_set(&se->load, 0);
7806 se->parent = parent;
7810 #ifdef CONFIG_RT_GROUP_SCHED
7811 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7812 struct sched_rt_entity *rt_se, int cpu,
7813 struct sched_rt_entity *parent)
7815 struct rq *rq = cpu_rq(cpu);
7817 tg->rt_rq[cpu] = rt_rq;
7818 init_rt_rq(rt_rq, rq);
7820 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7822 tg->rt_se[cpu] = rt_se;
7827 rt_se->rt_rq = &rq->rt;
7829 rt_se->rt_rq = parent->my_q;
7831 rt_se->my_q = rt_rq;
7832 rt_se->parent = parent;
7833 INIT_LIST_HEAD(&rt_se->run_list);
7837 void __init sched_init(void)
7840 unsigned long alloc_size = 0, ptr;
7842 #ifdef CONFIG_FAIR_GROUP_SCHED
7843 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7845 #ifdef CONFIG_RT_GROUP_SCHED
7846 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7848 #ifdef CONFIG_CPUMASK_OFFSTACK
7849 alloc_size += num_possible_cpus() * cpumask_size();
7852 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7854 #ifdef CONFIG_FAIR_GROUP_SCHED
7855 root_task_group.se = (struct sched_entity **)ptr;
7856 ptr += nr_cpu_ids * sizeof(void **);
7858 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7859 ptr += nr_cpu_ids * sizeof(void **);
7861 #endif /* CONFIG_FAIR_GROUP_SCHED */
7862 #ifdef CONFIG_RT_GROUP_SCHED
7863 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7864 ptr += nr_cpu_ids * sizeof(void **);
7866 root_task_group.rt_rq = (struct rt_rq **)ptr;
7867 ptr += nr_cpu_ids * sizeof(void **);
7869 #endif /* CONFIG_RT_GROUP_SCHED */
7870 #ifdef CONFIG_CPUMASK_OFFSTACK
7871 for_each_possible_cpu(i) {
7872 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
7873 ptr += cpumask_size();
7875 #endif /* CONFIG_CPUMASK_OFFSTACK */
7879 init_defrootdomain();
7882 init_rt_bandwidth(&def_rt_bandwidth,
7883 global_rt_period(), global_rt_runtime());
7885 #ifdef CONFIG_RT_GROUP_SCHED
7886 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7887 global_rt_period(), global_rt_runtime());
7888 #endif /* CONFIG_RT_GROUP_SCHED */
7890 #ifdef CONFIG_CGROUP_SCHED
7891 list_add(&root_task_group.list, &task_groups);
7892 INIT_LIST_HEAD(&root_task_group.children);
7893 autogroup_init(&init_task);
7894 #endif /* CONFIG_CGROUP_SCHED */
7896 for_each_possible_cpu(i) {
7900 raw_spin_lock_init(&rq->lock);
7902 rq->calc_load_active = 0;
7903 rq->calc_load_update = jiffies + LOAD_FREQ;
7904 init_cfs_rq(&rq->cfs, rq);
7905 init_rt_rq(&rq->rt, rq);
7906 #ifdef CONFIG_FAIR_GROUP_SCHED
7907 root_task_group.shares = root_task_group_load;
7908 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7910 * How much cpu bandwidth does root_task_group get?
7912 * In case of task-groups formed thr' the cgroup filesystem, it
7913 * gets 100% of the cpu resources in the system. This overall
7914 * system cpu resource is divided among the tasks of
7915 * root_task_group and its child task-groups in a fair manner,
7916 * based on each entity's (task or task-group's) weight
7917 * (se->load.weight).
7919 * In other words, if root_task_group has 10 tasks of weight
7920 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7921 * then A0's share of the cpu resource is:
7923 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7925 * We achieve this by letting root_task_group's tasks sit
7926 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7928 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7929 #endif /* CONFIG_FAIR_GROUP_SCHED */
7931 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7932 #ifdef CONFIG_RT_GROUP_SCHED
7933 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7934 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7937 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7938 rq->cpu_load[j] = 0;
7940 rq->last_load_update_tick = jiffies;
7945 rq->cpu_power = SCHED_POWER_SCALE;
7946 rq->post_schedule = 0;
7947 rq->active_balance = 0;
7948 rq->next_balance = jiffies;
7953 rq->avg_idle = 2*sysctl_sched_migration_cost;
7954 rq_attach_root(rq, &def_root_domain);
7956 rq->nohz_balance_kick = 0;
7957 init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb, i));
7961 atomic_set(&rq->nr_iowait, 0);
7964 set_load_weight(&init_task);
7966 #ifdef CONFIG_PREEMPT_NOTIFIERS
7967 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7971 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7974 #ifdef CONFIG_RT_MUTEXES
7975 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
7979 * The boot idle thread does lazy MMU switching as well:
7981 atomic_inc(&init_mm.mm_count);
7982 enter_lazy_tlb(&init_mm, current);
7985 * Make us the idle thread. Technically, schedule() should not be
7986 * called from this thread, however somewhere below it might be,
7987 * but because we are the idle thread, we just pick up running again
7988 * when this runqueue becomes "idle".
7990 init_idle(current, smp_processor_id());
7992 calc_load_update = jiffies + LOAD_FREQ;
7995 * During early bootup we pretend to be a normal task:
7997 current->sched_class = &fair_sched_class;
7999 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8000 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
8002 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
8004 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8005 alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT);
8006 atomic_set(&nohz.load_balancer, nr_cpu_ids);
8007 atomic_set(&nohz.first_pick_cpu, nr_cpu_ids);
8008 atomic_set(&nohz.second_pick_cpu, nr_cpu_ids);
8010 /* May be allocated at isolcpus cmdline parse time */
8011 if (cpu_isolated_map == NULL)
8012 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
8015 scheduler_running = 1;
8018 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8019 static inline int preempt_count_equals(int preempt_offset)
8021 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
8023 return (nested == preempt_offset);
8026 void __might_sleep(const char *file, int line, int preempt_offset)
8029 static unsigned long prev_jiffy; /* ratelimiting */
8031 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
8032 system_state != SYSTEM_RUNNING || oops_in_progress)
8034 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8036 prev_jiffy = jiffies;
8039 "BUG: sleeping function called from invalid context at %s:%d\n",
8042 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8043 in_atomic(), irqs_disabled(),
8044 current->pid, current->comm);
8046 debug_show_held_locks(current);
8047 if (irqs_disabled())
8048 print_irqtrace_events(current);
8052 EXPORT_SYMBOL(__might_sleep);
8055 #ifdef CONFIG_MAGIC_SYSRQ
8056 static void normalize_task(struct rq *rq, struct task_struct *p)
8058 const struct sched_class *prev_class = p->sched_class;
8059 int old_prio = p->prio;
8064 deactivate_task(rq, p, 0);
8065 __setscheduler(rq, p, SCHED_NORMAL, 0);
8067 activate_task(rq, p, 0);
8068 resched_task(rq->curr);
8071 check_class_changed(rq, p, prev_class, old_prio);
8074 void normalize_rt_tasks(void)
8076 struct task_struct *g, *p;
8077 unsigned long flags;
8080 read_lock_irqsave(&tasklist_lock, flags);
8081 do_each_thread(g, p) {
8083 * Only normalize user tasks:
8088 p->se.exec_start = 0;
8089 #ifdef CONFIG_SCHEDSTATS
8090 p->se.statistics.wait_start = 0;
8091 p->se.statistics.sleep_start = 0;
8092 p->se.statistics.block_start = 0;
8097 * Renice negative nice level userspace
8100 if (TASK_NICE(p) < 0 && p->mm)
8101 set_user_nice(p, 0);
8105 raw_spin_lock(&p->pi_lock);
8106 rq = __task_rq_lock(p);
8108 normalize_task(rq, p);
8110 __task_rq_unlock(rq);
8111 raw_spin_unlock(&p->pi_lock);
8112 } while_each_thread(g, p);
8114 read_unlock_irqrestore(&tasklist_lock, flags);
8117 #endif /* CONFIG_MAGIC_SYSRQ */
8119 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8121 * These functions are only useful for the IA64 MCA handling, or kdb.
8123 * They can only be called when the whole system has been
8124 * stopped - every CPU needs to be quiescent, and no scheduling
8125 * activity can take place. Using them for anything else would
8126 * be a serious bug, and as a result, they aren't even visible
8127 * under any other configuration.
8131 * curr_task - return the current task for a given cpu.
8132 * @cpu: the processor in question.
8134 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8136 struct task_struct *curr_task(int cpu)
8138 return cpu_curr(cpu);
8141 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8145 * set_curr_task - set the current task for a given cpu.
8146 * @cpu: the processor in question.
8147 * @p: the task pointer to set.
8149 * Description: This function must only be used when non-maskable interrupts
8150 * are serviced on a separate stack. It allows the architecture to switch the
8151 * notion of the current task on a cpu in a non-blocking manner. This function
8152 * must be called with all CPU's synchronized, and interrupts disabled, the
8153 * and caller must save the original value of the current task (see
8154 * curr_task() above) and restore that value before reenabling interrupts and
8155 * re-starting the system.
8157 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8159 void set_curr_task(int cpu, struct task_struct *p)
8166 #ifdef CONFIG_FAIR_GROUP_SCHED
8167 static void free_fair_sched_group(struct task_group *tg)
8171 for_each_possible_cpu(i) {
8173 kfree(tg->cfs_rq[i]);
8183 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8185 struct cfs_rq *cfs_rq;
8186 struct sched_entity *se;
8189 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8192 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8196 tg->shares = NICE_0_LOAD;
8198 for_each_possible_cpu(i) {
8199 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8200 GFP_KERNEL, cpu_to_node(i));
8204 se = kzalloc_node(sizeof(struct sched_entity),
8205 GFP_KERNEL, cpu_to_node(i));
8209 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8220 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8222 struct rq *rq = cpu_rq(cpu);
8223 unsigned long flags;
8226 * Only empty task groups can be destroyed; so we can speculatively
8227 * check on_list without danger of it being re-added.
8229 if (!tg->cfs_rq[cpu]->on_list)
8232 raw_spin_lock_irqsave(&rq->lock, flags);
8233 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8234 raw_spin_unlock_irqrestore(&rq->lock, flags);
8236 #else /* !CONFG_FAIR_GROUP_SCHED */
8237 static inline void free_fair_sched_group(struct task_group *tg)
8242 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8247 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8250 #endif /* CONFIG_FAIR_GROUP_SCHED */
8252 #ifdef CONFIG_RT_GROUP_SCHED
8253 static void free_rt_sched_group(struct task_group *tg)
8257 destroy_rt_bandwidth(&tg->rt_bandwidth);
8259 for_each_possible_cpu(i) {
8261 kfree(tg->rt_rq[i]);
8263 kfree(tg->rt_se[i]);
8271 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8273 struct rt_rq *rt_rq;
8274 struct sched_rt_entity *rt_se;
8277 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8280 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8284 init_rt_bandwidth(&tg->rt_bandwidth,
8285 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8287 for_each_possible_cpu(i) {
8288 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8289 GFP_KERNEL, cpu_to_node(i));
8293 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8294 GFP_KERNEL, cpu_to_node(i));
8298 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
8308 #else /* !CONFIG_RT_GROUP_SCHED */
8309 static inline void free_rt_sched_group(struct task_group *tg)
8314 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8318 #endif /* CONFIG_RT_GROUP_SCHED */
8320 #ifdef CONFIG_CGROUP_SCHED
8321 static void free_sched_group(struct task_group *tg)
8323 free_fair_sched_group(tg);
8324 free_rt_sched_group(tg);
8329 /* allocate runqueue etc for a new task group */
8330 struct task_group *sched_create_group(struct task_group *parent)
8332 struct task_group *tg;
8333 unsigned long flags;
8335 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8337 return ERR_PTR(-ENOMEM);
8339 if (!alloc_fair_sched_group(tg, parent))
8342 if (!alloc_rt_sched_group(tg, parent))
8345 spin_lock_irqsave(&task_group_lock, flags);
8346 list_add_rcu(&tg->list, &task_groups);
8348 WARN_ON(!parent); /* root should already exist */
8350 tg->parent = parent;
8351 INIT_LIST_HEAD(&tg->children);
8352 list_add_rcu(&tg->siblings, &parent->children);
8353 spin_unlock_irqrestore(&task_group_lock, flags);
8358 free_sched_group(tg);
8359 return ERR_PTR(-ENOMEM);
8362 /* rcu callback to free various structures associated with a task group */
8363 static void free_sched_group_rcu(struct rcu_head *rhp)
8365 /* now it should be safe to free those cfs_rqs */
8366 free_sched_group(container_of(rhp, struct task_group, rcu));
8369 /* Destroy runqueue etc associated with a task group */
8370 void sched_destroy_group(struct task_group *tg)
8372 unsigned long flags;
8375 /* end participation in shares distribution */
8376 for_each_possible_cpu(i)
8377 unregister_fair_sched_group(tg, i);
8379 spin_lock_irqsave(&task_group_lock, flags);
8380 list_del_rcu(&tg->list);
8381 list_del_rcu(&tg->siblings);
8382 spin_unlock_irqrestore(&task_group_lock, flags);
8384 /* wait for possible concurrent references to cfs_rqs complete */
8385 call_rcu(&tg->rcu, free_sched_group_rcu);
8388 /* change task's runqueue when it moves between groups.
8389 * The caller of this function should have put the task in its new group
8390 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8391 * reflect its new group.
8393 void sched_move_task(struct task_struct *tsk)
8396 unsigned long flags;
8399 rq = task_rq_lock(tsk, &flags);
8401 running = task_current(rq, tsk);
8405 dequeue_task(rq, tsk, 0);
8406 if (unlikely(running))
8407 tsk->sched_class->put_prev_task(rq, tsk);
8409 #ifdef CONFIG_FAIR_GROUP_SCHED
8410 if (tsk->sched_class->task_move_group)
8411 tsk->sched_class->task_move_group(tsk, on_rq);
8414 set_task_rq(tsk, task_cpu(tsk));
8416 if (unlikely(running))
8417 tsk->sched_class->set_curr_task(rq);
8419 enqueue_task(rq, tsk, 0);
8421 task_rq_unlock(rq, tsk, &flags);
8423 #endif /* CONFIG_CGROUP_SCHED */
8425 #ifdef CONFIG_FAIR_GROUP_SCHED
8426 static DEFINE_MUTEX(shares_mutex);
8428 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8431 unsigned long flags;
8434 * We can't change the weight of the root cgroup.
8439 if (shares < MIN_SHARES)
8440 shares = MIN_SHARES;
8441 else if (shares > MAX_SHARES)
8442 shares = MAX_SHARES;
8444 mutex_lock(&shares_mutex);
8445 if (tg->shares == shares)
8448 tg->shares = shares;
8449 for_each_possible_cpu(i) {
8450 struct rq *rq = cpu_rq(i);
8451 struct sched_entity *se;
8454 /* Propagate contribution to hierarchy */
8455 raw_spin_lock_irqsave(&rq->lock, flags);
8456 for_each_sched_entity(se)
8457 update_cfs_shares(group_cfs_rq(se));
8458 raw_spin_unlock_irqrestore(&rq->lock, flags);
8462 mutex_unlock(&shares_mutex);
8466 unsigned long sched_group_shares(struct task_group *tg)
8472 #ifdef CONFIG_RT_GROUP_SCHED
8474 * Ensure that the real time constraints are schedulable.
8476 static DEFINE_MUTEX(rt_constraints_mutex);
8478 static unsigned long to_ratio(u64 period, u64 runtime)
8480 if (runtime == RUNTIME_INF)
8483 return div64_u64(runtime << 20, period);
8486 /* Must be called with tasklist_lock held */
8487 static inline int tg_has_rt_tasks(struct task_group *tg)
8489 struct task_struct *g, *p;
8491 do_each_thread(g, p) {
8492 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8494 } while_each_thread(g, p);
8499 struct rt_schedulable_data {
8500 struct task_group *tg;
8505 static int tg_schedulable(struct task_group *tg, void *data)
8507 struct rt_schedulable_data *d = data;
8508 struct task_group *child;
8509 unsigned long total, sum = 0;
8510 u64 period, runtime;
8512 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8513 runtime = tg->rt_bandwidth.rt_runtime;
8516 period = d->rt_period;
8517 runtime = d->rt_runtime;
8521 * Cannot have more runtime than the period.
8523 if (runtime > period && runtime != RUNTIME_INF)
8527 * Ensure we don't starve existing RT tasks.
8529 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8532 total = to_ratio(period, runtime);
8535 * Nobody can have more than the global setting allows.
8537 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8541 * The sum of our children's runtime should not exceed our own.
8543 list_for_each_entry_rcu(child, &tg->children, siblings) {
8544 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8545 runtime = child->rt_bandwidth.rt_runtime;
8547 if (child == d->tg) {
8548 period = d->rt_period;
8549 runtime = d->rt_runtime;
8552 sum += to_ratio(period, runtime);
8561 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8563 struct rt_schedulable_data data = {
8565 .rt_period = period,
8566 .rt_runtime = runtime,
8569 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8572 static int tg_set_bandwidth(struct task_group *tg,
8573 u64 rt_period, u64 rt_runtime)
8577 mutex_lock(&rt_constraints_mutex);
8578 read_lock(&tasklist_lock);
8579 err = __rt_schedulable(tg, rt_period, rt_runtime);
8583 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8584 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8585 tg->rt_bandwidth.rt_runtime = rt_runtime;
8587 for_each_possible_cpu(i) {
8588 struct rt_rq *rt_rq = tg->rt_rq[i];
8590 raw_spin_lock(&rt_rq->rt_runtime_lock);
8591 rt_rq->rt_runtime = rt_runtime;
8592 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8594 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8596 read_unlock(&tasklist_lock);
8597 mutex_unlock(&rt_constraints_mutex);
8602 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8604 u64 rt_runtime, rt_period;
8606 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8607 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8608 if (rt_runtime_us < 0)
8609 rt_runtime = RUNTIME_INF;
8611 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8614 long sched_group_rt_runtime(struct task_group *tg)
8618 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8621 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8622 do_div(rt_runtime_us, NSEC_PER_USEC);
8623 return rt_runtime_us;
8626 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8628 u64 rt_runtime, rt_period;
8630 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8631 rt_runtime = tg->rt_bandwidth.rt_runtime;
8636 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8639 long sched_group_rt_period(struct task_group *tg)
8643 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8644 do_div(rt_period_us, NSEC_PER_USEC);
8645 return rt_period_us;
8648 static int sched_rt_global_constraints(void)
8650 u64 runtime, period;
8653 if (sysctl_sched_rt_period <= 0)
8656 runtime = global_rt_runtime();
8657 period = global_rt_period();
8660 * Sanity check on the sysctl variables.
8662 if (runtime > period && runtime != RUNTIME_INF)
8665 mutex_lock(&rt_constraints_mutex);
8666 read_lock(&tasklist_lock);
8667 ret = __rt_schedulable(NULL, 0, 0);
8668 read_unlock(&tasklist_lock);
8669 mutex_unlock(&rt_constraints_mutex);
8674 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8676 /* Don't accept realtime tasks when there is no way for them to run */
8677 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8683 #else /* !CONFIG_RT_GROUP_SCHED */
8684 static int sched_rt_global_constraints(void)
8686 unsigned long flags;
8689 if (sysctl_sched_rt_period <= 0)
8693 * There's always some RT tasks in the root group
8694 * -- migration, kstopmachine etc..
8696 if (sysctl_sched_rt_runtime == 0)
8699 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8700 for_each_possible_cpu(i) {
8701 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8703 raw_spin_lock(&rt_rq->rt_runtime_lock);
8704 rt_rq->rt_runtime = global_rt_runtime();
8705 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8707 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8711 #endif /* CONFIG_RT_GROUP_SCHED */
8713 int sched_rt_handler(struct ctl_table *table, int write,
8714 void __user *buffer, size_t *lenp,
8718 int old_period, old_runtime;
8719 static DEFINE_MUTEX(mutex);
8722 old_period = sysctl_sched_rt_period;
8723 old_runtime = sysctl_sched_rt_runtime;
8725 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8727 if (!ret && write) {
8728 ret = sched_rt_global_constraints();
8730 sysctl_sched_rt_period = old_period;
8731 sysctl_sched_rt_runtime = old_runtime;
8733 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8734 def_rt_bandwidth.rt_period =
8735 ns_to_ktime(global_rt_period());
8738 mutex_unlock(&mutex);
8743 #ifdef CONFIG_CGROUP_SCHED
8745 /* return corresponding task_group object of a cgroup */
8746 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8748 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8749 struct task_group, css);
8752 static struct cgroup_subsys_state *
8753 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8755 struct task_group *tg, *parent;
8757 if (!cgrp->parent) {
8758 /* This is early initialization for the top cgroup */
8759 return &root_task_group.css;
8762 parent = cgroup_tg(cgrp->parent);
8763 tg = sched_create_group(parent);
8765 return ERR_PTR(-ENOMEM);
8771 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8773 struct task_group *tg = cgroup_tg(cgrp);
8775 sched_destroy_group(tg);
8779 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8781 #ifdef CONFIG_RT_GROUP_SCHED
8782 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
8785 /* We don't support RT-tasks being in separate groups */
8786 if (tsk->sched_class != &fair_sched_class)
8793 cpu_cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8795 sched_move_task(tsk);
8799 cpu_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
8800 struct cgroup *old_cgrp, struct task_struct *task)
8803 * cgroup_exit() is called in the copy_process() failure path.
8804 * Ignore this case since the task hasn't ran yet, this avoids
8805 * trying to poke a half freed task state from generic code.
8807 if (!(task->flags & PF_EXITING))
8810 sched_move_task(task);
8813 #ifdef CONFIG_FAIR_GROUP_SCHED
8814 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8817 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
8820 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8822 struct task_group *tg = cgroup_tg(cgrp);
8824 return (u64) scale_load_down(tg->shares);
8826 #endif /* CONFIG_FAIR_GROUP_SCHED */
8828 #ifdef CONFIG_RT_GROUP_SCHED
8829 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8832 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8835 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8837 return sched_group_rt_runtime(cgroup_tg(cgrp));
8840 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8843 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8846 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8848 return sched_group_rt_period(cgroup_tg(cgrp));
8850 #endif /* CONFIG_RT_GROUP_SCHED */
8852 static struct cftype cpu_files[] = {
8853 #ifdef CONFIG_FAIR_GROUP_SCHED
8856 .read_u64 = cpu_shares_read_u64,
8857 .write_u64 = cpu_shares_write_u64,
8860 #ifdef CONFIG_RT_GROUP_SCHED
8862 .name = "rt_runtime_us",
8863 .read_s64 = cpu_rt_runtime_read,
8864 .write_s64 = cpu_rt_runtime_write,
8867 .name = "rt_period_us",
8868 .read_u64 = cpu_rt_period_read_uint,
8869 .write_u64 = cpu_rt_period_write_uint,
8874 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8876 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8879 struct cgroup_subsys cpu_cgroup_subsys = {
8881 .create = cpu_cgroup_create,
8882 .destroy = cpu_cgroup_destroy,
8883 .can_attach_task = cpu_cgroup_can_attach_task,
8884 .attach_task = cpu_cgroup_attach_task,
8885 .exit = cpu_cgroup_exit,
8886 .populate = cpu_cgroup_populate,
8887 .subsys_id = cpu_cgroup_subsys_id,
8891 #endif /* CONFIG_CGROUP_SCHED */
8893 #ifdef CONFIG_CGROUP_CPUACCT
8896 * CPU accounting code for task groups.
8898 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8899 * (balbir@in.ibm.com).
8902 /* track cpu usage of a group of tasks and its child groups */
8904 struct cgroup_subsys_state css;
8905 /* cpuusage holds pointer to a u64-type object on every cpu */
8906 u64 __percpu *cpuusage;
8907 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
8908 struct cpuacct *parent;
8911 struct cgroup_subsys cpuacct_subsys;
8913 /* return cpu accounting group corresponding to this container */
8914 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8916 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8917 struct cpuacct, css);
8920 /* return cpu accounting group to which this task belongs */
8921 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8923 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8924 struct cpuacct, css);
8927 /* create a new cpu accounting group */
8928 static struct cgroup_subsys_state *cpuacct_create(
8929 struct cgroup_subsys *ss, struct cgroup *cgrp)
8931 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8937 ca->cpuusage = alloc_percpu(u64);
8941 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8942 if (percpu_counter_init(&ca->cpustat[i], 0))
8943 goto out_free_counters;
8946 ca->parent = cgroup_ca(cgrp->parent);
8952 percpu_counter_destroy(&ca->cpustat[i]);
8953 free_percpu(ca->cpuusage);
8957 return ERR_PTR(-ENOMEM);
8960 /* destroy an existing cpu accounting group */
8962 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8964 struct cpuacct *ca = cgroup_ca(cgrp);
8967 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8968 percpu_counter_destroy(&ca->cpustat[i]);
8969 free_percpu(ca->cpuusage);
8973 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
8975 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8978 #ifndef CONFIG_64BIT
8980 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8982 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8984 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8992 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
8994 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8996 #ifndef CONFIG_64BIT
8998 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9000 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9002 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9008 /* return total cpu usage (in nanoseconds) of a group */
9009 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9011 struct cpuacct *ca = cgroup_ca(cgrp);
9012 u64 totalcpuusage = 0;
9015 for_each_present_cpu(i)
9016 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9018 return totalcpuusage;
9021 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9024 struct cpuacct *ca = cgroup_ca(cgrp);
9033 for_each_present_cpu(i)
9034 cpuacct_cpuusage_write(ca, i, 0);
9040 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9043 struct cpuacct *ca = cgroup_ca(cgroup);
9047 for_each_present_cpu(i) {
9048 percpu = cpuacct_cpuusage_read(ca, i);
9049 seq_printf(m, "%llu ", (unsigned long long) percpu);
9051 seq_printf(m, "\n");
9055 static const char *cpuacct_stat_desc[] = {
9056 [CPUACCT_STAT_USER] = "user",
9057 [CPUACCT_STAT_SYSTEM] = "system",
9060 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
9061 struct cgroup_map_cb *cb)
9063 struct cpuacct *ca = cgroup_ca(cgrp);
9066 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
9067 s64 val = percpu_counter_read(&ca->cpustat[i]);
9068 val = cputime64_to_clock_t(val);
9069 cb->fill(cb, cpuacct_stat_desc[i], val);
9074 static struct cftype files[] = {
9077 .read_u64 = cpuusage_read,
9078 .write_u64 = cpuusage_write,
9081 .name = "usage_percpu",
9082 .read_seq_string = cpuacct_percpu_seq_read,
9086 .read_map = cpuacct_stats_show,
9090 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9092 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9096 * charge this task's execution time to its accounting group.
9098 * called with rq->lock held.
9100 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9105 if (unlikely(!cpuacct_subsys.active))
9108 cpu = task_cpu(tsk);
9114 for (; ca; ca = ca->parent) {
9115 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9116 *cpuusage += cputime;
9123 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9124 * in cputime_t units. As a result, cpuacct_update_stats calls
9125 * percpu_counter_add with values large enough to always overflow the
9126 * per cpu batch limit causing bad SMP scalability.
9128 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9129 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9130 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9133 #define CPUACCT_BATCH \
9134 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9136 #define CPUACCT_BATCH 0
9140 * Charge the system/user time to the task's accounting group.
9142 static void cpuacct_update_stats(struct task_struct *tsk,
9143 enum cpuacct_stat_index idx, cputime_t val)
9146 int batch = CPUACCT_BATCH;
9148 if (unlikely(!cpuacct_subsys.active))
9155 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9161 struct cgroup_subsys cpuacct_subsys = {
9163 .create = cpuacct_create,
9164 .destroy = cpuacct_destroy,
9165 .populate = cpuacct_populate,
9166 .subsys_id = cpuacct_subsys_id,
9168 #endif /* CONFIG_CGROUP_CPUACCT */