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>
78 #ifdef CONFIG_PARAVIRT
79 #include <asm/paravirt.h>
82 #include "sched_cpupri.h"
83 #include "workqueue_sched.h"
84 #include "sched_autogroup.h"
86 #define CREATE_TRACE_POINTS
87 #include <trace/events/sched.h>
90 * Convert user-nice values [ -20 ... 0 ... 19 ]
91 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
94 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
95 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
96 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
99 * 'User priority' is the nice value converted to something we
100 * can work with better when scaling various scheduler parameters,
101 * it's a [ 0 ... 39 ] range.
103 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
104 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
105 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
108 * Helpers for converting nanosecond timing to jiffy resolution
110 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
112 #define NICE_0_LOAD SCHED_LOAD_SCALE
113 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
116 * These are the 'tuning knobs' of the scheduler:
118 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
119 * Timeslices get refilled after they expire.
121 #define DEF_TIMESLICE (100 * HZ / 1000)
124 * single value that denotes runtime == period, ie unlimited time.
126 #define RUNTIME_INF ((u64)~0ULL)
128 static inline int rt_policy(int policy)
130 if (policy == SCHED_FIFO || policy == SCHED_RR)
135 static inline int task_has_rt_policy(struct task_struct *p)
137 return rt_policy(p->policy);
141 * This is the priority-queue data structure of the RT scheduling class:
143 struct rt_prio_array {
144 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
145 struct list_head queue[MAX_RT_PRIO];
148 struct rt_bandwidth {
149 /* nests inside the rq lock: */
150 raw_spinlock_t rt_runtime_lock;
153 struct hrtimer rt_period_timer;
156 static struct rt_bandwidth def_rt_bandwidth;
158 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
160 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
162 struct rt_bandwidth *rt_b =
163 container_of(timer, struct rt_bandwidth, rt_period_timer);
169 now = hrtimer_cb_get_time(timer);
170 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
175 idle = do_sched_rt_period_timer(rt_b, overrun);
178 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
182 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
184 rt_b->rt_period = ns_to_ktime(period);
185 rt_b->rt_runtime = runtime;
187 raw_spin_lock_init(&rt_b->rt_runtime_lock);
189 hrtimer_init(&rt_b->rt_period_timer,
190 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
191 rt_b->rt_period_timer.function = sched_rt_period_timer;
194 static inline int rt_bandwidth_enabled(void)
196 return sysctl_sched_rt_runtime >= 0;
199 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
203 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
206 if (hrtimer_active(&rt_b->rt_period_timer))
209 raw_spin_lock(&rt_b->rt_runtime_lock);
214 if (hrtimer_active(&rt_b->rt_period_timer))
217 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
218 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
220 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
221 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
222 delta = ktime_to_ns(ktime_sub(hard, soft));
223 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
224 HRTIMER_MODE_ABS_PINNED, 0);
226 raw_spin_unlock(&rt_b->rt_runtime_lock);
229 #ifdef CONFIG_RT_GROUP_SCHED
230 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
232 hrtimer_cancel(&rt_b->rt_period_timer);
237 * sched_domains_mutex serializes calls to init_sched_domains,
238 * detach_destroy_domains and partition_sched_domains.
240 static DEFINE_MUTEX(sched_domains_mutex);
242 #ifdef CONFIG_CGROUP_SCHED
244 #include <linux/cgroup.h>
248 static LIST_HEAD(task_groups);
250 /* task group related information */
252 struct cgroup_subsys_state css;
254 #ifdef CONFIG_FAIR_GROUP_SCHED
255 /* schedulable entities of this group on each cpu */
256 struct sched_entity **se;
257 /* runqueue "owned" by this group on each cpu */
258 struct cfs_rq **cfs_rq;
259 unsigned long shares;
261 atomic_t load_weight;
264 #ifdef CONFIG_RT_GROUP_SCHED
265 struct sched_rt_entity **rt_se;
266 struct rt_rq **rt_rq;
268 struct rt_bandwidth rt_bandwidth;
272 struct list_head list;
274 struct task_group *parent;
275 struct list_head siblings;
276 struct list_head children;
278 #ifdef CONFIG_SCHED_AUTOGROUP
279 struct autogroup *autogroup;
283 /* task_group_lock serializes the addition/removal of task groups */
284 static DEFINE_SPINLOCK(task_group_lock);
286 #ifdef CONFIG_FAIR_GROUP_SCHED
288 # define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
291 * A weight of 0 or 1 can cause arithmetics problems.
292 * A weight of a cfs_rq is the sum of weights of which entities
293 * are queued on this cfs_rq, so a weight of a entity should not be
294 * too large, so as the shares value of a task group.
295 * (The default weight is 1024 - so there's no practical
296 * limitation from this.)
298 #define MIN_SHARES (1UL << 1)
299 #define MAX_SHARES (1UL << 18)
301 static int root_task_group_load = ROOT_TASK_GROUP_LOAD;
304 /* Default task group.
305 * Every task in system belong to this group at bootup.
307 struct task_group root_task_group;
309 #endif /* CONFIG_CGROUP_SCHED */
311 /* CFS-related fields in a runqueue */
313 struct load_weight load;
314 unsigned long nr_running, h_nr_running;
319 u64 min_vruntime_copy;
322 struct rb_root tasks_timeline;
323 struct rb_node *rb_leftmost;
325 struct list_head tasks;
326 struct list_head *balance_iterator;
329 * 'curr' points to currently running entity on this cfs_rq.
330 * It is set to NULL otherwise (i.e when none are currently running).
332 struct sched_entity *curr, *next, *last, *skip;
334 #ifdef CONFIG_SCHED_DEBUG
335 unsigned int nr_spread_over;
338 #ifdef CONFIG_FAIR_GROUP_SCHED
339 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
342 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
343 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
344 * (like users, containers etc.)
346 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
347 * list is used during load balance.
350 struct list_head leaf_cfs_rq_list;
351 struct task_group *tg; /* group that "owns" this runqueue */
355 * the part of load.weight contributed by tasks
357 unsigned long task_weight;
360 * h_load = weight * f(tg)
362 * Where f(tg) is the recursive weight fraction assigned to
365 unsigned long h_load;
368 * Maintaining per-cpu shares distribution for group scheduling
370 * load_stamp is the last time we updated the load average
371 * load_last is the last time we updated the load average and saw load
372 * load_unacc_exec_time is currently unaccounted execution time
376 u64 load_stamp, load_last, load_unacc_exec_time;
378 unsigned long load_contribution;
383 /* Real-Time classes' related field in a runqueue: */
385 struct rt_prio_array active;
386 unsigned long rt_nr_running;
387 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
389 int curr; /* highest queued rt task prio */
391 int next; /* next highest */
396 unsigned long rt_nr_migratory;
397 unsigned long rt_nr_total;
399 struct plist_head pushable_tasks;
404 /* Nests inside the rq lock: */
405 raw_spinlock_t rt_runtime_lock;
407 #ifdef CONFIG_RT_GROUP_SCHED
408 unsigned long rt_nr_boosted;
411 struct list_head leaf_rt_rq_list;
412 struct task_group *tg;
419 * We add the notion of a root-domain which will be used to define per-domain
420 * variables. Each exclusive cpuset essentially defines an island domain by
421 * fully partitioning the member cpus from any other cpuset. Whenever a new
422 * exclusive cpuset is created, we also create and attach a new root-domain
431 cpumask_var_t online;
434 * The "RT overload" flag: it gets set if a CPU has more than
435 * one runnable RT task.
437 cpumask_var_t rto_mask;
438 struct cpupri cpupri;
442 * By default the system creates a single root-domain with all cpus as
443 * members (mimicking the global state we have today).
445 static struct root_domain def_root_domain;
447 #endif /* CONFIG_SMP */
450 * This is the main, per-CPU runqueue data structure.
452 * Locking rule: those places that want to lock multiple runqueues
453 * (such as the load balancing or the thread migration code), lock
454 * acquire operations must be ordered by ascending &runqueue.
461 * nr_running and cpu_load should be in the same cacheline because
462 * remote CPUs use both these fields when doing load calculation.
464 unsigned long nr_running;
465 #define CPU_LOAD_IDX_MAX 5
466 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
467 unsigned long last_load_update_tick;
470 unsigned char nohz_balance_kick;
472 int skip_clock_update;
474 /* capture load from *all* tasks on this cpu: */
475 struct load_weight load;
476 unsigned long nr_load_updates;
482 #ifdef CONFIG_FAIR_GROUP_SCHED
483 /* list of leaf cfs_rq on this cpu: */
484 struct list_head leaf_cfs_rq_list;
486 #ifdef CONFIG_RT_GROUP_SCHED
487 struct list_head leaf_rt_rq_list;
491 * This is part of a global counter where only the total sum
492 * over all CPUs matters. A task can increase this counter on
493 * one CPU and if it got migrated afterwards it may decrease
494 * it on another CPU. Always updated under the runqueue lock:
496 unsigned long nr_uninterruptible;
498 struct task_struct *curr, *idle, *stop;
499 unsigned long next_balance;
500 struct mm_struct *prev_mm;
508 struct root_domain *rd;
509 struct sched_domain *sd;
511 unsigned long cpu_power;
513 unsigned char idle_at_tick;
514 /* For active balancing */
518 struct cpu_stop_work active_balance_work;
519 /* cpu of this runqueue: */
529 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
532 #ifdef CONFIG_PARAVIRT
535 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
536 u64 prev_steal_time_rq;
539 /* calc_load related fields */
540 unsigned long calc_load_update;
541 long calc_load_active;
543 #ifdef CONFIG_SCHED_HRTICK
545 int hrtick_csd_pending;
546 struct call_single_data hrtick_csd;
548 struct hrtimer hrtick_timer;
551 #ifdef CONFIG_SCHEDSTATS
553 struct sched_info rq_sched_info;
554 unsigned long long rq_cpu_time;
555 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
557 /* sys_sched_yield() stats */
558 unsigned int yld_count;
560 /* schedule() stats */
561 unsigned int sched_switch;
562 unsigned int sched_count;
563 unsigned int sched_goidle;
565 /* try_to_wake_up() stats */
566 unsigned int ttwu_count;
567 unsigned int ttwu_local;
571 struct task_struct *wake_list;
575 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
578 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags);
580 static inline int cpu_of(struct rq *rq)
589 #define rcu_dereference_check_sched_domain(p) \
590 rcu_dereference_check((p), \
591 lockdep_is_held(&sched_domains_mutex))
594 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
595 * See detach_destroy_domains: synchronize_sched for details.
597 * The domain tree of any CPU may only be accessed from within
598 * preempt-disabled sections.
600 #define for_each_domain(cpu, __sd) \
601 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
603 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
604 #define this_rq() (&__get_cpu_var(runqueues))
605 #define task_rq(p) cpu_rq(task_cpu(p))
606 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
607 #define raw_rq() (&__raw_get_cpu_var(runqueues))
609 #ifdef CONFIG_CGROUP_SCHED
612 * Return the group to which this tasks belongs.
614 * We use task_subsys_state_check() and extend the RCU verification with
615 * pi->lock and rq->lock because cpu_cgroup_attach() holds those locks for each
616 * task it moves into the cgroup. Therefore by holding either of those locks,
617 * we pin the task to the current cgroup.
619 static inline struct task_group *task_group(struct task_struct *p)
621 struct task_group *tg;
622 struct cgroup_subsys_state *css;
624 css = task_subsys_state_check(p, cpu_cgroup_subsys_id,
625 lockdep_is_held(&p->pi_lock) ||
626 lockdep_is_held(&task_rq(p)->lock));
627 tg = container_of(css, struct task_group, css);
629 return autogroup_task_group(p, tg);
632 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
633 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
635 #ifdef CONFIG_FAIR_GROUP_SCHED
636 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
637 p->se.parent = task_group(p)->se[cpu];
640 #ifdef CONFIG_RT_GROUP_SCHED
641 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
642 p->rt.parent = task_group(p)->rt_se[cpu];
646 #else /* CONFIG_CGROUP_SCHED */
648 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
649 static inline struct task_group *task_group(struct task_struct *p)
654 #endif /* CONFIG_CGROUP_SCHED */
656 static void update_rq_clock_task(struct rq *rq, s64 delta);
658 static void update_rq_clock(struct rq *rq)
662 if (rq->skip_clock_update > 0)
665 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
667 update_rq_clock_task(rq, delta);
671 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
673 #ifdef CONFIG_SCHED_DEBUG
674 # define const_debug __read_mostly
676 # define const_debug static const
680 * runqueue_is_locked - Returns true if the current cpu runqueue is locked
681 * @cpu: the processor in question.
683 * This interface allows printk to be called with the runqueue lock
684 * held and know whether or not it is OK to wake up the klogd.
686 int runqueue_is_locked(int cpu)
688 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
692 * Debugging: various feature bits
695 #define SCHED_FEAT(name, enabled) \
696 __SCHED_FEAT_##name ,
699 #include "sched_features.h"
704 #define SCHED_FEAT(name, enabled) \
705 (1UL << __SCHED_FEAT_##name) * enabled |
707 const_debug unsigned int sysctl_sched_features =
708 #include "sched_features.h"
713 #ifdef CONFIG_SCHED_DEBUG
714 #define SCHED_FEAT(name, enabled) \
717 static __read_mostly char *sched_feat_names[] = {
718 #include "sched_features.h"
724 static int sched_feat_show(struct seq_file *m, void *v)
728 for (i = 0; sched_feat_names[i]; i++) {
729 if (!(sysctl_sched_features & (1UL << i)))
731 seq_printf(m, "%s ", sched_feat_names[i]);
739 sched_feat_write(struct file *filp, const char __user *ubuf,
740 size_t cnt, loff_t *ppos)
750 if (copy_from_user(&buf, ubuf, cnt))
756 if (strncmp(cmp, "NO_", 3) == 0) {
761 for (i = 0; sched_feat_names[i]; i++) {
762 if (strcmp(cmp, sched_feat_names[i]) == 0) {
764 sysctl_sched_features &= ~(1UL << i);
766 sysctl_sched_features |= (1UL << i);
771 if (!sched_feat_names[i])
779 static int sched_feat_open(struct inode *inode, struct file *filp)
781 return single_open(filp, sched_feat_show, NULL);
784 static const struct file_operations sched_feat_fops = {
785 .open = sched_feat_open,
786 .write = sched_feat_write,
789 .release = single_release,
792 static __init int sched_init_debug(void)
794 debugfs_create_file("sched_features", 0644, NULL, NULL,
799 late_initcall(sched_init_debug);
803 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
806 * Number of tasks to iterate in a single balance run.
807 * Limited because this is done with IRQs disabled.
809 const_debug unsigned int sysctl_sched_nr_migrate = 32;
812 * period over which we average the RT time consumption, measured
817 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
820 * period over which we measure -rt task cpu usage in us.
823 unsigned int sysctl_sched_rt_period = 1000000;
825 static __read_mostly int scheduler_running;
828 * part of the period that we allow rt tasks to run in us.
831 int sysctl_sched_rt_runtime = 950000;
833 static inline u64 global_rt_period(void)
835 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
838 static inline u64 global_rt_runtime(void)
840 if (sysctl_sched_rt_runtime < 0)
843 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
846 #ifndef prepare_arch_switch
847 # define prepare_arch_switch(next) do { } while (0)
849 #ifndef finish_arch_switch
850 # define finish_arch_switch(prev) do { } while (0)
853 static inline int task_current(struct rq *rq, struct task_struct *p)
855 return rq->curr == p;
858 static inline int task_running(struct rq *rq, struct task_struct *p)
863 return task_current(rq, p);
867 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
868 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
872 * We can optimise this out completely for !SMP, because the
873 * SMP rebalancing from interrupt is the only thing that cares
880 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
884 * After ->on_cpu is cleared, the task can be moved to a different CPU.
885 * We must ensure this doesn't happen until the switch is completely
891 #ifdef CONFIG_DEBUG_SPINLOCK
892 /* this is a valid case when another task releases the spinlock */
893 rq->lock.owner = current;
896 * If we are tracking spinlock dependencies then we have to
897 * fix up the runqueue lock - which gets 'carried over' from
900 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
902 raw_spin_unlock_irq(&rq->lock);
905 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
906 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
910 * We can optimise this out completely for !SMP, because the
911 * SMP rebalancing from interrupt is the only thing that cares
916 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
917 raw_spin_unlock_irq(&rq->lock);
919 raw_spin_unlock(&rq->lock);
923 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
927 * After ->on_cpu is cleared, the task can be moved to a different CPU.
928 * We must ensure this doesn't happen until the switch is completely
934 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
938 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
941 * __task_rq_lock - lock the rq @p resides on.
943 static inline struct rq *__task_rq_lock(struct task_struct *p)
948 lockdep_assert_held(&p->pi_lock);
952 raw_spin_lock(&rq->lock);
953 if (likely(rq == task_rq(p)))
955 raw_spin_unlock(&rq->lock);
960 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
962 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
963 __acquires(p->pi_lock)
969 raw_spin_lock_irqsave(&p->pi_lock, *flags);
971 raw_spin_lock(&rq->lock);
972 if (likely(rq == task_rq(p)))
974 raw_spin_unlock(&rq->lock);
975 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
979 static void __task_rq_unlock(struct rq *rq)
982 raw_spin_unlock(&rq->lock);
986 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
988 __releases(p->pi_lock)
990 raw_spin_unlock(&rq->lock);
991 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
995 * this_rq_lock - lock this runqueue and disable interrupts.
997 static struct rq *this_rq_lock(void)
1002 local_irq_disable();
1004 raw_spin_lock(&rq->lock);
1009 #ifdef CONFIG_SCHED_HRTICK
1011 * Use HR-timers to deliver accurate preemption points.
1013 * Its all a bit involved since we cannot program an hrt while holding the
1014 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1017 * When we get rescheduled we reprogram the hrtick_timer outside of the
1023 * - enabled by features
1024 * - hrtimer is actually high res
1026 static inline int hrtick_enabled(struct rq *rq)
1028 if (!sched_feat(HRTICK))
1030 if (!cpu_active(cpu_of(rq)))
1032 return hrtimer_is_hres_active(&rq->hrtick_timer);
1035 static void hrtick_clear(struct rq *rq)
1037 if (hrtimer_active(&rq->hrtick_timer))
1038 hrtimer_cancel(&rq->hrtick_timer);
1042 * High-resolution timer tick.
1043 * Runs from hardirq context with interrupts disabled.
1045 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1047 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1049 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1051 raw_spin_lock(&rq->lock);
1052 update_rq_clock(rq);
1053 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1054 raw_spin_unlock(&rq->lock);
1056 return HRTIMER_NORESTART;
1061 * called from hardirq (IPI) context
1063 static void __hrtick_start(void *arg)
1065 struct rq *rq = arg;
1067 raw_spin_lock(&rq->lock);
1068 hrtimer_restart(&rq->hrtick_timer);
1069 rq->hrtick_csd_pending = 0;
1070 raw_spin_unlock(&rq->lock);
1074 * Called to set the hrtick timer state.
1076 * called with rq->lock held and irqs disabled
1078 static void hrtick_start(struct rq *rq, u64 delay)
1080 struct hrtimer *timer = &rq->hrtick_timer;
1081 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1083 hrtimer_set_expires(timer, time);
1085 if (rq == this_rq()) {
1086 hrtimer_restart(timer);
1087 } else if (!rq->hrtick_csd_pending) {
1088 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1089 rq->hrtick_csd_pending = 1;
1094 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1096 int cpu = (int)(long)hcpu;
1099 case CPU_UP_CANCELED:
1100 case CPU_UP_CANCELED_FROZEN:
1101 case CPU_DOWN_PREPARE:
1102 case CPU_DOWN_PREPARE_FROZEN:
1104 case CPU_DEAD_FROZEN:
1105 hrtick_clear(cpu_rq(cpu));
1112 static __init void init_hrtick(void)
1114 hotcpu_notifier(hotplug_hrtick, 0);
1118 * Called to set the hrtick timer state.
1120 * called with rq->lock held and irqs disabled
1122 static void hrtick_start(struct rq *rq, u64 delay)
1124 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1125 HRTIMER_MODE_REL_PINNED, 0);
1128 static inline void init_hrtick(void)
1131 #endif /* CONFIG_SMP */
1133 static void init_rq_hrtick(struct rq *rq)
1136 rq->hrtick_csd_pending = 0;
1138 rq->hrtick_csd.flags = 0;
1139 rq->hrtick_csd.func = __hrtick_start;
1140 rq->hrtick_csd.info = rq;
1143 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1144 rq->hrtick_timer.function = hrtick;
1146 #else /* CONFIG_SCHED_HRTICK */
1147 static inline void hrtick_clear(struct rq *rq)
1151 static inline void init_rq_hrtick(struct rq *rq)
1155 static inline void init_hrtick(void)
1158 #endif /* CONFIG_SCHED_HRTICK */
1161 * resched_task - mark a task 'to be rescheduled now'.
1163 * On UP this means the setting of the need_resched flag, on SMP it
1164 * might also involve a cross-CPU call to trigger the scheduler on
1169 #ifndef tsk_is_polling
1170 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1173 static void resched_task(struct task_struct *p)
1177 assert_raw_spin_locked(&task_rq(p)->lock);
1179 if (test_tsk_need_resched(p))
1182 set_tsk_need_resched(p);
1185 if (cpu == smp_processor_id())
1188 /* NEED_RESCHED must be visible before we test polling */
1190 if (!tsk_is_polling(p))
1191 smp_send_reschedule(cpu);
1194 static void resched_cpu(int cpu)
1196 struct rq *rq = cpu_rq(cpu);
1197 unsigned long flags;
1199 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1201 resched_task(cpu_curr(cpu));
1202 raw_spin_unlock_irqrestore(&rq->lock, flags);
1207 * In the semi idle case, use the nearest busy cpu for migrating timers
1208 * from an idle cpu. This is good for power-savings.
1210 * We don't do similar optimization for completely idle system, as
1211 * selecting an idle cpu will add more delays to the timers than intended
1212 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1214 int get_nohz_timer_target(void)
1216 int cpu = smp_processor_id();
1218 struct sched_domain *sd;
1221 for_each_domain(cpu, sd) {
1222 for_each_cpu(i, sched_domain_span(sd)) {
1234 * When add_timer_on() enqueues a timer into the timer wheel of an
1235 * idle CPU then this timer might expire before the next timer event
1236 * which is scheduled to wake up that CPU. In case of a completely
1237 * idle system the next event might even be infinite time into the
1238 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1239 * leaves the inner idle loop so the newly added timer is taken into
1240 * account when the CPU goes back to idle and evaluates the timer
1241 * wheel for the next timer event.
1243 void wake_up_idle_cpu(int cpu)
1245 struct rq *rq = cpu_rq(cpu);
1247 if (cpu == smp_processor_id())
1251 * This is safe, as this function is called with the timer
1252 * wheel base lock of (cpu) held. When the CPU is on the way
1253 * to idle and has not yet set rq->curr to idle then it will
1254 * be serialized on the timer wheel base lock and take the new
1255 * timer into account automatically.
1257 if (rq->curr != rq->idle)
1261 * We can set TIF_RESCHED on the idle task of the other CPU
1262 * lockless. The worst case is that the other CPU runs the
1263 * idle task through an additional NOOP schedule()
1265 set_tsk_need_resched(rq->idle);
1267 /* NEED_RESCHED must be visible before we test polling */
1269 if (!tsk_is_polling(rq->idle))
1270 smp_send_reschedule(cpu);
1273 #endif /* CONFIG_NO_HZ */
1275 static u64 sched_avg_period(void)
1277 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1280 static void sched_avg_update(struct rq *rq)
1282 s64 period = sched_avg_period();
1284 while ((s64)(rq->clock - rq->age_stamp) > period) {
1286 * Inline assembly required to prevent the compiler
1287 * optimising this loop into a divmod call.
1288 * See __iter_div_u64_rem() for another example of this.
1290 asm("" : "+rm" (rq->age_stamp));
1291 rq->age_stamp += period;
1296 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1298 rq->rt_avg += rt_delta;
1299 sched_avg_update(rq);
1302 #else /* !CONFIG_SMP */
1303 static void resched_task(struct task_struct *p)
1305 assert_raw_spin_locked(&task_rq(p)->lock);
1306 set_tsk_need_resched(p);
1309 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1313 static void sched_avg_update(struct rq *rq)
1316 #endif /* CONFIG_SMP */
1318 #if BITS_PER_LONG == 32
1319 # define WMULT_CONST (~0UL)
1321 # define WMULT_CONST (1UL << 32)
1324 #define WMULT_SHIFT 32
1327 * Shift right and round:
1329 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1332 * delta *= weight / lw
1334 static unsigned long
1335 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1336 struct load_weight *lw)
1341 * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched
1342 * entities since MIN_SHARES = 2. Treat weight as 1 if less than
1343 * 2^SCHED_LOAD_RESOLUTION.
1345 if (likely(weight > (1UL << SCHED_LOAD_RESOLUTION)))
1346 tmp = (u64)delta_exec * scale_load_down(weight);
1348 tmp = (u64)delta_exec;
1350 if (!lw->inv_weight) {
1351 unsigned long w = scale_load_down(lw->weight);
1353 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
1355 else if (unlikely(!w))
1356 lw->inv_weight = WMULT_CONST;
1358 lw->inv_weight = WMULT_CONST / w;
1362 * Check whether we'd overflow the 64-bit multiplication:
1364 if (unlikely(tmp > WMULT_CONST))
1365 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1368 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1370 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1373 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1379 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1385 static inline void update_load_set(struct load_weight *lw, unsigned long w)
1392 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1393 * of tasks with abnormal "nice" values across CPUs the contribution that
1394 * each task makes to its run queue's load is weighted according to its
1395 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1396 * scaled version of the new time slice allocation that they receive on time
1400 #define WEIGHT_IDLEPRIO 3
1401 #define WMULT_IDLEPRIO 1431655765
1404 * Nice levels are multiplicative, with a gentle 10% change for every
1405 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1406 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1407 * that remained on nice 0.
1409 * The "10% effect" is relative and cumulative: from _any_ nice level,
1410 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1411 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1412 * If a task goes up by ~10% and another task goes down by ~10% then
1413 * the relative distance between them is ~25%.)
1415 static const int prio_to_weight[40] = {
1416 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1417 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1418 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1419 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1420 /* 0 */ 1024, 820, 655, 526, 423,
1421 /* 5 */ 335, 272, 215, 172, 137,
1422 /* 10 */ 110, 87, 70, 56, 45,
1423 /* 15 */ 36, 29, 23, 18, 15,
1427 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1429 * In cases where the weight does not change often, we can use the
1430 * precalculated inverse to speed up arithmetics by turning divisions
1431 * into multiplications:
1433 static const u32 prio_to_wmult[40] = {
1434 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1435 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1436 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1437 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1438 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1439 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1440 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1441 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1444 /* Time spent by the tasks of the cpu accounting group executing in ... */
1445 enum cpuacct_stat_index {
1446 CPUACCT_STAT_USER, /* ... user mode */
1447 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1449 CPUACCT_STAT_NSTATS,
1452 #ifdef CONFIG_CGROUP_CPUACCT
1453 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1454 static void cpuacct_update_stats(struct task_struct *tsk,
1455 enum cpuacct_stat_index idx, cputime_t val);
1457 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1458 static inline void cpuacct_update_stats(struct task_struct *tsk,
1459 enum cpuacct_stat_index idx, cputime_t val) {}
1462 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1464 update_load_add(&rq->load, load);
1467 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1469 update_load_sub(&rq->load, load);
1472 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1473 typedef int (*tg_visitor)(struct task_group *, void *);
1476 * Iterate the full tree, calling @down when first entering a node and @up when
1477 * leaving it for the final time.
1479 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1481 struct task_group *parent, *child;
1485 parent = &root_task_group;
1487 ret = (*down)(parent, data);
1490 list_for_each_entry_rcu(child, &parent->children, siblings) {
1497 ret = (*up)(parent, data);
1502 parent = parent->parent;
1511 static int tg_nop(struct task_group *tg, void *data)
1518 /* Used instead of source_load when we know the type == 0 */
1519 static unsigned long weighted_cpuload(const int cpu)
1521 return cpu_rq(cpu)->load.weight;
1525 * Return a low guess at the load of a migration-source cpu weighted
1526 * according to the scheduling class and "nice" value.
1528 * We want to under-estimate the load of migration sources, to
1529 * balance conservatively.
1531 static unsigned long source_load(int cpu, int type)
1533 struct rq *rq = cpu_rq(cpu);
1534 unsigned long total = weighted_cpuload(cpu);
1536 if (type == 0 || !sched_feat(LB_BIAS))
1539 return min(rq->cpu_load[type-1], total);
1543 * Return a high guess at the load of a migration-target cpu weighted
1544 * according to the scheduling class and "nice" value.
1546 static unsigned long target_load(int cpu, int type)
1548 struct rq *rq = cpu_rq(cpu);
1549 unsigned long total = weighted_cpuload(cpu);
1551 if (type == 0 || !sched_feat(LB_BIAS))
1554 return max(rq->cpu_load[type-1], total);
1557 static unsigned long power_of(int cpu)
1559 return cpu_rq(cpu)->cpu_power;
1562 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1564 static unsigned long cpu_avg_load_per_task(int cpu)
1566 struct rq *rq = cpu_rq(cpu);
1567 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1570 return rq->load.weight / nr_running;
1575 #ifdef CONFIG_PREEMPT
1577 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1580 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1581 * way at the expense of forcing extra atomic operations in all
1582 * invocations. This assures that the double_lock is acquired using the
1583 * same underlying policy as the spinlock_t on this architecture, which
1584 * reduces latency compared to the unfair variant below. However, it
1585 * also adds more overhead and therefore may reduce throughput.
1587 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1588 __releases(this_rq->lock)
1589 __acquires(busiest->lock)
1590 __acquires(this_rq->lock)
1592 raw_spin_unlock(&this_rq->lock);
1593 double_rq_lock(this_rq, busiest);
1600 * Unfair double_lock_balance: Optimizes throughput at the expense of
1601 * latency by eliminating extra atomic operations when the locks are
1602 * already in proper order on entry. This favors lower cpu-ids and will
1603 * grant the double lock to lower cpus over higher ids under contention,
1604 * regardless of entry order into the function.
1606 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1607 __releases(this_rq->lock)
1608 __acquires(busiest->lock)
1609 __acquires(this_rq->lock)
1613 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1614 if (busiest < this_rq) {
1615 raw_spin_unlock(&this_rq->lock);
1616 raw_spin_lock(&busiest->lock);
1617 raw_spin_lock_nested(&this_rq->lock,
1618 SINGLE_DEPTH_NESTING);
1621 raw_spin_lock_nested(&busiest->lock,
1622 SINGLE_DEPTH_NESTING);
1627 #endif /* CONFIG_PREEMPT */
1630 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1632 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1634 if (unlikely(!irqs_disabled())) {
1635 /* printk() doesn't work good under rq->lock */
1636 raw_spin_unlock(&this_rq->lock);
1640 return _double_lock_balance(this_rq, busiest);
1643 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1644 __releases(busiest->lock)
1646 raw_spin_unlock(&busiest->lock);
1647 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1651 * double_rq_lock - safely lock two runqueues
1653 * Note this does not disable interrupts like task_rq_lock,
1654 * you need to do so manually before calling.
1656 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1657 __acquires(rq1->lock)
1658 __acquires(rq2->lock)
1660 BUG_ON(!irqs_disabled());
1662 raw_spin_lock(&rq1->lock);
1663 __acquire(rq2->lock); /* Fake it out ;) */
1666 raw_spin_lock(&rq1->lock);
1667 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1669 raw_spin_lock(&rq2->lock);
1670 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1676 * double_rq_unlock - safely unlock two runqueues
1678 * Note this does not restore interrupts like task_rq_unlock,
1679 * you need to do so manually after calling.
1681 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1682 __releases(rq1->lock)
1683 __releases(rq2->lock)
1685 raw_spin_unlock(&rq1->lock);
1687 raw_spin_unlock(&rq2->lock);
1689 __release(rq2->lock);
1692 #else /* CONFIG_SMP */
1695 * double_rq_lock - safely lock two runqueues
1697 * Note this does not disable interrupts like task_rq_lock,
1698 * you need to do so manually before calling.
1700 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1701 __acquires(rq1->lock)
1702 __acquires(rq2->lock)
1704 BUG_ON(!irqs_disabled());
1706 raw_spin_lock(&rq1->lock);
1707 __acquire(rq2->lock); /* Fake it out ;) */
1711 * double_rq_unlock - safely unlock two runqueues
1713 * Note this does not restore interrupts like task_rq_unlock,
1714 * you need to do so manually after calling.
1716 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1717 __releases(rq1->lock)
1718 __releases(rq2->lock)
1721 raw_spin_unlock(&rq1->lock);
1722 __release(rq2->lock);
1727 static void calc_load_account_idle(struct rq *this_rq);
1728 static void update_sysctl(void);
1729 static int get_update_sysctl_factor(void);
1730 static void update_cpu_load(struct rq *this_rq);
1732 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1734 set_task_rq(p, cpu);
1737 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1738 * successfuly executed on another CPU. We must ensure that updates of
1739 * per-task data have been completed by this moment.
1742 task_thread_info(p)->cpu = cpu;
1746 static const struct sched_class rt_sched_class;
1748 #define sched_class_highest (&stop_sched_class)
1749 #define for_each_class(class) \
1750 for (class = sched_class_highest; class; class = class->next)
1752 #include "sched_stats.h"
1754 static void inc_nr_running(struct rq *rq)
1759 static void dec_nr_running(struct rq *rq)
1764 static void set_load_weight(struct task_struct *p)
1766 int prio = p->static_prio - MAX_RT_PRIO;
1767 struct load_weight *load = &p->se.load;
1770 * SCHED_IDLE tasks get minimal weight:
1772 if (p->policy == SCHED_IDLE) {
1773 load->weight = scale_load(WEIGHT_IDLEPRIO);
1774 load->inv_weight = WMULT_IDLEPRIO;
1778 load->weight = scale_load(prio_to_weight[prio]);
1779 load->inv_weight = prio_to_wmult[prio];
1782 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1784 update_rq_clock(rq);
1785 sched_info_queued(p);
1786 p->sched_class->enqueue_task(rq, p, flags);
1789 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1791 update_rq_clock(rq);
1792 sched_info_dequeued(p);
1793 p->sched_class->dequeue_task(rq, p, flags);
1797 * activate_task - move a task to the runqueue.
1799 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1801 if (task_contributes_to_load(p))
1802 rq->nr_uninterruptible--;
1804 enqueue_task(rq, p, flags);
1808 * deactivate_task - remove a task from the runqueue.
1810 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1812 if (task_contributes_to_load(p))
1813 rq->nr_uninterruptible++;
1815 dequeue_task(rq, p, flags);
1818 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1821 * There are no locks covering percpu hardirq/softirq time.
1822 * They are only modified in account_system_vtime, on corresponding CPU
1823 * with interrupts disabled. So, writes are safe.
1824 * They are read and saved off onto struct rq in update_rq_clock().
1825 * This may result in other CPU reading this CPU's irq time and can
1826 * race with irq/account_system_vtime on this CPU. We would either get old
1827 * or new value with a side effect of accounting a slice of irq time to wrong
1828 * task when irq is in progress while we read rq->clock. That is a worthy
1829 * compromise in place of having locks on each irq in account_system_time.
1831 static DEFINE_PER_CPU(u64, cpu_hardirq_time);
1832 static DEFINE_PER_CPU(u64, cpu_softirq_time);
1834 static DEFINE_PER_CPU(u64, irq_start_time);
1835 static int sched_clock_irqtime;
1837 void enable_sched_clock_irqtime(void)
1839 sched_clock_irqtime = 1;
1842 void disable_sched_clock_irqtime(void)
1844 sched_clock_irqtime = 0;
1847 #ifndef CONFIG_64BIT
1848 static DEFINE_PER_CPU(seqcount_t, irq_time_seq);
1850 static inline void irq_time_write_begin(void)
1852 __this_cpu_inc(irq_time_seq.sequence);
1856 static inline void irq_time_write_end(void)
1859 __this_cpu_inc(irq_time_seq.sequence);
1862 static inline u64 irq_time_read(int cpu)
1868 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
1869 irq_time = per_cpu(cpu_softirq_time, cpu) +
1870 per_cpu(cpu_hardirq_time, cpu);
1871 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
1875 #else /* CONFIG_64BIT */
1876 static inline void irq_time_write_begin(void)
1880 static inline void irq_time_write_end(void)
1884 static inline u64 irq_time_read(int cpu)
1886 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
1888 #endif /* CONFIG_64BIT */
1891 * Called before incrementing preempt_count on {soft,}irq_enter
1892 * and before decrementing preempt_count on {soft,}irq_exit.
1894 void account_system_vtime(struct task_struct *curr)
1896 unsigned long flags;
1900 if (!sched_clock_irqtime)
1903 local_irq_save(flags);
1905 cpu = smp_processor_id();
1906 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
1907 __this_cpu_add(irq_start_time, delta);
1909 irq_time_write_begin();
1911 * We do not account for softirq time from ksoftirqd here.
1912 * We want to continue accounting softirq time to ksoftirqd thread
1913 * in that case, so as not to confuse scheduler with a special task
1914 * that do not consume any time, but still wants to run.
1916 if (hardirq_count())
1917 __this_cpu_add(cpu_hardirq_time, delta);
1918 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
1919 __this_cpu_add(cpu_softirq_time, delta);
1921 irq_time_write_end();
1922 local_irq_restore(flags);
1924 EXPORT_SYMBOL_GPL(account_system_vtime);
1926 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
1928 #ifdef CONFIG_PARAVIRT
1929 static inline u64 steal_ticks(u64 steal)
1931 if (unlikely(steal > NSEC_PER_SEC))
1932 return div_u64(steal, TICK_NSEC);
1934 return __iter_div_u64_rem(steal, TICK_NSEC, &steal);
1938 static void update_rq_clock_task(struct rq *rq, s64 delta)
1941 * In theory, the compile should just see 0 here, and optimize out the call
1942 * to sched_rt_avg_update. But I don't trust it...
1944 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
1945 s64 steal = 0, irq_delta = 0;
1947 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1948 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
1951 * Since irq_time is only updated on {soft,}irq_exit, we might run into
1952 * this case when a previous update_rq_clock() happened inside a
1953 * {soft,}irq region.
1955 * When this happens, we stop ->clock_task and only update the
1956 * prev_irq_time stamp to account for the part that fit, so that a next
1957 * update will consume the rest. This ensures ->clock_task is
1960 * It does however cause some slight miss-attribution of {soft,}irq
1961 * time, a more accurate solution would be to update the irq_time using
1962 * the current rq->clock timestamp, except that would require using
1965 if (irq_delta > delta)
1968 rq->prev_irq_time += irq_delta;
1971 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
1972 if (static_branch((¶virt_steal_rq_enabled))) {
1975 steal = paravirt_steal_clock(cpu_of(rq));
1976 steal -= rq->prev_steal_time_rq;
1978 if (unlikely(steal > delta))
1981 st = steal_ticks(steal);
1982 steal = st * TICK_NSEC;
1984 rq->prev_steal_time_rq += steal;
1990 rq->clock_task += delta;
1992 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
1993 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
1994 sched_rt_avg_update(rq, irq_delta + steal);
1998 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1999 static int irqtime_account_hi_update(void)
2001 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2002 unsigned long flags;
2006 local_irq_save(flags);
2007 latest_ns = this_cpu_read(cpu_hardirq_time);
2008 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->irq))
2010 local_irq_restore(flags);
2014 static int irqtime_account_si_update(void)
2016 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2017 unsigned long flags;
2021 local_irq_save(flags);
2022 latest_ns = this_cpu_read(cpu_softirq_time);
2023 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->softirq))
2025 local_irq_restore(flags);
2029 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
2031 #define sched_clock_irqtime (0)
2035 #include "sched_idletask.c"
2036 #include "sched_fair.c"
2037 #include "sched_rt.c"
2038 #include "sched_autogroup.c"
2039 #include "sched_stoptask.c"
2040 #ifdef CONFIG_SCHED_DEBUG
2041 # include "sched_debug.c"
2044 void sched_set_stop_task(int cpu, struct task_struct *stop)
2046 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
2047 struct task_struct *old_stop = cpu_rq(cpu)->stop;
2051 * Make it appear like a SCHED_FIFO task, its something
2052 * userspace knows about and won't get confused about.
2054 * Also, it will make PI more or less work without too
2055 * much confusion -- but then, stop work should not
2056 * rely on PI working anyway.
2058 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
2060 stop->sched_class = &stop_sched_class;
2063 cpu_rq(cpu)->stop = stop;
2067 * Reset it back to a normal scheduling class so that
2068 * it can die in pieces.
2070 old_stop->sched_class = &rt_sched_class;
2075 * __normal_prio - return the priority that is based on the static prio
2077 static inline int __normal_prio(struct task_struct *p)
2079 return p->static_prio;
2083 * Calculate the expected normal priority: i.e. priority
2084 * without taking RT-inheritance into account. Might be
2085 * boosted by interactivity modifiers. Changes upon fork,
2086 * setprio syscalls, and whenever the interactivity
2087 * estimator recalculates.
2089 static inline int normal_prio(struct task_struct *p)
2093 if (task_has_rt_policy(p))
2094 prio = MAX_RT_PRIO-1 - p->rt_priority;
2096 prio = __normal_prio(p);
2101 * Calculate the current priority, i.e. the priority
2102 * taken into account by the scheduler. This value might
2103 * be boosted by RT tasks, or might be boosted by
2104 * interactivity modifiers. Will be RT if the task got
2105 * RT-boosted. If not then it returns p->normal_prio.
2107 static int effective_prio(struct task_struct *p)
2109 p->normal_prio = normal_prio(p);
2111 * If we are RT tasks or we were boosted to RT priority,
2112 * keep the priority unchanged. Otherwise, update priority
2113 * to the normal priority:
2115 if (!rt_prio(p->prio))
2116 return p->normal_prio;
2121 * task_curr - is this task currently executing on a CPU?
2122 * @p: the task in question.
2124 inline int task_curr(const struct task_struct *p)
2126 return cpu_curr(task_cpu(p)) == p;
2129 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2130 const struct sched_class *prev_class,
2133 if (prev_class != p->sched_class) {
2134 if (prev_class->switched_from)
2135 prev_class->switched_from(rq, p);
2136 p->sched_class->switched_to(rq, p);
2137 } else if (oldprio != p->prio)
2138 p->sched_class->prio_changed(rq, p, oldprio);
2141 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2143 const struct sched_class *class;
2145 if (p->sched_class == rq->curr->sched_class) {
2146 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2148 for_each_class(class) {
2149 if (class == rq->curr->sched_class)
2151 if (class == p->sched_class) {
2152 resched_task(rq->curr);
2159 * A queue event has occurred, and we're going to schedule. In
2160 * this case, we can save a useless back to back clock update.
2162 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
2163 rq->skip_clock_update = 1;
2168 * Is this task likely cache-hot:
2171 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2175 if (p->sched_class != &fair_sched_class)
2178 if (unlikely(p->policy == SCHED_IDLE))
2182 * Buddy candidates are cache hot:
2184 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2185 (&p->se == cfs_rq_of(&p->se)->next ||
2186 &p->se == cfs_rq_of(&p->se)->last))
2189 if (sysctl_sched_migration_cost == -1)
2191 if (sysctl_sched_migration_cost == 0)
2194 delta = now - p->se.exec_start;
2196 return delta < (s64)sysctl_sched_migration_cost;
2199 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2201 #ifdef CONFIG_SCHED_DEBUG
2203 * We should never call set_task_cpu() on a blocked task,
2204 * ttwu() will sort out the placement.
2206 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2207 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2209 #ifdef CONFIG_LOCKDEP
2211 * The caller should hold either p->pi_lock or rq->lock, when changing
2212 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
2214 * sched_move_task() holds both and thus holding either pins the cgroup,
2215 * see set_task_rq().
2217 * Furthermore, all task_rq users should acquire both locks, see
2220 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
2221 lockdep_is_held(&task_rq(p)->lock)));
2225 trace_sched_migrate_task(p, new_cpu);
2227 if (task_cpu(p) != new_cpu) {
2228 p->se.nr_migrations++;
2229 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
2232 __set_task_cpu(p, new_cpu);
2235 struct migration_arg {
2236 struct task_struct *task;
2240 static int migration_cpu_stop(void *data);
2243 * wait_task_inactive - wait for a thread to unschedule.
2245 * If @match_state is nonzero, it's the @p->state value just checked and
2246 * not expected to change. If it changes, i.e. @p might have woken up,
2247 * then return zero. When we succeed in waiting for @p to be off its CPU,
2248 * we return a positive number (its total switch count). If a second call
2249 * a short while later returns the same number, the caller can be sure that
2250 * @p has remained unscheduled the whole time.
2252 * The caller must ensure that the task *will* unschedule sometime soon,
2253 * else this function might spin for a *long* time. This function can't
2254 * be called with interrupts off, or it may introduce deadlock with
2255 * smp_call_function() if an IPI is sent by the same process we are
2256 * waiting to become inactive.
2258 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2260 unsigned long flags;
2267 * We do the initial early heuristics without holding
2268 * any task-queue locks at all. We'll only try to get
2269 * the runqueue lock when things look like they will
2275 * If the task is actively running on another CPU
2276 * still, just relax and busy-wait without holding
2279 * NOTE! Since we don't hold any locks, it's not
2280 * even sure that "rq" stays as the right runqueue!
2281 * But we don't care, since "task_running()" will
2282 * return false if the runqueue has changed and p
2283 * is actually now running somewhere else!
2285 while (task_running(rq, p)) {
2286 if (match_state && unlikely(p->state != match_state))
2292 * Ok, time to look more closely! We need the rq
2293 * lock now, to be *sure*. If we're wrong, we'll
2294 * just go back and repeat.
2296 rq = task_rq_lock(p, &flags);
2297 trace_sched_wait_task(p);
2298 running = task_running(rq, p);
2301 if (!match_state || p->state == match_state)
2302 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2303 task_rq_unlock(rq, p, &flags);
2306 * If it changed from the expected state, bail out now.
2308 if (unlikely(!ncsw))
2312 * Was it really running after all now that we
2313 * checked with the proper locks actually held?
2315 * Oops. Go back and try again..
2317 if (unlikely(running)) {
2323 * It's not enough that it's not actively running,
2324 * it must be off the runqueue _entirely_, and not
2327 * So if it was still runnable (but just not actively
2328 * running right now), it's preempted, and we should
2329 * yield - it could be a while.
2331 if (unlikely(on_rq)) {
2332 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
2334 set_current_state(TASK_UNINTERRUPTIBLE);
2335 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
2340 * Ahh, all good. It wasn't running, and it wasn't
2341 * runnable, which means that it will never become
2342 * running in the future either. We're all done!
2351 * kick_process - kick a running thread to enter/exit the kernel
2352 * @p: the to-be-kicked thread
2354 * Cause a process which is running on another CPU to enter
2355 * kernel-mode, without any delay. (to get signals handled.)
2357 * NOTE: this function doesn't have to take the runqueue lock,
2358 * because all it wants to ensure is that the remote task enters
2359 * the kernel. If the IPI races and the task has been migrated
2360 * to another CPU then no harm is done and the purpose has been
2363 void kick_process(struct task_struct *p)
2369 if ((cpu != smp_processor_id()) && task_curr(p))
2370 smp_send_reschedule(cpu);
2373 EXPORT_SYMBOL_GPL(kick_process);
2374 #endif /* CONFIG_SMP */
2378 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
2380 static int select_fallback_rq(int cpu, struct task_struct *p)
2383 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2385 /* Look for allowed, online CPU in same node. */
2386 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2387 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2390 /* Any allowed, online CPU? */
2391 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2392 if (dest_cpu < nr_cpu_ids)
2395 /* No more Mr. Nice Guy. */
2396 dest_cpu = cpuset_cpus_allowed_fallback(p);
2398 * Don't tell them about moving exiting tasks or
2399 * kernel threads (both mm NULL), since they never
2402 if (p->mm && printk_ratelimit()) {
2403 printk(KERN_INFO "process %d (%s) no longer affine to cpu%d\n",
2404 task_pid_nr(p), p->comm, cpu);
2411 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
2414 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2416 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2419 * In order not to call set_task_cpu() on a blocking task we need
2420 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2423 * Since this is common to all placement strategies, this lives here.
2425 * [ this allows ->select_task() to simply return task_cpu(p) and
2426 * not worry about this generic constraint ]
2428 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2430 cpu = select_fallback_rq(task_cpu(p), p);
2435 static void update_avg(u64 *avg, u64 sample)
2437 s64 diff = sample - *avg;
2443 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
2445 #ifdef CONFIG_SCHEDSTATS
2446 struct rq *rq = this_rq();
2449 int this_cpu = smp_processor_id();
2451 if (cpu == this_cpu) {
2452 schedstat_inc(rq, ttwu_local);
2453 schedstat_inc(p, se.statistics.nr_wakeups_local);
2455 struct sched_domain *sd;
2457 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2459 for_each_domain(this_cpu, sd) {
2460 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2461 schedstat_inc(sd, ttwu_wake_remote);
2468 if (wake_flags & WF_MIGRATED)
2469 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2471 #endif /* CONFIG_SMP */
2473 schedstat_inc(rq, ttwu_count);
2474 schedstat_inc(p, se.statistics.nr_wakeups);
2476 if (wake_flags & WF_SYNC)
2477 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2479 #endif /* CONFIG_SCHEDSTATS */
2482 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
2484 activate_task(rq, p, en_flags);
2487 /* if a worker is waking up, notify workqueue */
2488 if (p->flags & PF_WQ_WORKER)
2489 wq_worker_waking_up(p, cpu_of(rq));
2493 * Mark the task runnable and perform wakeup-preemption.
2496 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
2498 trace_sched_wakeup(p, true);
2499 check_preempt_curr(rq, p, wake_flags);
2501 p->state = TASK_RUNNING;
2503 if (p->sched_class->task_woken)
2504 p->sched_class->task_woken(rq, p);
2506 if (rq->idle_stamp) {
2507 u64 delta = rq->clock - rq->idle_stamp;
2508 u64 max = 2*sysctl_sched_migration_cost;
2513 update_avg(&rq->avg_idle, delta);
2520 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
2523 if (p->sched_contributes_to_load)
2524 rq->nr_uninterruptible--;
2527 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
2528 ttwu_do_wakeup(rq, p, wake_flags);
2532 * Called in case the task @p isn't fully descheduled from its runqueue,
2533 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
2534 * since all we need to do is flip p->state to TASK_RUNNING, since
2535 * the task is still ->on_rq.
2537 static int ttwu_remote(struct task_struct *p, int wake_flags)
2542 rq = __task_rq_lock(p);
2544 ttwu_do_wakeup(rq, p, wake_flags);
2547 __task_rq_unlock(rq);
2553 static void sched_ttwu_do_pending(struct task_struct *list)
2555 struct rq *rq = this_rq();
2557 raw_spin_lock(&rq->lock);
2560 struct task_struct *p = list;
2561 list = list->wake_entry;
2562 ttwu_do_activate(rq, p, 0);
2565 raw_spin_unlock(&rq->lock);
2568 #ifdef CONFIG_HOTPLUG_CPU
2570 static void sched_ttwu_pending(void)
2572 struct rq *rq = this_rq();
2573 struct task_struct *list = xchg(&rq->wake_list, NULL);
2578 sched_ttwu_do_pending(list);
2581 #endif /* CONFIG_HOTPLUG_CPU */
2583 void scheduler_ipi(void)
2585 struct rq *rq = this_rq();
2586 struct task_struct *list = xchg(&rq->wake_list, NULL);
2592 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
2593 * traditionally all their work was done from the interrupt return
2594 * path. Now that we actually do some work, we need to make sure
2597 * Some archs already do call them, luckily irq_enter/exit nest
2600 * Arguably we should visit all archs and update all handlers,
2601 * however a fair share of IPIs are still resched only so this would
2602 * somewhat pessimize the simple resched case.
2605 sched_ttwu_do_pending(list);
2609 static void ttwu_queue_remote(struct task_struct *p, int cpu)
2611 struct rq *rq = cpu_rq(cpu);
2612 struct task_struct *next = rq->wake_list;
2615 struct task_struct *old = next;
2617 p->wake_entry = next;
2618 next = cmpxchg(&rq->wake_list, old, p);
2624 smp_send_reschedule(cpu);
2627 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2628 static int ttwu_activate_remote(struct task_struct *p, int wake_flags)
2633 rq = __task_rq_lock(p);
2635 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2636 ttwu_do_wakeup(rq, p, wake_flags);
2639 __task_rq_unlock(rq);
2644 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2645 #endif /* CONFIG_SMP */
2647 static void ttwu_queue(struct task_struct *p, int cpu)
2649 struct rq *rq = cpu_rq(cpu);
2651 #if defined(CONFIG_SMP)
2652 if (sched_feat(TTWU_QUEUE) && cpu != smp_processor_id()) {
2653 sched_clock_cpu(cpu); /* sync clocks x-cpu */
2654 ttwu_queue_remote(p, cpu);
2659 raw_spin_lock(&rq->lock);
2660 ttwu_do_activate(rq, p, 0);
2661 raw_spin_unlock(&rq->lock);
2665 * try_to_wake_up - wake up a thread
2666 * @p: the thread to be awakened
2667 * @state: the mask of task states that can be woken
2668 * @wake_flags: wake modifier flags (WF_*)
2670 * Put it on the run-queue if it's not already there. The "current"
2671 * thread is always on the run-queue (except when the actual
2672 * re-schedule is in progress), and as such you're allowed to do
2673 * the simpler "current->state = TASK_RUNNING" to mark yourself
2674 * runnable without the overhead of this.
2676 * Returns %true if @p was woken up, %false if it was already running
2677 * or @state didn't match @p's state.
2680 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
2682 unsigned long flags;
2683 int cpu, success = 0;
2686 raw_spin_lock_irqsave(&p->pi_lock, flags);
2687 if (!(p->state & state))
2690 success = 1; /* we're going to change ->state */
2693 if (p->on_rq && ttwu_remote(p, wake_flags))
2698 * If the owning (remote) cpu is still in the middle of schedule() with
2699 * this task as prev, wait until its done referencing the task.
2702 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2704 * In case the architecture enables interrupts in
2705 * context_switch(), we cannot busy wait, since that
2706 * would lead to deadlocks when an interrupt hits and
2707 * tries to wake up @prev. So bail and do a complete
2710 if (ttwu_activate_remote(p, wake_flags))
2717 * Pairs with the smp_wmb() in finish_lock_switch().
2721 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2722 p->state = TASK_WAKING;
2724 if (p->sched_class->task_waking)
2725 p->sched_class->task_waking(p);
2727 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2728 if (task_cpu(p) != cpu) {
2729 wake_flags |= WF_MIGRATED;
2730 set_task_cpu(p, cpu);
2732 #endif /* CONFIG_SMP */
2736 ttwu_stat(p, cpu, wake_flags);
2738 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2744 * try_to_wake_up_local - try to wake up a local task with rq lock held
2745 * @p: the thread to be awakened
2747 * Put @p on the run-queue if it's not already there. The caller must
2748 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2751 static void try_to_wake_up_local(struct task_struct *p)
2753 struct rq *rq = task_rq(p);
2755 BUG_ON(rq != this_rq());
2756 BUG_ON(p == current);
2757 lockdep_assert_held(&rq->lock);
2759 if (!raw_spin_trylock(&p->pi_lock)) {
2760 raw_spin_unlock(&rq->lock);
2761 raw_spin_lock(&p->pi_lock);
2762 raw_spin_lock(&rq->lock);
2765 if (!(p->state & TASK_NORMAL))
2769 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2771 ttwu_do_wakeup(rq, p, 0);
2772 ttwu_stat(p, smp_processor_id(), 0);
2774 raw_spin_unlock(&p->pi_lock);
2778 * wake_up_process - Wake up a specific process
2779 * @p: The process to be woken up.
2781 * Attempt to wake up the nominated process and move it to the set of runnable
2782 * processes. Returns 1 if the process was woken up, 0 if it was already
2785 * It may be assumed that this function implies a write memory barrier before
2786 * changing the task state if and only if any tasks are woken up.
2788 int wake_up_process(struct task_struct *p)
2790 return try_to_wake_up(p, TASK_ALL, 0);
2792 EXPORT_SYMBOL(wake_up_process);
2794 int wake_up_state(struct task_struct *p, unsigned int state)
2796 return try_to_wake_up(p, state, 0);
2800 * Perform scheduler related setup for a newly forked process p.
2801 * p is forked by current.
2803 * __sched_fork() is basic setup used by init_idle() too:
2805 static void __sched_fork(struct task_struct *p)
2810 p->se.exec_start = 0;
2811 p->se.sum_exec_runtime = 0;
2812 p->se.prev_sum_exec_runtime = 0;
2813 p->se.nr_migrations = 0;
2815 INIT_LIST_HEAD(&p->se.group_node);
2817 #ifdef CONFIG_SCHEDSTATS
2818 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2821 INIT_LIST_HEAD(&p->rt.run_list);
2823 #ifdef CONFIG_PREEMPT_NOTIFIERS
2824 INIT_HLIST_HEAD(&p->preempt_notifiers);
2829 * fork()/clone()-time setup:
2831 void sched_fork(struct task_struct *p)
2833 unsigned long flags;
2834 int cpu = get_cpu();
2838 * We mark the process as running here. This guarantees that
2839 * nobody will actually run it, and a signal or other external
2840 * event cannot wake it up and insert it on the runqueue either.
2842 p->state = TASK_RUNNING;
2845 * Make sure we do not leak PI boosting priority to the child.
2847 p->prio = current->normal_prio;
2850 * Revert to default priority/policy on fork if requested.
2852 if (unlikely(p->sched_reset_on_fork)) {
2853 if (task_has_rt_policy(p)) {
2854 p->policy = SCHED_NORMAL;
2855 p->static_prio = NICE_TO_PRIO(0);
2857 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2858 p->static_prio = NICE_TO_PRIO(0);
2860 p->prio = p->normal_prio = __normal_prio(p);
2864 * We don't need the reset flag anymore after the fork. It has
2865 * fulfilled its duty:
2867 p->sched_reset_on_fork = 0;
2870 if (!rt_prio(p->prio))
2871 p->sched_class = &fair_sched_class;
2873 if (p->sched_class->task_fork)
2874 p->sched_class->task_fork(p);
2877 * The child is not yet in the pid-hash so no cgroup attach races,
2878 * and the cgroup is pinned to this child due to cgroup_fork()
2879 * is ran before sched_fork().
2881 * Silence PROVE_RCU.
2883 raw_spin_lock_irqsave(&p->pi_lock, flags);
2884 set_task_cpu(p, cpu);
2885 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2887 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2888 if (likely(sched_info_on()))
2889 memset(&p->sched_info, 0, sizeof(p->sched_info));
2891 #if defined(CONFIG_SMP)
2894 #ifdef CONFIG_PREEMPT_COUNT
2895 /* Want to start with kernel preemption disabled. */
2896 task_thread_info(p)->preempt_count = 1;
2899 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2906 * wake_up_new_task - wake up a newly created task for the first time.
2908 * This function will do some initial scheduler statistics housekeeping
2909 * that must be done for every newly created context, then puts the task
2910 * on the runqueue and wakes it.
2912 void wake_up_new_task(struct task_struct *p)
2914 unsigned long flags;
2917 raw_spin_lock_irqsave(&p->pi_lock, flags);
2920 * Fork balancing, do it here and not earlier because:
2921 * - cpus_allowed can change in the fork path
2922 * - any previously selected cpu might disappear through hotplug
2924 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
2927 rq = __task_rq_lock(p);
2928 activate_task(rq, p, 0);
2930 trace_sched_wakeup_new(p, true);
2931 check_preempt_curr(rq, p, WF_FORK);
2933 if (p->sched_class->task_woken)
2934 p->sched_class->task_woken(rq, p);
2936 task_rq_unlock(rq, p, &flags);
2939 #ifdef CONFIG_PREEMPT_NOTIFIERS
2942 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2943 * @notifier: notifier struct to register
2945 void preempt_notifier_register(struct preempt_notifier *notifier)
2947 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2949 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2952 * preempt_notifier_unregister - no longer interested in preemption notifications
2953 * @notifier: notifier struct to unregister
2955 * This is safe to call from within a preemption notifier.
2957 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2959 hlist_del(¬ifier->link);
2961 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2963 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2965 struct preempt_notifier *notifier;
2966 struct hlist_node *node;
2968 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2969 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2973 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2974 struct task_struct *next)
2976 struct preempt_notifier *notifier;
2977 struct hlist_node *node;
2979 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2980 notifier->ops->sched_out(notifier, next);
2983 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2985 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2990 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2991 struct task_struct *next)
2995 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2998 * prepare_task_switch - prepare to switch tasks
2999 * @rq: the runqueue preparing to switch
3000 * @prev: the current task that is being switched out
3001 * @next: the task we are going to switch to.
3003 * This is called with the rq lock held and interrupts off. It must
3004 * be paired with a subsequent finish_task_switch after the context
3007 * prepare_task_switch sets up locking and calls architecture specific
3011 prepare_task_switch(struct rq *rq, struct task_struct *prev,
3012 struct task_struct *next)
3014 sched_info_switch(prev, next);
3015 perf_event_task_sched_out(prev, next);
3016 fire_sched_out_preempt_notifiers(prev, next);
3017 prepare_lock_switch(rq, next);
3018 prepare_arch_switch(next);
3019 trace_sched_switch(prev, next);
3023 * finish_task_switch - clean up after a task-switch
3024 * @rq: runqueue associated with task-switch
3025 * @prev: the thread we just switched away from.
3027 * finish_task_switch must be called after the context switch, paired
3028 * with a prepare_task_switch call before the context switch.
3029 * finish_task_switch will reconcile locking set up by prepare_task_switch,
3030 * and do any other architecture-specific cleanup actions.
3032 * Note that we may have delayed dropping an mm in context_switch(). If
3033 * so, we finish that here outside of the runqueue lock. (Doing it
3034 * with the lock held can cause deadlocks; see schedule() for
3037 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
3038 __releases(rq->lock)
3040 struct mm_struct *mm = rq->prev_mm;
3046 * A task struct has one reference for the use as "current".
3047 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
3048 * schedule one last time. The schedule call will never return, and
3049 * the scheduled task must drop that reference.
3050 * The test for TASK_DEAD must occur while the runqueue locks are
3051 * still held, otherwise prev could be scheduled on another cpu, die
3052 * there before we look at prev->state, and then the reference would
3054 * Manfred Spraul <manfred@colorfullife.com>
3056 prev_state = prev->state;
3057 finish_arch_switch(prev);
3058 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3059 local_irq_disable();
3060 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3061 perf_event_task_sched_in(current);
3062 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3064 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3065 finish_lock_switch(rq, prev);
3067 fire_sched_in_preempt_notifiers(current);
3070 if (unlikely(prev_state == TASK_DEAD)) {
3072 * Remove function-return probe instances associated with this
3073 * task and put them back on the free list.
3075 kprobe_flush_task(prev);
3076 put_task_struct(prev);
3082 /* assumes rq->lock is held */
3083 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
3085 if (prev->sched_class->pre_schedule)
3086 prev->sched_class->pre_schedule(rq, prev);
3089 /* rq->lock is NOT held, but preemption is disabled */
3090 static inline void post_schedule(struct rq *rq)
3092 if (rq->post_schedule) {
3093 unsigned long flags;
3095 raw_spin_lock_irqsave(&rq->lock, flags);
3096 if (rq->curr->sched_class->post_schedule)
3097 rq->curr->sched_class->post_schedule(rq);
3098 raw_spin_unlock_irqrestore(&rq->lock, flags);
3100 rq->post_schedule = 0;
3106 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
3110 static inline void post_schedule(struct rq *rq)
3117 * schedule_tail - first thing a freshly forked thread must call.
3118 * @prev: the thread we just switched away from.
3120 asmlinkage void schedule_tail(struct task_struct *prev)
3121 __releases(rq->lock)
3123 struct rq *rq = this_rq();
3125 finish_task_switch(rq, prev);
3128 * FIXME: do we need to worry about rq being invalidated by the
3133 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
3134 /* In this case, finish_task_switch does not reenable preemption */
3137 if (current->set_child_tid)
3138 put_user(task_pid_vnr(current), current->set_child_tid);
3142 * context_switch - switch to the new MM and the new
3143 * thread's register state.
3146 context_switch(struct rq *rq, struct task_struct *prev,
3147 struct task_struct *next)
3149 struct mm_struct *mm, *oldmm;
3151 prepare_task_switch(rq, prev, next);
3154 oldmm = prev->active_mm;
3156 * For paravirt, this is coupled with an exit in switch_to to
3157 * combine the page table reload and the switch backend into
3160 arch_start_context_switch(prev);
3163 next->active_mm = oldmm;
3164 atomic_inc(&oldmm->mm_count);
3165 enter_lazy_tlb(oldmm, next);
3167 switch_mm(oldmm, mm, next);
3170 prev->active_mm = NULL;
3171 rq->prev_mm = oldmm;
3174 * Since the runqueue lock will be released by the next
3175 * task (which is an invalid locking op but in the case
3176 * of the scheduler it's an obvious special-case), so we
3177 * do an early lockdep release here:
3179 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
3180 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3183 /* Here we just switch the register state and the stack. */
3184 switch_to(prev, next, prev);
3188 * this_rq must be evaluated again because prev may have moved
3189 * CPUs since it called schedule(), thus the 'rq' on its stack
3190 * frame will be invalid.
3192 finish_task_switch(this_rq(), prev);
3196 * nr_running, nr_uninterruptible and nr_context_switches:
3198 * externally visible scheduler statistics: current number of runnable
3199 * threads, current number of uninterruptible-sleeping threads, total
3200 * number of context switches performed since bootup.
3202 unsigned long nr_running(void)
3204 unsigned long i, sum = 0;
3206 for_each_online_cpu(i)
3207 sum += cpu_rq(i)->nr_running;
3212 unsigned long nr_uninterruptible(void)
3214 unsigned long i, sum = 0;
3216 for_each_possible_cpu(i)
3217 sum += cpu_rq(i)->nr_uninterruptible;
3220 * Since we read the counters lockless, it might be slightly
3221 * inaccurate. Do not allow it to go below zero though:
3223 if (unlikely((long)sum < 0))
3229 unsigned long long nr_context_switches(void)
3232 unsigned long long sum = 0;
3234 for_each_possible_cpu(i)
3235 sum += cpu_rq(i)->nr_switches;
3240 unsigned long nr_iowait(void)
3242 unsigned long i, sum = 0;
3244 for_each_possible_cpu(i)
3245 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3250 unsigned long nr_iowait_cpu(int cpu)
3252 struct rq *this = cpu_rq(cpu);
3253 return atomic_read(&this->nr_iowait);
3256 unsigned long this_cpu_load(void)
3258 struct rq *this = this_rq();
3259 return this->cpu_load[0];
3263 /* Variables and functions for calc_load */
3264 static atomic_long_t calc_load_tasks;
3265 static unsigned long calc_load_update;
3266 unsigned long avenrun[3];
3267 EXPORT_SYMBOL(avenrun);
3269 static long calc_load_fold_active(struct rq *this_rq)
3271 long nr_active, delta = 0;
3273 nr_active = this_rq->nr_running;
3274 nr_active += (long) this_rq->nr_uninterruptible;
3276 if (nr_active != this_rq->calc_load_active) {
3277 delta = nr_active - this_rq->calc_load_active;
3278 this_rq->calc_load_active = nr_active;
3284 static unsigned long
3285 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3288 load += active * (FIXED_1 - exp);
3289 load += 1UL << (FSHIFT - 1);
3290 return load >> FSHIFT;
3295 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
3297 * When making the ILB scale, we should try to pull this in as well.
3299 static atomic_long_t calc_load_tasks_idle;
3301 static void calc_load_account_idle(struct rq *this_rq)
3305 delta = calc_load_fold_active(this_rq);
3307 atomic_long_add(delta, &calc_load_tasks_idle);
3310 static long calc_load_fold_idle(void)
3315 * Its got a race, we don't care...
3317 if (atomic_long_read(&calc_load_tasks_idle))
3318 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
3324 * fixed_power_int - compute: x^n, in O(log n) time
3326 * @x: base of the power
3327 * @frac_bits: fractional bits of @x
3328 * @n: power to raise @x to.
3330 * By exploiting the relation between the definition of the natural power
3331 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
3332 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
3333 * (where: n_i \elem {0, 1}, the binary vector representing n),
3334 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
3335 * of course trivially computable in O(log_2 n), the length of our binary
3338 static unsigned long
3339 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
3341 unsigned long result = 1UL << frac_bits;
3346 result += 1UL << (frac_bits - 1);
3347 result >>= frac_bits;
3353 x += 1UL << (frac_bits - 1);
3361 * a1 = a0 * e + a * (1 - e)
3363 * a2 = a1 * e + a * (1 - e)
3364 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
3365 * = a0 * e^2 + a * (1 - e) * (1 + e)
3367 * a3 = a2 * e + a * (1 - e)
3368 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
3369 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
3373 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
3374 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
3375 * = a0 * e^n + a * (1 - e^n)
3377 * [1] application of the geometric series:
3380 * S_n := \Sum x^i = -------------
3383 static unsigned long
3384 calc_load_n(unsigned long load, unsigned long exp,
3385 unsigned long active, unsigned int n)
3388 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
3392 * NO_HZ can leave us missing all per-cpu ticks calling
3393 * calc_load_account_active(), but since an idle CPU folds its delta into
3394 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
3395 * in the pending idle delta if our idle period crossed a load cycle boundary.
3397 * Once we've updated the global active value, we need to apply the exponential
3398 * weights adjusted to the number of cycles missed.
3400 static void calc_global_nohz(unsigned long ticks)
3402 long delta, active, n;
3404 if (time_before(jiffies, calc_load_update))
3408 * If we crossed a calc_load_update boundary, make sure to fold
3409 * any pending idle changes, the respective CPUs might have
3410 * missed the tick driven calc_load_account_active() update
3413 delta = calc_load_fold_idle();
3415 atomic_long_add(delta, &calc_load_tasks);
3418 * If we were idle for multiple load cycles, apply them.
3420 if (ticks >= LOAD_FREQ) {
3421 n = ticks / LOAD_FREQ;
3423 active = atomic_long_read(&calc_load_tasks);
3424 active = active > 0 ? active * FIXED_1 : 0;
3426 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
3427 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
3428 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
3430 calc_load_update += n * LOAD_FREQ;
3434 * Its possible the remainder of the above division also crosses
3435 * a LOAD_FREQ period, the regular check in calc_global_load()
3436 * which comes after this will take care of that.
3438 * Consider us being 11 ticks before a cycle completion, and us
3439 * sleeping for 4*LOAD_FREQ + 22 ticks, then the above code will
3440 * age us 4 cycles, and the test in calc_global_load() will
3441 * pick up the final one.
3445 static void calc_load_account_idle(struct rq *this_rq)
3449 static inline long calc_load_fold_idle(void)
3454 static void calc_global_nohz(unsigned long ticks)
3460 * get_avenrun - get the load average array
3461 * @loads: pointer to dest load array
3462 * @offset: offset to add
3463 * @shift: shift count to shift the result left
3465 * These values are estimates at best, so no need for locking.
3467 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3469 loads[0] = (avenrun[0] + offset) << shift;
3470 loads[1] = (avenrun[1] + offset) << shift;
3471 loads[2] = (avenrun[2] + offset) << shift;
3475 * calc_load - update the avenrun load estimates 10 ticks after the
3476 * CPUs have updated calc_load_tasks.
3478 void calc_global_load(unsigned long ticks)
3482 calc_global_nohz(ticks);
3484 if (time_before(jiffies, calc_load_update + 10))
3487 active = atomic_long_read(&calc_load_tasks);
3488 active = active > 0 ? active * FIXED_1 : 0;
3490 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3491 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3492 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3494 calc_load_update += LOAD_FREQ;
3498 * Called from update_cpu_load() to periodically update this CPU's
3501 static void calc_load_account_active(struct rq *this_rq)
3505 if (time_before(jiffies, this_rq->calc_load_update))
3508 delta = calc_load_fold_active(this_rq);
3509 delta += calc_load_fold_idle();
3511 atomic_long_add(delta, &calc_load_tasks);
3513 this_rq->calc_load_update += LOAD_FREQ;
3517 * The exact cpuload at various idx values, calculated at every tick would be
3518 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3520 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3521 * on nth tick when cpu may be busy, then we have:
3522 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3523 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3525 * decay_load_missed() below does efficient calculation of
3526 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3527 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3529 * The calculation is approximated on a 128 point scale.
3530 * degrade_zero_ticks is the number of ticks after which load at any
3531 * particular idx is approximated to be zero.
3532 * degrade_factor is a precomputed table, a row for each load idx.
3533 * Each column corresponds to degradation factor for a power of two ticks,
3534 * based on 128 point scale.
3536 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3537 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3539 * With this power of 2 load factors, we can degrade the load n times
3540 * by looking at 1 bits in n and doing as many mult/shift instead of
3541 * n mult/shifts needed by the exact degradation.
3543 #define DEGRADE_SHIFT 7
3544 static const unsigned char
3545 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3546 static const unsigned char
3547 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3548 {0, 0, 0, 0, 0, 0, 0, 0},
3549 {64, 32, 8, 0, 0, 0, 0, 0},
3550 {96, 72, 40, 12, 1, 0, 0},
3551 {112, 98, 75, 43, 15, 1, 0},
3552 {120, 112, 98, 76, 45, 16, 2} };
3555 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3556 * would be when CPU is idle and so we just decay the old load without
3557 * adding any new load.
3559 static unsigned long
3560 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3564 if (!missed_updates)
3567 if (missed_updates >= degrade_zero_ticks[idx])
3571 return load >> missed_updates;
3573 while (missed_updates) {
3574 if (missed_updates % 2)
3575 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3577 missed_updates >>= 1;
3584 * Update rq->cpu_load[] statistics. This function is usually called every
3585 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3586 * every tick. We fix it up based on jiffies.
3588 static void update_cpu_load(struct rq *this_rq)
3590 unsigned long this_load = this_rq->load.weight;
3591 unsigned long curr_jiffies = jiffies;
3592 unsigned long pending_updates;
3595 this_rq->nr_load_updates++;
3597 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3598 if (curr_jiffies == this_rq->last_load_update_tick)
3601 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3602 this_rq->last_load_update_tick = curr_jiffies;
3604 /* Update our load: */
3605 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3606 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3607 unsigned long old_load, new_load;
3609 /* scale is effectively 1 << i now, and >> i divides by scale */
3611 old_load = this_rq->cpu_load[i];
3612 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3613 new_load = this_load;
3615 * Round up the averaging division if load is increasing. This
3616 * prevents us from getting stuck on 9 if the load is 10, for
3619 if (new_load > old_load)
3620 new_load += scale - 1;
3622 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3625 sched_avg_update(this_rq);
3628 static void update_cpu_load_active(struct rq *this_rq)
3630 update_cpu_load(this_rq);
3632 calc_load_account_active(this_rq);
3638 * sched_exec - execve() is a valuable balancing opportunity, because at
3639 * this point the task has the smallest effective memory and cache footprint.
3641 void sched_exec(void)
3643 struct task_struct *p = current;
3644 unsigned long flags;
3647 raw_spin_lock_irqsave(&p->pi_lock, flags);
3648 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
3649 if (dest_cpu == smp_processor_id())
3652 if (likely(cpu_active(dest_cpu))) {
3653 struct migration_arg arg = { p, dest_cpu };
3655 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3656 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
3660 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3665 DEFINE_PER_CPU(struct kernel_stat, kstat);
3667 EXPORT_PER_CPU_SYMBOL(kstat);
3670 * Return any ns on the sched_clock that have not yet been accounted in
3671 * @p in case that task is currently running.
3673 * Called with task_rq_lock() held on @rq.
3675 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3679 if (task_current(rq, p)) {
3680 update_rq_clock(rq);
3681 ns = rq->clock_task - p->se.exec_start;
3689 unsigned long long task_delta_exec(struct task_struct *p)
3691 unsigned long flags;
3695 rq = task_rq_lock(p, &flags);
3696 ns = do_task_delta_exec(p, rq);
3697 task_rq_unlock(rq, p, &flags);
3703 * Return accounted runtime for the task.
3704 * In case the task is currently running, return the runtime plus current's
3705 * pending runtime that have not been accounted yet.
3707 unsigned long long task_sched_runtime(struct task_struct *p)
3709 unsigned long flags;
3713 rq = task_rq_lock(p, &flags);
3714 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3715 task_rq_unlock(rq, p, &flags);
3721 * Return sum_exec_runtime for the thread group.
3722 * In case the task is currently running, return the sum plus current's
3723 * pending runtime that have not been accounted yet.
3725 * Note that the thread group might have other running tasks as well,
3726 * so the return value not includes other pending runtime that other
3727 * running tasks might have.
3729 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3731 struct task_cputime totals;
3732 unsigned long flags;
3736 rq = task_rq_lock(p, &flags);
3737 thread_group_cputime(p, &totals);
3738 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3739 task_rq_unlock(rq, p, &flags);
3745 * Account user cpu time to a process.
3746 * @p: the process that the cpu time gets accounted to
3747 * @cputime: the cpu time spent in user space since the last update
3748 * @cputime_scaled: cputime scaled by cpu frequency
3750 void account_user_time(struct task_struct *p, cputime_t cputime,
3751 cputime_t cputime_scaled)
3753 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3756 /* Add user time to process. */
3757 p->utime = cputime_add(p->utime, cputime);
3758 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3759 account_group_user_time(p, cputime);
3761 /* Add user time to cpustat. */
3762 tmp = cputime_to_cputime64(cputime);
3763 if (TASK_NICE(p) > 0)
3764 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3766 cpustat->user = cputime64_add(cpustat->user, tmp);
3768 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3769 /* Account for user time used */
3770 acct_update_integrals(p);
3774 * Account guest cpu time to a process.
3775 * @p: the process that the cpu time gets accounted to
3776 * @cputime: the cpu time spent in virtual machine since the last update
3777 * @cputime_scaled: cputime scaled by cpu frequency
3779 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3780 cputime_t cputime_scaled)
3783 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3785 tmp = cputime_to_cputime64(cputime);
3787 /* Add guest time to process. */
3788 p->utime = cputime_add(p->utime, cputime);
3789 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3790 account_group_user_time(p, cputime);
3791 p->gtime = cputime_add(p->gtime, cputime);
3793 /* Add guest time to cpustat. */
3794 if (TASK_NICE(p) > 0) {
3795 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3796 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3798 cpustat->user = cputime64_add(cpustat->user, tmp);
3799 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3804 * Account system cpu time to a process and desired cpustat field
3805 * @p: the process that the cpu time gets accounted to
3806 * @cputime: the cpu time spent in kernel space since the last update
3807 * @cputime_scaled: cputime scaled by cpu frequency
3808 * @target_cputime64: pointer to cpustat field that has to be updated
3811 void __account_system_time(struct task_struct *p, cputime_t cputime,
3812 cputime_t cputime_scaled, cputime64_t *target_cputime64)
3814 cputime64_t tmp = cputime_to_cputime64(cputime);
3816 /* Add system time to process. */
3817 p->stime = cputime_add(p->stime, cputime);
3818 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3819 account_group_system_time(p, cputime);
3821 /* Add system time to cpustat. */
3822 *target_cputime64 = cputime64_add(*target_cputime64, tmp);
3823 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3825 /* Account for system time used */
3826 acct_update_integrals(p);
3830 * Account system cpu time to a process.
3831 * @p: the process that the cpu time gets accounted to
3832 * @hardirq_offset: the offset to subtract from hardirq_count()
3833 * @cputime: the cpu time spent in kernel space since the last update
3834 * @cputime_scaled: cputime scaled by cpu frequency
3836 void account_system_time(struct task_struct *p, int hardirq_offset,
3837 cputime_t cputime, cputime_t cputime_scaled)
3839 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3840 cputime64_t *target_cputime64;
3842 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3843 account_guest_time(p, cputime, cputime_scaled);
3847 if (hardirq_count() - hardirq_offset)
3848 target_cputime64 = &cpustat->irq;
3849 else if (in_serving_softirq())
3850 target_cputime64 = &cpustat->softirq;
3852 target_cputime64 = &cpustat->system;
3854 __account_system_time(p, cputime, cputime_scaled, target_cputime64);
3858 * Account for involuntary wait time.
3859 * @cputime: the cpu time spent in involuntary wait
3861 void account_steal_time(cputime_t cputime)
3863 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3864 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3866 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3870 * Account for idle time.
3871 * @cputime: the cpu time spent in idle wait
3873 void account_idle_time(cputime_t cputime)
3875 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3876 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3877 struct rq *rq = this_rq();
3879 if (atomic_read(&rq->nr_iowait) > 0)
3880 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3882 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3885 static __always_inline bool steal_account_process_tick(void)
3887 #ifdef CONFIG_PARAVIRT
3888 if (static_branch(¶virt_steal_enabled)) {
3891 steal = paravirt_steal_clock(smp_processor_id());
3892 steal -= this_rq()->prev_steal_time;
3894 st = steal_ticks(steal);
3895 this_rq()->prev_steal_time += st * TICK_NSEC;
3897 account_steal_time(st);
3904 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3906 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
3908 * Account a tick to a process and cpustat
3909 * @p: the process that the cpu time gets accounted to
3910 * @user_tick: is the tick from userspace
3911 * @rq: the pointer to rq
3913 * Tick demultiplexing follows the order
3914 * - pending hardirq update
3915 * - pending softirq update
3919 * - check for guest_time
3920 * - else account as system_time
3922 * Check for hardirq is done both for system and user time as there is
3923 * no timer going off while we are on hardirq and hence we may never get an
3924 * opportunity to update it solely in system time.
3925 * p->stime and friends are only updated on system time and not on irq
3926 * softirq as those do not count in task exec_runtime any more.
3928 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
3931 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3932 cputime64_t tmp = cputime_to_cputime64(cputime_one_jiffy);
3933 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3935 if (steal_account_process_tick())
3938 if (irqtime_account_hi_update()) {
3939 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3940 } else if (irqtime_account_si_update()) {
3941 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3942 } else if (this_cpu_ksoftirqd() == p) {
3944 * ksoftirqd time do not get accounted in cpu_softirq_time.
3945 * So, we have to handle it separately here.
3946 * Also, p->stime needs to be updated for ksoftirqd.
3948 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
3950 } else if (user_tick) {
3951 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3952 } else if (p == rq->idle) {
3953 account_idle_time(cputime_one_jiffy);
3954 } else if (p->flags & PF_VCPU) { /* System time or guest time */
3955 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled);
3957 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
3962 static void irqtime_account_idle_ticks(int ticks)
3965 struct rq *rq = this_rq();
3967 for (i = 0; i < ticks; i++)
3968 irqtime_account_process_tick(current, 0, rq);
3970 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
3971 static void irqtime_account_idle_ticks(int ticks) {}
3972 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
3974 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
3977 * Account a single tick of cpu time.
3978 * @p: the process that the cpu time gets accounted to
3979 * @user_tick: indicates if the tick is a user or a system tick
3981 void account_process_tick(struct task_struct *p, int user_tick)
3983 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3984 struct rq *rq = this_rq();
3986 if (sched_clock_irqtime) {
3987 irqtime_account_process_tick(p, user_tick, rq);
3991 if (steal_account_process_tick())
3995 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3996 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3997 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
4000 account_idle_time(cputime_one_jiffy);
4004 * Account multiple ticks of steal time.
4005 * @p: the process from which the cpu time has been stolen
4006 * @ticks: number of stolen ticks
4008 void account_steal_ticks(unsigned long ticks)
4010 account_steal_time(jiffies_to_cputime(ticks));
4014 * Account multiple ticks of idle time.
4015 * @ticks: number of stolen ticks
4017 void account_idle_ticks(unsigned long ticks)
4020 if (sched_clock_irqtime) {
4021 irqtime_account_idle_ticks(ticks);
4025 account_idle_time(jiffies_to_cputime(ticks));
4031 * Use precise platform statistics if available:
4033 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4034 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4040 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4042 struct task_cputime cputime;
4044 thread_group_cputime(p, &cputime);
4046 *ut = cputime.utime;
4047 *st = cputime.stime;
4051 #ifndef nsecs_to_cputime
4052 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
4055 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4057 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
4060 * Use CFS's precise accounting:
4062 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
4068 do_div(temp, total);
4069 utime = (cputime_t)temp;
4074 * Compare with previous values, to keep monotonicity:
4076 p->prev_utime = max(p->prev_utime, utime);
4077 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
4079 *ut = p->prev_utime;
4080 *st = p->prev_stime;
4084 * Must be called with siglock held.
4086 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4088 struct signal_struct *sig = p->signal;
4089 struct task_cputime cputime;
4090 cputime_t rtime, utime, total;
4092 thread_group_cputime(p, &cputime);
4094 total = cputime_add(cputime.utime, cputime.stime);
4095 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
4100 temp *= cputime.utime;
4101 do_div(temp, total);
4102 utime = (cputime_t)temp;
4106 sig->prev_utime = max(sig->prev_utime, utime);
4107 sig->prev_stime = max(sig->prev_stime,
4108 cputime_sub(rtime, sig->prev_utime));
4110 *ut = sig->prev_utime;
4111 *st = sig->prev_stime;
4116 * This function gets called by the timer code, with HZ frequency.
4117 * We call it with interrupts disabled.
4119 void scheduler_tick(void)
4121 int cpu = smp_processor_id();
4122 struct rq *rq = cpu_rq(cpu);
4123 struct task_struct *curr = rq->curr;
4127 raw_spin_lock(&rq->lock);
4128 update_rq_clock(rq);
4129 update_cpu_load_active(rq);
4130 curr->sched_class->task_tick(rq, curr, 0);
4131 raw_spin_unlock(&rq->lock);
4133 perf_event_task_tick();
4136 rq->idle_at_tick = idle_cpu(cpu);
4137 trigger_load_balance(rq, cpu);
4141 notrace unsigned long get_parent_ip(unsigned long addr)
4143 if (in_lock_functions(addr)) {
4144 addr = CALLER_ADDR2;
4145 if (in_lock_functions(addr))
4146 addr = CALLER_ADDR3;
4151 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4152 defined(CONFIG_PREEMPT_TRACER))
4154 void __kprobes add_preempt_count(int val)
4156 #ifdef CONFIG_DEBUG_PREEMPT
4160 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4163 preempt_count() += val;
4164 #ifdef CONFIG_DEBUG_PREEMPT
4166 * Spinlock count overflowing soon?
4168 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4171 if (preempt_count() == val)
4172 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4174 EXPORT_SYMBOL(add_preempt_count);
4176 void __kprobes sub_preempt_count(int val)
4178 #ifdef CONFIG_DEBUG_PREEMPT
4182 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4185 * Is the spinlock portion underflowing?
4187 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4188 !(preempt_count() & PREEMPT_MASK)))
4192 if (preempt_count() == val)
4193 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4194 preempt_count() -= val;
4196 EXPORT_SYMBOL(sub_preempt_count);
4201 * Print scheduling while atomic bug:
4203 static noinline void __schedule_bug(struct task_struct *prev)
4205 struct pt_regs *regs = get_irq_regs();
4207 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4208 prev->comm, prev->pid, preempt_count());
4210 debug_show_held_locks(prev);
4212 if (irqs_disabled())
4213 print_irqtrace_events(prev);
4222 * Various schedule()-time debugging checks and statistics:
4224 static inline void schedule_debug(struct task_struct *prev)
4227 * Test if we are atomic. Since do_exit() needs to call into
4228 * schedule() atomically, we ignore that path for now.
4229 * Otherwise, whine if we are scheduling when we should not be.
4231 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4232 __schedule_bug(prev);
4234 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4236 schedstat_inc(this_rq(), sched_count);
4239 static void put_prev_task(struct rq *rq, struct task_struct *prev)
4241 if (prev->on_rq || rq->skip_clock_update < 0)
4242 update_rq_clock(rq);
4243 prev->sched_class->put_prev_task(rq, prev);
4247 * Pick up the highest-prio task:
4249 static inline struct task_struct *
4250 pick_next_task(struct rq *rq)
4252 const struct sched_class *class;
4253 struct task_struct *p;
4256 * Optimization: we know that if all tasks are in
4257 * the fair class we can call that function directly:
4259 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
4260 p = fair_sched_class.pick_next_task(rq);
4265 for_each_class(class) {
4266 p = class->pick_next_task(rq);
4271 BUG(); /* the idle class will always have a runnable task */
4275 * schedule() is the main scheduler function.
4277 asmlinkage void __sched schedule(void)
4279 struct task_struct *prev, *next;
4280 unsigned long *switch_count;
4286 cpu = smp_processor_id();
4288 rcu_note_context_switch(cpu);
4291 schedule_debug(prev);
4293 if (sched_feat(HRTICK))
4296 raw_spin_lock_irq(&rq->lock);
4298 switch_count = &prev->nivcsw;
4299 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4300 if (unlikely(signal_pending_state(prev->state, prev))) {
4301 prev->state = TASK_RUNNING;
4303 deactivate_task(rq, prev, DEQUEUE_SLEEP);
4307 * If a worker went to sleep, notify and ask workqueue
4308 * whether it wants to wake up a task to maintain
4311 if (prev->flags & PF_WQ_WORKER) {
4312 struct task_struct *to_wakeup;
4314 to_wakeup = wq_worker_sleeping(prev, cpu);
4316 try_to_wake_up_local(to_wakeup);
4320 * If we are going to sleep and we have plugged IO
4321 * queued, make sure to submit it to avoid deadlocks.
4323 if (blk_needs_flush_plug(prev)) {
4324 raw_spin_unlock(&rq->lock);
4325 blk_schedule_flush_plug(prev);
4326 raw_spin_lock(&rq->lock);
4329 switch_count = &prev->nvcsw;
4332 pre_schedule(rq, prev);
4334 if (unlikely(!rq->nr_running))
4335 idle_balance(cpu, rq);
4337 put_prev_task(rq, prev);
4338 next = pick_next_task(rq);
4339 clear_tsk_need_resched(prev);
4340 rq->skip_clock_update = 0;
4342 if (likely(prev != next)) {
4347 context_switch(rq, prev, next); /* unlocks the rq */
4349 * The context switch have flipped the stack from under us
4350 * and restored the local variables which were saved when
4351 * this task called schedule() in the past. prev == current
4352 * is still correct, but it can be moved to another cpu/rq.
4354 cpu = smp_processor_id();
4357 raw_spin_unlock_irq(&rq->lock);
4361 preempt_enable_no_resched();
4365 EXPORT_SYMBOL(schedule);
4367 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
4369 static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
4371 if (lock->owner != owner)
4375 * Ensure we emit the owner->on_cpu, dereference _after_ checking
4376 * lock->owner still matches owner, if that fails, owner might
4377 * point to free()d memory, if it still matches, the rcu_read_lock()
4378 * ensures the memory stays valid.
4382 return owner->on_cpu;
4386 * Look out! "owner" is an entirely speculative pointer
4387 * access and not reliable.
4389 int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
4391 if (!sched_feat(OWNER_SPIN))
4395 while (owner_running(lock, owner)) {
4399 arch_mutex_cpu_relax();
4404 * We break out the loop above on need_resched() and when the
4405 * owner changed, which is a sign for heavy contention. Return
4406 * success only when lock->owner is NULL.
4408 return lock->owner == NULL;
4412 #ifdef CONFIG_PREEMPT
4414 * this is the entry point to schedule() from in-kernel preemption
4415 * off of preempt_enable. Kernel preemptions off return from interrupt
4416 * occur there and call schedule directly.
4418 asmlinkage void __sched notrace preempt_schedule(void)
4420 struct thread_info *ti = current_thread_info();
4423 * If there is a non-zero preempt_count or interrupts are disabled,
4424 * we do not want to preempt the current task. Just return..
4426 if (likely(ti->preempt_count || irqs_disabled()))
4430 add_preempt_count_notrace(PREEMPT_ACTIVE);
4432 sub_preempt_count_notrace(PREEMPT_ACTIVE);
4435 * Check again in case we missed a preemption opportunity
4436 * between schedule and now.
4439 } while (need_resched());
4441 EXPORT_SYMBOL(preempt_schedule);
4444 * this is the entry point to schedule() from kernel preemption
4445 * off of irq context.
4446 * Note, that this is called and return with irqs disabled. This will
4447 * protect us against recursive calling from irq.
4449 asmlinkage void __sched preempt_schedule_irq(void)
4451 struct thread_info *ti = current_thread_info();
4453 /* Catch callers which need to be fixed */
4454 BUG_ON(ti->preempt_count || !irqs_disabled());
4457 add_preempt_count(PREEMPT_ACTIVE);
4460 local_irq_disable();
4461 sub_preempt_count(PREEMPT_ACTIVE);
4464 * Check again in case we missed a preemption opportunity
4465 * between schedule and now.
4468 } while (need_resched());
4471 #endif /* CONFIG_PREEMPT */
4473 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
4476 return try_to_wake_up(curr->private, mode, wake_flags);
4478 EXPORT_SYMBOL(default_wake_function);
4481 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4482 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4483 * number) then we wake all the non-exclusive tasks and one exclusive task.
4485 * There are circumstances in which we can try to wake a task which has already
4486 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4487 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4489 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4490 int nr_exclusive, int wake_flags, void *key)
4492 wait_queue_t *curr, *next;
4494 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4495 unsigned flags = curr->flags;
4497 if (curr->func(curr, mode, wake_flags, key) &&
4498 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4504 * __wake_up - wake up threads blocked on a waitqueue.
4506 * @mode: which threads
4507 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4508 * @key: is directly passed to the wakeup function
4510 * It may be assumed that this function implies a write memory barrier before
4511 * changing the task state if and only if any tasks are woken up.
4513 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4514 int nr_exclusive, void *key)
4516 unsigned long flags;
4518 spin_lock_irqsave(&q->lock, flags);
4519 __wake_up_common(q, mode, nr_exclusive, 0, key);
4520 spin_unlock_irqrestore(&q->lock, flags);
4522 EXPORT_SYMBOL(__wake_up);
4525 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4527 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4529 __wake_up_common(q, mode, 1, 0, NULL);
4531 EXPORT_SYMBOL_GPL(__wake_up_locked);
4533 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
4535 __wake_up_common(q, mode, 1, 0, key);
4537 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
4540 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4542 * @mode: which threads
4543 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4544 * @key: opaque value to be passed to wakeup targets
4546 * The sync wakeup differs that the waker knows that it will schedule
4547 * away soon, so while the target thread will be woken up, it will not
4548 * be migrated to another CPU - ie. the two threads are 'synchronized'
4549 * with each other. This can prevent needless bouncing between CPUs.
4551 * On UP it can prevent extra preemption.
4553 * It may be assumed that this function implies a write memory barrier before
4554 * changing the task state if and only if any tasks are woken up.
4556 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4557 int nr_exclusive, void *key)
4559 unsigned long flags;
4560 int wake_flags = WF_SYNC;
4565 if (unlikely(!nr_exclusive))
4568 spin_lock_irqsave(&q->lock, flags);
4569 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4570 spin_unlock_irqrestore(&q->lock, flags);
4572 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4575 * __wake_up_sync - see __wake_up_sync_key()
4577 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4579 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4581 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4584 * complete: - signals a single thread waiting on this completion
4585 * @x: holds the state of this particular completion
4587 * This will wake up a single thread waiting on this completion. Threads will be
4588 * awakened in the same order in which they were queued.
4590 * See also complete_all(), wait_for_completion() and related routines.
4592 * It may be assumed that this function implies a write memory barrier before
4593 * changing the task state if and only if any tasks are woken up.
4595 void complete(struct completion *x)
4597 unsigned long flags;
4599 spin_lock_irqsave(&x->wait.lock, flags);
4601 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4602 spin_unlock_irqrestore(&x->wait.lock, flags);
4604 EXPORT_SYMBOL(complete);
4607 * complete_all: - signals all threads waiting on this completion
4608 * @x: holds the state of this particular completion
4610 * This will wake up all threads waiting on this particular completion event.
4612 * It may be assumed that this function implies a write memory barrier before
4613 * changing the task state if and only if any tasks are woken up.
4615 void complete_all(struct completion *x)
4617 unsigned long flags;
4619 spin_lock_irqsave(&x->wait.lock, flags);
4620 x->done += UINT_MAX/2;
4621 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4622 spin_unlock_irqrestore(&x->wait.lock, flags);
4624 EXPORT_SYMBOL(complete_all);
4626 static inline long __sched
4627 do_wait_for_common(struct completion *x, long timeout, int state)
4630 DECLARE_WAITQUEUE(wait, current);
4632 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4634 if (signal_pending_state(state, current)) {
4635 timeout = -ERESTARTSYS;
4638 __set_current_state(state);
4639 spin_unlock_irq(&x->wait.lock);
4640 timeout = schedule_timeout(timeout);
4641 spin_lock_irq(&x->wait.lock);
4642 } while (!x->done && timeout);
4643 __remove_wait_queue(&x->wait, &wait);
4648 return timeout ?: 1;
4652 wait_for_common(struct completion *x, long timeout, int state)
4656 spin_lock_irq(&x->wait.lock);
4657 timeout = do_wait_for_common(x, timeout, state);
4658 spin_unlock_irq(&x->wait.lock);
4663 * wait_for_completion: - waits for completion of a task
4664 * @x: holds the state of this particular completion
4666 * This waits to be signaled for completion of a specific task. It is NOT
4667 * interruptible and there is no timeout.
4669 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4670 * and interrupt capability. Also see complete().
4672 void __sched wait_for_completion(struct completion *x)
4674 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4676 EXPORT_SYMBOL(wait_for_completion);
4679 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4680 * @x: holds the state of this particular completion
4681 * @timeout: timeout value in jiffies
4683 * This waits for either a completion of a specific task to be signaled or for a
4684 * specified timeout to expire. The timeout is in jiffies. It is not
4687 unsigned long __sched
4688 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4690 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4692 EXPORT_SYMBOL(wait_for_completion_timeout);
4695 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4696 * @x: holds the state of this particular completion
4698 * This waits for completion of a specific task to be signaled. It is
4701 int __sched wait_for_completion_interruptible(struct completion *x)
4703 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4704 if (t == -ERESTARTSYS)
4708 EXPORT_SYMBOL(wait_for_completion_interruptible);
4711 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4712 * @x: holds the state of this particular completion
4713 * @timeout: timeout value in jiffies
4715 * This waits for either a completion of a specific task to be signaled or for a
4716 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4719 wait_for_completion_interruptible_timeout(struct completion *x,
4720 unsigned long timeout)
4722 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4724 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4727 * wait_for_completion_killable: - waits for completion of a task (killable)
4728 * @x: holds the state of this particular completion
4730 * This waits to be signaled for completion of a specific task. It can be
4731 * interrupted by a kill signal.
4733 int __sched wait_for_completion_killable(struct completion *x)
4735 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4736 if (t == -ERESTARTSYS)
4740 EXPORT_SYMBOL(wait_for_completion_killable);
4743 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4744 * @x: holds the state of this particular completion
4745 * @timeout: timeout value in jiffies
4747 * This waits for either a completion of a specific task to be
4748 * signaled or for a specified timeout to expire. It can be
4749 * interrupted by a kill signal. The timeout is in jiffies.
4752 wait_for_completion_killable_timeout(struct completion *x,
4753 unsigned long timeout)
4755 return wait_for_common(x, timeout, TASK_KILLABLE);
4757 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
4760 * try_wait_for_completion - try to decrement a completion without blocking
4761 * @x: completion structure
4763 * Returns: 0 if a decrement cannot be done without blocking
4764 * 1 if a decrement succeeded.
4766 * If a completion is being used as a counting completion,
4767 * attempt to decrement the counter without blocking. This
4768 * enables us to avoid waiting if the resource the completion
4769 * is protecting is not available.
4771 bool try_wait_for_completion(struct completion *x)
4773 unsigned long flags;
4776 spin_lock_irqsave(&x->wait.lock, flags);
4781 spin_unlock_irqrestore(&x->wait.lock, flags);
4784 EXPORT_SYMBOL(try_wait_for_completion);
4787 * completion_done - Test to see if a completion has any waiters
4788 * @x: completion structure
4790 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4791 * 1 if there are no waiters.
4794 bool completion_done(struct completion *x)
4796 unsigned long flags;
4799 spin_lock_irqsave(&x->wait.lock, flags);
4802 spin_unlock_irqrestore(&x->wait.lock, flags);
4805 EXPORT_SYMBOL(completion_done);
4808 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4810 unsigned long flags;
4813 init_waitqueue_entry(&wait, current);
4815 __set_current_state(state);
4817 spin_lock_irqsave(&q->lock, flags);
4818 __add_wait_queue(q, &wait);
4819 spin_unlock(&q->lock);
4820 timeout = schedule_timeout(timeout);
4821 spin_lock_irq(&q->lock);
4822 __remove_wait_queue(q, &wait);
4823 spin_unlock_irqrestore(&q->lock, flags);
4828 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4830 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4832 EXPORT_SYMBOL(interruptible_sleep_on);
4835 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4837 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4839 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4841 void __sched sleep_on(wait_queue_head_t *q)
4843 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4845 EXPORT_SYMBOL(sleep_on);
4847 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4849 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4851 EXPORT_SYMBOL(sleep_on_timeout);
4853 #ifdef CONFIG_RT_MUTEXES
4856 * rt_mutex_setprio - set the current priority of a task
4858 * @prio: prio value (kernel-internal form)
4860 * This function changes the 'effective' priority of a task. It does
4861 * not touch ->normal_prio like __setscheduler().
4863 * Used by the rt_mutex code to implement priority inheritance logic.
4865 void rt_mutex_setprio(struct task_struct *p, int prio)
4867 int oldprio, on_rq, running;
4869 const struct sched_class *prev_class;
4871 BUG_ON(prio < 0 || prio > MAX_PRIO);
4873 rq = __task_rq_lock(p);
4875 trace_sched_pi_setprio(p, prio);
4877 prev_class = p->sched_class;
4879 running = task_current(rq, p);
4881 dequeue_task(rq, p, 0);
4883 p->sched_class->put_prev_task(rq, p);
4886 p->sched_class = &rt_sched_class;
4888 p->sched_class = &fair_sched_class;
4893 p->sched_class->set_curr_task(rq);
4895 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4897 check_class_changed(rq, p, prev_class, oldprio);
4898 __task_rq_unlock(rq);
4903 void set_user_nice(struct task_struct *p, long nice)
4905 int old_prio, delta, on_rq;
4906 unsigned long flags;
4909 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4912 * We have to be careful, if called from sys_setpriority(),
4913 * the task might be in the middle of scheduling on another CPU.
4915 rq = task_rq_lock(p, &flags);
4917 * The RT priorities are set via sched_setscheduler(), but we still
4918 * allow the 'normal' nice value to be set - but as expected
4919 * it wont have any effect on scheduling until the task is
4920 * SCHED_FIFO/SCHED_RR:
4922 if (task_has_rt_policy(p)) {
4923 p->static_prio = NICE_TO_PRIO(nice);
4928 dequeue_task(rq, p, 0);
4930 p->static_prio = NICE_TO_PRIO(nice);
4933 p->prio = effective_prio(p);
4934 delta = p->prio - old_prio;
4937 enqueue_task(rq, p, 0);
4939 * If the task increased its priority or is running and
4940 * lowered its priority, then reschedule its CPU:
4942 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4943 resched_task(rq->curr);
4946 task_rq_unlock(rq, p, &flags);
4948 EXPORT_SYMBOL(set_user_nice);
4951 * can_nice - check if a task can reduce its nice value
4955 int can_nice(const struct task_struct *p, const int nice)
4957 /* convert nice value [19,-20] to rlimit style value [1,40] */
4958 int nice_rlim = 20 - nice;
4960 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4961 capable(CAP_SYS_NICE));
4964 #ifdef __ARCH_WANT_SYS_NICE
4967 * sys_nice - change the priority of the current process.
4968 * @increment: priority increment
4970 * sys_setpriority is a more generic, but much slower function that
4971 * does similar things.
4973 SYSCALL_DEFINE1(nice, int, increment)
4978 * Setpriority might change our priority at the same moment.
4979 * We don't have to worry. Conceptually one call occurs first
4980 * and we have a single winner.
4982 if (increment < -40)
4987 nice = TASK_NICE(current) + increment;
4993 if (increment < 0 && !can_nice(current, nice))
4996 retval = security_task_setnice(current, nice);
5000 set_user_nice(current, nice);
5007 * task_prio - return the priority value of a given task.
5008 * @p: the task in question.
5010 * This is the priority value as seen by users in /proc.
5011 * RT tasks are offset by -200. Normal tasks are centered
5012 * around 0, value goes from -16 to +15.
5014 int task_prio(const struct task_struct *p)
5016 return p->prio - MAX_RT_PRIO;
5020 * task_nice - return the nice value of a given task.
5021 * @p: the task in question.
5023 int task_nice(const struct task_struct *p)
5025 return TASK_NICE(p);
5027 EXPORT_SYMBOL(task_nice);
5030 * idle_cpu - is a given cpu idle currently?
5031 * @cpu: the processor in question.
5033 int idle_cpu(int cpu)
5035 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5039 * idle_task - return the idle task for a given cpu.
5040 * @cpu: the processor in question.
5042 struct task_struct *idle_task(int cpu)
5044 return cpu_rq(cpu)->idle;
5048 * find_process_by_pid - find a process with a matching PID value.
5049 * @pid: the pid in question.
5051 static struct task_struct *find_process_by_pid(pid_t pid)
5053 return pid ? find_task_by_vpid(pid) : current;
5056 /* Actually do priority change: must hold rq lock. */
5058 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5061 p->rt_priority = prio;
5062 p->normal_prio = normal_prio(p);
5063 /* we are holding p->pi_lock already */
5064 p->prio = rt_mutex_getprio(p);
5065 if (rt_prio(p->prio))
5066 p->sched_class = &rt_sched_class;
5068 p->sched_class = &fair_sched_class;
5073 * check the target process has a UID that matches the current process's
5075 static bool check_same_owner(struct task_struct *p)
5077 const struct cred *cred = current_cred(), *pcred;
5081 pcred = __task_cred(p);
5082 if (cred->user->user_ns == pcred->user->user_ns)
5083 match = (cred->euid == pcred->euid ||
5084 cred->euid == pcred->uid);
5091 static int __sched_setscheduler(struct task_struct *p, int policy,
5092 const struct sched_param *param, bool user)
5094 int retval, oldprio, oldpolicy = -1, on_rq, running;
5095 unsigned long flags;
5096 const struct sched_class *prev_class;
5100 /* may grab non-irq protected spin_locks */
5101 BUG_ON(in_interrupt());
5103 /* double check policy once rq lock held */
5105 reset_on_fork = p->sched_reset_on_fork;
5106 policy = oldpolicy = p->policy;
5108 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
5109 policy &= ~SCHED_RESET_ON_FORK;
5111 if (policy != SCHED_FIFO && policy != SCHED_RR &&
5112 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5113 policy != SCHED_IDLE)
5118 * Valid priorities for SCHED_FIFO and SCHED_RR are
5119 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5120 * SCHED_BATCH and SCHED_IDLE is 0.
5122 if (param->sched_priority < 0 ||
5123 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5124 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5126 if (rt_policy(policy) != (param->sched_priority != 0))
5130 * Allow unprivileged RT tasks to decrease priority:
5132 if (user && !capable(CAP_SYS_NICE)) {
5133 if (rt_policy(policy)) {
5134 unsigned long rlim_rtprio =
5135 task_rlimit(p, RLIMIT_RTPRIO);
5137 /* can't set/change the rt policy */
5138 if (policy != p->policy && !rlim_rtprio)
5141 /* can't increase priority */
5142 if (param->sched_priority > p->rt_priority &&
5143 param->sched_priority > rlim_rtprio)
5148 * Treat SCHED_IDLE as nice 20. Only allow a switch to
5149 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
5151 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
5152 if (!can_nice(p, TASK_NICE(p)))
5156 /* can't change other user's priorities */
5157 if (!check_same_owner(p))
5160 /* Normal users shall not reset the sched_reset_on_fork flag */
5161 if (p->sched_reset_on_fork && !reset_on_fork)
5166 retval = security_task_setscheduler(p);
5172 * make sure no PI-waiters arrive (or leave) while we are
5173 * changing the priority of the task:
5175 * To be able to change p->policy safely, the appropriate
5176 * runqueue lock must be held.
5178 rq = task_rq_lock(p, &flags);
5181 * Changing the policy of the stop threads its a very bad idea
5183 if (p == rq->stop) {
5184 task_rq_unlock(rq, p, &flags);
5189 * If not changing anything there's no need to proceed further:
5191 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
5192 param->sched_priority == p->rt_priority))) {
5194 __task_rq_unlock(rq);
5195 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5199 #ifdef CONFIG_RT_GROUP_SCHED
5202 * Do not allow realtime tasks into groups that have no runtime
5205 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5206 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
5207 !task_group_is_autogroup(task_group(p))) {
5208 task_rq_unlock(rq, p, &flags);
5214 /* recheck policy now with rq lock held */
5215 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5216 policy = oldpolicy = -1;
5217 task_rq_unlock(rq, p, &flags);
5221 running = task_current(rq, p);
5223 deactivate_task(rq, p, 0);
5225 p->sched_class->put_prev_task(rq, p);
5227 p->sched_reset_on_fork = reset_on_fork;
5230 prev_class = p->sched_class;
5231 __setscheduler(rq, p, policy, param->sched_priority);
5234 p->sched_class->set_curr_task(rq);
5236 activate_task(rq, p, 0);
5238 check_class_changed(rq, p, prev_class, oldprio);
5239 task_rq_unlock(rq, p, &flags);
5241 rt_mutex_adjust_pi(p);
5247 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5248 * @p: the task in question.
5249 * @policy: new policy.
5250 * @param: structure containing the new RT priority.
5252 * NOTE that the task may be already dead.
5254 int sched_setscheduler(struct task_struct *p, int policy,
5255 const struct sched_param *param)
5257 return __sched_setscheduler(p, policy, param, true);
5259 EXPORT_SYMBOL_GPL(sched_setscheduler);
5262 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5263 * @p: the task in question.
5264 * @policy: new policy.
5265 * @param: structure containing the new RT priority.
5267 * Just like sched_setscheduler, only don't bother checking if the
5268 * current context has permission. For example, this is needed in
5269 * stop_machine(): we create temporary high priority worker threads,
5270 * but our caller might not have that capability.
5272 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5273 const struct sched_param *param)
5275 return __sched_setscheduler(p, policy, param, false);
5279 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5281 struct sched_param lparam;
5282 struct task_struct *p;
5285 if (!param || pid < 0)
5287 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5292 p = find_process_by_pid(pid);
5294 retval = sched_setscheduler(p, policy, &lparam);
5301 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5302 * @pid: the pid in question.
5303 * @policy: new policy.
5304 * @param: structure containing the new RT priority.
5306 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
5307 struct sched_param __user *, param)
5309 /* negative values for policy are not valid */
5313 return do_sched_setscheduler(pid, policy, param);
5317 * sys_sched_setparam - set/change the RT priority of a thread
5318 * @pid: the pid in question.
5319 * @param: structure containing the new RT priority.
5321 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5323 return do_sched_setscheduler(pid, -1, param);
5327 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5328 * @pid: the pid in question.
5330 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5332 struct task_struct *p;
5340 p = find_process_by_pid(pid);
5342 retval = security_task_getscheduler(p);
5345 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
5352 * sys_sched_getparam - get the RT priority of a thread
5353 * @pid: the pid in question.
5354 * @param: structure containing the RT priority.
5356 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5358 struct sched_param lp;
5359 struct task_struct *p;
5362 if (!param || pid < 0)
5366 p = find_process_by_pid(pid);
5371 retval = security_task_getscheduler(p);
5375 lp.sched_priority = p->rt_priority;
5379 * This one might sleep, we cannot do it with a spinlock held ...
5381 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5390 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5392 cpumask_var_t cpus_allowed, new_mask;
5393 struct task_struct *p;
5399 p = find_process_by_pid(pid);
5406 /* Prevent p going away */
5410 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5414 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5416 goto out_free_cpus_allowed;
5419 if (!check_same_owner(p) && !task_ns_capable(p, CAP_SYS_NICE))
5422 retval = security_task_setscheduler(p);
5426 cpuset_cpus_allowed(p, cpus_allowed);
5427 cpumask_and(new_mask, in_mask, cpus_allowed);
5429 retval = set_cpus_allowed_ptr(p, new_mask);
5432 cpuset_cpus_allowed(p, cpus_allowed);
5433 if (!cpumask_subset(new_mask, cpus_allowed)) {
5435 * We must have raced with a concurrent cpuset
5436 * update. Just reset the cpus_allowed to the
5437 * cpuset's cpus_allowed
5439 cpumask_copy(new_mask, cpus_allowed);
5444 free_cpumask_var(new_mask);
5445 out_free_cpus_allowed:
5446 free_cpumask_var(cpus_allowed);
5453 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5454 struct cpumask *new_mask)
5456 if (len < cpumask_size())
5457 cpumask_clear(new_mask);
5458 else if (len > cpumask_size())
5459 len = cpumask_size();
5461 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5465 * sys_sched_setaffinity - set the cpu affinity of a process
5466 * @pid: pid of the process
5467 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5468 * @user_mask_ptr: user-space pointer to the new cpu mask
5470 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5471 unsigned long __user *, user_mask_ptr)
5473 cpumask_var_t new_mask;
5476 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5479 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5481 retval = sched_setaffinity(pid, new_mask);
5482 free_cpumask_var(new_mask);
5486 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5488 struct task_struct *p;
5489 unsigned long flags;
5496 p = find_process_by_pid(pid);
5500 retval = security_task_getscheduler(p);
5504 raw_spin_lock_irqsave(&p->pi_lock, flags);
5505 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5506 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5516 * sys_sched_getaffinity - get the cpu affinity of a process
5517 * @pid: pid of the process
5518 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5519 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5521 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5522 unsigned long __user *, user_mask_ptr)
5527 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5529 if (len & (sizeof(unsigned long)-1))
5532 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5535 ret = sched_getaffinity(pid, mask);
5537 size_t retlen = min_t(size_t, len, cpumask_size());
5539 if (copy_to_user(user_mask_ptr, mask, retlen))
5544 free_cpumask_var(mask);
5550 * sys_sched_yield - yield the current processor to other threads.
5552 * This function yields the current CPU to other tasks. If there are no
5553 * other threads running on this CPU then this function will return.
5555 SYSCALL_DEFINE0(sched_yield)
5557 struct rq *rq = this_rq_lock();
5559 schedstat_inc(rq, yld_count);
5560 current->sched_class->yield_task(rq);
5563 * Since we are going to call schedule() anyway, there's
5564 * no need to preempt or enable interrupts:
5566 __release(rq->lock);
5567 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5568 do_raw_spin_unlock(&rq->lock);
5569 preempt_enable_no_resched();
5576 static inline int should_resched(void)
5578 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5581 static void __cond_resched(void)
5583 add_preempt_count(PREEMPT_ACTIVE);
5585 sub_preempt_count(PREEMPT_ACTIVE);
5588 int __sched _cond_resched(void)
5590 if (should_resched()) {
5596 EXPORT_SYMBOL(_cond_resched);
5599 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5600 * call schedule, and on return reacquire the lock.
5602 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5603 * operations here to prevent schedule() from being called twice (once via
5604 * spin_unlock(), once by hand).
5606 int __cond_resched_lock(spinlock_t *lock)
5608 int resched = should_resched();
5611 lockdep_assert_held(lock);
5613 if (spin_needbreak(lock) || resched) {
5624 EXPORT_SYMBOL(__cond_resched_lock);
5626 int __sched __cond_resched_softirq(void)
5628 BUG_ON(!in_softirq());
5630 if (should_resched()) {
5638 EXPORT_SYMBOL(__cond_resched_softirq);
5641 * yield - yield the current processor to other threads.
5643 * This is a shortcut for kernel-space yielding - it marks the
5644 * thread runnable and calls sys_sched_yield().
5646 void __sched yield(void)
5648 set_current_state(TASK_RUNNING);
5651 EXPORT_SYMBOL(yield);
5654 * yield_to - yield the current processor to another thread in
5655 * your thread group, or accelerate that thread toward the
5656 * processor it's on.
5658 * @preempt: whether task preemption is allowed or not
5660 * It's the caller's job to ensure that the target task struct
5661 * can't go away on us before we can do any checks.
5663 * Returns true if we indeed boosted the target task.
5665 bool __sched yield_to(struct task_struct *p, bool preempt)
5667 struct task_struct *curr = current;
5668 struct rq *rq, *p_rq;
5669 unsigned long flags;
5672 local_irq_save(flags);
5677 double_rq_lock(rq, p_rq);
5678 while (task_rq(p) != p_rq) {
5679 double_rq_unlock(rq, p_rq);
5683 if (!curr->sched_class->yield_to_task)
5686 if (curr->sched_class != p->sched_class)
5689 if (task_running(p_rq, p) || p->state)
5692 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5694 schedstat_inc(rq, yld_count);
5696 * Make p's CPU reschedule; pick_next_entity takes care of
5699 if (preempt && rq != p_rq)
5700 resched_task(p_rq->curr);
5704 double_rq_unlock(rq, p_rq);
5705 local_irq_restore(flags);
5712 EXPORT_SYMBOL_GPL(yield_to);
5715 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5716 * that process accounting knows that this is a task in IO wait state.
5718 void __sched io_schedule(void)
5720 struct rq *rq = raw_rq();
5722 delayacct_blkio_start();
5723 atomic_inc(&rq->nr_iowait);
5724 blk_flush_plug(current);
5725 current->in_iowait = 1;
5727 current->in_iowait = 0;
5728 atomic_dec(&rq->nr_iowait);
5729 delayacct_blkio_end();
5731 EXPORT_SYMBOL(io_schedule);
5733 long __sched io_schedule_timeout(long timeout)
5735 struct rq *rq = raw_rq();
5738 delayacct_blkio_start();
5739 atomic_inc(&rq->nr_iowait);
5740 blk_flush_plug(current);
5741 current->in_iowait = 1;
5742 ret = schedule_timeout(timeout);
5743 current->in_iowait = 0;
5744 atomic_dec(&rq->nr_iowait);
5745 delayacct_blkio_end();
5750 * sys_sched_get_priority_max - return maximum RT priority.
5751 * @policy: scheduling class.
5753 * this syscall returns the maximum rt_priority that can be used
5754 * by a given scheduling class.
5756 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5763 ret = MAX_USER_RT_PRIO-1;
5775 * sys_sched_get_priority_min - return minimum RT priority.
5776 * @policy: scheduling class.
5778 * this syscall returns the minimum rt_priority that can be used
5779 * by a given scheduling class.
5781 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5799 * sys_sched_rr_get_interval - return the default timeslice of a process.
5800 * @pid: pid of the process.
5801 * @interval: userspace pointer to the timeslice value.
5803 * this syscall writes the default timeslice value of a given process
5804 * into the user-space timespec buffer. A value of '0' means infinity.
5806 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5807 struct timespec __user *, interval)
5809 struct task_struct *p;
5810 unsigned int time_slice;
5811 unsigned long flags;
5821 p = find_process_by_pid(pid);
5825 retval = security_task_getscheduler(p);
5829 rq = task_rq_lock(p, &flags);
5830 time_slice = p->sched_class->get_rr_interval(rq, p);
5831 task_rq_unlock(rq, p, &flags);
5834 jiffies_to_timespec(time_slice, &t);
5835 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5843 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5845 void sched_show_task(struct task_struct *p)
5847 unsigned long free = 0;
5850 state = p->state ? __ffs(p->state) + 1 : 0;
5851 printk(KERN_INFO "%-15.15s %c", p->comm,
5852 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5853 #if BITS_PER_LONG == 32
5854 if (state == TASK_RUNNING)
5855 printk(KERN_CONT " running ");
5857 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5859 if (state == TASK_RUNNING)
5860 printk(KERN_CONT " running task ");
5862 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5864 #ifdef CONFIG_DEBUG_STACK_USAGE
5865 free = stack_not_used(p);
5867 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5868 task_pid_nr(p), task_pid_nr(p->real_parent),
5869 (unsigned long)task_thread_info(p)->flags);
5871 show_stack(p, NULL);
5874 void show_state_filter(unsigned long state_filter)
5876 struct task_struct *g, *p;
5878 #if BITS_PER_LONG == 32
5880 " task PC stack pid father\n");
5883 " task PC stack pid father\n");
5885 read_lock(&tasklist_lock);
5886 do_each_thread(g, p) {
5888 * reset the NMI-timeout, listing all files on a slow
5889 * console might take a lot of time:
5891 touch_nmi_watchdog();
5892 if (!state_filter || (p->state & state_filter))
5894 } while_each_thread(g, p);
5896 touch_all_softlockup_watchdogs();
5898 #ifdef CONFIG_SCHED_DEBUG
5899 sysrq_sched_debug_show();
5901 read_unlock(&tasklist_lock);
5903 * Only show locks if all tasks are dumped:
5906 debug_show_all_locks();
5909 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5911 idle->sched_class = &idle_sched_class;
5915 * init_idle - set up an idle thread for a given CPU
5916 * @idle: task in question
5917 * @cpu: cpu the idle task belongs to
5919 * NOTE: this function does not set the idle thread's NEED_RESCHED
5920 * flag, to make booting more robust.
5922 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5924 struct rq *rq = cpu_rq(cpu);
5925 unsigned long flags;
5927 raw_spin_lock_irqsave(&rq->lock, flags);
5930 idle->state = TASK_RUNNING;
5931 idle->se.exec_start = sched_clock();
5933 do_set_cpus_allowed(idle, cpumask_of(cpu));
5935 * We're having a chicken and egg problem, even though we are
5936 * holding rq->lock, the cpu isn't yet set to this cpu so the
5937 * lockdep check in task_group() will fail.
5939 * Similar case to sched_fork(). / Alternatively we could
5940 * use task_rq_lock() here and obtain the other rq->lock.
5945 __set_task_cpu(idle, cpu);
5948 rq->curr = rq->idle = idle;
5949 #if defined(CONFIG_SMP)
5952 raw_spin_unlock_irqrestore(&rq->lock, flags);
5954 /* Set the preempt count _outside_ the spinlocks! */
5955 task_thread_info(idle)->preempt_count = 0;
5958 * The idle tasks have their own, simple scheduling class:
5960 idle->sched_class = &idle_sched_class;
5961 ftrace_graph_init_idle_task(idle, cpu);
5965 * In a system that switches off the HZ timer nohz_cpu_mask
5966 * indicates which cpus entered this state. This is used
5967 * in the rcu update to wait only for active cpus. For system
5968 * which do not switch off the HZ timer nohz_cpu_mask should
5969 * always be CPU_BITS_NONE.
5971 cpumask_var_t nohz_cpu_mask;
5974 * Increase the granularity value when there are more CPUs,
5975 * because with more CPUs the 'effective latency' as visible
5976 * to users decreases. But the relationship is not linear,
5977 * so pick a second-best guess by going with the log2 of the
5980 * This idea comes from the SD scheduler of Con Kolivas:
5982 static int get_update_sysctl_factor(void)
5984 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5985 unsigned int factor;
5987 switch (sysctl_sched_tunable_scaling) {
5988 case SCHED_TUNABLESCALING_NONE:
5991 case SCHED_TUNABLESCALING_LINEAR:
5994 case SCHED_TUNABLESCALING_LOG:
5996 factor = 1 + ilog2(cpus);
6003 static void update_sysctl(void)
6005 unsigned int factor = get_update_sysctl_factor();
6007 #define SET_SYSCTL(name) \
6008 (sysctl_##name = (factor) * normalized_sysctl_##name)
6009 SET_SYSCTL(sched_min_granularity);
6010 SET_SYSCTL(sched_latency);
6011 SET_SYSCTL(sched_wakeup_granularity);
6015 static inline void sched_init_granularity(void)
6021 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
6023 if (p->sched_class && p->sched_class->set_cpus_allowed)
6024 p->sched_class->set_cpus_allowed(p, new_mask);
6026 cpumask_copy(&p->cpus_allowed, new_mask);
6027 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
6032 * This is how migration works:
6034 * 1) we invoke migration_cpu_stop() on the target CPU using
6036 * 2) stopper starts to run (implicitly forcing the migrated thread
6038 * 3) it checks whether the migrated task is still in the wrong runqueue.
6039 * 4) if it's in the wrong runqueue then the migration thread removes
6040 * it and puts it into the right queue.
6041 * 5) stopper completes and stop_one_cpu() returns and the migration
6046 * Change a given task's CPU affinity. Migrate the thread to a
6047 * proper CPU and schedule it away if the CPU it's executing on
6048 * is removed from the allowed bitmask.
6050 * NOTE: the caller must have a valid reference to the task, the
6051 * task must not exit() & deallocate itself prematurely. The
6052 * call is not atomic; no spinlocks may be held.
6054 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
6056 unsigned long flags;
6058 unsigned int dest_cpu;
6061 rq = task_rq_lock(p, &flags);
6063 if (cpumask_equal(&p->cpus_allowed, new_mask))
6066 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
6071 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
6076 do_set_cpus_allowed(p, new_mask);
6078 /* Can the task run on the task's current CPU? If so, we're done */
6079 if (cpumask_test_cpu(task_cpu(p), new_mask))
6082 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
6084 struct migration_arg arg = { p, dest_cpu };
6085 /* Need help from migration thread: drop lock and wait. */
6086 task_rq_unlock(rq, p, &flags);
6087 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
6088 tlb_migrate_finish(p->mm);
6092 task_rq_unlock(rq, p, &flags);
6096 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
6099 * Move (not current) task off this cpu, onto dest cpu. We're doing
6100 * this because either it can't run here any more (set_cpus_allowed()
6101 * away from this CPU, or CPU going down), or because we're
6102 * attempting to rebalance this task on exec (sched_exec).
6104 * So we race with normal scheduler movements, but that's OK, as long
6105 * as the task is no longer on this CPU.
6107 * Returns non-zero if task was successfully migrated.
6109 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6111 struct rq *rq_dest, *rq_src;
6114 if (unlikely(!cpu_active(dest_cpu)))
6117 rq_src = cpu_rq(src_cpu);
6118 rq_dest = cpu_rq(dest_cpu);
6120 raw_spin_lock(&p->pi_lock);
6121 double_rq_lock(rq_src, rq_dest);
6122 /* Already moved. */
6123 if (task_cpu(p) != src_cpu)
6125 /* Affinity changed (again). */
6126 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6130 * If we're not on a rq, the next wake-up will ensure we're
6134 deactivate_task(rq_src, p, 0);
6135 set_task_cpu(p, dest_cpu);
6136 activate_task(rq_dest, p, 0);
6137 check_preempt_curr(rq_dest, p, 0);
6142 double_rq_unlock(rq_src, rq_dest);
6143 raw_spin_unlock(&p->pi_lock);
6148 * migration_cpu_stop - this will be executed by a highprio stopper thread
6149 * and performs thread migration by bumping thread off CPU then
6150 * 'pushing' onto another runqueue.
6152 static int migration_cpu_stop(void *data)
6154 struct migration_arg *arg = data;
6157 * The original target cpu might have gone down and we might
6158 * be on another cpu but it doesn't matter.
6160 local_irq_disable();
6161 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
6166 #ifdef CONFIG_HOTPLUG_CPU
6169 * Ensures that the idle task is using init_mm right before its cpu goes
6172 void idle_task_exit(void)
6174 struct mm_struct *mm = current->active_mm;
6176 BUG_ON(cpu_online(smp_processor_id()));
6179 switch_mm(mm, &init_mm, current);
6184 * While a dead CPU has no uninterruptible tasks queued at this point,
6185 * it might still have a nonzero ->nr_uninterruptible counter, because
6186 * for performance reasons the counter is not stricly tracking tasks to
6187 * their home CPUs. So we just add the counter to another CPU's counter,
6188 * to keep the global sum constant after CPU-down:
6190 static void migrate_nr_uninterruptible(struct rq *rq_src)
6192 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
6194 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6195 rq_src->nr_uninterruptible = 0;
6199 * remove the tasks which were accounted by rq from calc_load_tasks.
6201 static void calc_global_load_remove(struct rq *rq)
6203 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
6204 rq->calc_load_active = 0;
6208 * Migrate all tasks from the rq, sleeping tasks will be migrated by
6209 * try_to_wake_up()->select_task_rq().
6211 * Called with rq->lock held even though we'er in stop_machine() and
6212 * there's no concurrency possible, we hold the required locks anyway
6213 * because of lock validation efforts.
6215 static void migrate_tasks(unsigned int dead_cpu)
6217 struct rq *rq = cpu_rq(dead_cpu);
6218 struct task_struct *next, *stop = rq->stop;
6222 * Fudge the rq selection such that the below task selection loop
6223 * doesn't get stuck on the currently eligible stop task.
6225 * We're currently inside stop_machine() and the rq is either stuck
6226 * in the stop_machine_cpu_stop() loop, or we're executing this code,
6227 * either way we should never end up calling schedule() until we're
6234 * There's this thread running, bail when that's the only
6237 if (rq->nr_running == 1)
6240 next = pick_next_task(rq);
6242 next->sched_class->put_prev_task(rq, next);
6244 /* Find suitable destination for @next, with force if needed. */
6245 dest_cpu = select_fallback_rq(dead_cpu, next);
6246 raw_spin_unlock(&rq->lock);
6248 __migrate_task(next, dead_cpu, dest_cpu);
6250 raw_spin_lock(&rq->lock);
6256 #endif /* CONFIG_HOTPLUG_CPU */
6258 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6260 static struct ctl_table sd_ctl_dir[] = {
6262 .procname = "sched_domain",
6268 static struct ctl_table sd_ctl_root[] = {
6270 .procname = "kernel",
6272 .child = sd_ctl_dir,
6277 static struct ctl_table *sd_alloc_ctl_entry(int n)
6279 struct ctl_table *entry =
6280 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6285 static void sd_free_ctl_entry(struct ctl_table **tablep)
6287 struct ctl_table *entry;
6290 * In the intermediate directories, both the child directory and
6291 * procname are dynamically allocated and could fail but the mode
6292 * will always be set. In the lowest directory the names are
6293 * static strings and all have proc handlers.
6295 for (entry = *tablep; entry->mode; entry++) {
6297 sd_free_ctl_entry(&entry->child);
6298 if (entry->proc_handler == NULL)
6299 kfree(entry->procname);
6307 set_table_entry(struct ctl_table *entry,
6308 const char *procname, void *data, int maxlen,
6309 mode_t mode, proc_handler *proc_handler)
6311 entry->procname = procname;
6313 entry->maxlen = maxlen;
6315 entry->proc_handler = proc_handler;
6318 static struct ctl_table *
6319 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6321 struct ctl_table *table = sd_alloc_ctl_entry(13);
6326 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6327 sizeof(long), 0644, proc_doulongvec_minmax);
6328 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6329 sizeof(long), 0644, proc_doulongvec_minmax);
6330 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6331 sizeof(int), 0644, proc_dointvec_minmax);
6332 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6333 sizeof(int), 0644, proc_dointvec_minmax);
6334 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6335 sizeof(int), 0644, proc_dointvec_minmax);
6336 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6337 sizeof(int), 0644, proc_dointvec_minmax);
6338 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6339 sizeof(int), 0644, proc_dointvec_minmax);
6340 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6341 sizeof(int), 0644, proc_dointvec_minmax);
6342 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6343 sizeof(int), 0644, proc_dointvec_minmax);
6344 set_table_entry(&table[9], "cache_nice_tries",
6345 &sd->cache_nice_tries,
6346 sizeof(int), 0644, proc_dointvec_minmax);
6347 set_table_entry(&table[10], "flags", &sd->flags,
6348 sizeof(int), 0644, proc_dointvec_minmax);
6349 set_table_entry(&table[11], "name", sd->name,
6350 CORENAME_MAX_SIZE, 0444, proc_dostring);
6351 /* &table[12] is terminator */
6356 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6358 struct ctl_table *entry, *table;
6359 struct sched_domain *sd;
6360 int domain_num = 0, i;
6363 for_each_domain(cpu, sd)
6365 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6370 for_each_domain(cpu, sd) {
6371 snprintf(buf, 32, "domain%d", i);
6372 entry->procname = kstrdup(buf, GFP_KERNEL);
6374 entry->child = sd_alloc_ctl_domain_table(sd);
6381 static struct ctl_table_header *sd_sysctl_header;
6382 static void register_sched_domain_sysctl(void)
6384 int i, cpu_num = num_possible_cpus();
6385 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6388 WARN_ON(sd_ctl_dir[0].child);
6389 sd_ctl_dir[0].child = entry;
6394 for_each_possible_cpu(i) {
6395 snprintf(buf, 32, "cpu%d", i);
6396 entry->procname = kstrdup(buf, GFP_KERNEL);
6398 entry->child = sd_alloc_ctl_cpu_table(i);
6402 WARN_ON(sd_sysctl_header);
6403 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6406 /* may be called multiple times per register */
6407 static void unregister_sched_domain_sysctl(void)
6409 if (sd_sysctl_header)
6410 unregister_sysctl_table(sd_sysctl_header);
6411 sd_sysctl_header = NULL;
6412 if (sd_ctl_dir[0].child)
6413 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6416 static void register_sched_domain_sysctl(void)
6419 static void unregister_sched_domain_sysctl(void)
6424 static void set_rq_online(struct rq *rq)
6427 const struct sched_class *class;
6429 cpumask_set_cpu(rq->cpu, rq->rd->online);
6432 for_each_class(class) {
6433 if (class->rq_online)
6434 class->rq_online(rq);
6439 static void set_rq_offline(struct rq *rq)
6442 const struct sched_class *class;
6444 for_each_class(class) {
6445 if (class->rq_offline)
6446 class->rq_offline(rq);
6449 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6455 * migration_call - callback that gets triggered when a CPU is added.
6456 * Here we can start up the necessary migration thread for the new CPU.
6458 static int __cpuinit
6459 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6461 int cpu = (long)hcpu;
6462 unsigned long flags;
6463 struct rq *rq = cpu_rq(cpu);
6465 switch (action & ~CPU_TASKS_FROZEN) {
6467 case CPU_UP_PREPARE:
6468 rq->calc_load_update = calc_load_update;
6472 /* Update our root-domain */
6473 raw_spin_lock_irqsave(&rq->lock, flags);
6475 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6479 raw_spin_unlock_irqrestore(&rq->lock, flags);
6482 #ifdef CONFIG_HOTPLUG_CPU
6484 sched_ttwu_pending();
6485 /* Update our root-domain */
6486 raw_spin_lock_irqsave(&rq->lock, flags);
6488 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6492 BUG_ON(rq->nr_running != 1); /* the migration thread */
6493 raw_spin_unlock_irqrestore(&rq->lock, flags);
6495 migrate_nr_uninterruptible(rq);
6496 calc_global_load_remove(rq);
6501 update_max_interval();
6507 * Register at high priority so that task migration (migrate_all_tasks)
6508 * happens before everything else. This has to be lower priority than
6509 * the notifier in the perf_event subsystem, though.
6511 static struct notifier_block __cpuinitdata migration_notifier = {
6512 .notifier_call = migration_call,
6513 .priority = CPU_PRI_MIGRATION,
6516 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
6517 unsigned long action, void *hcpu)
6519 switch (action & ~CPU_TASKS_FROZEN) {
6521 case CPU_DOWN_FAILED:
6522 set_cpu_active((long)hcpu, true);
6529 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
6530 unsigned long action, void *hcpu)
6532 switch (action & ~CPU_TASKS_FROZEN) {
6533 case CPU_DOWN_PREPARE:
6534 set_cpu_active((long)hcpu, false);
6541 static int __init migration_init(void)
6543 void *cpu = (void *)(long)smp_processor_id();
6546 /* Initialize migration for the boot CPU */
6547 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6548 BUG_ON(err == NOTIFY_BAD);
6549 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6550 register_cpu_notifier(&migration_notifier);
6552 /* Register cpu active notifiers */
6553 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
6554 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
6558 early_initcall(migration_init);
6563 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
6565 #ifdef CONFIG_SCHED_DEBUG
6567 static __read_mostly int sched_domain_debug_enabled;
6569 static int __init sched_domain_debug_setup(char *str)
6571 sched_domain_debug_enabled = 1;
6575 early_param("sched_debug", sched_domain_debug_setup);
6577 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6578 struct cpumask *groupmask)
6580 struct sched_group *group = sd->groups;
6583 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6584 cpumask_clear(groupmask);
6586 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6588 if (!(sd->flags & SD_LOAD_BALANCE)) {
6589 printk("does not load-balance\n");
6591 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6596 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6598 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6599 printk(KERN_ERR "ERROR: domain->span does not contain "
6602 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6603 printk(KERN_ERR "ERROR: domain->groups does not contain"
6607 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6611 printk(KERN_ERR "ERROR: group is NULL\n");
6615 if (!group->sgp->power) {
6616 printk(KERN_CONT "\n");
6617 printk(KERN_ERR "ERROR: domain->cpu_power not "
6622 if (!cpumask_weight(sched_group_cpus(group))) {
6623 printk(KERN_CONT "\n");
6624 printk(KERN_ERR "ERROR: empty group\n");
6628 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6629 printk(KERN_CONT "\n");
6630 printk(KERN_ERR "ERROR: repeated CPUs\n");
6634 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6636 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6638 printk(KERN_CONT " %s", str);
6639 if (group->sgp->power != SCHED_POWER_SCALE) {
6640 printk(KERN_CONT " (cpu_power = %d)",
6644 group = group->next;
6645 } while (group != sd->groups);
6646 printk(KERN_CONT "\n");
6648 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6649 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6652 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6653 printk(KERN_ERR "ERROR: parent span is not a superset "
6654 "of domain->span\n");
6658 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6662 if (!sched_domain_debug_enabled)
6666 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6670 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6673 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
6681 #else /* !CONFIG_SCHED_DEBUG */
6682 # define sched_domain_debug(sd, cpu) do { } while (0)
6683 #endif /* CONFIG_SCHED_DEBUG */
6685 static int sd_degenerate(struct sched_domain *sd)
6687 if (cpumask_weight(sched_domain_span(sd)) == 1)
6690 /* Following flags need at least 2 groups */
6691 if (sd->flags & (SD_LOAD_BALANCE |
6692 SD_BALANCE_NEWIDLE |
6696 SD_SHARE_PKG_RESOURCES)) {
6697 if (sd->groups != sd->groups->next)
6701 /* Following flags don't use groups */
6702 if (sd->flags & (SD_WAKE_AFFINE))
6709 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6711 unsigned long cflags = sd->flags, pflags = parent->flags;
6713 if (sd_degenerate(parent))
6716 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6719 /* Flags needing groups don't count if only 1 group in parent */
6720 if (parent->groups == parent->groups->next) {
6721 pflags &= ~(SD_LOAD_BALANCE |
6722 SD_BALANCE_NEWIDLE |
6726 SD_SHARE_PKG_RESOURCES);
6727 if (nr_node_ids == 1)
6728 pflags &= ~SD_SERIALIZE;
6730 if (~cflags & pflags)
6736 static void free_rootdomain(struct rcu_head *rcu)
6738 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
6740 cpupri_cleanup(&rd->cpupri);
6741 free_cpumask_var(rd->rto_mask);
6742 free_cpumask_var(rd->online);
6743 free_cpumask_var(rd->span);
6747 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6749 struct root_domain *old_rd = NULL;
6750 unsigned long flags;
6752 raw_spin_lock_irqsave(&rq->lock, flags);
6757 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6760 cpumask_clear_cpu(rq->cpu, old_rd->span);
6763 * If we dont want to free the old_rt yet then
6764 * set old_rd to NULL to skip the freeing later
6767 if (!atomic_dec_and_test(&old_rd->refcount))
6771 atomic_inc(&rd->refcount);
6774 cpumask_set_cpu(rq->cpu, rd->span);
6775 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6778 raw_spin_unlock_irqrestore(&rq->lock, flags);
6781 call_rcu_sched(&old_rd->rcu, free_rootdomain);
6784 static int init_rootdomain(struct root_domain *rd)
6786 memset(rd, 0, sizeof(*rd));
6788 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
6790 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
6792 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6795 if (cpupri_init(&rd->cpupri) != 0)
6800 free_cpumask_var(rd->rto_mask);
6802 free_cpumask_var(rd->online);
6804 free_cpumask_var(rd->span);
6809 static void init_defrootdomain(void)
6811 init_rootdomain(&def_root_domain);
6813 atomic_set(&def_root_domain.refcount, 1);
6816 static struct root_domain *alloc_rootdomain(void)
6818 struct root_domain *rd;
6820 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6824 if (init_rootdomain(rd) != 0) {
6832 static void free_sched_groups(struct sched_group *sg, int free_sgp)
6834 struct sched_group *tmp, *first;
6843 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
6848 } while (sg != first);
6851 static void free_sched_domain(struct rcu_head *rcu)
6853 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
6856 * If its an overlapping domain it has private groups, iterate and
6859 if (sd->flags & SD_OVERLAP) {
6860 free_sched_groups(sd->groups, 1);
6861 } else if (atomic_dec_and_test(&sd->groups->ref)) {
6862 kfree(sd->groups->sgp);
6868 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
6870 call_rcu(&sd->rcu, free_sched_domain);
6873 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
6875 for (; sd; sd = sd->parent)
6876 destroy_sched_domain(sd, cpu);
6880 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6881 * hold the hotplug lock.
6884 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6886 struct rq *rq = cpu_rq(cpu);
6887 struct sched_domain *tmp;
6889 /* Remove the sched domains which do not contribute to scheduling. */
6890 for (tmp = sd; tmp; ) {
6891 struct sched_domain *parent = tmp->parent;
6895 if (sd_parent_degenerate(tmp, parent)) {
6896 tmp->parent = parent->parent;
6898 parent->parent->child = tmp;
6899 destroy_sched_domain(parent, cpu);
6904 if (sd && sd_degenerate(sd)) {
6907 destroy_sched_domain(tmp, cpu);
6912 sched_domain_debug(sd, cpu);
6914 rq_attach_root(rq, rd);
6916 rcu_assign_pointer(rq->sd, sd);
6917 destroy_sched_domains(tmp, cpu);
6920 /* cpus with isolated domains */
6921 static cpumask_var_t cpu_isolated_map;
6923 /* Setup the mask of cpus configured for isolated domains */
6924 static int __init isolated_cpu_setup(char *str)
6926 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6927 cpulist_parse(str, cpu_isolated_map);
6931 __setup("isolcpus=", isolated_cpu_setup);
6933 #define SD_NODES_PER_DOMAIN 16
6938 * find_next_best_node - find the next node to include in a sched_domain
6939 * @node: node whose sched_domain we're building
6940 * @used_nodes: nodes already in the sched_domain
6942 * Find the next node to include in a given scheduling domain. Simply
6943 * finds the closest node not already in the @used_nodes map.
6945 * Should use nodemask_t.
6947 static int find_next_best_node(int node, nodemask_t *used_nodes)
6949 int i, n, val, min_val, best_node = -1;
6953 for (i = 0; i < nr_node_ids; i++) {
6954 /* Start at @node */
6955 n = (node + i) % nr_node_ids;
6957 if (!nr_cpus_node(n))
6960 /* Skip already used nodes */
6961 if (node_isset(n, *used_nodes))
6964 /* Simple min distance search */
6965 val = node_distance(node, n);
6967 if (val < min_val) {
6973 if (best_node != -1)
6974 node_set(best_node, *used_nodes);
6979 * sched_domain_node_span - get a cpumask for a node's sched_domain
6980 * @node: node whose cpumask we're constructing
6981 * @span: resulting cpumask
6983 * Given a node, construct a good cpumask for its sched_domain to span. It
6984 * should be one that prevents unnecessary balancing, but also spreads tasks
6987 static void sched_domain_node_span(int node, struct cpumask *span)
6989 nodemask_t used_nodes;
6992 cpumask_clear(span);
6993 nodes_clear(used_nodes);
6995 cpumask_or(span, span, cpumask_of_node(node));
6996 node_set(node, used_nodes);
6998 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6999 int next_node = find_next_best_node(node, &used_nodes);
7002 cpumask_or(span, span, cpumask_of_node(next_node));
7006 static const struct cpumask *cpu_node_mask(int cpu)
7008 lockdep_assert_held(&sched_domains_mutex);
7010 sched_domain_node_span(cpu_to_node(cpu), sched_domains_tmpmask);
7012 return sched_domains_tmpmask;
7015 static const struct cpumask *cpu_allnodes_mask(int cpu)
7017 return cpu_possible_mask;
7019 #endif /* CONFIG_NUMA */
7021 static const struct cpumask *cpu_cpu_mask(int cpu)
7023 return cpumask_of_node(cpu_to_node(cpu));
7026 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
7029 struct sched_domain **__percpu sd;
7030 struct sched_group **__percpu sg;
7031 struct sched_group_power **__percpu sgp;
7035 struct sched_domain ** __percpu sd;
7036 struct root_domain *rd;
7046 struct sched_domain_topology_level;
7048 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
7049 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
7051 #define SDTL_OVERLAP 0x01
7053 struct sched_domain_topology_level {
7054 sched_domain_init_f init;
7055 sched_domain_mask_f mask;
7057 struct sd_data data;
7061 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
7063 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
7064 const struct cpumask *span = sched_domain_span(sd);
7065 struct cpumask *covered = sched_domains_tmpmask;
7066 struct sd_data *sdd = sd->private;
7067 struct sched_domain *child;
7070 cpumask_clear(covered);
7072 for_each_cpu(i, span) {
7073 struct cpumask *sg_span;
7075 if (cpumask_test_cpu(i, covered))
7078 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7079 GFP_KERNEL, cpu_to_node(i));
7084 sg_span = sched_group_cpus(sg);
7086 child = *per_cpu_ptr(sdd->sd, i);
7088 child = child->child;
7089 cpumask_copy(sg_span, sched_domain_span(child));
7091 cpumask_set_cpu(i, sg_span);
7093 cpumask_or(covered, covered, sg_span);
7095 sg->sgp = *per_cpu_ptr(sdd->sgp, cpumask_first(sg_span));
7096 atomic_inc(&sg->sgp->ref);
7098 if (cpumask_test_cpu(cpu, sg_span))
7108 sd->groups = groups;
7113 free_sched_groups(first, 0);
7118 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
7120 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
7121 struct sched_domain *child = sd->child;
7124 cpu = cpumask_first(sched_domain_span(child));
7127 *sg = *per_cpu_ptr(sdd->sg, cpu);
7128 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
7129 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
7136 * build_sched_groups will build a circular linked list of the groups
7137 * covered by the given span, and will set each group's ->cpumask correctly,
7138 * and ->cpu_power to 0.
7140 * Assumes the sched_domain tree is fully constructed
7143 build_sched_groups(struct sched_domain *sd, int cpu)
7145 struct sched_group *first = NULL, *last = NULL;
7146 struct sd_data *sdd = sd->private;
7147 const struct cpumask *span = sched_domain_span(sd);
7148 struct cpumask *covered;
7151 get_group(cpu, sdd, &sd->groups);
7152 atomic_inc(&sd->groups->ref);
7154 if (cpu != cpumask_first(sched_domain_span(sd)))
7157 lockdep_assert_held(&sched_domains_mutex);
7158 covered = sched_domains_tmpmask;
7160 cpumask_clear(covered);
7162 for_each_cpu(i, span) {
7163 struct sched_group *sg;
7164 int group = get_group(i, sdd, &sg);
7167 if (cpumask_test_cpu(i, covered))
7170 cpumask_clear(sched_group_cpus(sg));
7173 for_each_cpu(j, span) {
7174 if (get_group(j, sdd, NULL) != group)
7177 cpumask_set_cpu(j, covered);
7178 cpumask_set_cpu(j, sched_group_cpus(sg));
7193 * Initialize sched groups cpu_power.
7195 * cpu_power indicates the capacity of sched group, which is used while
7196 * distributing the load between different sched groups in a sched domain.
7197 * Typically cpu_power for all the groups in a sched domain will be same unless
7198 * there are asymmetries in the topology. If there are asymmetries, group
7199 * having more cpu_power will pickup more load compared to the group having
7202 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7204 struct sched_group *sg = sd->groups;
7206 WARN_ON(!sd || !sg);
7209 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
7211 } while (sg != sd->groups);
7213 if (cpu != group_first_cpu(sg))
7216 update_group_power(sd, cpu);
7220 * Initializers for schedule domains
7221 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7224 #ifdef CONFIG_SCHED_DEBUG
7225 # define SD_INIT_NAME(sd, type) sd->name = #type
7227 # define SD_INIT_NAME(sd, type) do { } while (0)
7230 #define SD_INIT_FUNC(type) \
7231 static noinline struct sched_domain * \
7232 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
7234 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
7235 *sd = SD_##type##_INIT; \
7236 SD_INIT_NAME(sd, type); \
7237 sd->private = &tl->data; \
7243 SD_INIT_FUNC(ALLNODES)
7246 #ifdef CONFIG_SCHED_SMT
7247 SD_INIT_FUNC(SIBLING)
7249 #ifdef CONFIG_SCHED_MC
7252 #ifdef CONFIG_SCHED_BOOK
7256 static int default_relax_domain_level = -1;
7257 int sched_domain_level_max;
7259 static int __init setup_relax_domain_level(char *str)
7263 val = simple_strtoul(str, NULL, 0);
7264 if (val < sched_domain_level_max)
7265 default_relax_domain_level = val;
7269 __setup("relax_domain_level=", setup_relax_domain_level);
7271 static void set_domain_attribute(struct sched_domain *sd,
7272 struct sched_domain_attr *attr)
7276 if (!attr || attr->relax_domain_level < 0) {
7277 if (default_relax_domain_level < 0)
7280 request = default_relax_domain_level;
7282 request = attr->relax_domain_level;
7283 if (request < sd->level) {
7284 /* turn off idle balance on this domain */
7285 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7287 /* turn on idle balance on this domain */
7288 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7292 static void __sdt_free(const struct cpumask *cpu_map);
7293 static int __sdt_alloc(const struct cpumask *cpu_map);
7295 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
7296 const struct cpumask *cpu_map)
7300 if (!atomic_read(&d->rd->refcount))
7301 free_rootdomain(&d->rd->rcu); /* fall through */
7303 free_percpu(d->sd); /* fall through */
7305 __sdt_free(cpu_map); /* fall through */
7311 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
7312 const struct cpumask *cpu_map)
7314 memset(d, 0, sizeof(*d));
7316 if (__sdt_alloc(cpu_map))
7317 return sa_sd_storage;
7318 d->sd = alloc_percpu(struct sched_domain *);
7320 return sa_sd_storage;
7321 d->rd = alloc_rootdomain();
7324 return sa_rootdomain;
7328 * NULL the sd_data elements we've used to build the sched_domain and
7329 * sched_group structure so that the subsequent __free_domain_allocs()
7330 * will not free the data we're using.
7332 static void claim_allocations(int cpu, struct sched_domain *sd)
7334 struct sd_data *sdd = sd->private;
7336 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
7337 *per_cpu_ptr(sdd->sd, cpu) = NULL;
7339 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
7340 *per_cpu_ptr(sdd->sg, cpu) = NULL;
7342 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
7343 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
7346 #ifdef CONFIG_SCHED_SMT
7347 static const struct cpumask *cpu_smt_mask(int cpu)
7349 return topology_thread_cpumask(cpu);
7354 * Topology list, bottom-up.
7356 static struct sched_domain_topology_level default_topology[] = {
7357 #ifdef CONFIG_SCHED_SMT
7358 { sd_init_SIBLING, cpu_smt_mask, },
7360 #ifdef CONFIG_SCHED_MC
7361 { sd_init_MC, cpu_coregroup_mask, },
7363 #ifdef CONFIG_SCHED_BOOK
7364 { sd_init_BOOK, cpu_book_mask, },
7366 { sd_init_CPU, cpu_cpu_mask, },
7368 { sd_init_NODE, cpu_node_mask, SDTL_OVERLAP, },
7369 { sd_init_ALLNODES, cpu_allnodes_mask, },
7374 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
7376 static int __sdt_alloc(const struct cpumask *cpu_map)
7378 struct sched_domain_topology_level *tl;
7381 for (tl = sched_domain_topology; tl->init; tl++) {
7382 struct sd_data *sdd = &tl->data;
7384 sdd->sd = alloc_percpu(struct sched_domain *);
7388 sdd->sg = alloc_percpu(struct sched_group *);
7392 sdd->sgp = alloc_percpu(struct sched_group_power *);
7396 for_each_cpu(j, cpu_map) {
7397 struct sched_domain *sd;
7398 struct sched_group *sg;
7399 struct sched_group_power *sgp;
7401 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
7402 GFP_KERNEL, cpu_to_node(j));
7406 *per_cpu_ptr(sdd->sd, j) = sd;
7408 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7409 GFP_KERNEL, cpu_to_node(j));
7413 *per_cpu_ptr(sdd->sg, j) = sg;
7415 sgp = kzalloc_node(sizeof(struct sched_group_power),
7416 GFP_KERNEL, cpu_to_node(j));
7420 *per_cpu_ptr(sdd->sgp, j) = sgp;
7427 static void __sdt_free(const struct cpumask *cpu_map)
7429 struct sched_domain_topology_level *tl;
7432 for (tl = sched_domain_topology; tl->init; tl++) {
7433 struct sd_data *sdd = &tl->data;
7435 for_each_cpu(j, cpu_map) {
7436 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, j);
7437 if (sd && (sd->flags & SD_OVERLAP))
7438 free_sched_groups(sd->groups, 0);
7439 kfree(*per_cpu_ptr(sdd->sg, j));
7440 kfree(*per_cpu_ptr(sdd->sgp, j));
7442 free_percpu(sdd->sd);
7443 free_percpu(sdd->sg);
7444 free_percpu(sdd->sgp);
7448 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
7449 struct s_data *d, const struct cpumask *cpu_map,
7450 struct sched_domain_attr *attr, struct sched_domain *child,
7453 struct sched_domain *sd = tl->init(tl, cpu);
7457 set_domain_attribute(sd, attr);
7458 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
7460 sd->level = child->level + 1;
7461 sched_domain_level_max = max(sched_domain_level_max, sd->level);
7470 * Build sched domains for a given set of cpus and attach the sched domains
7471 * to the individual cpus
7473 static int build_sched_domains(const struct cpumask *cpu_map,
7474 struct sched_domain_attr *attr)
7476 enum s_alloc alloc_state = sa_none;
7477 struct sched_domain *sd;
7479 int i, ret = -ENOMEM;
7481 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7482 if (alloc_state != sa_rootdomain)
7485 /* Set up domains for cpus specified by the cpu_map. */
7486 for_each_cpu(i, cpu_map) {
7487 struct sched_domain_topology_level *tl;
7490 for (tl = sched_domain_topology; tl->init; tl++) {
7491 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
7492 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
7493 sd->flags |= SD_OVERLAP;
7494 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
7501 *per_cpu_ptr(d.sd, i) = sd;
7504 /* Build the groups for the domains */
7505 for_each_cpu(i, cpu_map) {
7506 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7507 sd->span_weight = cpumask_weight(sched_domain_span(sd));
7508 if (sd->flags & SD_OVERLAP) {
7509 if (build_overlap_sched_groups(sd, i))
7512 if (build_sched_groups(sd, i))
7518 /* Calculate CPU power for physical packages and nodes */
7519 for (i = nr_cpumask_bits-1; i >= 0; i--) {
7520 if (!cpumask_test_cpu(i, cpu_map))
7523 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7524 claim_allocations(i, sd);
7525 init_sched_groups_power(i, sd);
7529 /* Attach the domains */
7531 for_each_cpu(i, cpu_map) {
7532 sd = *per_cpu_ptr(d.sd, i);
7533 cpu_attach_domain(sd, d.rd, i);
7539 __free_domain_allocs(&d, alloc_state, cpu_map);
7543 static cpumask_var_t *doms_cur; /* current sched domains */
7544 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7545 static struct sched_domain_attr *dattr_cur;
7546 /* attribues of custom domains in 'doms_cur' */
7549 * Special case: If a kmalloc of a doms_cur partition (array of
7550 * cpumask) fails, then fallback to a single sched domain,
7551 * as determined by the single cpumask fallback_doms.
7553 static cpumask_var_t fallback_doms;
7556 * arch_update_cpu_topology lets virtualized architectures update the
7557 * cpu core maps. It is supposed to return 1 if the topology changed
7558 * or 0 if it stayed the same.
7560 int __attribute__((weak)) arch_update_cpu_topology(void)
7565 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7568 cpumask_var_t *doms;
7570 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7573 for (i = 0; i < ndoms; i++) {
7574 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7575 free_sched_domains(doms, i);
7582 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7585 for (i = 0; i < ndoms; i++)
7586 free_cpumask_var(doms[i]);
7591 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7592 * For now this just excludes isolated cpus, but could be used to
7593 * exclude other special cases in the future.
7595 static int init_sched_domains(const struct cpumask *cpu_map)
7599 arch_update_cpu_topology();
7601 doms_cur = alloc_sched_domains(ndoms_cur);
7603 doms_cur = &fallback_doms;
7604 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7606 err = build_sched_domains(doms_cur[0], NULL);
7607 register_sched_domain_sysctl();
7613 * Detach sched domains from a group of cpus specified in cpu_map
7614 * These cpus will now be attached to the NULL domain
7616 static void detach_destroy_domains(const struct cpumask *cpu_map)
7621 for_each_cpu(i, cpu_map)
7622 cpu_attach_domain(NULL, &def_root_domain, i);
7626 /* handle null as "default" */
7627 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7628 struct sched_domain_attr *new, int idx_new)
7630 struct sched_domain_attr tmp;
7637 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7638 new ? (new + idx_new) : &tmp,
7639 sizeof(struct sched_domain_attr));
7643 * Partition sched domains as specified by the 'ndoms_new'
7644 * cpumasks in the array doms_new[] of cpumasks. This compares
7645 * doms_new[] to the current sched domain partitioning, doms_cur[].
7646 * It destroys each deleted domain and builds each new domain.
7648 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7649 * The masks don't intersect (don't overlap.) We should setup one
7650 * sched domain for each mask. CPUs not in any of the cpumasks will
7651 * not be load balanced. If the same cpumask appears both in the
7652 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7655 * The passed in 'doms_new' should be allocated using
7656 * alloc_sched_domains. This routine takes ownership of it and will
7657 * free_sched_domains it when done with it. If the caller failed the
7658 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7659 * and partition_sched_domains() will fallback to the single partition
7660 * 'fallback_doms', it also forces the domains to be rebuilt.
7662 * If doms_new == NULL it will be replaced with cpu_online_mask.
7663 * ndoms_new == 0 is a special case for destroying existing domains,
7664 * and it will not create the default domain.
7666 * Call with hotplug lock held
7668 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7669 struct sched_domain_attr *dattr_new)
7674 mutex_lock(&sched_domains_mutex);
7676 /* always unregister in case we don't destroy any domains */
7677 unregister_sched_domain_sysctl();
7679 /* Let architecture update cpu core mappings. */
7680 new_topology = arch_update_cpu_topology();
7682 n = doms_new ? ndoms_new : 0;
7684 /* Destroy deleted domains */
7685 for (i = 0; i < ndoms_cur; i++) {
7686 for (j = 0; j < n && !new_topology; j++) {
7687 if (cpumask_equal(doms_cur[i], doms_new[j])
7688 && dattrs_equal(dattr_cur, i, dattr_new, j))
7691 /* no match - a current sched domain not in new doms_new[] */
7692 detach_destroy_domains(doms_cur[i]);
7697 if (doms_new == NULL) {
7699 doms_new = &fallback_doms;
7700 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7701 WARN_ON_ONCE(dattr_new);
7704 /* Build new domains */
7705 for (i = 0; i < ndoms_new; i++) {
7706 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7707 if (cpumask_equal(doms_new[i], doms_cur[j])
7708 && dattrs_equal(dattr_new, i, dattr_cur, j))
7711 /* no match - add a new doms_new */
7712 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7717 /* Remember the new sched domains */
7718 if (doms_cur != &fallback_doms)
7719 free_sched_domains(doms_cur, ndoms_cur);
7720 kfree(dattr_cur); /* kfree(NULL) is safe */
7721 doms_cur = doms_new;
7722 dattr_cur = dattr_new;
7723 ndoms_cur = ndoms_new;
7725 register_sched_domain_sysctl();
7727 mutex_unlock(&sched_domains_mutex);
7730 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7731 static void reinit_sched_domains(void)
7735 /* Destroy domains first to force the rebuild */
7736 partition_sched_domains(0, NULL, NULL);
7738 rebuild_sched_domains();
7742 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7744 unsigned int level = 0;
7746 if (sscanf(buf, "%u", &level) != 1)
7750 * level is always be positive so don't check for
7751 * level < POWERSAVINGS_BALANCE_NONE which is 0
7752 * What happens on 0 or 1 byte write,
7753 * need to check for count as well?
7756 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7760 sched_smt_power_savings = level;
7762 sched_mc_power_savings = level;
7764 reinit_sched_domains();
7769 #ifdef CONFIG_SCHED_MC
7770 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7771 struct sysdev_class_attribute *attr,
7774 return sprintf(page, "%u\n", sched_mc_power_savings);
7776 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7777 struct sysdev_class_attribute *attr,
7778 const char *buf, size_t count)
7780 return sched_power_savings_store(buf, count, 0);
7782 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7783 sched_mc_power_savings_show,
7784 sched_mc_power_savings_store);
7787 #ifdef CONFIG_SCHED_SMT
7788 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7789 struct sysdev_class_attribute *attr,
7792 return sprintf(page, "%u\n", sched_smt_power_savings);
7794 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7795 struct sysdev_class_attribute *attr,
7796 const char *buf, size_t count)
7798 return sched_power_savings_store(buf, count, 1);
7800 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7801 sched_smt_power_savings_show,
7802 sched_smt_power_savings_store);
7805 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7809 #ifdef CONFIG_SCHED_SMT
7811 err = sysfs_create_file(&cls->kset.kobj,
7812 &attr_sched_smt_power_savings.attr);
7814 #ifdef CONFIG_SCHED_MC
7815 if (!err && mc_capable())
7816 err = sysfs_create_file(&cls->kset.kobj,
7817 &attr_sched_mc_power_savings.attr);
7821 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7824 * Update cpusets according to cpu_active mask. If cpusets are
7825 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7826 * around partition_sched_domains().
7828 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7831 switch (action & ~CPU_TASKS_FROZEN) {
7833 case CPU_DOWN_FAILED:
7834 cpuset_update_active_cpus();
7841 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7844 switch (action & ~CPU_TASKS_FROZEN) {
7845 case CPU_DOWN_PREPARE:
7846 cpuset_update_active_cpus();
7853 static int update_runtime(struct notifier_block *nfb,
7854 unsigned long action, void *hcpu)
7856 int cpu = (int)(long)hcpu;
7859 case CPU_DOWN_PREPARE:
7860 case CPU_DOWN_PREPARE_FROZEN:
7861 disable_runtime(cpu_rq(cpu));
7864 case CPU_DOWN_FAILED:
7865 case CPU_DOWN_FAILED_FROZEN:
7867 case CPU_ONLINE_FROZEN:
7868 enable_runtime(cpu_rq(cpu));
7876 void __init sched_init_smp(void)
7878 cpumask_var_t non_isolated_cpus;
7880 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7881 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7884 mutex_lock(&sched_domains_mutex);
7885 init_sched_domains(cpu_active_mask);
7886 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7887 if (cpumask_empty(non_isolated_cpus))
7888 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7889 mutex_unlock(&sched_domains_mutex);
7892 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7893 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7895 /* RT runtime code needs to handle some hotplug events */
7896 hotcpu_notifier(update_runtime, 0);
7900 /* Move init over to a non-isolated CPU */
7901 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7903 sched_init_granularity();
7904 free_cpumask_var(non_isolated_cpus);
7906 init_sched_rt_class();
7909 void __init sched_init_smp(void)
7911 sched_init_granularity();
7913 #endif /* CONFIG_SMP */
7915 const_debug unsigned int sysctl_timer_migration = 1;
7917 int in_sched_functions(unsigned long addr)
7919 return in_lock_functions(addr) ||
7920 (addr >= (unsigned long)__sched_text_start
7921 && addr < (unsigned long)__sched_text_end);
7924 static void init_cfs_rq(struct cfs_rq *cfs_rq)
7926 cfs_rq->tasks_timeline = RB_ROOT;
7927 INIT_LIST_HEAD(&cfs_rq->tasks);
7928 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7929 #ifndef CONFIG_64BIT
7930 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
7934 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7936 struct rt_prio_array *array;
7939 array = &rt_rq->active;
7940 for (i = 0; i < MAX_RT_PRIO; i++) {
7941 INIT_LIST_HEAD(array->queue + i);
7942 __clear_bit(i, array->bitmap);
7944 /* delimiter for bitsearch: */
7945 __set_bit(MAX_RT_PRIO, array->bitmap);
7947 #if defined CONFIG_SMP
7948 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7949 rt_rq->highest_prio.next = MAX_RT_PRIO;
7950 rt_rq->rt_nr_migratory = 0;
7951 rt_rq->overloaded = 0;
7952 plist_head_init(&rt_rq->pushable_tasks);
7956 rt_rq->rt_throttled = 0;
7957 rt_rq->rt_runtime = 0;
7958 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7961 #ifdef CONFIG_FAIR_GROUP_SCHED
7962 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7963 struct sched_entity *se, int cpu,
7964 struct sched_entity *parent)
7966 struct rq *rq = cpu_rq(cpu);
7971 /* allow initial update_cfs_load() to truncate */
7972 cfs_rq->load_stamp = 1;
7975 tg->cfs_rq[cpu] = cfs_rq;
7978 /* se could be NULL for root_task_group */
7983 se->cfs_rq = &rq->cfs;
7985 se->cfs_rq = parent->my_q;
7988 update_load_set(&se->load, 0);
7989 se->parent = parent;
7993 #ifdef CONFIG_RT_GROUP_SCHED
7994 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7995 struct sched_rt_entity *rt_se, int cpu,
7996 struct sched_rt_entity *parent)
7998 struct rq *rq = cpu_rq(cpu);
8000 rt_rq->highest_prio.curr = MAX_RT_PRIO;
8001 rt_rq->rt_nr_boosted = 0;
8005 tg->rt_rq[cpu] = rt_rq;
8006 tg->rt_se[cpu] = rt_se;
8012 rt_se->rt_rq = &rq->rt;
8014 rt_se->rt_rq = parent->my_q;
8016 rt_se->my_q = rt_rq;
8017 rt_se->parent = parent;
8018 INIT_LIST_HEAD(&rt_se->run_list);
8022 void __init sched_init(void)
8025 unsigned long alloc_size = 0, ptr;
8027 #ifdef CONFIG_FAIR_GROUP_SCHED
8028 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8030 #ifdef CONFIG_RT_GROUP_SCHED
8031 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8033 #ifdef CONFIG_CPUMASK_OFFSTACK
8034 alloc_size += num_possible_cpus() * cpumask_size();
8037 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
8039 #ifdef CONFIG_FAIR_GROUP_SCHED
8040 root_task_group.se = (struct sched_entity **)ptr;
8041 ptr += nr_cpu_ids * sizeof(void **);
8043 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8044 ptr += nr_cpu_ids * sizeof(void **);
8046 #endif /* CONFIG_FAIR_GROUP_SCHED */
8047 #ifdef CONFIG_RT_GROUP_SCHED
8048 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8049 ptr += nr_cpu_ids * sizeof(void **);
8051 root_task_group.rt_rq = (struct rt_rq **)ptr;
8052 ptr += nr_cpu_ids * sizeof(void **);
8054 #endif /* CONFIG_RT_GROUP_SCHED */
8055 #ifdef CONFIG_CPUMASK_OFFSTACK
8056 for_each_possible_cpu(i) {
8057 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
8058 ptr += cpumask_size();
8060 #endif /* CONFIG_CPUMASK_OFFSTACK */
8064 init_defrootdomain();
8067 init_rt_bandwidth(&def_rt_bandwidth,
8068 global_rt_period(), global_rt_runtime());
8070 #ifdef CONFIG_RT_GROUP_SCHED
8071 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8072 global_rt_period(), global_rt_runtime());
8073 #endif /* CONFIG_RT_GROUP_SCHED */
8075 #ifdef CONFIG_CGROUP_SCHED
8076 list_add(&root_task_group.list, &task_groups);
8077 INIT_LIST_HEAD(&root_task_group.children);
8078 autogroup_init(&init_task);
8079 #endif /* CONFIG_CGROUP_SCHED */
8081 for_each_possible_cpu(i) {
8085 raw_spin_lock_init(&rq->lock);
8087 rq->calc_load_active = 0;
8088 rq->calc_load_update = jiffies + LOAD_FREQ;
8089 init_cfs_rq(&rq->cfs);
8090 init_rt_rq(&rq->rt, rq);
8091 #ifdef CONFIG_FAIR_GROUP_SCHED
8092 root_task_group.shares = root_task_group_load;
8093 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8095 * How much cpu bandwidth does root_task_group get?
8097 * In case of task-groups formed thr' the cgroup filesystem, it
8098 * gets 100% of the cpu resources in the system. This overall
8099 * system cpu resource is divided among the tasks of
8100 * root_task_group and its child task-groups in a fair manner,
8101 * based on each entity's (task or task-group's) weight
8102 * (se->load.weight).
8104 * In other words, if root_task_group has 10 tasks of weight
8105 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8106 * then A0's share of the cpu resource is:
8108 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8110 * We achieve this by letting root_task_group's tasks sit
8111 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
8113 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
8114 #endif /* CONFIG_FAIR_GROUP_SCHED */
8116 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8117 #ifdef CONFIG_RT_GROUP_SCHED
8118 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8119 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
8122 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8123 rq->cpu_load[j] = 0;
8125 rq->last_load_update_tick = jiffies;
8130 rq->cpu_power = SCHED_POWER_SCALE;
8131 rq->post_schedule = 0;
8132 rq->active_balance = 0;
8133 rq->next_balance = jiffies;
8138 rq->avg_idle = 2*sysctl_sched_migration_cost;
8139 rq_attach_root(rq, &def_root_domain);
8141 rq->nohz_balance_kick = 0;
8142 init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb, i));
8146 atomic_set(&rq->nr_iowait, 0);
8149 set_load_weight(&init_task);
8151 #ifdef CONFIG_PREEMPT_NOTIFIERS
8152 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8156 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8159 #ifdef CONFIG_RT_MUTEXES
8160 plist_head_init(&init_task.pi_waiters);
8164 * The boot idle thread does lazy MMU switching as well:
8166 atomic_inc(&init_mm.mm_count);
8167 enter_lazy_tlb(&init_mm, current);
8170 * Make us the idle thread. Technically, schedule() should not be
8171 * called from this thread, however somewhere below it might be,
8172 * but because we are the idle thread, we just pick up running again
8173 * when this runqueue becomes "idle".
8175 init_idle(current, smp_processor_id());
8177 calc_load_update = jiffies + LOAD_FREQ;
8180 * During early bootup we pretend to be a normal task:
8182 current->sched_class = &fair_sched_class;
8184 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8185 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
8187 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
8189 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8190 alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT);
8191 atomic_set(&nohz.load_balancer, nr_cpu_ids);
8192 atomic_set(&nohz.first_pick_cpu, nr_cpu_ids);
8193 atomic_set(&nohz.second_pick_cpu, nr_cpu_ids);
8195 /* May be allocated at isolcpus cmdline parse time */
8196 if (cpu_isolated_map == NULL)
8197 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
8200 scheduler_running = 1;
8203 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
8204 static inline int preempt_count_equals(int preempt_offset)
8206 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
8208 return (nested == preempt_offset);
8211 void __might_sleep(const char *file, int line, int preempt_offset)
8213 static unsigned long prev_jiffy; /* ratelimiting */
8215 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
8216 system_state != SYSTEM_RUNNING || oops_in_progress)
8218 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8220 prev_jiffy = jiffies;
8223 "BUG: sleeping function called from invalid context at %s:%d\n",
8226 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8227 in_atomic(), irqs_disabled(),
8228 current->pid, current->comm);
8230 debug_show_held_locks(current);
8231 if (irqs_disabled())
8232 print_irqtrace_events(current);
8235 EXPORT_SYMBOL(__might_sleep);
8238 #ifdef CONFIG_MAGIC_SYSRQ
8239 static void normalize_task(struct rq *rq, struct task_struct *p)
8241 const struct sched_class *prev_class = p->sched_class;
8242 int old_prio = p->prio;
8247 deactivate_task(rq, p, 0);
8248 __setscheduler(rq, p, SCHED_NORMAL, 0);
8250 activate_task(rq, p, 0);
8251 resched_task(rq->curr);
8254 check_class_changed(rq, p, prev_class, old_prio);
8257 void normalize_rt_tasks(void)
8259 struct task_struct *g, *p;
8260 unsigned long flags;
8263 read_lock_irqsave(&tasklist_lock, flags);
8264 do_each_thread(g, p) {
8266 * Only normalize user tasks:
8271 p->se.exec_start = 0;
8272 #ifdef CONFIG_SCHEDSTATS
8273 p->se.statistics.wait_start = 0;
8274 p->se.statistics.sleep_start = 0;
8275 p->se.statistics.block_start = 0;
8280 * Renice negative nice level userspace
8283 if (TASK_NICE(p) < 0 && p->mm)
8284 set_user_nice(p, 0);
8288 raw_spin_lock(&p->pi_lock);
8289 rq = __task_rq_lock(p);
8291 normalize_task(rq, p);
8293 __task_rq_unlock(rq);
8294 raw_spin_unlock(&p->pi_lock);
8295 } while_each_thread(g, p);
8297 read_unlock_irqrestore(&tasklist_lock, flags);
8300 #endif /* CONFIG_MAGIC_SYSRQ */
8302 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8304 * These functions are only useful for the IA64 MCA handling, or kdb.
8306 * They can only be called when the whole system has been
8307 * stopped - every CPU needs to be quiescent, and no scheduling
8308 * activity can take place. Using them for anything else would
8309 * be a serious bug, and as a result, they aren't even visible
8310 * under any other configuration.
8314 * curr_task - return the current task for a given cpu.
8315 * @cpu: the processor in question.
8317 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8319 struct task_struct *curr_task(int cpu)
8321 return cpu_curr(cpu);
8324 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8328 * set_curr_task - set the current task for a given cpu.
8329 * @cpu: the processor in question.
8330 * @p: the task pointer to set.
8332 * Description: This function must only be used when non-maskable interrupts
8333 * are serviced on a separate stack. It allows the architecture to switch the
8334 * notion of the current task on a cpu in a non-blocking manner. This function
8335 * must be called with all CPU's synchronized, and interrupts disabled, the
8336 * and caller must save the original value of the current task (see
8337 * curr_task() above) and restore that value before reenabling interrupts and
8338 * re-starting the system.
8340 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8342 void set_curr_task(int cpu, struct task_struct *p)
8349 #ifdef CONFIG_FAIR_GROUP_SCHED
8350 static void free_fair_sched_group(struct task_group *tg)
8354 for_each_possible_cpu(i) {
8356 kfree(tg->cfs_rq[i]);
8366 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8368 struct cfs_rq *cfs_rq;
8369 struct sched_entity *se;
8372 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8375 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8379 tg->shares = NICE_0_LOAD;
8381 for_each_possible_cpu(i) {
8382 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8383 GFP_KERNEL, cpu_to_node(i));
8387 se = kzalloc_node(sizeof(struct sched_entity),
8388 GFP_KERNEL, cpu_to_node(i));
8392 init_cfs_rq(cfs_rq);
8393 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8404 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8406 struct rq *rq = cpu_rq(cpu);
8407 unsigned long flags;
8410 * Only empty task groups can be destroyed; so we can speculatively
8411 * check on_list without danger of it being re-added.
8413 if (!tg->cfs_rq[cpu]->on_list)
8416 raw_spin_lock_irqsave(&rq->lock, flags);
8417 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8418 raw_spin_unlock_irqrestore(&rq->lock, flags);
8420 #else /* !CONFIG_FAIR_GROUP_SCHED */
8421 static inline void free_fair_sched_group(struct task_group *tg)
8426 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8431 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8434 #endif /* CONFIG_FAIR_GROUP_SCHED */
8436 #ifdef CONFIG_RT_GROUP_SCHED
8437 static void free_rt_sched_group(struct task_group *tg)
8442 destroy_rt_bandwidth(&tg->rt_bandwidth);
8444 for_each_possible_cpu(i) {
8446 kfree(tg->rt_rq[i]);
8448 kfree(tg->rt_se[i]);
8456 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8458 struct rt_rq *rt_rq;
8459 struct sched_rt_entity *rt_se;
8462 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8465 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8469 init_rt_bandwidth(&tg->rt_bandwidth,
8470 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8472 for_each_possible_cpu(i) {
8473 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8474 GFP_KERNEL, cpu_to_node(i));
8478 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8479 GFP_KERNEL, cpu_to_node(i));
8483 init_rt_rq(rt_rq, cpu_rq(i));
8484 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8485 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
8495 #else /* !CONFIG_RT_GROUP_SCHED */
8496 static inline void free_rt_sched_group(struct task_group *tg)
8501 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8505 #endif /* CONFIG_RT_GROUP_SCHED */
8507 #ifdef CONFIG_CGROUP_SCHED
8508 static void free_sched_group(struct task_group *tg)
8510 free_fair_sched_group(tg);
8511 free_rt_sched_group(tg);
8516 /* allocate runqueue etc for a new task group */
8517 struct task_group *sched_create_group(struct task_group *parent)
8519 struct task_group *tg;
8520 unsigned long flags;
8522 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8524 return ERR_PTR(-ENOMEM);
8526 if (!alloc_fair_sched_group(tg, parent))
8529 if (!alloc_rt_sched_group(tg, parent))
8532 spin_lock_irqsave(&task_group_lock, flags);
8533 list_add_rcu(&tg->list, &task_groups);
8535 WARN_ON(!parent); /* root should already exist */
8537 tg->parent = parent;
8538 INIT_LIST_HEAD(&tg->children);
8539 list_add_rcu(&tg->siblings, &parent->children);
8540 spin_unlock_irqrestore(&task_group_lock, flags);
8545 free_sched_group(tg);
8546 return ERR_PTR(-ENOMEM);
8549 /* rcu callback to free various structures associated with a task group */
8550 static void free_sched_group_rcu(struct rcu_head *rhp)
8552 /* now it should be safe to free those cfs_rqs */
8553 free_sched_group(container_of(rhp, struct task_group, rcu));
8556 /* Destroy runqueue etc associated with a task group */
8557 void sched_destroy_group(struct task_group *tg)
8559 unsigned long flags;
8562 /* end participation in shares distribution */
8563 for_each_possible_cpu(i)
8564 unregister_fair_sched_group(tg, i);
8566 spin_lock_irqsave(&task_group_lock, flags);
8567 list_del_rcu(&tg->list);
8568 list_del_rcu(&tg->siblings);
8569 spin_unlock_irqrestore(&task_group_lock, flags);
8571 /* wait for possible concurrent references to cfs_rqs complete */
8572 call_rcu(&tg->rcu, free_sched_group_rcu);
8575 /* change task's runqueue when it moves between groups.
8576 * The caller of this function should have put the task in its new group
8577 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8578 * reflect its new group.
8580 void sched_move_task(struct task_struct *tsk)
8583 unsigned long flags;
8586 rq = task_rq_lock(tsk, &flags);
8588 running = task_current(rq, tsk);
8592 dequeue_task(rq, tsk, 0);
8593 if (unlikely(running))
8594 tsk->sched_class->put_prev_task(rq, tsk);
8596 #ifdef CONFIG_FAIR_GROUP_SCHED
8597 if (tsk->sched_class->task_move_group)
8598 tsk->sched_class->task_move_group(tsk, on_rq);
8601 set_task_rq(tsk, task_cpu(tsk));
8603 if (unlikely(running))
8604 tsk->sched_class->set_curr_task(rq);
8606 enqueue_task(rq, tsk, 0);
8608 task_rq_unlock(rq, tsk, &flags);
8610 #endif /* CONFIG_CGROUP_SCHED */
8612 #ifdef CONFIG_FAIR_GROUP_SCHED
8613 static DEFINE_MUTEX(shares_mutex);
8615 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8618 unsigned long flags;
8621 * We can't change the weight of the root cgroup.
8626 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
8628 mutex_lock(&shares_mutex);
8629 if (tg->shares == shares)
8632 tg->shares = shares;
8633 for_each_possible_cpu(i) {
8634 struct rq *rq = cpu_rq(i);
8635 struct sched_entity *se;
8638 /* Propagate contribution to hierarchy */
8639 raw_spin_lock_irqsave(&rq->lock, flags);
8640 for_each_sched_entity(se)
8641 update_cfs_shares(group_cfs_rq(se));
8642 raw_spin_unlock_irqrestore(&rq->lock, flags);
8646 mutex_unlock(&shares_mutex);
8650 unsigned long sched_group_shares(struct task_group *tg)
8656 #ifdef CONFIG_RT_GROUP_SCHED
8658 * Ensure that the real time constraints are schedulable.
8660 static DEFINE_MUTEX(rt_constraints_mutex);
8662 static unsigned long to_ratio(u64 period, u64 runtime)
8664 if (runtime == RUNTIME_INF)
8667 return div64_u64(runtime << 20, period);
8670 /* Must be called with tasklist_lock held */
8671 static inline int tg_has_rt_tasks(struct task_group *tg)
8673 struct task_struct *g, *p;
8675 do_each_thread(g, p) {
8676 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8678 } while_each_thread(g, p);
8683 struct rt_schedulable_data {
8684 struct task_group *tg;
8689 static int tg_schedulable(struct task_group *tg, void *data)
8691 struct rt_schedulable_data *d = data;
8692 struct task_group *child;
8693 unsigned long total, sum = 0;
8694 u64 period, runtime;
8696 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8697 runtime = tg->rt_bandwidth.rt_runtime;
8700 period = d->rt_period;
8701 runtime = d->rt_runtime;
8705 * Cannot have more runtime than the period.
8707 if (runtime > period && runtime != RUNTIME_INF)
8711 * Ensure we don't starve existing RT tasks.
8713 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8716 total = to_ratio(period, runtime);
8719 * Nobody can have more than the global setting allows.
8721 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8725 * The sum of our children's runtime should not exceed our own.
8727 list_for_each_entry_rcu(child, &tg->children, siblings) {
8728 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8729 runtime = child->rt_bandwidth.rt_runtime;
8731 if (child == d->tg) {
8732 period = d->rt_period;
8733 runtime = d->rt_runtime;
8736 sum += to_ratio(period, runtime);
8745 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8747 struct rt_schedulable_data data = {
8749 .rt_period = period,
8750 .rt_runtime = runtime,
8753 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8756 static int tg_set_bandwidth(struct task_group *tg,
8757 u64 rt_period, u64 rt_runtime)
8761 mutex_lock(&rt_constraints_mutex);
8762 read_lock(&tasklist_lock);
8763 err = __rt_schedulable(tg, rt_period, rt_runtime);
8767 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8768 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8769 tg->rt_bandwidth.rt_runtime = rt_runtime;
8771 for_each_possible_cpu(i) {
8772 struct rt_rq *rt_rq = tg->rt_rq[i];
8774 raw_spin_lock(&rt_rq->rt_runtime_lock);
8775 rt_rq->rt_runtime = rt_runtime;
8776 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8778 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8780 read_unlock(&tasklist_lock);
8781 mutex_unlock(&rt_constraints_mutex);
8786 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8788 u64 rt_runtime, rt_period;
8790 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8791 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8792 if (rt_runtime_us < 0)
8793 rt_runtime = RUNTIME_INF;
8795 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8798 long sched_group_rt_runtime(struct task_group *tg)
8802 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8805 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8806 do_div(rt_runtime_us, NSEC_PER_USEC);
8807 return rt_runtime_us;
8810 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8812 u64 rt_runtime, rt_period;
8814 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8815 rt_runtime = tg->rt_bandwidth.rt_runtime;
8820 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8823 long sched_group_rt_period(struct task_group *tg)
8827 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8828 do_div(rt_period_us, NSEC_PER_USEC);
8829 return rt_period_us;
8832 static int sched_rt_global_constraints(void)
8834 u64 runtime, period;
8837 if (sysctl_sched_rt_period <= 0)
8840 runtime = global_rt_runtime();
8841 period = global_rt_period();
8844 * Sanity check on the sysctl variables.
8846 if (runtime > period && runtime != RUNTIME_INF)
8849 mutex_lock(&rt_constraints_mutex);
8850 read_lock(&tasklist_lock);
8851 ret = __rt_schedulable(NULL, 0, 0);
8852 read_unlock(&tasklist_lock);
8853 mutex_unlock(&rt_constraints_mutex);
8858 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8860 /* Don't accept realtime tasks when there is no way for them to run */
8861 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8867 #else /* !CONFIG_RT_GROUP_SCHED */
8868 static int sched_rt_global_constraints(void)
8870 unsigned long flags;
8873 if (sysctl_sched_rt_period <= 0)
8877 * There's always some RT tasks in the root group
8878 * -- migration, kstopmachine etc..
8880 if (sysctl_sched_rt_runtime == 0)
8883 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8884 for_each_possible_cpu(i) {
8885 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8887 raw_spin_lock(&rt_rq->rt_runtime_lock);
8888 rt_rq->rt_runtime = global_rt_runtime();
8889 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8891 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8895 #endif /* CONFIG_RT_GROUP_SCHED */
8897 int sched_rt_handler(struct ctl_table *table, int write,
8898 void __user *buffer, size_t *lenp,
8902 int old_period, old_runtime;
8903 static DEFINE_MUTEX(mutex);
8906 old_period = sysctl_sched_rt_period;
8907 old_runtime = sysctl_sched_rt_runtime;
8909 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8911 if (!ret && write) {
8912 ret = sched_rt_global_constraints();
8914 sysctl_sched_rt_period = old_period;
8915 sysctl_sched_rt_runtime = old_runtime;
8917 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8918 def_rt_bandwidth.rt_period =
8919 ns_to_ktime(global_rt_period());
8922 mutex_unlock(&mutex);
8927 #ifdef CONFIG_CGROUP_SCHED
8929 /* return corresponding task_group object of a cgroup */
8930 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8932 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8933 struct task_group, css);
8936 static struct cgroup_subsys_state *
8937 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8939 struct task_group *tg, *parent;
8941 if (!cgrp->parent) {
8942 /* This is early initialization for the top cgroup */
8943 return &root_task_group.css;
8946 parent = cgroup_tg(cgrp->parent);
8947 tg = sched_create_group(parent);
8949 return ERR_PTR(-ENOMEM);
8955 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8957 struct task_group *tg = cgroup_tg(cgrp);
8959 sched_destroy_group(tg);
8963 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8965 #ifdef CONFIG_RT_GROUP_SCHED
8966 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
8969 /* We don't support RT-tasks being in separate groups */
8970 if (tsk->sched_class != &fair_sched_class)
8977 cpu_cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8979 sched_move_task(tsk);
8983 cpu_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
8984 struct cgroup *old_cgrp, struct task_struct *task)
8987 * cgroup_exit() is called in the copy_process() failure path.
8988 * Ignore this case since the task hasn't ran yet, this avoids
8989 * trying to poke a half freed task state from generic code.
8991 if (!(task->flags & PF_EXITING))
8994 sched_move_task(task);
8997 #ifdef CONFIG_FAIR_GROUP_SCHED
8998 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9001 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
9004 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9006 struct task_group *tg = cgroup_tg(cgrp);
9008 return (u64) scale_load_down(tg->shares);
9010 #endif /* CONFIG_FAIR_GROUP_SCHED */
9012 #ifdef CONFIG_RT_GROUP_SCHED
9013 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9016 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9019 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9021 return sched_group_rt_runtime(cgroup_tg(cgrp));
9024 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9027 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9030 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9032 return sched_group_rt_period(cgroup_tg(cgrp));
9034 #endif /* CONFIG_RT_GROUP_SCHED */
9036 static struct cftype cpu_files[] = {
9037 #ifdef CONFIG_FAIR_GROUP_SCHED
9040 .read_u64 = cpu_shares_read_u64,
9041 .write_u64 = cpu_shares_write_u64,
9044 #ifdef CONFIG_RT_GROUP_SCHED
9046 .name = "rt_runtime_us",
9047 .read_s64 = cpu_rt_runtime_read,
9048 .write_s64 = cpu_rt_runtime_write,
9051 .name = "rt_period_us",
9052 .read_u64 = cpu_rt_period_read_uint,
9053 .write_u64 = cpu_rt_period_write_uint,
9058 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9060 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9063 struct cgroup_subsys cpu_cgroup_subsys = {
9065 .create = cpu_cgroup_create,
9066 .destroy = cpu_cgroup_destroy,
9067 .can_attach_task = cpu_cgroup_can_attach_task,
9068 .attach_task = cpu_cgroup_attach_task,
9069 .exit = cpu_cgroup_exit,
9070 .populate = cpu_cgroup_populate,
9071 .subsys_id = cpu_cgroup_subsys_id,
9075 #endif /* CONFIG_CGROUP_SCHED */
9077 #ifdef CONFIG_CGROUP_CPUACCT
9080 * CPU accounting code for task groups.
9082 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9083 * (balbir@in.ibm.com).
9086 /* track cpu usage of a group of tasks and its child groups */
9088 struct cgroup_subsys_state css;
9089 /* cpuusage holds pointer to a u64-type object on every cpu */
9090 u64 __percpu *cpuusage;
9091 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
9092 struct cpuacct *parent;
9095 struct cgroup_subsys cpuacct_subsys;
9097 /* return cpu accounting group corresponding to this container */
9098 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9100 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9101 struct cpuacct, css);
9104 /* return cpu accounting group to which this task belongs */
9105 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9107 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9108 struct cpuacct, css);
9111 /* create a new cpu accounting group */
9112 static struct cgroup_subsys_state *cpuacct_create(
9113 struct cgroup_subsys *ss, struct cgroup *cgrp)
9115 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9121 ca->cpuusage = alloc_percpu(u64);
9125 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9126 if (percpu_counter_init(&ca->cpustat[i], 0))
9127 goto out_free_counters;
9130 ca->parent = cgroup_ca(cgrp->parent);
9136 percpu_counter_destroy(&ca->cpustat[i]);
9137 free_percpu(ca->cpuusage);
9141 return ERR_PTR(-ENOMEM);
9144 /* destroy an existing cpu accounting group */
9146 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9148 struct cpuacct *ca = cgroup_ca(cgrp);
9151 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9152 percpu_counter_destroy(&ca->cpustat[i]);
9153 free_percpu(ca->cpuusage);
9157 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
9159 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9162 #ifndef CONFIG_64BIT
9164 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9166 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9168 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9176 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9178 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9180 #ifndef CONFIG_64BIT
9182 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9184 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9186 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9192 /* return total cpu usage (in nanoseconds) of a group */
9193 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9195 struct cpuacct *ca = cgroup_ca(cgrp);
9196 u64 totalcpuusage = 0;
9199 for_each_present_cpu(i)
9200 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9202 return totalcpuusage;
9205 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9208 struct cpuacct *ca = cgroup_ca(cgrp);
9217 for_each_present_cpu(i)
9218 cpuacct_cpuusage_write(ca, i, 0);
9224 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9227 struct cpuacct *ca = cgroup_ca(cgroup);
9231 for_each_present_cpu(i) {
9232 percpu = cpuacct_cpuusage_read(ca, i);
9233 seq_printf(m, "%llu ", (unsigned long long) percpu);
9235 seq_printf(m, "\n");
9239 static const char *cpuacct_stat_desc[] = {
9240 [CPUACCT_STAT_USER] = "user",
9241 [CPUACCT_STAT_SYSTEM] = "system",
9244 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
9245 struct cgroup_map_cb *cb)
9247 struct cpuacct *ca = cgroup_ca(cgrp);
9250 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
9251 s64 val = percpu_counter_read(&ca->cpustat[i]);
9252 val = cputime64_to_clock_t(val);
9253 cb->fill(cb, cpuacct_stat_desc[i], val);
9258 static struct cftype files[] = {
9261 .read_u64 = cpuusage_read,
9262 .write_u64 = cpuusage_write,
9265 .name = "usage_percpu",
9266 .read_seq_string = cpuacct_percpu_seq_read,
9270 .read_map = cpuacct_stats_show,
9274 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9276 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9280 * charge this task's execution time to its accounting group.
9282 * called with rq->lock held.
9284 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9289 if (unlikely(!cpuacct_subsys.active))
9292 cpu = task_cpu(tsk);
9298 for (; ca; ca = ca->parent) {
9299 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9300 *cpuusage += cputime;
9307 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9308 * in cputime_t units. As a result, cpuacct_update_stats calls
9309 * percpu_counter_add with values large enough to always overflow the
9310 * per cpu batch limit causing bad SMP scalability.
9312 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9313 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9314 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9317 #define CPUACCT_BATCH \
9318 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9320 #define CPUACCT_BATCH 0
9324 * Charge the system/user time to the task's accounting group.
9326 static void cpuacct_update_stats(struct task_struct *tsk,
9327 enum cpuacct_stat_index idx, cputime_t val)
9330 int batch = CPUACCT_BATCH;
9332 if (unlikely(!cpuacct_subsys.active))
9339 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9345 struct cgroup_subsys cpuacct_subsys = {
9347 .create = cpuacct_create,
9348 .destroy = cpuacct_destroy,
9349 .populate = cpuacct_populate,
9350 .subsys_id = cpuacct_subsys_id,
9352 #endif /* CONFIG_CGROUP_CPUACCT */