4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/stop_machine.h>
60 #include <linux/sysctl.h>
61 #include <linux/syscalls.h>
62 #include <linux/times.h>
63 #include <linux/tsacct_kern.h>
64 #include <linux/kprobes.h>
65 #include <linux/delayacct.h>
66 #include <linux/unistd.h>
67 #include <linux/pagemap.h>
68 #include <linux/hrtimer.h>
69 #include <linux/tick.h>
70 #include <linux/debugfs.h>
71 #include <linux/ctype.h>
72 #include <linux/ftrace.h>
73 #include <linux/slab.h>
76 #include <asm/irq_regs.h>
77 #include <asm/mutex.h>
79 #include "sched_cpupri.h"
80 #include "workqueue_sched.h"
81 #include "sched_autogroup.h"
83 #define CREATE_TRACE_POINTS
84 #include <trace/events/sched.h>
87 * Convert user-nice values [ -20 ... 0 ... 19 ]
88 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
91 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
92 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
93 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
96 * 'User priority' is the nice value converted to something we
97 * can work with better when scaling various scheduler parameters,
98 * it's a [ 0 ... 39 ] range.
100 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
101 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
102 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
105 * Helpers for converting nanosecond timing to jiffy resolution
107 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
109 #define NICE_0_LOAD SCHED_LOAD_SCALE
110 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
113 * These are the 'tuning knobs' of the scheduler:
115 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
116 * Timeslices get refilled after they expire.
118 #define DEF_TIMESLICE (100 * HZ / 1000)
121 * single value that denotes runtime == period, ie unlimited time.
123 #define RUNTIME_INF ((u64)~0ULL)
125 static inline int rt_policy(int policy)
127 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
132 static inline int task_has_rt_policy(struct task_struct *p)
134 return rt_policy(p->policy);
138 * This is the priority-queue data structure of the RT scheduling class:
140 struct rt_prio_array {
141 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
142 struct list_head queue[MAX_RT_PRIO];
145 struct rt_bandwidth {
146 /* nests inside the rq lock: */
147 raw_spinlock_t rt_runtime_lock;
150 struct hrtimer rt_period_timer;
153 static struct rt_bandwidth def_rt_bandwidth;
155 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
157 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
159 struct rt_bandwidth *rt_b =
160 container_of(timer, struct rt_bandwidth, rt_period_timer);
166 now = hrtimer_cb_get_time(timer);
167 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
172 idle = do_sched_rt_period_timer(rt_b, overrun);
175 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
179 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
181 rt_b->rt_period = ns_to_ktime(period);
182 rt_b->rt_runtime = runtime;
184 raw_spin_lock_init(&rt_b->rt_runtime_lock);
186 hrtimer_init(&rt_b->rt_period_timer,
187 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
188 rt_b->rt_period_timer.function = sched_rt_period_timer;
191 static inline int rt_bandwidth_enabled(void)
193 return sysctl_sched_rt_runtime >= 0;
196 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
200 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
203 if (hrtimer_active(&rt_b->rt_period_timer))
206 raw_spin_lock(&rt_b->rt_runtime_lock);
211 if (hrtimer_active(&rt_b->rt_period_timer))
214 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
215 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
217 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
218 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
219 delta = ktime_to_ns(ktime_sub(hard, soft));
220 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
221 HRTIMER_MODE_ABS_PINNED, 0);
223 raw_spin_unlock(&rt_b->rt_runtime_lock);
226 #ifdef CONFIG_RT_GROUP_SCHED
227 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
229 hrtimer_cancel(&rt_b->rt_period_timer);
234 * sched_domains_mutex serializes calls to arch_init_sched_domains,
235 * detach_destroy_domains and partition_sched_domains.
237 static DEFINE_MUTEX(sched_domains_mutex);
239 #ifdef CONFIG_CGROUP_SCHED
241 #include <linux/cgroup.h>
245 static LIST_HEAD(task_groups);
247 /* task group related information */
249 struct cgroup_subsys_state css;
251 #ifdef CONFIG_FAIR_GROUP_SCHED
252 /* schedulable entities of this group on each cpu */
253 struct sched_entity **se;
254 /* runqueue "owned" by this group on each cpu */
255 struct cfs_rq **cfs_rq;
256 unsigned long shares;
258 atomic_t load_weight;
261 #ifdef CONFIG_RT_GROUP_SCHED
262 struct sched_rt_entity **rt_se;
263 struct rt_rq **rt_rq;
265 struct rt_bandwidth rt_bandwidth;
269 struct list_head list;
271 struct task_group *parent;
272 struct list_head siblings;
273 struct list_head children;
275 #ifdef CONFIG_SCHED_AUTOGROUP
276 struct autogroup *autogroup;
280 /* task_group_lock serializes the addition/removal of task groups */
281 static DEFINE_SPINLOCK(task_group_lock);
283 #ifdef CONFIG_FAIR_GROUP_SCHED
285 # define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
288 * A weight of 0 or 1 can cause arithmetics problems.
289 * A weight of a cfs_rq is the sum of weights of which entities
290 * are queued on this cfs_rq, so a weight of a entity should not be
291 * too large, so as the shares value of a task group.
292 * (The default weight is 1024 - so there's no practical
293 * limitation from this.)
296 #define MAX_SHARES (1UL << 18)
298 static int root_task_group_load = ROOT_TASK_GROUP_LOAD;
301 /* Default task group.
302 * Every task in system belong to this group at bootup.
304 struct task_group root_task_group;
306 #endif /* CONFIG_CGROUP_SCHED */
308 /* CFS-related fields in a runqueue */
310 struct load_weight load;
311 unsigned long nr_running;
316 u64 min_vruntime_copy;
319 struct rb_root tasks_timeline;
320 struct rb_node *rb_leftmost;
322 struct list_head tasks;
323 struct list_head *balance_iterator;
326 * 'curr' points to currently running entity on this cfs_rq.
327 * It is set to NULL otherwise (i.e when none are currently running).
329 struct sched_entity *curr, *next, *last, *skip;
331 unsigned int nr_spread_over;
333 #ifdef CONFIG_FAIR_GROUP_SCHED
334 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
337 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
338 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
339 * (like users, containers etc.)
341 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
342 * list is used during load balance.
345 struct list_head leaf_cfs_rq_list;
346 struct task_group *tg; /* group that "owns" this runqueue */
350 * the part of load.weight contributed by tasks
352 unsigned long task_weight;
355 * h_load = weight * f(tg)
357 * Where f(tg) is the recursive weight fraction assigned to
360 unsigned long h_load;
363 * Maintaining per-cpu shares distribution for group scheduling
365 * load_stamp is the last time we updated the load average
366 * load_last is the last time we updated the load average and saw load
367 * load_unacc_exec_time is currently unaccounted execution time
371 u64 load_stamp, load_last, load_unacc_exec_time;
373 unsigned long load_contribution;
378 /* Real-Time classes' related field in a runqueue: */
380 struct rt_prio_array active;
381 unsigned long rt_nr_running;
382 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
384 int curr; /* highest queued rt task prio */
386 int next; /* next highest */
391 unsigned long rt_nr_migratory;
392 unsigned long rt_nr_total;
394 struct plist_head pushable_tasks;
399 /* Nests inside the rq lock: */
400 raw_spinlock_t rt_runtime_lock;
402 #ifdef CONFIG_RT_GROUP_SCHED
403 unsigned long rt_nr_boosted;
406 struct list_head leaf_rt_rq_list;
407 struct task_group *tg;
414 * We add the notion of a root-domain which will be used to define per-domain
415 * variables. Each exclusive cpuset essentially defines an island domain by
416 * fully partitioning the member cpus from any other cpuset. Whenever a new
417 * exclusive cpuset is created, we also create and attach a new root-domain
424 cpumask_var_t online;
427 * The "RT overload" flag: it gets set if a CPU has more than
428 * one runnable RT task.
430 cpumask_var_t rto_mask;
432 struct cpupri cpupri;
436 * By default the system creates a single root-domain with all cpus as
437 * members (mimicking the global state we have today).
439 static struct root_domain def_root_domain;
441 #endif /* CONFIG_SMP */
444 * This is the main, per-CPU runqueue data structure.
446 * Locking rule: those places that want to lock multiple runqueues
447 * (such as the load balancing or the thread migration code), lock
448 * acquire operations must be ordered by ascending &runqueue.
455 * nr_running and cpu_load should be in the same cacheline because
456 * remote CPUs use both these fields when doing load calculation.
458 unsigned long nr_running;
459 #define CPU_LOAD_IDX_MAX 5
460 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
461 unsigned long last_load_update_tick;
464 unsigned char nohz_balance_kick;
466 unsigned int skip_clock_update;
468 /* capture load from *all* tasks on this cpu: */
469 struct load_weight load;
470 unsigned long nr_load_updates;
476 #ifdef CONFIG_FAIR_GROUP_SCHED
477 /* list of leaf cfs_rq on this cpu: */
478 struct list_head leaf_cfs_rq_list;
480 #ifdef CONFIG_RT_GROUP_SCHED
481 struct list_head leaf_rt_rq_list;
485 * This is part of a global counter where only the total sum
486 * over all CPUs matters. A task can increase this counter on
487 * one CPU and if it got migrated afterwards it may decrease
488 * it on another CPU. Always updated under the runqueue lock:
490 unsigned long nr_uninterruptible;
492 struct task_struct *curr, *idle, *stop;
493 unsigned long next_balance;
494 struct mm_struct *prev_mm;
502 struct root_domain *rd;
503 struct sched_domain *sd;
505 unsigned long cpu_power;
507 unsigned char idle_at_tick;
508 /* For active balancing */
512 struct cpu_stop_work active_balance_work;
513 /* cpu of this runqueue: */
517 unsigned long avg_load_per_task;
525 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
529 /* calc_load related fields */
530 unsigned long calc_load_update;
531 long calc_load_active;
533 #ifdef CONFIG_SCHED_HRTICK
535 int hrtick_csd_pending;
536 struct call_single_data hrtick_csd;
538 struct hrtimer hrtick_timer;
541 #ifdef CONFIG_SCHEDSTATS
543 struct sched_info rq_sched_info;
544 unsigned long long rq_cpu_time;
545 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
547 /* sys_sched_yield() stats */
548 unsigned int yld_count;
550 /* schedule() stats */
551 unsigned int sched_switch;
552 unsigned int sched_count;
553 unsigned int sched_goidle;
555 /* try_to_wake_up() stats */
556 unsigned int ttwu_count;
557 unsigned int ttwu_local;
561 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
564 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags);
566 static inline int cpu_of(struct rq *rq)
575 #define rcu_dereference_check_sched_domain(p) \
576 rcu_dereference_check((p), \
577 rcu_read_lock_sched_held() || \
578 lockdep_is_held(&sched_domains_mutex))
581 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
582 * See detach_destroy_domains: synchronize_sched for details.
584 * The domain tree of any CPU may only be accessed from within
585 * preempt-disabled sections.
587 #define for_each_domain(cpu, __sd) \
588 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
590 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
591 #define this_rq() (&__get_cpu_var(runqueues))
592 #define task_rq(p) cpu_rq(task_cpu(p))
593 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
594 #define raw_rq() (&__raw_get_cpu_var(runqueues))
596 #ifdef CONFIG_CGROUP_SCHED
599 * Return the group to which this tasks belongs.
601 * We use task_subsys_state_check() and extend the RCU verification
602 * with lockdep_is_held(&p->pi_lock) because cpu_cgroup_attach()
603 * holds that lock for each task it moves into the cgroup. Therefore
604 * by holding that lock, we pin the task to the current cgroup.
606 static inline struct task_group *task_group(struct task_struct *p)
608 struct task_group *tg;
609 struct cgroup_subsys_state *css;
611 css = task_subsys_state_check(p, cpu_cgroup_subsys_id,
612 lockdep_is_held(&p->pi_lock));
613 tg = container_of(css, struct task_group, css);
615 return autogroup_task_group(p, tg);
618 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
619 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
621 #ifdef CONFIG_FAIR_GROUP_SCHED
622 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
623 p->se.parent = task_group(p)->se[cpu];
626 #ifdef CONFIG_RT_GROUP_SCHED
627 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
628 p->rt.parent = task_group(p)->rt_se[cpu];
632 #else /* CONFIG_CGROUP_SCHED */
634 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
635 static inline struct task_group *task_group(struct task_struct *p)
640 #endif /* CONFIG_CGROUP_SCHED */
642 static void update_rq_clock_task(struct rq *rq, s64 delta);
644 static void update_rq_clock(struct rq *rq)
648 if (rq->skip_clock_update)
651 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
653 update_rq_clock_task(rq, delta);
657 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
659 #ifdef CONFIG_SCHED_DEBUG
660 # define const_debug __read_mostly
662 # define const_debug static const
666 * runqueue_is_locked - Returns true if the current cpu runqueue is locked
667 * @cpu: the processor in question.
669 * This interface allows printk to be called with the runqueue lock
670 * held and know whether or not it is OK to wake up the klogd.
672 int runqueue_is_locked(int cpu)
674 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
678 * Debugging: various feature bits
681 #define SCHED_FEAT(name, enabled) \
682 __SCHED_FEAT_##name ,
685 #include "sched_features.h"
690 #define SCHED_FEAT(name, enabled) \
691 (1UL << __SCHED_FEAT_##name) * enabled |
693 const_debug unsigned int sysctl_sched_features =
694 #include "sched_features.h"
699 #ifdef CONFIG_SCHED_DEBUG
700 #define SCHED_FEAT(name, enabled) \
703 static __read_mostly char *sched_feat_names[] = {
704 #include "sched_features.h"
710 static int sched_feat_show(struct seq_file *m, void *v)
714 for (i = 0; sched_feat_names[i]; i++) {
715 if (!(sysctl_sched_features & (1UL << i)))
717 seq_printf(m, "%s ", sched_feat_names[i]);
725 sched_feat_write(struct file *filp, const char __user *ubuf,
726 size_t cnt, loff_t *ppos)
736 if (copy_from_user(&buf, ubuf, cnt))
742 if (strncmp(cmp, "NO_", 3) == 0) {
747 for (i = 0; sched_feat_names[i]; i++) {
748 if (strcmp(cmp, sched_feat_names[i]) == 0) {
750 sysctl_sched_features &= ~(1UL << i);
752 sysctl_sched_features |= (1UL << i);
757 if (!sched_feat_names[i])
765 static int sched_feat_open(struct inode *inode, struct file *filp)
767 return single_open(filp, sched_feat_show, NULL);
770 static const struct file_operations sched_feat_fops = {
771 .open = sched_feat_open,
772 .write = sched_feat_write,
775 .release = single_release,
778 static __init int sched_init_debug(void)
780 debugfs_create_file("sched_features", 0644, NULL, NULL,
785 late_initcall(sched_init_debug);
789 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
792 * Number of tasks to iterate in a single balance run.
793 * Limited because this is done with IRQs disabled.
795 const_debug unsigned int sysctl_sched_nr_migrate = 32;
798 * period over which we average the RT time consumption, measured
803 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
806 * period over which we measure -rt task cpu usage in us.
809 unsigned int sysctl_sched_rt_period = 1000000;
811 static __read_mostly int scheduler_running;
814 * part of the period that we allow rt tasks to run in us.
817 int sysctl_sched_rt_runtime = 950000;
819 static inline u64 global_rt_period(void)
821 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
824 static inline u64 global_rt_runtime(void)
826 if (sysctl_sched_rt_runtime < 0)
829 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
832 #ifndef prepare_arch_switch
833 # define prepare_arch_switch(next) do { } while (0)
835 #ifndef finish_arch_switch
836 # define finish_arch_switch(prev) do { } while (0)
839 static inline int task_current(struct rq *rq, struct task_struct *p)
841 return rq->curr == p;
844 static inline int task_running(struct rq *rq, struct task_struct *p)
849 return task_current(rq, p);
853 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
854 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
858 * We can optimise this out completely for !SMP, because the
859 * SMP rebalancing from interrupt is the only thing that cares
866 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
870 * After ->on_cpu is cleared, the task can be moved to a different CPU.
871 * We must ensure this doesn't happen until the switch is completely
877 #ifdef CONFIG_DEBUG_SPINLOCK
878 /* this is a valid case when another task releases the spinlock */
879 rq->lock.owner = current;
882 * If we are tracking spinlock dependencies then we have to
883 * fix up the runqueue lock - which gets 'carried over' from
886 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
888 raw_spin_unlock_irq(&rq->lock);
891 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
892 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
896 * We can optimise this out completely for !SMP, because the
897 * SMP rebalancing from interrupt is the only thing that cares
902 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
903 raw_spin_unlock_irq(&rq->lock);
905 raw_spin_unlock(&rq->lock);
909 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
913 * After ->on_cpu is cleared, the task can be moved to a different CPU.
914 * We must ensure this doesn't happen until the switch is completely
920 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
924 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
927 * __task_rq_lock - lock the rq @p resides on.
929 static inline struct rq *__task_rq_lock(struct task_struct *p)
934 lockdep_assert_held(&p->pi_lock);
938 raw_spin_lock(&rq->lock);
939 if (likely(rq == task_rq(p)))
941 raw_spin_unlock(&rq->lock);
946 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
948 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
949 __acquires(p->pi_lock)
955 raw_spin_lock_irqsave(&p->pi_lock, *flags);
957 raw_spin_lock(&rq->lock);
958 if (likely(rq == task_rq(p)))
960 raw_spin_unlock(&rq->lock);
961 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
965 static void __task_rq_unlock(struct rq *rq)
968 raw_spin_unlock(&rq->lock);
972 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
974 __releases(p->pi_lock)
976 raw_spin_unlock(&rq->lock);
977 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
981 * this_rq_lock - lock this runqueue and disable interrupts.
983 static struct rq *this_rq_lock(void)
990 raw_spin_lock(&rq->lock);
995 #ifdef CONFIG_SCHED_HRTICK
997 * Use HR-timers to deliver accurate preemption points.
999 * Its all a bit involved since we cannot program an hrt while holding the
1000 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1003 * When we get rescheduled we reprogram the hrtick_timer outside of the
1009 * - enabled by features
1010 * - hrtimer is actually high res
1012 static inline int hrtick_enabled(struct rq *rq)
1014 if (!sched_feat(HRTICK))
1016 if (!cpu_active(cpu_of(rq)))
1018 return hrtimer_is_hres_active(&rq->hrtick_timer);
1021 static void hrtick_clear(struct rq *rq)
1023 if (hrtimer_active(&rq->hrtick_timer))
1024 hrtimer_cancel(&rq->hrtick_timer);
1028 * High-resolution timer tick.
1029 * Runs from hardirq context with interrupts disabled.
1031 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1033 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1035 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1037 raw_spin_lock(&rq->lock);
1038 update_rq_clock(rq);
1039 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1040 raw_spin_unlock(&rq->lock);
1042 return HRTIMER_NORESTART;
1047 * called from hardirq (IPI) context
1049 static void __hrtick_start(void *arg)
1051 struct rq *rq = arg;
1053 raw_spin_lock(&rq->lock);
1054 hrtimer_restart(&rq->hrtick_timer);
1055 rq->hrtick_csd_pending = 0;
1056 raw_spin_unlock(&rq->lock);
1060 * Called to set the hrtick timer state.
1062 * called with rq->lock held and irqs disabled
1064 static void hrtick_start(struct rq *rq, u64 delay)
1066 struct hrtimer *timer = &rq->hrtick_timer;
1067 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1069 hrtimer_set_expires(timer, time);
1071 if (rq == this_rq()) {
1072 hrtimer_restart(timer);
1073 } else if (!rq->hrtick_csd_pending) {
1074 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1075 rq->hrtick_csd_pending = 1;
1080 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1082 int cpu = (int)(long)hcpu;
1085 case CPU_UP_CANCELED:
1086 case CPU_UP_CANCELED_FROZEN:
1087 case CPU_DOWN_PREPARE:
1088 case CPU_DOWN_PREPARE_FROZEN:
1090 case CPU_DEAD_FROZEN:
1091 hrtick_clear(cpu_rq(cpu));
1098 static __init void init_hrtick(void)
1100 hotcpu_notifier(hotplug_hrtick, 0);
1104 * Called to set the hrtick timer state.
1106 * called with rq->lock held and irqs disabled
1108 static void hrtick_start(struct rq *rq, u64 delay)
1110 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1111 HRTIMER_MODE_REL_PINNED, 0);
1114 static inline void init_hrtick(void)
1117 #endif /* CONFIG_SMP */
1119 static void init_rq_hrtick(struct rq *rq)
1122 rq->hrtick_csd_pending = 0;
1124 rq->hrtick_csd.flags = 0;
1125 rq->hrtick_csd.func = __hrtick_start;
1126 rq->hrtick_csd.info = rq;
1129 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1130 rq->hrtick_timer.function = hrtick;
1132 #else /* CONFIG_SCHED_HRTICK */
1133 static inline void hrtick_clear(struct rq *rq)
1137 static inline void init_rq_hrtick(struct rq *rq)
1141 static inline void init_hrtick(void)
1144 #endif /* CONFIG_SCHED_HRTICK */
1147 * resched_task - mark a task 'to be rescheduled now'.
1149 * On UP this means the setting of the need_resched flag, on SMP it
1150 * might also involve a cross-CPU call to trigger the scheduler on
1155 #ifndef tsk_is_polling
1156 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1159 static void resched_task(struct task_struct *p)
1163 assert_raw_spin_locked(&task_rq(p)->lock);
1165 if (test_tsk_need_resched(p))
1168 set_tsk_need_resched(p);
1171 if (cpu == smp_processor_id())
1174 /* NEED_RESCHED must be visible before we test polling */
1176 if (!tsk_is_polling(p))
1177 smp_send_reschedule(cpu);
1180 static void resched_cpu(int cpu)
1182 struct rq *rq = cpu_rq(cpu);
1183 unsigned long flags;
1185 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1187 resched_task(cpu_curr(cpu));
1188 raw_spin_unlock_irqrestore(&rq->lock, flags);
1193 * In the semi idle case, use the nearest busy cpu for migrating timers
1194 * from an idle cpu. This is good for power-savings.
1196 * We don't do similar optimization for completely idle system, as
1197 * selecting an idle cpu will add more delays to the timers than intended
1198 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1200 int get_nohz_timer_target(void)
1202 int cpu = smp_processor_id();
1204 struct sched_domain *sd;
1206 for_each_domain(cpu, sd) {
1207 for_each_cpu(i, sched_domain_span(sd))
1214 * When add_timer_on() enqueues a timer into the timer wheel of an
1215 * idle CPU then this timer might expire before the next timer event
1216 * which is scheduled to wake up that CPU. In case of a completely
1217 * idle system the next event might even be infinite time into the
1218 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1219 * leaves the inner idle loop so the newly added timer is taken into
1220 * account when the CPU goes back to idle and evaluates the timer
1221 * wheel for the next timer event.
1223 void wake_up_idle_cpu(int cpu)
1225 struct rq *rq = cpu_rq(cpu);
1227 if (cpu == smp_processor_id())
1231 * This is safe, as this function is called with the timer
1232 * wheel base lock of (cpu) held. When the CPU is on the way
1233 * to idle and has not yet set rq->curr to idle then it will
1234 * be serialized on the timer wheel base lock and take the new
1235 * timer into account automatically.
1237 if (rq->curr != rq->idle)
1241 * We can set TIF_RESCHED on the idle task of the other CPU
1242 * lockless. The worst case is that the other CPU runs the
1243 * idle task through an additional NOOP schedule()
1245 set_tsk_need_resched(rq->idle);
1247 /* NEED_RESCHED must be visible before we test polling */
1249 if (!tsk_is_polling(rq->idle))
1250 smp_send_reschedule(cpu);
1253 #endif /* CONFIG_NO_HZ */
1255 static u64 sched_avg_period(void)
1257 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1260 static void sched_avg_update(struct rq *rq)
1262 s64 period = sched_avg_period();
1264 while ((s64)(rq->clock - rq->age_stamp) > period) {
1266 * Inline assembly required to prevent the compiler
1267 * optimising this loop into a divmod call.
1268 * See __iter_div_u64_rem() for another example of this.
1270 asm("" : "+rm" (rq->age_stamp));
1271 rq->age_stamp += period;
1276 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1278 rq->rt_avg += rt_delta;
1279 sched_avg_update(rq);
1282 #else /* !CONFIG_SMP */
1283 static void resched_task(struct task_struct *p)
1285 assert_raw_spin_locked(&task_rq(p)->lock);
1286 set_tsk_need_resched(p);
1289 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1293 static void sched_avg_update(struct rq *rq)
1296 #endif /* CONFIG_SMP */
1298 #if BITS_PER_LONG == 32
1299 # define WMULT_CONST (~0UL)
1301 # define WMULT_CONST (1UL << 32)
1304 #define WMULT_SHIFT 32
1307 * Shift right and round:
1309 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1312 * delta *= weight / lw
1314 static unsigned long
1315 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1316 struct load_weight *lw)
1320 if (!lw->inv_weight) {
1321 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1324 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1328 tmp = (u64)delta_exec * weight;
1330 * Check whether we'd overflow the 64-bit multiplication:
1332 if (unlikely(tmp > WMULT_CONST))
1333 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1336 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1338 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1341 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1347 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1353 static inline void update_load_set(struct load_weight *lw, unsigned long w)
1360 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1361 * of tasks with abnormal "nice" values across CPUs the contribution that
1362 * each task makes to its run queue's load is weighted according to its
1363 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1364 * scaled version of the new time slice allocation that they receive on time
1368 #define WEIGHT_IDLEPRIO 3
1369 #define WMULT_IDLEPRIO 1431655765
1372 * Nice levels are multiplicative, with a gentle 10% change for every
1373 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1374 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1375 * that remained on nice 0.
1377 * The "10% effect" is relative and cumulative: from _any_ nice level,
1378 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1379 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1380 * If a task goes up by ~10% and another task goes down by ~10% then
1381 * the relative distance between them is ~25%.)
1383 static const int prio_to_weight[40] = {
1384 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1385 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1386 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1387 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1388 /* 0 */ 1024, 820, 655, 526, 423,
1389 /* 5 */ 335, 272, 215, 172, 137,
1390 /* 10 */ 110, 87, 70, 56, 45,
1391 /* 15 */ 36, 29, 23, 18, 15,
1395 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1397 * In cases where the weight does not change often, we can use the
1398 * precalculated inverse to speed up arithmetics by turning divisions
1399 * into multiplications:
1401 static const u32 prio_to_wmult[40] = {
1402 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1403 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1404 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1405 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1406 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1407 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1408 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1409 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1412 /* Time spent by the tasks of the cpu accounting group executing in ... */
1413 enum cpuacct_stat_index {
1414 CPUACCT_STAT_USER, /* ... user mode */
1415 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1417 CPUACCT_STAT_NSTATS,
1420 #ifdef CONFIG_CGROUP_CPUACCT
1421 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1422 static void cpuacct_update_stats(struct task_struct *tsk,
1423 enum cpuacct_stat_index idx, cputime_t val);
1425 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1426 static inline void cpuacct_update_stats(struct task_struct *tsk,
1427 enum cpuacct_stat_index idx, cputime_t val) {}
1430 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1432 update_load_add(&rq->load, load);
1435 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1437 update_load_sub(&rq->load, load);
1440 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1441 typedef int (*tg_visitor)(struct task_group *, void *);
1444 * Iterate the full tree, calling @down when first entering a node and @up when
1445 * leaving it for the final time.
1447 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1449 struct task_group *parent, *child;
1453 parent = &root_task_group;
1455 ret = (*down)(parent, data);
1458 list_for_each_entry_rcu(child, &parent->children, siblings) {
1465 ret = (*up)(parent, data);
1470 parent = parent->parent;
1479 static int tg_nop(struct task_group *tg, void *data)
1486 /* Used instead of source_load when we know the type == 0 */
1487 static unsigned long weighted_cpuload(const int cpu)
1489 return cpu_rq(cpu)->load.weight;
1493 * Return a low guess at the load of a migration-source cpu weighted
1494 * according to the scheduling class and "nice" value.
1496 * We want to under-estimate the load of migration sources, to
1497 * balance conservatively.
1499 static unsigned long source_load(int cpu, int type)
1501 struct rq *rq = cpu_rq(cpu);
1502 unsigned long total = weighted_cpuload(cpu);
1504 if (type == 0 || !sched_feat(LB_BIAS))
1507 return min(rq->cpu_load[type-1], total);
1511 * Return a high guess at the load of a migration-target cpu weighted
1512 * according to the scheduling class and "nice" value.
1514 static unsigned long target_load(int cpu, int type)
1516 struct rq *rq = cpu_rq(cpu);
1517 unsigned long total = weighted_cpuload(cpu);
1519 if (type == 0 || !sched_feat(LB_BIAS))
1522 return max(rq->cpu_load[type-1], total);
1525 static unsigned long power_of(int cpu)
1527 return cpu_rq(cpu)->cpu_power;
1530 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1532 static unsigned long cpu_avg_load_per_task(int cpu)
1534 struct rq *rq = cpu_rq(cpu);
1535 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1538 rq->avg_load_per_task = rq->load.weight / nr_running;
1540 rq->avg_load_per_task = 0;
1542 return rq->avg_load_per_task;
1545 #ifdef CONFIG_FAIR_GROUP_SCHED
1548 * Compute the cpu's hierarchical load factor for each task group.
1549 * This needs to be done in a top-down fashion because the load of a child
1550 * group is a fraction of its parents load.
1552 static int tg_load_down(struct task_group *tg, void *data)
1555 long cpu = (long)data;
1558 load = cpu_rq(cpu)->load.weight;
1560 load = tg->parent->cfs_rq[cpu]->h_load;
1561 load *= tg->se[cpu]->load.weight;
1562 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1565 tg->cfs_rq[cpu]->h_load = load;
1570 static void update_h_load(long cpu)
1572 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1577 #ifdef CONFIG_PREEMPT
1579 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1582 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1583 * way at the expense of forcing extra atomic operations in all
1584 * invocations. This assures that the double_lock is acquired using the
1585 * same underlying policy as the spinlock_t on this architecture, which
1586 * reduces latency compared to the unfair variant below. However, it
1587 * also adds more overhead and therefore may reduce throughput.
1589 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1590 __releases(this_rq->lock)
1591 __acquires(busiest->lock)
1592 __acquires(this_rq->lock)
1594 raw_spin_unlock(&this_rq->lock);
1595 double_rq_lock(this_rq, busiest);
1602 * Unfair double_lock_balance: Optimizes throughput at the expense of
1603 * latency by eliminating extra atomic operations when the locks are
1604 * already in proper order on entry. This favors lower cpu-ids and will
1605 * grant the double lock to lower cpus over higher ids under contention,
1606 * regardless of entry order into the function.
1608 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1609 __releases(this_rq->lock)
1610 __acquires(busiest->lock)
1611 __acquires(this_rq->lock)
1615 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1616 if (busiest < this_rq) {
1617 raw_spin_unlock(&this_rq->lock);
1618 raw_spin_lock(&busiest->lock);
1619 raw_spin_lock_nested(&this_rq->lock,
1620 SINGLE_DEPTH_NESTING);
1623 raw_spin_lock_nested(&busiest->lock,
1624 SINGLE_DEPTH_NESTING);
1629 #endif /* CONFIG_PREEMPT */
1632 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1634 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1636 if (unlikely(!irqs_disabled())) {
1637 /* printk() doesn't work good under rq->lock */
1638 raw_spin_unlock(&this_rq->lock);
1642 return _double_lock_balance(this_rq, busiest);
1645 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1646 __releases(busiest->lock)
1648 raw_spin_unlock(&busiest->lock);
1649 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1653 * double_rq_lock - safely lock two runqueues
1655 * Note this does not disable interrupts like task_rq_lock,
1656 * you need to do so manually before calling.
1658 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1659 __acquires(rq1->lock)
1660 __acquires(rq2->lock)
1662 BUG_ON(!irqs_disabled());
1664 raw_spin_lock(&rq1->lock);
1665 __acquire(rq2->lock); /* Fake it out ;) */
1668 raw_spin_lock(&rq1->lock);
1669 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1671 raw_spin_lock(&rq2->lock);
1672 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1678 * double_rq_unlock - safely unlock two runqueues
1680 * Note this does not restore interrupts like task_rq_unlock,
1681 * you need to do so manually after calling.
1683 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1684 __releases(rq1->lock)
1685 __releases(rq2->lock)
1687 raw_spin_unlock(&rq1->lock);
1689 raw_spin_unlock(&rq2->lock);
1691 __release(rq2->lock);
1694 #else /* CONFIG_SMP */
1697 * double_rq_lock - safely lock two runqueues
1699 * Note this does not disable interrupts like task_rq_lock,
1700 * you need to do so manually before calling.
1702 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1703 __acquires(rq1->lock)
1704 __acquires(rq2->lock)
1706 BUG_ON(!irqs_disabled());
1708 raw_spin_lock(&rq1->lock);
1709 __acquire(rq2->lock); /* Fake it out ;) */
1713 * double_rq_unlock - safely unlock two runqueues
1715 * Note this does not restore interrupts like task_rq_unlock,
1716 * you need to do so manually after calling.
1718 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1719 __releases(rq1->lock)
1720 __releases(rq2->lock)
1723 raw_spin_unlock(&rq1->lock);
1724 __release(rq2->lock);
1729 static void calc_load_account_idle(struct rq *this_rq);
1730 static void update_sysctl(void);
1731 static int get_update_sysctl_factor(void);
1732 static void update_cpu_load(struct rq *this_rq);
1734 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1736 set_task_rq(p, cpu);
1739 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1740 * successfuly executed on another CPU. We must ensure that updates of
1741 * per-task data have been completed by this moment.
1744 task_thread_info(p)->cpu = cpu;
1748 static const struct sched_class rt_sched_class;
1750 #define sched_class_highest (&stop_sched_class)
1751 #define for_each_class(class) \
1752 for (class = sched_class_highest; class; class = class->next)
1754 #include "sched_stats.h"
1756 static void inc_nr_running(struct rq *rq)
1761 static void dec_nr_running(struct rq *rq)
1766 static void set_load_weight(struct task_struct *p)
1769 * SCHED_IDLE tasks get minimal weight:
1771 if (p->policy == SCHED_IDLE) {
1772 p->se.load.weight = WEIGHT_IDLEPRIO;
1773 p->se.load.inv_weight = WMULT_IDLEPRIO;
1777 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1778 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1781 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1783 update_rq_clock(rq);
1784 sched_info_queued(p);
1785 p->sched_class->enqueue_task(rq, p, flags);
1788 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1790 update_rq_clock(rq);
1791 sched_info_dequeued(p);
1792 p->sched_class->dequeue_task(rq, p, flags);
1796 * activate_task - move a task to the runqueue.
1798 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1800 if (task_contributes_to_load(p))
1801 rq->nr_uninterruptible--;
1803 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);
1819 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1822 * There are no locks covering percpu hardirq/softirq time.
1823 * They are only modified in account_system_vtime, on corresponding CPU
1824 * with interrupts disabled. So, writes are safe.
1825 * They are read and saved off onto struct rq in update_rq_clock().
1826 * This may result in other CPU reading this CPU's irq time and can
1827 * race with irq/account_system_vtime on this CPU. We would either get old
1828 * or new value with a side effect of accounting a slice of irq time to wrong
1829 * task when irq is in progress while we read rq->clock. That is a worthy
1830 * compromise in place of having locks on each irq in account_system_time.
1832 static DEFINE_PER_CPU(u64, cpu_hardirq_time);
1833 static DEFINE_PER_CPU(u64, cpu_softirq_time);
1835 static DEFINE_PER_CPU(u64, irq_start_time);
1836 static int sched_clock_irqtime;
1838 void enable_sched_clock_irqtime(void)
1840 sched_clock_irqtime = 1;
1843 void disable_sched_clock_irqtime(void)
1845 sched_clock_irqtime = 0;
1848 #ifndef CONFIG_64BIT
1849 static DEFINE_PER_CPU(seqcount_t, irq_time_seq);
1851 static inline void irq_time_write_begin(void)
1853 __this_cpu_inc(irq_time_seq.sequence);
1857 static inline void irq_time_write_end(void)
1860 __this_cpu_inc(irq_time_seq.sequence);
1863 static inline u64 irq_time_read(int cpu)
1869 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
1870 irq_time = per_cpu(cpu_softirq_time, cpu) +
1871 per_cpu(cpu_hardirq_time, cpu);
1872 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
1876 #else /* CONFIG_64BIT */
1877 static inline void irq_time_write_begin(void)
1881 static inline void irq_time_write_end(void)
1885 static inline u64 irq_time_read(int cpu)
1887 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
1889 #endif /* CONFIG_64BIT */
1892 * Called before incrementing preempt_count on {soft,}irq_enter
1893 * and before decrementing preempt_count on {soft,}irq_exit.
1895 void account_system_vtime(struct task_struct *curr)
1897 unsigned long flags;
1901 if (!sched_clock_irqtime)
1904 local_irq_save(flags);
1906 cpu = smp_processor_id();
1907 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
1908 __this_cpu_add(irq_start_time, delta);
1910 irq_time_write_begin();
1912 * We do not account for softirq time from ksoftirqd here.
1913 * We want to continue accounting softirq time to ksoftirqd thread
1914 * in that case, so as not to confuse scheduler with a special task
1915 * that do not consume any time, but still wants to run.
1917 if (hardirq_count())
1918 __this_cpu_add(cpu_hardirq_time, delta);
1919 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
1920 __this_cpu_add(cpu_softirq_time, delta);
1922 irq_time_write_end();
1923 local_irq_restore(flags);
1925 EXPORT_SYMBOL_GPL(account_system_vtime);
1927 static void update_rq_clock_task(struct rq *rq, s64 delta)
1931 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
1934 * Since irq_time is only updated on {soft,}irq_exit, we might run into
1935 * this case when a previous update_rq_clock() happened inside a
1936 * {soft,}irq region.
1938 * When this happens, we stop ->clock_task and only update the
1939 * prev_irq_time stamp to account for the part that fit, so that a next
1940 * update will consume the rest. This ensures ->clock_task is
1943 * It does however cause some slight miss-attribution of {soft,}irq
1944 * time, a more accurate solution would be to update the irq_time using
1945 * the current rq->clock timestamp, except that would require using
1948 if (irq_delta > delta)
1951 rq->prev_irq_time += irq_delta;
1953 rq->clock_task += delta;
1955 if (irq_delta && sched_feat(NONIRQ_POWER))
1956 sched_rt_avg_update(rq, irq_delta);
1959 static int irqtime_account_hi_update(void)
1961 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
1962 unsigned long flags;
1966 local_irq_save(flags);
1967 latest_ns = this_cpu_read(cpu_hardirq_time);
1968 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->irq))
1970 local_irq_restore(flags);
1974 static int irqtime_account_si_update(void)
1976 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
1977 unsigned long flags;
1981 local_irq_save(flags);
1982 latest_ns = this_cpu_read(cpu_softirq_time);
1983 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->softirq))
1985 local_irq_restore(flags);
1989 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
1991 #define sched_clock_irqtime (0)
1993 static void update_rq_clock_task(struct rq *rq, s64 delta)
1995 rq->clock_task += delta;
1998 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2000 #include "sched_idletask.c"
2001 #include "sched_fair.c"
2002 #include "sched_rt.c"
2003 #include "sched_autogroup.c"
2004 #include "sched_stoptask.c"
2005 #ifdef CONFIG_SCHED_DEBUG
2006 # include "sched_debug.c"
2009 void sched_set_stop_task(int cpu, struct task_struct *stop)
2011 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
2012 struct task_struct *old_stop = cpu_rq(cpu)->stop;
2016 * Make it appear like a SCHED_FIFO task, its something
2017 * userspace knows about and won't get confused about.
2019 * Also, it will make PI more or less work without too
2020 * much confusion -- but then, stop work should not
2021 * rely on PI working anyway.
2023 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
2025 stop->sched_class = &stop_sched_class;
2028 cpu_rq(cpu)->stop = stop;
2032 * Reset it back to a normal scheduling class so that
2033 * it can die in pieces.
2035 old_stop->sched_class = &rt_sched_class;
2040 * __normal_prio - return the priority that is based on the static prio
2042 static inline int __normal_prio(struct task_struct *p)
2044 return p->static_prio;
2048 * Calculate the expected normal priority: i.e. priority
2049 * without taking RT-inheritance into account. Might be
2050 * boosted by interactivity modifiers. Changes upon fork,
2051 * setprio syscalls, and whenever the interactivity
2052 * estimator recalculates.
2054 static inline int normal_prio(struct task_struct *p)
2058 if (task_has_rt_policy(p))
2059 prio = MAX_RT_PRIO-1 - p->rt_priority;
2061 prio = __normal_prio(p);
2066 * Calculate the current priority, i.e. the priority
2067 * taken into account by the scheduler. This value might
2068 * be boosted by RT tasks, or might be boosted by
2069 * interactivity modifiers. Will be RT if the task got
2070 * RT-boosted. If not then it returns p->normal_prio.
2072 static int effective_prio(struct task_struct *p)
2074 p->normal_prio = normal_prio(p);
2076 * If we are RT tasks or we were boosted to RT priority,
2077 * keep the priority unchanged. Otherwise, update priority
2078 * to the normal priority:
2080 if (!rt_prio(p->prio))
2081 return p->normal_prio;
2086 * task_curr - is this task currently executing on a CPU?
2087 * @p: the task in question.
2089 inline int task_curr(const struct task_struct *p)
2091 return cpu_curr(task_cpu(p)) == p;
2094 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2095 const struct sched_class *prev_class,
2098 if (prev_class != p->sched_class) {
2099 if (prev_class->switched_from)
2100 prev_class->switched_from(rq, p);
2101 p->sched_class->switched_to(rq, p);
2102 } else if (oldprio != p->prio)
2103 p->sched_class->prio_changed(rq, p, oldprio);
2106 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2108 const struct sched_class *class;
2110 if (p->sched_class == rq->curr->sched_class) {
2111 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2113 for_each_class(class) {
2114 if (class == rq->curr->sched_class)
2116 if (class == p->sched_class) {
2117 resched_task(rq->curr);
2124 * A queue event has occurred, and we're going to schedule. In
2125 * this case, we can save a useless back to back clock update.
2127 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
2128 rq->skip_clock_update = 1;
2133 * Is this task likely cache-hot:
2136 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2140 if (p->sched_class != &fair_sched_class)
2143 if (unlikely(p->policy == SCHED_IDLE))
2147 * Buddy candidates are cache hot:
2149 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2150 (&p->se == cfs_rq_of(&p->se)->next ||
2151 &p->se == cfs_rq_of(&p->se)->last))
2154 if (sysctl_sched_migration_cost == -1)
2156 if (sysctl_sched_migration_cost == 0)
2159 delta = now - p->se.exec_start;
2161 return delta < (s64)sysctl_sched_migration_cost;
2164 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2166 #ifdef CONFIG_SCHED_DEBUG
2168 * We should never call set_task_cpu() on a blocked task,
2169 * ttwu() will sort out the placement.
2171 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2172 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2174 #ifdef CONFIG_LOCKDEP
2175 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
2176 lockdep_is_held(&task_rq(p)->lock)));
2180 trace_sched_migrate_task(p, new_cpu);
2182 if (task_cpu(p) != new_cpu) {
2183 p->se.nr_migrations++;
2184 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2187 __set_task_cpu(p, new_cpu);
2190 struct migration_arg {
2191 struct task_struct *task;
2195 static int migration_cpu_stop(void *data);
2198 * The task's runqueue lock must be held.
2199 * Returns true if you have to wait for migration thread.
2201 static bool need_migrate_task(struct task_struct *p)
2204 * If the task is not on a runqueue (and not running), then
2205 * the next wake-up will properly place the task.
2207 bool running = p->on_rq || p->on_cpu;
2208 smp_rmb(); /* finish_lock_switch() */
2213 * wait_task_inactive - wait for a thread to unschedule.
2215 * If @match_state is nonzero, it's the @p->state value just checked and
2216 * not expected to change. If it changes, i.e. @p might have woken up,
2217 * then return zero. When we succeed in waiting for @p to be off its CPU,
2218 * we return a positive number (its total switch count). If a second call
2219 * a short while later returns the same number, the caller can be sure that
2220 * @p has remained unscheduled the whole time.
2222 * The caller must ensure that the task *will* unschedule sometime soon,
2223 * else this function might spin for a *long* time. This function can't
2224 * be called with interrupts off, or it may introduce deadlock with
2225 * smp_call_function() if an IPI is sent by the same process we are
2226 * waiting to become inactive.
2228 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2230 unsigned long flags;
2237 * We do the initial early heuristics without holding
2238 * any task-queue locks at all. We'll only try to get
2239 * the runqueue lock when things look like they will
2245 * If the task is actively running on another CPU
2246 * still, just relax and busy-wait without holding
2249 * NOTE! Since we don't hold any locks, it's not
2250 * even sure that "rq" stays as the right runqueue!
2251 * But we don't care, since "task_running()" will
2252 * return false if the runqueue has changed and p
2253 * is actually now running somewhere else!
2255 while (task_running(rq, p)) {
2256 if (match_state && unlikely(p->state != match_state))
2262 * Ok, time to look more closely! We need the rq
2263 * lock now, to be *sure*. If we're wrong, we'll
2264 * just go back and repeat.
2266 rq = task_rq_lock(p, &flags);
2267 trace_sched_wait_task(p);
2268 running = task_running(rq, p);
2271 if (!match_state || p->state == match_state)
2272 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2273 task_rq_unlock(rq, p, &flags);
2276 * If it changed from the expected state, bail out now.
2278 if (unlikely(!ncsw))
2282 * Was it really running after all now that we
2283 * checked with the proper locks actually held?
2285 * Oops. Go back and try again..
2287 if (unlikely(running)) {
2293 * It's not enough that it's not actively running,
2294 * it must be off the runqueue _entirely_, and not
2297 * So if it was still runnable (but just not actively
2298 * running right now), it's preempted, and we should
2299 * yield - it could be a while.
2301 if (unlikely(on_rq)) {
2302 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
2304 set_current_state(TASK_UNINTERRUPTIBLE);
2305 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
2310 * Ahh, all good. It wasn't running, and it wasn't
2311 * runnable, which means that it will never become
2312 * running in the future either. We're all done!
2321 * kick_process - kick a running thread to enter/exit the kernel
2322 * @p: the to-be-kicked thread
2324 * Cause a process which is running on another CPU to enter
2325 * kernel-mode, without any delay. (to get signals handled.)
2327 * NOTE: this function doesn't have to take the runqueue lock,
2328 * because all it wants to ensure is that the remote task enters
2329 * the kernel. If the IPI races and the task has been migrated
2330 * to another CPU then no harm is done and the purpose has been
2333 void kick_process(struct task_struct *p)
2339 if ((cpu != smp_processor_id()) && task_curr(p))
2340 smp_send_reschedule(cpu);
2343 EXPORT_SYMBOL_GPL(kick_process);
2344 #endif /* CONFIG_SMP */
2348 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
2350 static int select_fallback_rq(int cpu, struct task_struct *p)
2353 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2355 /* Look for allowed, online CPU in same node. */
2356 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2357 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2360 /* Any allowed, online CPU? */
2361 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2362 if (dest_cpu < nr_cpu_ids)
2365 /* No more Mr. Nice Guy. */
2366 dest_cpu = cpuset_cpus_allowed_fallback(p);
2368 * Don't tell them about moving exiting tasks or
2369 * kernel threads (both mm NULL), since they never
2372 if (p->mm && printk_ratelimit()) {
2373 printk(KERN_INFO "process %d (%s) no longer affine to cpu%d\n",
2374 task_pid_nr(p), p->comm, cpu);
2381 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
2384 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2386 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2389 * In order not to call set_task_cpu() on a blocking task we need
2390 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2393 * Since this is common to all placement strategies, this lives here.
2395 * [ this allows ->select_task() to simply return task_cpu(p) and
2396 * not worry about this generic constraint ]
2398 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2400 cpu = select_fallback_rq(task_cpu(p), p);
2405 static void update_avg(u64 *avg, u64 sample)
2407 s64 diff = sample - *avg;
2413 ttwu_stat(struct rq *rq, struct task_struct *p, int cpu, int wake_flags)
2415 #ifdef CONFIG_SCHEDSTATS
2417 int this_cpu = smp_processor_id();
2419 if (cpu == this_cpu) {
2420 schedstat_inc(rq, ttwu_local);
2421 schedstat_inc(p, se.statistics.nr_wakeups_local);
2423 struct sched_domain *sd;
2425 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2426 for_each_domain(this_cpu, sd) {
2427 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2428 schedstat_inc(sd, ttwu_wake_remote);
2433 #endif /* CONFIG_SMP */
2435 schedstat_inc(rq, ttwu_count);
2436 schedstat_inc(p, se.statistics.nr_wakeups);
2438 if (wake_flags & WF_SYNC)
2439 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2441 if (cpu != task_cpu(p))
2442 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2444 #endif /* CONFIG_SCHEDSTATS */
2447 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
2449 activate_task(rq, p, en_flags);
2452 /* if a worker is waking up, notify workqueue */
2453 if (p->flags & PF_WQ_WORKER)
2454 wq_worker_waking_up(p, cpu_of(rq));
2458 ttwu_post_activation(struct task_struct *p, struct rq *rq, int wake_flags)
2460 trace_sched_wakeup(p, true);
2461 check_preempt_curr(rq, p, wake_flags);
2463 p->state = TASK_RUNNING;
2465 if (p->sched_class->task_woken)
2466 p->sched_class->task_woken(rq, p);
2468 if (unlikely(rq->idle_stamp)) {
2469 u64 delta = rq->clock - rq->idle_stamp;
2470 u64 max = 2*sysctl_sched_migration_cost;
2475 update_avg(&rq->avg_idle, delta);
2482 * try_to_wake_up - wake up a thread
2483 * @p: the thread to be awakened
2484 * @state: the mask of task states that can be woken
2485 * @wake_flags: wake modifier flags (WF_*)
2487 * Put it on the run-queue if it's not already there. The "current"
2488 * thread is always on the run-queue (except when the actual
2489 * re-schedule is in progress), and as such you're allowed to do
2490 * the simpler "current->state = TASK_RUNNING" to mark yourself
2491 * runnable without the overhead of this.
2493 * Returns %true if @p was woken up, %false if it was already running
2494 * or @state didn't match @p's state.
2496 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2499 int cpu, orig_cpu, this_cpu, success = 0;
2500 unsigned long flags;
2501 unsigned long en_flags = ENQUEUE_WAKEUP;
2504 this_cpu = get_cpu();
2507 raw_spin_lock_irqsave(&p->pi_lock, flags);
2508 rq = __task_rq_lock(p);
2509 if (!(p->state & state))
2519 if (unlikely(task_running(rq, p)))
2522 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2523 p->state = TASK_WAKING;
2525 if (p->sched_class->task_waking) {
2526 p->sched_class->task_waking(p);
2527 en_flags |= ENQUEUE_WAKING;
2530 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2531 if (cpu != orig_cpu)
2532 set_task_cpu(p, cpu);
2533 __task_rq_unlock(rq);
2536 raw_spin_lock(&rq->lock);
2539 * We migrated the task without holding either rq->lock, however
2540 * since the task is not on the task list itself, nobody else
2541 * will try and migrate the task, hence the rq should match the
2542 * cpu we just moved it to.
2544 WARN_ON(task_cpu(p) != cpu);
2545 WARN_ON(p->state != TASK_WAKING);
2547 if (p->sched_contributes_to_load)
2548 rq->nr_uninterruptible--;
2551 #endif /* CONFIG_SMP */
2552 ttwu_activate(rq, p, en_flags);
2554 ttwu_post_activation(p, rq, wake_flags);
2555 ttwu_stat(rq, p, cpu, wake_flags);
2558 __task_rq_unlock(rq);
2559 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2566 * try_to_wake_up_local - try to wake up a local task with rq lock held
2567 * @p: the thread to be awakened
2569 * Put @p on the run-queue if it's not already there. The caller must
2570 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2573 static void try_to_wake_up_local(struct task_struct *p)
2575 struct rq *rq = task_rq(p);
2577 BUG_ON(rq != this_rq());
2578 BUG_ON(p == current);
2579 lockdep_assert_held(&rq->lock);
2581 if (!raw_spin_trylock(&p->pi_lock)) {
2582 raw_spin_unlock(&rq->lock);
2583 raw_spin_lock(&p->pi_lock);
2584 raw_spin_lock(&rq->lock);
2587 if (!(p->state & TASK_NORMAL))
2591 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2593 ttwu_post_activation(p, rq, 0);
2594 ttwu_stat(rq, p, smp_processor_id(), 0);
2596 raw_spin_unlock(&p->pi_lock);
2600 * wake_up_process - Wake up a specific process
2601 * @p: The process to be woken up.
2603 * Attempt to wake up the nominated process and move it to the set of runnable
2604 * processes. Returns 1 if the process was woken up, 0 if it was already
2607 * It may be assumed that this function implies a write memory barrier before
2608 * changing the task state if and only if any tasks are woken up.
2610 int wake_up_process(struct task_struct *p)
2612 return try_to_wake_up(p, TASK_ALL, 0);
2614 EXPORT_SYMBOL(wake_up_process);
2616 int wake_up_state(struct task_struct *p, unsigned int state)
2618 return try_to_wake_up(p, state, 0);
2622 * Perform scheduler related setup for a newly forked process p.
2623 * p is forked by current.
2625 * __sched_fork() is basic setup used by init_idle() too:
2627 static void __sched_fork(struct task_struct *p)
2632 p->se.exec_start = 0;
2633 p->se.sum_exec_runtime = 0;
2634 p->se.prev_sum_exec_runtime = 0;
2635 p->se.nr_migrations = 0;
2637 INIT_LIST_HEAD(&p->se.group_node);
2639 #ifdef CONFIG_SCHEDSTATS
2640 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2643 INIT_LIST_HEAD(&p->rt.run_list);
2645 #ifdef CONFIG_PREEMPT_NOTIFIERS
2646 INIT_HLIST_HEAD(&p->preempt_notifiers);
2651 * fork()/clone()-time setup:
2653 void sched_fork(struct task_struct *p, int clone_flags)
2655 unsigned long flags;
2656 int cpu = get_cpu();
2660 * We mark the process as running here. This guarantees that
2661 * nobody will actually run it, and a signal or other external
2662 * event cannot wake it up and insert it on the runqueue either.
2664 p->state = TASK_RUNNING;
2667 * Revert to default priority/policy on fork if requested.
2669 if (unlikely(p->sched_reset_on_fork)) {
2670 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2671 p->policy = SCHED_NORMAL;
2672 p->normal_prio = p->static_prio;
2675 if (PRIO_TO_NICE(p->static_prio) < 0) {
2676 p->static_prio = NICE_TO_PRIO(0);
2677 p->normal_prio = p->static_prio;
2682 * We don't need the reset flag anymore after the fork. It has
2683 * fulfilled its duty:
2685 p->sched_reset_on_fork = 0;
2689 * Make sure we do not leak PI boosting priority to the child.
2691 p->prio = current->normal_prio;
2693 if (!rt_prio(p->prio))
2694 p->sched_class = &fair_sched_class;
2696 if (p->sched_class->task_fork)
2697 p->sched_class->task_fork(p);
2700 * The child is not yet in the pid-hash so no cgroup attach races,
2701 * and the cgroup is pinned to this child due to cgroup_fork()
2702 * is ran before sched_fork().
2704 * Silence PROVE_RCU.
2706 raw_spin_lock_irqsave(&p->pi_lock, flags);
2707 set_task_cpu(p, cpu);
2708 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2710 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2711 if (likely(sched_info_on()))
2712 memset(&p->sched_info, 0, sizeof(p->sched_info));
2714 #if defined(CONFIG_SMP)
2717 #ifdef CONFIG_PREEMPT
2718 /* Want to start with kernel preemption disabled. */
2719 task_thread_info(p)->preempt_count = 1;
2722 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2729 * wake_up_new_task - wake up a newly created task for the first time.
2731 * This function will do some initial scheduler statistics housekeeping
2732 * that must be done for every newly created context, then puts the task
2733 * on the runqueue and wakes it.
2735 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2737 unsigned long flags;
2740 raw_spin_lock_irqsave(&p->pi_lock, flags);
2743 * Fork balancing, do it here and not earlier because:
2744 * - cpus_allowed can change in the fork path
2745 * - any previously selected cpu might disappear through hotplug
2747 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
2750 rq = __task_rq_lock(p);
2751 activate_task(rq, p, 0);
2753 trace_sched_wakeup_new(p, true);
2754 check_preempt_curr(rq, p, WF_FORK);
2756 if (p->sched_class->task_woken)
2757 p->sched_class->task_woken(rq, p);
2759 task_rq_unlock(rq, p, &flags);
2762 #ifdef CONFIG_PREEMPT_NOTIFIERS
2765 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2766 * @notifier: notifier struct to register
2768 void preempt_notifier_register(struct preempt_notifier *notifier)
2770 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2772 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2775 * preempt_notifier_unregister - no longer interested in preemption notifications
2776 * @notifier: notifier struct to unregister
2778 * This is safe to call from within a preemption notifier.
2780 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2782 hlist_del(¬ifier->link);
2784 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2786 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2788 struct preempt_notifier *notifier;
2789 struct hlist_node *node;
2791 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2792 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2796 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2797 struct task_struct *next)
2799 struct preempt_notifier *notifier;
2800 struct hlist_node *node;
2802 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2803 notifier->ops->sched_out(notifier, next);
2806 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2808 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2813 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2814 struct task_struct *next)
2818 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2821 * prepare_task_switch - prepare to switch tasks
2822 * @rq: the runqueue preparing to switch
2823 * @prev: the current task that is being switched out
2824 * @next: the task we are going to switch to.
2826 * This is called with the rq lock held and interrupts off. It must
2827 * be paired with a subsequent finish_task_switch after the context
2830 * prepare_task_switch sets up locking and calls architecture specific
2834 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2835 struct task_struct *next)
2837 sched_info_switch(prev, next);
2838 perf_event_task_sched_out(prev, next);
2839 fire_sched_out_preempt_notifiers(prev, next);
2840 prepare_lock_switch(rq, next);
2841 prepare_arch_switch(next);
2842 trace_sched_switch(prev, next);
2846 * finish_task_switch - clean up after a task-switch
2847 * @rq: runqueue associated with task-switch
2848 * @prev: the thread we just switched away from.
2850 * finish_task_switch must be called after the context switch, paired
2851 * with a prepare_task_switch call before the context switch.
2852 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2853 * and do any other architecture-specific cleanup actions.
2855 * Note that we may have delayed dropping an mm in context_switch(). If
2856 * so, we finish that here outside of the runqueue lock. (Doing it
2857 * with the lock held can cause deadlocks; see schedule() for
2860 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2861 __releases(rq->lock)
2863 struct mm_struct *mm = rq->prev_mm;
2869 * A task struct has one reference for the use as "current".
2870 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2871 * schedule one last time. The schedule call will never return, and
2872 * the scheduled task must drop that reference.
2873 * The test for TASK_DEAD must occur while the runqueue locks are
2874 * still held, otherwise prev could be scheduled on another cpu, die
2875 * there before we look at prev->state, and then the reference would
2877 * Manfred Spraul <manfred@colorfullife.com>
2879 prev_state = prev->state;
2880 finish_arch_switch(prev);
2881 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2882 local_irq_disable();
2883 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2884 perf_event_task_sched_in(current);
2885 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2887 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2888 finish_lock_switch(rq, prev);
2890 fire_sched_in_preempt_notifiers(current);
2893 if (unlikely(prev_state == TASK_DEAD)) {
2895 * Remove function-return probe instances associated with this
2896 * task and put them back on the free list.
2898 kprobe_flush_task(prev);
2899 put_task_struct(prev);
2905 /* assumes rq->lock is held */
2906 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2908 if (prev->sched_class->pre_schedule)
2909 prev->sched_class->pre_schedule(rq, prev);
2912 /* rq->lock is NOT held, but preemption is disabled */
2913 static inline void post_schedule(struct rq *rq)
2915 if (rq->post_schedule) {
2916 unsigned long flags;
2918 raw_spin_lock_irqsave(&rq->lock, flags);
2919 if (rq->curr->sched_class->post_schedule)
2920 rq->curr->sched_class->post_schedule(rq);
2921 raw_spin_unlock_irqrestore(&rq->lock, flags);
2923 rq->post_schedule = 0;
2929 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2933 static inline void post_schedule(struct rq *rq)
2940 * schedule_tail - first thing a freshly forked thread must call.
2941 * @prev: the thread we just switched away from.
2943 asmlinkage void schedule_tail(struct task_struct *prev)
2944 __releases(rq->lock)
2946 struct rq *rq = this_rq();
2948 finish_task_switch(rq, prev);
2951 * FIXME: do we need to worry about rq being invalidated by the
2956 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2957 /* In this case, finish_task_switch does not reenable preemption */
2960 if (current->set_child_tid)
2961 put_user(task_pid_vnr(current), current->set_child_tid);
2965 * context_switch - switch to the new MM and the new
2966 * thread's register state.
2969 context_switch(struct rq *rq, struct task_struct *prev,
2970 struct task_struct *next)
2972 struct mm_struct *mm, *oldmm;
2974 prepare_task_switch(rq, prev, next);
2977 oldmm = prev->active_mm;
2979 * For paravirt, this is coupled with an exit in switch_to to
2980 * combine the page table reload and the switch backend into
2983 arch_start_context_switch(prev);
2986 next->active_mm = oldmm;
2987 atomic_inc(&oldmm->mm_count);
2988 enter_lazy_tlb(oldmm, next);
2990 switch_mm(oldmm, mm, next);
2993 prev->active_mm = NULL;
2994 rq->prev_mm = oldmm;
2997 * Since the runqueue lock will be released by the next
2998 * task (which is an invalid locking op but in the case
2999 * of the scheduler it's an obvious special-case), so we
3000 * do an early lockdep release here:
3002 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
3003 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3006 /* Here we just switch the register state and the stack. */
3007 switch_to(prev, next, prev);
3011 * this_rq must be evaluated again because prev may have moved
3012 * CPUs since it called schedule(), thus the 'rq' on its stack
3013 * frame will be invalid.
3015 finish_task_switch(this_rq(), prev);
3019 * nr_running, nr_uninterruptible and nr_context_switches:
3021 * externally visible scheduler statistics: current number of runnable
3022 * threads, current number of uninterruptible-sleeping threads, total
3023 * number of context switches performed since bootup.
3025 unsigned long nr_running(void)
3027 unsigned long i, sum = 0;
3029 for_each_online_cpu(i)
3030 sum += cpu_rq(i)->nr_running;
3035 unsigned long nr_uninterruptible(void)
3037 unsigned long i, sum = 0;
3039 for_each_possible_cpu(i)
3040 sum += cpu_rq(i)->nr_uninterruptible;
3043 * Since we read the counters lockless, it might be slightly
3044 * inaccurate. Do not allow it to go below zero though:
3046 if (unlikely((long)sum < 0))
3052 unsigned long long nr_context_switches(void)
3055 unsigned long long sum = 0;
3057 for_each_possible_cpu(i)
3058 sum += cpu_rq(i)->nr_switches;
3063 unsigned long nr_iowait(void)
3065 unsigned long i, sum = 0;
3067 for_each_possible_cpu(i)
3068 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3073 unsigned long nr_iowait_cpu(int cpu)
3075 struct rq *this = cpu_rq(cpu);
3076 return atomic_read(&this->nr_iowait);
3079 unsigned long this_cpu_load(void)
3081 struct rq *this = this_rq();
3082 return this->cpu_load[0];
3086 /* Variables and functions for calc_load */
3087 static atomic_long_t calc_load_tasks;
3088 static unsigned long calc_load_update;
3089 unsigned long avenrun[3];
3090 EXPORT_SYMBOL(avenrun);
3092 static long calc_load_fold_active(struct rq *this_rq)
3094 long nr_active, delta = 0;
3096 nr_active = this_rq->nr_running;
3097 nr_active += (long) this_rq->nr_uninterruptible;
3099 if (nr_active != this_rq->calc_load_active) {
3100 delta = nr_active - this_rq->calc_load_active;
3101 this_rq->calc_load_active = nr_active;
3107 static unsigned long
3108 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3111 load += active * (FIXED_1 - exp);
3112 load += 1UL << (FSHIFT - 1);
3113 return load >> FSHIFT;
3118 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
3120 * When making the ILB scale, we should try to pull this in as well.
3122 static atomic_long_t calc_load_tasks_idle;
3124 static void calc_load_account_idle(struct rq *this_rq)
3128 delta = calc_load_fold_active(this_rq);
3130 atomic_long_add(delta, &calc_load_tasks_idle);
3133 static long calc_load_fold_idle(void)
3138 * Its got a race, we don't care...
3140 if (atomic_long_read(&calc_load_tasks_idle))
3141 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
3147 * fixed_power_int - compute: x^n, in O(log n) time
3149 * @x: base of the power
3150 * @frac_bits: fractional bits of @x
3151 * @n: power to raise @x to.
3153 * By exploiting the relation between the definition of the natural power
3154 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
3155 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
3156 * (where: n_i \elem {0, 1}, the binary vector representing n),
3157 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
3158 * of course trivially computable in O(log_2 n), the length of our binary
3161 static unsigned long
3162 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
3164 unsigned long result = 1UL << frac_bits;
3169 result += 1UL << (frac_bits - 1);
3170 result >>= frac_bits;
3176 x += 1UL << (frac_bits - 1);
3184 * a1 = a0 * e + a * (1 - e)
3186 * a2 = a1 * e + a * (1 - e)
3187 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
3188 * = a0 * e^2 + a * (1 - e) * (1 + e)
3190 * a3 = a2 * e + a * (1 - e)
3191 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
3192 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
3196 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
3197 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
3198 * = a0 * e^n + a * (1 - e^n)
3200 * [1] application of the geometric series:
3203 * S_n := \Sum x^i = -------------
3206 static unsigned long
3207 calc_load_n(unsigned long load, unsigned long exp,
3208 unsigned long active, unsigned int n)
3211 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
3215 * NO_HZ can leave us missing all per-cpu ticks calling
3216 * calc_load_account_active(), but since an idle CPU folds its delta into
3217 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
3218 * in the pending idle delta if our idle period crossed a load cycle boundary.
3220 * Once we've updated the global active value, we need to apply the exponential
3221 * weights adjusted to the number of cycles missed.
3223 static void calc_global_nohz(unsigned long ticks)
3225 long delta, active, n;
3227 if (time_before(jiffies, calc_load_update))
3231 * If we crossed a calc_load_update boundary, make sure to fold
3232 * any pending idle changes, the respective CPUs might have
3233 * missed the tick driven calc_load_account_active() update
3236 delta = calc_load_fold_idle();
3238 atomic_long_add(delta, &calc_load_tasks);
3241 * If we were idle for multiple load cycles, apply them.
3243 if (ticks >= LOAD_FREQ) {
3244 n = ticks / LOAD_FREQ;
3246 active = atomic_long_read(&calc_load_tasks);
3247 active = active > 0 ? active * FIXED_1 : 0;
3249 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
3250 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
3251 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
3253 calc_load_update += n * LOAD_FREQ;
3257 * Its possible the remainder of the above division also crosses
3258 * a LOAD_FREQ period, the regular check in calc_global_load()
3259 * which comes after this will take care of that.
3261 * Consider us being 11 ticks before a cycle completion, and us
3262 * sleeping for 4*LOAD_FREQ + 22 ticks, then the above code will
3263 * age us 4 cycles, and the test in calc_global_load() will
3264 * pick up the final one.
3268 static void calc_load_account_idle(struct rq *this_rq)
3272 static inline long calc_load_fold_idle(void)
3277 static void calc_global_nohz(unsigned long ticks)
3283 * get_avenrun - get the load average array
3284 * @loads: pointer to dest load array
3285 * @offset: offset to add
3286 * @shift: shift count to shift the result left
3288 * These values are estimates at best, so no need for locking.
3290 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3292 loads[0] = (avenrun[0] + offset) << shift;
3293 loads[1] = (avenrun[1] + offset) << shift;
3294 loads[2] = (avenrun[2] + offset) << shift;
3298 * calc_load - update the avenrun load estimates 10 ticks after the
3299 * CPUs have updated calc_load_tasks.
3301 void calc_global_load(unsigned long ticks)
3305 calc_global_nohz(ticks);
3307 if (time_before(jiffies, calc_load_update + 10))
3310 active = atomic_long_read(&calc_load_tasks);
3311 active = active > 0 ? active * FIXED_1 : 0;
3313 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3314 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3315 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3317 calc_load_update += LOAD_FREQ;
3321 * Called from update_cpu_load() to periodically update this CPU's
3324 static void calc_load_account_active(struct rq *this_rq)
3328 if (time_before(jiffies, this_rq->calc_load_update))
3331 delta = calc_load_fold_active(this_rq);
3332 delta += calc_load_fold_idle();
3334 atomic_long_add(delta, &calc_load_tasks);
3336 this_rq->calc_load_update += LOAD_FREQ;
3340 * The exact cpuload at various idx values, calculated at every tick would be
3341 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3343 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3344 * on nth tick when cpu may be busy, then we have:
3345 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3346 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3348 * decay_load_missed() below does efficient calculation of
3349 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3350 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3352 * The calculation is approximated on a 128 point scale.
3353 * degrade_zero_ticks is the number of ticks after which load at any
3354 * particular idx is approximated to be zero.
3355 * degrade_factor is a precomputed table, a row for each load idx.
3356 * Each column corresponds to degradation factor for a power of two ticks,
3357 * based on 128 point scale.
3359 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3360 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3362 * With this power of 2 load factors, we can degrade the load n times
3363 * by looking at 1 bits in n and doing as many mult/shift instead of
3364 * n mult/shifts needed by the exact degradation.
3366 #define DEGRADE_SHIFT 7
3367 static const unsigned char
3368 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3369 static const unsigned char
3370 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3371 {0, 0, 0, 0, 0, 0, 0, 0},
3372 {64, 32, 8, 0, 0, 0, 0, 0},
3373 {96, 72, 40, 12, 1, 0, 0},
3374 {112, 98, 75, 43, 15, 1, 0},
3375 {120, 112, 98, 76, 45, 16, 2} };
3378 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3379 * would be when CPU is idle and so we just decay the old load without
3380 * adding any new load.
3382 static unsigned long
3383 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3387 if (!missed_updates)
3390 if (missed_updates >= degrade_zero_ticks[idx])
3394 return load >> missed_updates;
3396 while (missed_updates) {
3397 if (missed_updates % 2)
3398 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3400 missed_updates >>= 1;
3407 * Update rq->cpu_load[] statistics. This function is usually called every
3408 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3409 * every tick. We fix it up based on jiffies.
3411 static void update_cpu_load(struct rq *this_rq)
3413 unsigned long this_load = this_rq->load.weight;
3414 unsigned long curr_jiffies = jiffies;
3415 unsigned long pending_updates;
3418 this_rq->nr_load_updates++;
3420 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3421 if (curr_jiffies == this_rq->last_load_update_tick)
3424 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3425 this_rq->last_load_update_tick = curr_jiffies;
3427 /* Update our load: */
3428 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3429 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3430 unsigned long old_load, new_load;
3432 /* scale is effectively 1 << i now, and >> i divides by scale */
3434 old_load = this_rq->cpu_load[i];
3435 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3436 new_load = this_load;
3438 * Round up the averaging division if load is increasing. This
3439 * prevents us from getting stuck on 9 if the load is 10, for
3442 if (new_load > old_load)
3443 new_load += scale - 1;
3445 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3448 sched_avg_update(this_rq);
3451 static void update_cpu_load_active(struct rq *this_rq)
3453 update_cpu_load(this_rq);
3455 calc_load_account_active(this_rq);
3461 * sched_exec - execve() is a valuable balancing opportunity, because at
3462 * this point the task has the smallest effective memory and cache footprint.
3464 void sched_exec(void)
3466 struct task_struct *p = current;
3467 unsigned long flags;
3471 rq = task_rq_lock(p, &flags);
3472 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
3473 if (dest_cpu == smp_processor_id())
3477 * select_task_rq() can race against ->cpus_allowed
3479 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed) &&
3480 likely(cpu_active(dest_cpu)) && need_migrate_task(p)) {
3481 struct migration_arg arg = { p, dest_cpu };
3483 task_rq_unlock(rq, p, &flags);
3484 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
3488 task_rq_unlock(rq, p, &flags);
3493 DEFINE_PER_CPU(struct kernel_stat, kstat);
3495 EXPORT_PER_CPU_SYMBOL(kstat);
3498 * Return any ns on the sched_clock that have not yet been accounted in
3499 * @p in case that task is currently running.
3501 * Called with task_rq_lock() held on @rq.
3503 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3507 if (task_current(rq, p)) {
3508 update_rq_clock(rq);
3509 ns = rq->clock_task - p->se.exec_start;
3517 unsigned long long task_delta_exec(struct task_struct *p)
3519 unsigned long flags;
3523 rq = task_rq_lock(p, &flags);
3524 ns = do_task_delta_exec(p, rq);
3525 task_rq_unlock(rq, p, &flags);
3531 * Return accounted runtime for the task.
3532 * In case the task is currently running, return the runtime plus current's
3533 * pending runtime that have not been accounted yet.
3535 unsigned long long task_sched_runtime(struct task_struct *p)
3537 unsigned long flags;
3541 rq = task_rq_lock(p, &flags);
3542 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3543 task_rq_unlock(rq, p, &flags);
3549 * Return sum_exec_runtime for the thread group.
3550 * In case the task is currently running, return the sum plus current's
3551 * pending runtime that have not been accounted yet.
3553 * Note that the thread group might have other running tasks as well,
3554 * so the return value not includes other pending runtime that other
3555 * running tasks might have.
3557 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3559 struct task_cputime totals;
3560 unsigned long flags;
3564 rq = task_rq_lock(p, &flags);
3565 thread_group_cputime(p, &totals);
3566 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3567 task_rq_unlock(rq, p, &flags);
3573 * Account user cpu time to a process.
3574 * @p: the process that the cpu time gets accounted to
3575 * @cputime: the cpu time spent in user space since the last update
3576 * @cputime_scaled: cputime scaled by cpu frequency
3578 void account_user_time(struct task_struct *p, cputime_t cputime,
3579 cputime_t cputime_scaled)
3581 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3584 /* Add user time to process. */
3585 p->utime = cputime_add(p->utime, cputime);
3586 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3587 account_group_user_time(p, cputime);
3589 /* Add user time to cpustat. */
3590 tmp = cputime_to_cputime64(cputime);
3591 if (TASK_NICE(p) > 0)
3592 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3594 cpustat->user = cputime64_add(cpustat->user, tmp);
3596 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3597 /* Account for user time used */
3598 acct_update_integrals(p);
3602 * Account guest cpu time to a process.
3603 * @p: the process that the cpu time gets accounted to
3604 * @cputime: the cpu time spent in virtual machine since the last update
3605 * @cputime_scaled: cputime scaled by cpu frequency
3607 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3608 cputime_t cputime_scaled)
3611 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3613 tmp = cputime_to_cputime64(cputime);
3615 /* Add guest time to process. */
3616 p->utime = cputime_add(p->utime, cputime);
3617 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3618 account_group_user_time(p, cputime);
3619 p->gtime = cputime_add(p->gtime, cputime);
3621 /* Add guest time to cpustat. */
3622 if (TASK_NICE(p) > 0) {
3623 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3624 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3626 cpustat->user = cputime64_add(cpustat->user, tmp);
3627 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3632 * Account system cpu time to a process and desired cpustat field
3633 * @p: the process that the cpu time gets accounted to
3634 * @cputime: the cpu time spent in kernel space since the last update
3635 * @cputime_scaled: cputime scaled by cpu frequency
3636 * @target_cputime64: pointer to cpustat field that has to be updated
3639 void __account_system_time(struct task_struct *p, cputime_t cputime,
3640 cputime_t cputime_scaled, cputime64_t *target_cputime64)
3642 cputime64_t tmp = cputime_to_cputime64(cputime);
3644 /* Add system time to process. */
3645 p->stime = cputime_add(p->stime, cputime);
3646 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3647 account_group_system_time(p, cputime);
3649 /* Add system time to cpustat. */
3650 *target_cputime64 = cputime64_add(*target_cputime64, tmp);
3651 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3653 /* Account for system time used */
3654 acct_update_integrals(p);
3658 * Account system cpu time to a process.
3659 * @p: the process that the cpu time gets accounted to
3660 * @hardirq_offset: the offset to subtract from hardirq_count()
3661 * @cputime: the cpu time spent in kernel space since the last update
3662 * @cputime_scaled: cputime scaled by cpu frequency
3664 void account_system_time(struct task_struct *p, int hardirq_offset,
3665 cputime_t cputime, cputime_t cputime_scaled)
3667 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3668 cputime64_t *target_cputime64;
3670 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3671 account_guest_time(p, cputime, cputime_scaled);
3675 if (hardirq_count() - hardirq_offset)
3676 target_cputime64 = &cpustat->irq;
3677 else if (in_serving_softirq())
3678 target_cputime64 = &cpustat->softirq;
3680 target_cputime64 = &cpustat->system;
3682 __account_system_time(p, cputime, cputime_scaled, target_cputime64);
3686 * Account for involuntary wait time.
3687 * @cputime: the cpu time spent in involuntary wait
3689 void account_steal_time(cputime_t cputime)
3691 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3692 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3694 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3698 * Account for idle time.
3699 * @cputime: the cpu time spent in idle wait
3701 void account_idle_time(cputime_t cputime)
3703 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3704 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3705 struct rq *rq = this_rq();
3707 if (atomic_read(&rq->nr_iowait) > 0)
3708 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3710 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3713 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3715 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
3717 * Account a tick to a process and cpustat
3718 * @p: the process that the cpu time gets accounted to
3719 * @user_tick: is the tick from userspace
3720 * @rq: the pointer to rq
3722 * Tick demultiplexing follows the order
3723 * - pending hardirq update
3724 * - pending softirq update
3728 * - check for guest_time
3729 * - else account as system_time
3731 * Check for hardirq is done both for system and user time as there is
3732 * no timer going off while we are on hardirq and hence we may never get an
3733 * opportunity to update it solely in system time.
3734 * p->stime and friends are only updated on system time and not on irq
3735 * softirq as those do not count in task exec_runtime any more.
3737 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
3740 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3741 cputime64_t tmp = cputime_to_cputime64(cputime_one_jiffy);
3742 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3744 if (irqtime_account_hi_update()) {
3745 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3746 } else if (irqtime_account_si_update()) {
3747 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3748 } else if (this_cpu_ksoftirqd() == p) {
3750 * ksoftirqd time do not get accounted in cpu_softirq_time.
3751 * So, we have to handle it separately here.
3752 * Also, p->stime needs to be updated for ksoftirqd.
3754 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
3756 } else if (user_tick) {
3757 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3758 } else if (p == rq->idle) {
3759 account_idle_time(cputime_one_jiffy);
3760 } else if (p->flags & PF_VCPU) { /* System time or guest time */
3761 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled);
3763 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
3768 static void irqtime_account_idle_ticks(int ticks)
3771 struct rq *rq = this_rq();
3773 for (i = 0; i < ticks; i++)
3774 irqtime_account_process_tick(current, 0, rq);
3776 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
3777 static void irqtime_account_idle_ticks(int ticks) {}
3778 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
3780 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
3783 * Account a single tick of cpu time.
3784 * @p: the process that the cpu time gets accounted to
3785 * @user_tick: indicates if the tick is a user or a system tick
3787 void account_process_tick(struct task_struct *p, int user_tick)
3789 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3790 struct rq *rq = this_rq();
3792 if (sched_clock_irqtime) {
3793 irqtime_account_process_tick(p, user_tick, rq);
3798 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3799 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3800 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3803 account_idle_time(cputime_one_jiffy);
3807 * Account multiple ticks of steal time.
3808 * @p: the process from which the cpu time has been stolen
3809 * @ticks: number of stolen ticks
3811 void account_steal_ticks(unsigned long ticks)
3813 account_steal_time(jiffies_to_cputime(ticks));
3817 * Account multiple ticks of idle time.
3818 * @ticks: number of stolen ticks
3820 void account_idle_ticks(unsigned long ticks)
3823 if (sched_clock_irqtime) {
3824 irqtime_account_idle_ticks(ticks);
3828 account_idle_time(jiffies_to_cputime(ticks));
3834 * Use precise platform statistics if available:
3836 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3837 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3843 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3845 struct task_cputime cputime;
3847 thread_group_cputime(p, &cputime);
3849 *ut = cputime.utime;
3850 *st = cputime.stime;
3854 #ifndef nsecs_to_cputime
3855 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3858 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3860 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
3863 * Use CFS's precise accounting:
3865 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3871 do_div(temp, total);
3872 utime = (cputime_t)temp;
3877 * Compare with previous values, to keep monotonicity:
3879 p->prev_utime = max(p->prev_utime, utime);
3880 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
3882 *ut = p->prev_utime;
3883 *st = p->prev_stime;
3887 * Must be called with siglock held.
3889 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3891 struct signal_struct *sig = p->signal;
3892 struct task_cputime cputime;
3893 cputime_t rtime, utime, total;
3895 thread_group_cputime(p, &cputime);
3897 total = cputime_add(cputime.utime, cputime.stime);
3898 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3903 temp *= cputime.utime;
3904 do_div(temp, total);
3905 utime = (cputime_t)temp;
3909 sig->prev_utime = max(sig->prev_utime, utime);
3910 sig->prev_stime = max(sig->prev_stime,
3911 cputime_sub(rtime, sig->prev_utime));
3913 *ut = sig->prev_utime;
3914 *st = sig->prev_stime;
3919 * This function gets called by the timer code, with HZ frequency.
3920 * We call it with interrupts disabled.
3922 * It also gets called by the fork code, when changing the parent's
3925 void scheduler_tick(void)
3927 int cpu = smp_processor_id();
3928 struct rq *rq = cpu_rq(cpu);
3929 struct task_struct *curr = rq->curr;
3933 raw_spin_lock(&rq->lock);
3934 update_rq_clock(rq);
3935 update_cpu_load_active(rq);
3936 curr->sched_class->task_tick(rq, curr, 0);
3937 raw_spin_unlock(&rq->lock);
3939 perf_event_task_tick();
3942 rq->idle_at_tick = idle_cpu(cpu);
3943 trigger_load_balance(rq, cpu);
3947 notrace unsigned long get_parent_ip(unsigned long addr)
3949 if (in_lock_functions(addr)) {
3950 addr = CALLER_ADDR2;
3951 if (in_lock_functions(addr))
3952 addr = CALLER_ADDR3;
3957 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3958 defined(CONFIG_PREEMPT_TRACER))
3960 void __kprobes add_preempt_count(int val)
3962 #ifdef CONFIG_DEBUG_PREEMPT
3966 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3969 preempt_count() += val;
3970 #ifdef CONFIG_DEBUG_PREEMPT
3972 * Spinlock count overflowing soon?
3974 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3977 if (preempt_count() == val)
3978 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3980 EXPORT_SYMBOL(add_preempt_count);
3982 void __kprobes sub_preempt_count(int val)
3984 #ifdef CONFIG_DEBUG_PREEMPT
3988 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3991 * Is the spinlock portion underflowing?
3993 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3994 !(preempt_count() & PREEMPT_MASK)))
3998 if (preempt_count() == val)
3999 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4000 preempt_count() -= val;
4002 EXPORT_SYMBOL(sub_preempt_count);
4007 * Print scheduling while atomic bug:
4009 static noinline void __schedule_bug(struct task_struct *prev)
4011 struct pt_regs *regs = get_irq_regs();
4013 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4014 prev->comm, prev->pid, preempt_count());
4016 debug_show_held_locks(prev);
4018 if (irqs_disabled())
4019 print_irqtrace_events(prev);
4028 * Various schedule()-time debugging checks and statistics:
4030 static inline void schedule_debug(struct task_struct *prev)
4033 * Test if we are atomic. Since do_exit() needs to call into
4034 * schedule() atomically, we ignore that path for now.
4035 * Otherwise, whine if we are scheduling when we should not be.
4037 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4038 __schedule_bug(prev);
4040 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4042 schedstat_inc(this_rq(), sched_count);
4043 #ifdef CONFIG_SCHEDSTATS
4044 if (unlikely(prev->lock_depth >= 0)) {
4045 schedstat_inc(this_rq(), rq_sched_info.bkl_count);
4046 schedstat_inc(prev, sched_info.bkl_count);
4051 static void put_prev_task(struct rq *rq, struct task_struct *prev)
4054 update_rq_clock(rq);
4055 prev->sched_class->put_prev_task(rq, prev);
4059 * Pick up the highest-prio task:
4061 static inline struct task_struct *
4062 pick_next_task(struct rq *rq)
4064 const struct sched_class *class;
4065 struct task_struct *p;
4068 * Optimization: we know that if all tasks are in
4069 * the fair class we can call that function directly:
4071 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4072 p = fair_sched_class.pick_next_task(rq);
4077 for_each_class(class) {
4078 p = class->pick_next_task(rq);
4083 BUG(); /* the idle class will always have a runnable task */
4087 * schedule() is the main scheduler function.
4089 asmlinkage void __sched schedule(void)
4091 struct task_struct *prev, *next;
4092 unsigned long *switch_count;
4098 cpu = smp_processor_id();
4100 rcu_note_context_switch(cpu);
4103 schedule_debug(prev);
4105 if (sched_feat(HRTICK))
4108 raw_spin_lock_irq(&rq->lock);
4110 switch_count = &prev->nivcsw;
4111 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4112 if (unlikely(signal_pending_state(prev->state, prev))) {
4113 prev->state = TASK_RUNNING;
4115 deactivate_task(rq, prev, DEQUEUE_SLEEP);
4119 * If a worker went to sleep, notify and ask workqueue
4120 * whether it wants to wake up a task to maintain
4123 if (prev->flags & PF_WQ_WORKER) {
4124 struct task_struct *to_wakeup;
4126 to_wakeup = wq_worker_sleeping(prev, cpu);
4128 try_to_wake_up_local(to_wakeup);
4132 * If we are going to sleep and we have plugged IO
4133 * queued, make sure to submit it to avoid deadlocks.
4135 if (blk_needs_flush_plug(prev)) {
4136 raw_spin_unlock(&rq->lock);
4137 blk_flush_plug(prev);
4138 raw_spin_lock(&rq->lock);
4141 switch_count = &prev->nvcsw;
4144 pre_schedule(rq, prev);
4146 if (unlikely(!rq->nr_running))
4147 idle_balance(cpu, rq);
4149 put_prev_task(rq, prev);
4150 next = pick_next_task(rq);
4151 clear_tsk_need_resched(prev);
4152 rq->skip_clock_update = 0;
4154 if (likely(prev != next)) {
4159 context_switch(rq, prev, next); /* unlocks the rq */
4161 * The context switch have flipped the stack from under us
4162 * and restored the local variables which were saved when
4163 * this task called schedule() in the past. prev == current
4164 * is still correct, but it can be moved to another cpu/rq.
4166 cpu = smp_processor_id();
4169 raw_spin_unlock_irq(&rq->lock);
4173 preempt_enable_no_resched();
4177 EXPORT_SYMBOL(schedule);
4179 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
4181 static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
4186 if (lock->owner != owner)
4190 * Ensure we emit the owner->on_cpu, dereference _after_ checking
4191 * lock->owner still matches owner, if that fails, owner might
4192 * point to free()d memory, if it still matches, the rcu_read_lock()
4193 * ensures the memory stays valid.
4197 ret = owner->on_cpu;
4205 * Look out! "owner" is an entirely speculative pointer
4206 * access and not reliable.
4208 int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
4210 if (!sched_feat(OWNER_SPIN))
4213 while (owner_running(lock, owner)) {
4217 arch_mutex_cpu_relax();
4221 * If the owner changed to another task there is likely
4222 * heavy contention, stop spinning.
4231 #ifdef CONFIG_PREEMPT
4233 * this is the entry point to schedule() from in-kernel preemption
4234 * off of preempt_enable. Kernel preemptions off return from interrupt
4235 * occur there and call schedule directly.
4237 asmlinkage void __sched notrace preempt_schedule(void)
4239 struct thread_info *ti = current_thread_info();
4242 * If there is a non-zero preempt_count or interrupts are disabled,
4243 * we do not want to preempt the current task. Just return..
4245 if (likely(ti->preempt_count || irqs_disabled()))
4249 add_preempt_count_notrace(PREEMPT_ACTIVE);
4251 sub_preempt_count_notrace(PREEMPT_ACTIVE);
4254 * Check again in case we missed a preemption opportunity
4255 * between schedule and now.
4258 } while (need_resched());
4260 EXPORT_SYMBOL(preempt_schedule);
4263 * this is the entry point to schedule() from kernel preemption
4264 * off of irq context.
4265 * Note, that this is called and return with irqs disabled. This will
4266 * protect us against recursive calling from irq.
4268 asmlinkage void __sched preempt_schedule_irq(void)
4270 struct thread_info *ti = current_thread_info();
4272 /* Catch callers which need to be fixed */
4273 BUG_ON(ti->preempt_count || !irqs_disabled());
4276 add_preempt_count(PREEMPT_ACTIVE);
4279 local_irq_disable();
4280 sub_preempt_count(PREEMPT_ACTIVE);
4283 * Check again in case we missed a preemption opportunity
4284 * between schedule and now.
4287 } while (need_resched());
4290 #endif /* CONFIG_PREEMPT */
4292 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
4295 return try_to_wake_up(curr->private, mode, wake_flags);
4297 EXPORT_SYMBOL(default_wake_function);
4300 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4301 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4302 * number) then we wake all the non-exclusive tasks and one exclusive task.
4304 * There are circumstances in which we can try to wake a task which has already
4305 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4306 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4308 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4309 int nr_exclusive, int wake_flags, void *key)
4311 wait_queue_t *curr, *next;
4313 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4314 unsigned flags = curr->flags;
4316 if (curr->func(curr, mode, wake_flags, key) &&
4317 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4323 * __wake_up - wake up threads blocked on a waitqueue.
4325 * @mode: which threads
4326 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4327 * @key: is directly passed to the wakeup function
4329 * It may be assumed that this function implies a write memory barrier before
4330 * changing the task state if and only if any tasks are woken up.
4332 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4333 int nr_exclusive, void *key)
4335 unsigned long flags;
4337 spin_lock_irqsave(&q->lock, flags);
4338 __wake_up_common(q, mode, nr_exclusive, 0, key);
4339 spin_unlock_irqrestore(&q->lock, flags);
4341 EXPORT_SYMBOL(__wake_up);
4344 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4346 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4348 __wake_up_common(q, mode, 1, 0, NULL);
4350 EXPORT_SYMBOL_GPL(__wake_up_locked);
4352 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
4354 __wake_up_common(q, mode, 1, 0, key);
4356 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
4359 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4361 * @mode: which threads
4362 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4363 * @key: opaque value to be passed to wakeup targets
4365 * The sync wakeup differs that the waker knows that it will schedule
4366 * away soon, so while the target thread will be woken up, it will not
4367 * be migrated to another CPU - ie. the two threads are 'synchronized'
4368 * with each other. This can prevent needless bouncing between CPUs.
4370 * On UP it can prevent extra preemption.
4372 * It may be assumed that this function implies a write memory barrier before
4373 * changing the task state if and only if any tasks are woken up.
4375 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4376 int nr_exclusive, void *key)
4378 unsigned long flags;
4379 int wake_flags = WF_SYNC;
4384 if (unlikely(!nr_exclusive))
4387 spin_lock_irqsave(&q->lock, flags);
4388 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4389 spin_unlock_irqrestore(&q->lock, flags);
4391 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4394 * __wake_up_sync - see __wake_up_sync_key()
4396 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4398 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4400 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4403 * complete: - signals a single thread waiting on this completion
4404 * @x: holds the state of this particular completion
4406 * This will wake up a single thread waiting on this completion. Threads will be
4407 * awakened in the same order in which they were queued.
4409 * See also complete_all(), wait_for_completion() and related routines.
4411 * It may be assumed that this function implies a write memory barrier before
4412 * changing the task state if and only if any tasks are woken up.
4414 void complete(struct completion *x)
4416 unsigned long flags;
4418 spin_lock_irqsave(&x->wait.lock, flags);
4420 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4421 spin_unlock_irqrestore(&x->wait.lock, flags);
4423 EXPORT_SYMBOL(complete);
4426 * complete_all: - signals all threads waiting on this completion
4427 * @x: holds the state of this particular completion
4429 * This will wake up all threads waiting on this particular completion event.
4431 * It may be assumed that this function implies a write memory barrier before
4432 * changing the task state if and only if any tasks are woken up.
4434 void complete_all(struct completion *x)
4436 unsigned long flags;
4438 spin_lock_irqsave(&x->wait.lock, flags);
4439 x->done += UINT_MAX/2;
4440 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4441 spin_unlock_irqrestore(&x->wait.lock, flags);
4443 EXPORT_SYMBOL(complete_all);
4445 static inline long __sched
4446 do_wait_for_common(struct completion *x, long timeout, int state)
4449 DECLARE_WAITQUEUE(wait, current);
4451 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4453 if (signal_pending_state(state, current)) {
4454 timeout = -ERESTARTSYS;
4457 __set_current_state(state);
4458 spin_unlock_irq(&x->wait.lock);
4459 timeout = schedule_timeout(timeout);
4460 spin_lock_irq(&x->wait.lock);
4461 } while (!x->done && timeout);
4462 __remove_wait_queue(&x->wait, &wait);
4467 return timeout ?: 1;
4471 wait_for_common(struct completion *x, long timeout, int state)
4475 spin_lock_irq(&x->wait.lock);
4476 timeout = do_wait_for_common(x, timeout, state);
4477 spin_unlock_irq(&x->wait.lock);
4482 * wait_for_completion: - waits for completion of a task
4483 * @x: holds the state of this particular completion
4485 * This waits to be signaled for completion of a specific task. It is NOT
4486 * interruptible and there is no timeout.
4488 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4489 * and interrupt capability. Also see complete().
4491 void __sched wait_for_completion(struct completion *x)
4493 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4495 EXPORT_SYMBOL(wait_for_completion);
4498 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4499 * @x: holds the state of this particular completion
4500 * @timeout: timeout value in jiffies
4502 * This waits for either a completion of a specific task to be signaled or for a
4503 * specified timeout to expire. The timeout is in jiffies. It is not
4506 unsigned long __sched
4507 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4509 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4511 EXPORT_SYMBOL(wait_for_completion_timeout);
4514 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4515 * @x: holds the state of this particular completion
4517 * This waits for completion of a specific task to be signaled. It is
4520 int __sched wait_for_completion_interruptible(struct completion *x)
4522 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4523 if (t == -ERESTARTSYS)
4527 EXPORT_SYMBOL(wait_for_completion_interruptible);
4530 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4531 * @x: holds the state of this particular completion
4532 * @timeout: timeout value in jiffies
4534 * This waits for either a completion of a specific task to be signaled or for a
4535 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4538 wait_for_completion_interruptible_timeout(struct completion *x,
4539 unsigned long timeout)
4541 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4543 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4546 * wait_for_completion_killable: - waits for completion of a task (killable)
4547 * @x: holds the state of this particular completion
4549 * This waits to be signaled for completion of a specific task. It can be
4550 * interrupted by a kill signal.
4552 int __sched wait_for_completion_killable(struct completion *x)
4554 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4555 if (t == -ERESTARTSYS)
4559 EXPORT_SYMBOL(wait_for_completion_killable);
4562 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4563 * @x: holds the state of this particular completion
4564 * @timeout: timeout value in jiffies
4566 * This waits for either a completion of a specific task to be
4567 * signaled or for a specified timeout to expire. It can be
4568 * interrupted by a kill signal. The timeout is in jiffies.
4571 wait_for_completion_killable_timeout(struct completion *x,
4572 unsigned long timeout)
4574 return wait_for_common(x, timeout, TASK_KILLABLE);
4576 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
4579 * try_wait_for_completion - try to decrement a completion without blocking
4580 * @x: completion structure
4582 * Returns: 0 if a decrement cannot be done without blocking
4583 * 1 if a decrement succeeded.
4585 * If a completion is being used as a counting completion,
4586 * attempt to decrement the counter without blocking. This
4587 * enables us to avoid waiting if the resource the completion
4588 * is protecting is not available.
4590 bool try_wait_for_completion(struct completion *x)
4592 unsigned long flags;
4595 spin_lock_irqsave(&x->wait.lock, flags);
4600 spin_unlock_irqrestore(&x->wait.lock, flags);
4603 EXPORT_SYMBOL(try_wait_for_completion);
4606 * completion_done - Test to see if a completion has any waiters
4607 * @x: completion structure
4609 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4610 * 1 if there are no waiters.
4613 bool completion_done(struct completion *x)
4615 unsigned long flags;
4618 spin_lock_irqsave(&x->wait.lock, flags);
4621 spin_unlock_irqrestore(&x->wait.lock, flags);
4624 EXPORT_SYMBOL(completion_done);
4627 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4629 unsigned long flags;
4632 init_waitqueue_entry(&wait, current);
4634 __set_current_state(state);
4636 spin_lock_irqsave(&q->lock, flags);
4637 __add_wait_queue(q, &wait);
4638 spin_unlock(&q->lock);
4639 timeout = schedule_timeout(timeout);
4640 spin_lock_irq(&q->lock);
4641 __remove_wait_queue(q, &wait);
4642 spin_unlock_irqrestore(&q->lock, flags);
4647 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4649 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4651 EXPORT_SYMBOL(interruptible_sleep_on);
4654 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4656 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4658 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4660 void __sched sleep_on(wait_queue_head_t *q)
4662 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4664 EXPORT_SYMBOL(sleep_on);
4666 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4668 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4670 EXPORT_SYMBOL(sleep_on_timeout);
4672 #ifdef CONFIG_RT_MUTEXES
4675 * rt_mutex_setprio - set the current priority of a task
4677 * @prio: prio value (kernel-internal form)
4679 * This function changes the 'effective' priority of a task. It does
4680 * not touch ->normal_prio like __setscheduler().
4682 * Used by the rt_mutex code to implement priority inheritance logic.
4684 void rt_mutex_setprio(struct task_struct *p, int prio)
4686 int oldprio, on_rq, running;
4688 const struct sched_class *prev_class;
4690 BUG_ON(prio < 0 || prio > MAX_PRIO);
4692 rq = __task_rq_lock(p);
4694 trace_sched_pi_setprio(p, prio);
4696 prev_class = p->sched_class;
4698 running = task_current(rq, p);
4700 dequeue_task(rq, p, 0);
4702 p->sched_class->put_prev_task(rq, p);
4705 p->sched_class = &rt_sched_class;
4707 p->sched_class = &fair_sched_class;
4712 p->sched_class->set_curr_task(rq);
4714 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4716 check_class_changed(rq, p, prev_class, oldprio);
4717 __task_rq_unlock(rq);
4722 void set_user_nice(struct task_struct *p, long nice)
4724 int old_prio, delta, on_rq;
4725 unsigned long flags;
4728 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4731 * We have to be careful, if called from sys_setpriority(),
4732 * the task might be in the middle of scheduling on another CPU.
4734 rq = task_rq_lock(p, &flags);
4736 * The RT priorities are set via sched_setscheduler(), but we still
4737 * allow the 'normal' nice value to be set - but as expected
4738 * it wont have any effect on scheduling until the task is
4739 * SCHED_FIFO/SCHED_RR:
4741 if (task_has_rt_policy(p)) {
4742 p->static_prio = NICE_TO_PRIO(nice);
4747 dequeue_task(rq, p, 0);
4749 p->static_prio = NICE_TO_PRIO(nice);
4752 p->prio = effective_prio(p);
4753 delta = p->prio - old_prio;
4756 enqueue_task(rq, p, 0);
4758 * If the task increased its priority or is running and
4759 * lowered its priority, then reschedule its CPU:
4761 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4762 resched_task(rq->curr);
4765 task_rq_unlock(rq, p, &flags);
4767 EXPORT_SYMBOL(set_user_nice);
4770 * can_nice - check if a task can reduce its nice value
4774 int can_nice(const struct task_struct *p, const int nice)
4776 /* convert nice value [19,-20] to rlimit style value [1,40] */
4777 int nice_rlim = 20 - nice;
4779 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4780 capable(CAP_SYS_NICE));
4783 #ifdef __ARCH_WANT_SYS_NICE
4786 * sys_nice - change the priority of the current process.
4787 * @increment: priority increment
4789 * sys_setpriority is a more generic, but much slower function that
4790 * does similar things.
4792 SYSCALL_DEFINE1(nice, int, increment)
4797 * Setpriority might change our priority at the same moment.
4798 * We don't have to worry. Conceptually one call occurs first
4799 * and we have a single winner.
4801 if (increment < -40)
4806 nice = TASK_NICE(current) + increment;
4812 if (increment < 0 && !can_nice(current, nice))
4815 retval = security_task_setnice(current, nice);
4819 set_user_nice(current, nice);
4826 * task_prio - return the priority value of a given task.
4827 * @p: the task in question.
4829 * This is the priority value as seen by users in /proc.
4830 * RT tasks are offset by -200. Normal tasks are centered
4831 * around 0, value goes from -16 to +15.
4833 int task_prio(const struct task_struct *p)
4835 return p->prio - MAX_RT_PRIO;
4839 * task_nice - return the nice value of a given task.
4840 * @p: the task in question.
4842 int task_nice(const struct task_struct *p)
4844 return TASK_NICE(p);
4846 EXPORT_SYMBOL(task_nice);
4849 * idle_cpu - is a given cpu idle currently?
4850 * @cpu: the processor in question.
4852 int idle_cpu(int cpu)
4854 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4858 * idle_task - return the idle task for a given cpu.
4859 * @cpu: the processor in question.
4861 struct task_struct *idle_task(int cpu)
4863 return cpu_rq(cpu)->idle;
4867 * find_process_by_pid - find a process with a matching PID value.
4868 * @pid: the pid in question.
4870 static struct task_struct *find_process_by_pid(pid_t pid)
4872 return pid ? find_task_by_vpid(pid) : current;
4875 /* Actually do priority change: must hold rq lock. */
4877 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4880 p->rt_priority = prio;
4881 p->normal_prio = normal_prio(p);
4882 /* we are holding p->pi_lock already */
4883 p->prio = rt_mutex_getprio(p);
4884 if (rt_prio(p->prio))
4885 p->sched_class = &rt_sched_class;
4887 p->sched_class = &fair_sched_class;
4892 * check the target process has a UID that matches the current process's
4894 static bool check_same_owner(struct task_struct *p)
4896 const struct cred *cred = current_cred(), *pcred;
4900 pcred = __task_cred(p);
4901 if (cred->user->user_ns == pcred->user->user_ns)
4902 match = (cred->euid == pcred->euid ||
4903 cred->euid == pcred->uid);
4910 static int __sched_setscheduler(struct task_struct *p, int policy,
4911 const struct sched_param *param, bool user)
4913 int retval, oldprio, oldpolicy = -1, on_rq, running;
4914 unsigned long flags;
4915 const struct sched_class *prev_class;
4919 /* may grab non-irq protected spin_locks */
4920 BUG_ON(in_interrupt());
4922 /* double check policy once rq lock held */
4924 reset_on_fork = p->sched_reset_on_fork;
4925 policy = oldpolicy = p->policy;
4927 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4928 policy &= ~SCHED_RESET_ON_FORK;
4930 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4931 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4932 policy != SCHED_IDLE)
4937 * Valid priorities for SCHED_FIFO and SCHED_RR are
4938 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4939 * SCHED_BATCH and SCHED_IDLE is 0.
4941 if (param->sched_priority < 0 ||
4942 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4943 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4945 if (rt_policy(policy) != (param->sched_priority != 0))
4949 * Allow unprivileged RT tasks to decrease priority:
4951 if (user && !capable(CAP_SYS_NICE)) {
4952 if (rt_policy(policy)) {
4953 unsigned long rlim_rtprio =
4954 task_rlimit(p, RLIMIT_RTPRIO);
4956 /* can't set/change the rt policy */
4957 if (policy != p->policy && !rlim_rtprio)
4960 /* can't increase priority */
4961 if (param->sched_priority > p->rt_priority &&
4962 param->sched_priority > rlim_rtprio)
4967 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4968 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4970 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
4971 if (!can_nice(p, TASK_NICE(p)))
4975 /* can't change other user's priorities */
4976 if (!check_same_owner(p))
4979 /* Normal users shall not reset the sched_reset_on_fork flag */
4980 if (p->sched_reset_on_fork && !reset_on_fork)
4985 retval = security_task_setscheduler(p);
4991 * make sure no PI-waiters arrive (or leave) while we are
4992 * changing the priority of the task:
4994 * To be able to change p->policy safely, the appropriate
4995 * runqueue lock must be held.
4997 rq = task_rq_lock(p, &flags);
5000 * Changing the policy of the stop threads its a very bad idea
5002 if (p == rq->stop) {
5003 task_rq_unlock(rq, p, &flags);
5008 * If not changing anything there's no need to proceed further:
5010 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
5011 param->sched_priority == p->rt_priority))) {
5013 __task_rq_unlock(rq);
5014 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5018 #ifdef CONFIG_RT_GROUP_SCHED
5021 * Do not allow realtime tasks into groups that have no runtime
5024 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5025 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
5026 !task_group_is_autogroup(task_group(p))) {
5027 task_rq_unlock(rq, p, &flags);
5033 /* recheck policy now with rq lock held */
5034 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5035 policy = oldpolicy = -1;
5036 task_rq_unlock(rq, p, &flags);
5040 running = task_current(rq, p);
5042 deactivate_task(rq, p, 0);
5044 p->sched_class->put_prev_task(rq, p);
5046 p->sched_reset_on_fork = reset_on_fork;
5049 prev_class = p->sched_class;
5050 __setscheduler(rq, p, policy, param->sched_priority);
5053 p->sched_class->set_curr_task(rq);
5055 activate_task(rq, p, 0);
5057 check_class_changed(rq, p, prev_class, oldprio);
5058 task_rq_unlock(rq, p, &flags);
5060 rt_mutex_adjust_pi(p);
5066 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5067 * @p: the task in question.
5068 * @policy: new policy.
5069 * @param: structure containing the new RT priority.
5071 * NOTE that the task may be already dead.
5073 int sched_setscheduler(struct task_struct *p, int policy,
5074 const struct sched_param *param)
5076 return __sched_setscheduler(p, policy, param, true);
5078 EXPORT_SYMBOL_GPL(sched_setscheduler);
5081 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5082 * @p: the task in question.
5083 * @policy: new policy.
5084 * @param: structure containing the new RT priority.
5086 * Just like sched_setscheduler, only don't bother checking if the
5087 * current context has permission. For example, this is needed in
5088 * stop_machine(): we create temporary high priority worker threads,
5089 * but our caller might not have that capability.
5091 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5092 const struct sched_param *param)
5094 return __sched_setscheduler(p, policy, param, false);
5098 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5100 struct sched_param lparam;
5101 struct task_struct *p;
5104 if (!param || pid < 0)
5106 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5111 p = find_process_by_pid(pid);
5113 retval = sched_setscheduler(p, policy, &lparam);
5120 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5121 * @pid: the pid in question.
5122 * @policy: new policy.
5123 * @param: structure containing the new RT priority.
5125 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
5126 struct sched_param __user *, param)
5128 /* negative values for policy are not valid */
5132 return do_sched_setscheduler(pid, policy, param);
5136 * sys_sched_setparam - set/change the RT priority of a thread
5137 * @pid: the pid in question.
5138 * @param: structure containing the new RT priority.
5140 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5142 return do_sched_setscheduler(pid, -1, param);
5146 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5147 * @pid: the pid in question.
5149 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5151 struct task_struct *p;
5159 p = find_process_by_pid(pid);
5161 retval = security_task_getscheduler(p);
5164 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
5171 * sys_sched_getparam - get the RT priority of a thread
5172 * @pid: the pid in question.
5173 * @param: structure containing the RT priority.
5175 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5177 struct sched_param lp;
5178 struct task_struct *p;
5181 if (!param || pid < 0)
5185 p = find_process_by_pid(pid);
5190 retval = security_task_getscheduler(p);
5194 lp.sched_priority = p->rt_priority;
5198 * This one might sleep, we cannot do it with a spinlock held ...
5200 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5209 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5211 cpumask_var_t cpus_allowed, new_mask;
5212 struct task_struct *p;
5218 p = find_process_by_pid(pid);
5225 /* Prevent p going away */
5229 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5233 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5235 goto out_free_cpus_allowed;
5238 if (!check_same_owner(p) && !task_ns_capable(p, CAP_SYS_NICE))
5241 retval = security_task_setscheduler(p);
5245 cpuset_cpus_allowed(p, cpus_allowed);
5246 cpumask_and(new_mask, in_mask, cpus_allowed);
5248 retval = set_cpus_allowed_ptr(p, new_mask);
5251 cpuset_cpus_allowed(p, cpus_allowed);
5252 if (!cpumask_subset(new_mask, cpus_allowed)) {
5254 * We must have raced with a concurrent cpuset
5255 * update. Just reset the cpus_allowed to the
5256 * cpuset's cpus_allowed
5258 cpumask_copy(new_mask, cpus_allowed);
5263 free_cpumask_var(new_mask);
5264 out_free_cpus_allowed:
5265 free_cpumask_var(cpus_allowed);
5272 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5273 struct cpumask *new_mask)
5275 if (len < cpumask_size())
5276 cpumask_clear(new_mask);
5277 else if (len > cpumask_size())
5278 len = cpumask_size();
5280 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5284 * sys_sched_setaffinity - set the cpu affinity of a process
5285 * @pid: pid of the process
5286 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5287 * @user_mask_ptr: user-space pointer to the new cpu mask
5289 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5290 unsigned long __user *, user_mask_ptr)
5292 cpumask_var_t new_mask;
5295 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5298 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5300 retval = sched_setaffinity(pid, new_mask);
5301 free_cpumask_var(new_mask);
5305 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5307 struct task_struct *p;
5308 unsigned long flags;
5315 p = find_process_by_pid(pid);
5319 retval = security_task_getscheduler(p);
5323 raw_spin_lock_irqsave(&p->pi_lock, flags);
5324 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5325 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5335 * sys_sched_getaffinity - get the cpu affinity of a process
5336 * @pid: pid of the process
5337 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5338 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5340 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5341 unsigned long __user *, user_mask_ptr)
5346 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5348 if (len & (sizeof(unsigned long)-1))
5351 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5354 ret = sched_getaffinity(pid, mask);
5356 size_t retlen = min_t(size_t, len, cpumask_size());
5358 if (copy_to_user(user_mask_ptr, mask, retlen))
5363 free_cpumask_var(mask);
5369 * sys_sched_yield - yield the current processor to other threads.
5371 * This function yields the current CPU to other tasks. If there are no
5372 * other threads running on this CPU then this function will return.
5374 SYSCALL_DEFINE0(sched_yield)
5376 struct rq *rq = this_rq_lock();
5378 schedstat_inc(rq, yld_count);
5379 current->sched_class->yield_task(rq);
5382 * Since we are going to call schedule() anyway, there's
5383 * no need to preempt or enable interrupts:
5385 __release(rq->lock);
5386 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5387 do_raw_spin_unlock(&rq->lock);
5388 preempt_enable_no_resched();
5395 static inline int should_resched(void)
5397 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5400 static void __cond_resched(void)
5402 add_preempt_count(PREEMPT_ACTIVE);
5404 sub_preempt_count(PREEMPT_ACTIVE);
5407 int __sched _cond_resched(void)
5409 if (should_resched()) {
5415 EXPORT_SYMBOL(_cond_resched);
5418 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5419 * call schedule, and on return reacquire the lock.
5421 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5422 * operations here to prevent schedule() from being called twice (once via
5423 * spin_unlock(), once by hand).
5425 int __cond_resched_lock(spinlock_t *lock)
5427 int resched = should_resched();
5430 lockdep_assert_held(lock);
5432 if (spin_needbreak(lock) || resched) {
5443 EXPORT_SYMBOL(__cond_resched_lock);
5445 int __sched __cond_resched_softirq(void)
5447 BUG_ON(!in_softirq());
5449 if (should_resched()) {
5457 EXPORT_SYMBOL(__cond_resched_softirq);
5460 * yield - yield the current processor to other threads.
5462 * This is a shortcut for kernel-space yielding - it marks the
5463 * thread runnable and calls sys_sched_yield().
5465 void __sched yield(void)
5467 set_current_state(TASK_RUNNING);
5470 EXPORT_SYMBOL(yield);
5473 * yield_to - yield the current processor to another thread in
5474 * your thread group, or accelerate that thread toward the
5475 * processor it's on.
5477 * @preempt: whether task preemption is allowed or not
5479 * It's the caller's job to ensure that the target task struct
5480 * can't go away on us before we can do any checks.
5482 * Returns true if we indeed boosted the target task.
5484 bool __sched yield_to(struct task_struct *p, bool preempt)
5486 struct task_struct *curr = current;
5487 struct rq *rq, *p_rq;
5488 unsigned long flags;
5491 local_irq_save(flags);
5496 double_rq_lock(rq, p_rq);
5497 while (task_rq(p) != p_rq) {
5498 double_rq_unlock(rq, p_rq);
5502 if (!curr->sched_class->yield_to_task)
5505 if (curr->sched_class != p->sched_class)
5508 if (task_running(p_rq, p) || p->state)
5511 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5513 schedstat_inc(rq, yld_count);
5515 * Make p's CPU reschedule; pick_next_entity takes care of
5518 if (preempt && rq != p_rq)
5519 resched_task(p_rq->curr);
5523 double_rq_unlock(rq, p_rq);
5524 local_irq_restore(flags);
5531 EXPORT_SYMBOL_GPL(yield_to);
5534 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5535 * that process accounting knows that this is a task in IO wait state.
5537 void __sched io_schedule(void)
5539 struct rq *rq = raw_rq();
5541 delayacct_blkio_start();
5542 atomic_inc(&rq->nr_iowait);
5543 blk_flush_plug(current);
5544 current->in_iowait = 1;
5546 current->in_iowait = 0;
5547 atomic_dec(&rq->nr_iowait);
5548 delayacct_blkio_end();
5550 EXPORT_SYMBOL(io_schedule);
5552 long __sched io_schedule_timeout(long timeout)
5554 struct rq *rq = raw_rq();
5557 delayacct_blkio_start();
5558 atomic_inc(&rq->nr_iowait);
5559 blk_flush_plug(current);
5560 current->in_iowait = 1;
5561 ret = schedule_timeout(timeout);
5562 current->in_iowait = 0;
5563 atomic_dec(&rq->nr_iowait);
5564 delayacct_blkio_end();
5569 * sys_sched_get_priority_max - return maximum RT priority.
5570 * @policy: scheduling class.
5572 * this syscall returns the maximum rt_priority that can be used
5573 * by a given scheduling class.
5575 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5582 ret = MAX_USER_RT_PRIO-1;
5594 * sys_sched_get_priority_min - return minimum RT priority.
5595 * @policy: scheduling class.
5597 * this syscall returns the minimum rt_priority that can be used
5598 * by a given scheduling class.
5600 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5618 * sys_sched_rr_get_interval - return the default timeslice of a process.
5619 * @pid: pid of the process.
5620 * @interval: userspace pointer to the timeslice value.
5622 * this syscall writes the default timeslice value of a given process
5623 * into the user-space timespec buffer. A value of '0' means infinity.
5625 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5626 struct timespec __user *, interval)
5628 struct task_struct *p;
5629 unsigned int time_slice;
5630 unsigned long flags;
5640 p = find_process_by_pid(pid);
5644 retval = security_task_getscheduler(p);
5648 rq = task_rq_lock(p, &flags);
5649 time_slice = p->sched_class->get_rr_interval(rq, p);
5650 task_rq_unlock(rq, p, &flags);
5653 jiffies_to_timespec(time_slice, &t);
5654 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5662 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5664 void sched_show_task(struct task_struct *p)
5666 unsigned long free = 0;
5669 state = p->state ? __ffs(p->state) + 1 : 0;
5670 printk(KERN_INFO "%-15.15s %c", p->comm,
5671 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5672 #if BITS_PER_LONG == 32
5673 if (state == TASK_RUNNING)
5674 printk(KERN_CONT " running ");
5676 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5678 if (state == TASK_RUNNING)
5679 printk(KERN_CONT " running task ");
5681 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5683 #ifdef CONFIG_DEBUG_STACK_USAGE
5684 free = stack_not_used(p);
5686 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5687 task_pid_nr(p), task_pid_nr(p->real_parent),
5688 (unsigned long)task_thread_info(p)->flags);
5690 show_stack(p, NULL);
5693 void show_state_filter(unsigned long state_filter)
5695 struct task_struct *g, *p;
5697 #if BITS_PER_LONG == 32
5699 " task PC stack pid father\n");
5702 " task PC stack pid father\n");
5704 read_lock(&tasklist_lock);
5705 do_each_thread(g, p) {
5707 * reset the NMI-timeout, listing all files on a slow
5708 * console might take a lot of time:
5710 touch_nmi_watchdog();
5711 if (!state_filter || (p->state & state_filter))
5713 } while_each_thread(g, p);
5715 touch_all_softlockup_watchdogs();
5717 #ifdef CONFIG_SCHED_DEBUG
5718 sysrq_sched_debug_show();
5720 read_unlock(&tasklist_lock);
5722 * Only show locks if all tasks are dumped:
5725 debug_show_all_locks();
5728 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5730 idle->sched_class = &idle_sched_class;
5734 * init_idle - set up an idle thread for a given CPU
5735 * @idle: task in question
5736 * @cpu: cpu the idle task belongs to
5738 * NOTE: this function does not set the idle thread's NEED_RESCHED
5739 * flag, to make booting more robust.
5741 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5743 struct rq *rq = cpu_rq(cpu);
5744 unsigned long flags;
5746 raw_spin_lock_irqsave(&rq->lock, flags);
5749 idle->state = TASK_RUNNING;
5750 idle->se.exec_start = sched_clock();
5752 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
5754 * We're having a chicken and egg problem, even though we are
5755 * holding rq->lock, the cpu isn't yet set to this cpu so the
5756 * lockdep check in task_group() will fail.
5758 * Similar case to sched_fork(). / Alternatively we could
5759 * use task_rq_lock() here and obtain the other rq->lock.
5764 __set_task_cpu(idle, cpu);
5767 rq->curr = rq->idle = idle;
5768 #if defined(CONFIG_SMP)
5771 raw_spin_unlock_irqrestore(&rq->lock, flags);
5773 /* Set the preempt count _outside_ the spinlocks! */
5774 #if defined(CONFIG_PREEMPT)
5775 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5777 task_thread_info(idle)->preempt_count = 0;
5780 * The idle tasks have their own, simple scheduling class:
5782 idle->sched_class = &idle_sched_class;
5783 ftrace_graph_init_idle_task(idle, cpu);
5787 * In a system that switches off the HZ timer nohz_cpu_mask
5788 * indicates which cpus entered this state. This is used
5789 * in the rcu update to wait only for active cpus. For system
5790 * which do not switch off the HZ timer nohz_cpu_mask should
5791 * always be CPU_BITS_NONE.
5793 cpumask_var_t nohz_cpu_mask;
5796 * Increase the granularity value when there are more CPUs,
5797 * because with more CPUs the 'effective latency' as visible
5798 * to users decreases. But the relationship is not linear,
5799 * so pick a second-best guess by going with the log2 of the
5802 * This idea comes from the SD scheduler of Con Kolivas:
5804 static int get_update_sysctl_factor(void)
5806 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5807 unsigned int factor;
5809 switch (sysctl_sched_tunable_scaling) {
5810 case SCHED_TUNABLESCALING_NONE:
5813 case SCHED_TUNABLESCALING_LINEAR:
5816 case SCHED_TUNABLESCALING_LOG:
5818 factor = 1 + ilog2(cpus);
5825 static void update_sysctl(void)
5827 unsigned int factor = get_update_sysctl_factor();
5829 #define SET_SYSCTL(name) \
5830 (sysctl_##name = (factor) * normalized_sysctl_##name)
5831 SET_SYSCTL(sched_min_granularity);
5832 SET_SYSCTL(sched_latency);
5833 SET_SYSCTL(sched_wakeup_granularity);
5837 static inline void sched_init_granularity(void)
5844 * This is how migration works:
5846 * 1) we invoke migration_cpu_stop() on the target CPU using
5848 * 2) stopper starts to run (implicitly forcing the migrated thread
5850 * 3) it checks whether the migrated task is still in the wrong runqueue.
5851 * 4) if it's in the wrong runqueue then the migration thread removes
5852 * it and puts it into the right queue.
5853 * 5) stopper completes and stop_one_cpu() returns and the migration
5858 * Change a given task's CPU affinity. Migrate the thread to a
5859 * proper CPU and schedule it away if the CPU it's executing on
5860 * is removed from the allowed bitmask.
5862 * NOTE: the caller must have a valid reference to the task, the
5863 * task must not exit() & deallocate itself prematurely. The
5864 * call is not atomic; no spinlocks may be held.
5866 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5868 unsigned long flags;
5870 unsigned int dest_cpu;
5873 rq = task_rq_lock(p, &flags);
5875 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5880 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5881 !cpumask_equal(&p->cpus_allowed, new_mask))) {
5886 if (p->sched_class->set_cpus_allowed)
5887 p->sched_class->set_cpus_allowed(p, new_mask);
5889 cpumask_copy(&p->cpus_allowed, new_mask);
5890 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5893 /* Can the task run on the task's current CPU? If so, we're done */
5894 if (cpumask_test_cpu(task_cpu(p), new_mask))
5897 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
5898 if (need_migrate_task(p)) {
5899 struct migration_arg arg = { p, dest_cpu };
5900 /* Need help from migration thread: drop lock and wait. */
5901 task_rq_unlock(rq, p, &flags);
5902 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
5903 tlb_migrate_finish(p->mm);
5907 task_rq_unlock(rq, p, &flags);
5911 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5914 * Move (not current) task off this cpu, onto dest cpu. We're doing
5915 * this because either it can't run here any more (set_cpus_allowed()
5916 * away from this CPU, or CPU going down), or because we're
5917 * attempting to rebalance this task on exec (sched_exec).
5919 * So we race with normal scheduler movements, but that's OK, as long
5920 * as the task is no longer on this CPU.
5922 * Returns non-zero if task was successfully migrated.
5924 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5926 struct rq *rq_dest, *rq_src;
5929 if (unlikely(!cpu_active(dest_cpu)))
5932 rq_src = cpu_rq(src_cpu);
5933 rq_dest = cpu_rq(dest_cpu);
5935 raw_spin_lock(&p->pi_lock);
5936 double_rq_lock(rq_src, rq_dest);
5937 /* Already moved. */
5938 if (task_cpu(p) != src_cpu)
5940 /* Affinity changed (again). */
5941 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
5945 * If we're not on a rq, the next wake-up will ensure we're
5949 deactivate_task(rq_src, p, 0);
5950 set_task_cpu(p, dest_cpu);
5951 activate_task(rq_dest, p, 0);
5952 check_preempt_curr(rq_dest, p, 0);
5957 double_rq_unlock(rq_src, rq_dest);
5958 raw_spin_unlock(&p->pi_lock);
5963 * migration_cpu_stop - this will be executed by a highprio stopper thread
5964 * and performs thread migration by bumping thread off CPU then
5965 * 'pushing' onto another runqueue.
5967 static int migration_cpu_stop(void *data)
5969 struct migration_arg *arg = data;
5972 * The original target cpu might have gone down and we might
5973 * be on another cpu but it doesn't matter.
5975 local_irq_disable();
5976 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
5981 #ifdef CONFIG_HOTPLUG_CPU
5984 * Ensures that the idle task is using init_mm right before its cpu goes
5987 void idle_task_exit(void)
5989 struct mm_struct *mm = current->active_mm;
5991 BUG_ON(cpu_online(smp_processor_id()));
5994 switch_mm(mm, &init_mm, current);
5999 * While a dead CPU has no uninterruptible tasks queued at this point,
6000 * it might still have a nonzero ->nr_uninterruptible counter, because
6001 * for performance reasons the counter is not stricly tracking tasks to
6002 * their home CPUs. So we just add the counter to another CPU's counter,
6003 * to keep the global sum constant after CPU-down:
6005 static void migrate_nr_uninterruptible(struct rq *rq_src)
6007 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
6009 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6010 rq_src->nr_uninterruptible = 0;
6014 * remove the tasks which were accounted by rq from calc_load_tasks.
6016 static void calc_global_load_remove(struct rq *rq)
6018 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
6019 rq->calc_load_active = 0;
6023 * Migrate all tasks from the rq, sleeping tasks will be migrated by
6024 * try_to_wake_up()->select_task_rq().
6026 * Called with rq->lock held even though we'er in stop_machine() and
6027 * there's no concurrency possible, we hold the required locks anyway
6028 * because of lock validation efforts.
6030 static void migrate_tasks(unsigned int dead_cpu)
6032 struct rq *rq = cpu_rq(dead_cpu);
6033 struct task_struct *next, *stop = rq->stop;
6037 * Fudge the rq selection such that the below task selection loop
6038 * doesn't get stuck on the currently eligible stop task.
6040 * We're currently inside stop_machine() and the rq is either stuck
6041 * in the stop_machine_cpu_stop() loop, or we're executing this code,
6042 * either way we should never end up calling schedule() until we're
6049 * There's this thread running, bail when that's the only
6052 if (rq->nr_running == 1)
6055 next = pick_next_task(rq);
6057 next->sched_class->put_prev_task(rq, next);
6059 /* Find suitable destination for @next, with force if needed. */
6060 dest_cpu = select_fallback_rq(dead_cpu, next);
6061 raw_spin_unlock(&rq->lock);
6063 __migrate_task(next, dead_cpu, dest_cpu);
6065 raw_spin_lock(&rq->lock);
6071 #endif /* CONFIG_HOTPLUG_CPU */
6073 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6075 static struct ctl_table sd_ctl_dir[] = {
6077 .procname = "sched_domain",
6083 static struct ctl_table sd_ctl_root[] = {
6085 .procname = "kernel",
6087 .child = sd_ctl_dir,
6092 static struct ctl_table *sd_alloc_ctl_entry(int n)
6094 struct ctl_table *entry =
6095 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6100 static void sd_free_ctl_entry(struct ctl_table **tablep)
6102 struct ctl_table *entry;
6105 * In the intermediate directories, both the child directory and
6106 * procname are dynamically allocated and could fail but the mode
6107 * will always be set. In the lowest directory the names are
6108 * static strings and all have proc handlers.
6110 for (entry = *tablep; entry->mode; entry++) {
6112 sd_free_ctl_entry(&entry->child);
6113 if (entry->proc_handler == NULL)
6114 kfree(entry->procname);
6122 set_table_entry(struct ctl_table *entry,
6123 const char *procname, void *data, int maxlen,
6124 mode_t mode, proc_handler *proc_handler)
6126 entry->procname = procname;
6128 entry->maxlen = maxlen;
6130 entry->proc_handler = proc_handler;
6133 static struct ctl_table *
6134 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6136 struct ctl_table *table = sd_alloc_ctl_entry(13);
6141 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6142 sizeof(long), 0644, proc_doulongvec_minmax);
6143 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6144 sizeof(long), 0644, proc_doulongvec_minmax);
6145 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6146 sizeof(int), 0644, proc_dointvec_minmax);
6147 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6148 sizeof(int), 0644, proc_dointvec_minmax);
6149 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6150 sizeof(int), 0644, proc_dointvec_minmax);
6151 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6152 sizeof(int), 0644, proc_dointvec_minmax);
6153 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6154 sizeof(int), 0644, proc_dointvec_minmax);
6155 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6156 sizeof(int), 0644, proc_dointvec_minmax);
6157 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6158 sizeof(int), 0644, proc_dointvec_minmax);
6159 set_table_entry(&table[9], "cache_nice_tries",
6160 &sd->cache_nice_tries,
6161 sizeof(int), 0644, proc_dointvec_minmax);
6162 set_table_entry(&table[10], "flags", &sd->flags,
6163 sizeof(int), 0644, proc_dointvec_minmax);
6164 set_table_entry(&table[11], "name", sd->name,
6165 CORENAME_MAX_SIZE, 0444, proc_dostring);
6166 /* &table[12] is terminator */
6171 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6173 struct ctl_table *entry, *table;
6174 struct sched_domain *sd;
6175 int domain_num = 0, i;
6178 for_each_domain(cpu, sd)
6180 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6185 for_each_domain(cpu, sd) {
6186 snprintf(buf, 32, "domain%d", i);
6187 entry->procname = kstrdup(buf, GFP_KERNEL);
6189 entry->child = sd_alloc_ctl_domain_table(sd);
6196 static struct ctl_table_header *sd_sysctl_header;
6197 static void register_sched_domain_sysctl(void)
6199 int i, cpu_num = num_possible_cpus();
6200 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6203 WARN_ON(sd_ctl_dir[0].child);
6204 sd_ctl_dir[0].child = entry;
6209 for_each_possible_cpu(i) {
6210 snprintf(buf, 32, "cpu%d", i);
6211 entry->procname = kstrdup(buf, GFP_KERNEL);
6213 entry->child = sd_alloc_ctl_cpu_table(i);
6217 WARN_ON(sd_sysctl_header);
6218 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6221 /* may be called multiple times per register */
6222 static void unregister_sched_domain_sysctl(void)
6224 if (sd_sysctl_header)
6225 unregister_sysctl_table(sd_sysctl_header);
6226 sd_sysctl_header = NULL;
6227 if (sd_ctl_dir[0].child)
6228 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6231 static void register_sched_domain_sysctl(void)
6234 static void unregister_sched_domain_sysctl(void)
6239 static void set_rq_online(struct rq *rq)
6242 const struct sched_class *class;
6244 cpumask_set_cpu(rq->cpu, rq->rd->online);
6247 for_each_class(class) {
6248 if (class->rq_online)
6249 class->rq_online(rq);
6254 static void set_rq_offline(struct rq *rq)
6257 const struct sched_class *class;
6259 for_each_class(class) {
6260 if (class->rq_offline)
6261 class->rq_offline(rq);
6264 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6270 * migration_call - callback that gets triggered when a CPU is added.
6271 * Here we can start up the necessary migration thread for the new CPU.
6273 static int __cpuinit
6274 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6276 int cpu = (long)hcpu;
6277 unsigned long flags;
6278 struct rq *rq = cpu_rq(cpu);
6280 switch (action & ~CPU_TASKS_FROZEN) {
6282 case CPU_UP_PREPARE:
6283 rq->calc_load_update = calc_load_update;
6287 /* Update our root-domain */
6288 raw_spin_lock_irqsave(&rq->lock, flags);
6290 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6294 raw_spin_unlock_irqrestore(&rq->lock, flags);
6297 #ifdef CONFIG_HOTPLUG_CPU
6299 /* Update our root-domain */
6300 raw_spin_lock_irqsave(&rq->lock, flags);
6302 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6306 BUG_ON(rq->nr_running != 1); /* the migration thread */
6307 raw_spin_unlock_irqrestore(&rq->lock, flags);
6309 migrate_nr_uninterruptible(rq);
6310 calc_global_load_remove(rq);
6315 update_max_interval();
6321 * Register at high priority so that task migration (migrate_all_tasks)
6322 * happens before everything else. This has to be lower priority than
6323 * the notifier in the perf_event subsystem, though.
6325 static struct notifier_block __cpuinitdata migration_notifier = {
6326 .notifier_call = migration_call,
6327 .priority = CPU_PRI_MIGRATION,
6330 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
6331 unsigned long action, void *hcpu)
6333 switch (action & ~CPU_TASKS_FROZEN) {
6335 case CPU_DOWN_FAILED:
6336 set_cpu_active((long)hcpu, true);
6343 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
6344 unsigned long action, void *hcpu)
6346 switch (action & ~CPU_TASKS_FROZEN) {
6347 case CPU_DOWN_PREPARE:
6348 set_cpu_active((long)hcpu, false);
6355 static int __init migration_init(void)
6357 void *cpu = (void *)(long)smp_processor_id();
6360 /* Initialize migration for the boot CPU */
6361 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6362 BUG_ON(err == NOTIFY_BAD);
6363 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6364 register_cpu_notifier(&migration_notifier);
6366 /* Register cpu active notifiers */
6367 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
6368 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
6372 early_initcall(migration_init);
6377 #ifdef CONFIG_SCHED_DEBUG
6379 static __read_mostly int sched_domain_debug_enabled;
6381 static int __init sched_domain_debug_setup(char *str)
6383 sched_domain_debug_enabled = 1;
6387 early_param("sched_debug", sched_domain_debug_setup);
6389 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6390 struct cpumask *groupmask)
6392 struct sched_group *group = sd->groups;
6395 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6396 cpumask_clear(groupmask);
6398 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6400 if (!(sd->flags & SD_LOAD_BALANCE)) {
6401 printk("does not load-balance\n");
6403 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6408 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6410 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6411 printk(KERN_ERR "ERROR: domain->span does not contain "
6414 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6415 printk(KERN_ERR "ERROR: domain->groups does not contain"
6419 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6423 printk(KERN_ERR "ERROR: group is NULL\n");
6427 if (!group->cpu_power) {
6428 printk(KERN_CONT "\n");
6429 printk(KERN_ERR "ERROR: domain->cpu_power not "
6434 if (!cpumask_weight(sched_group_cpus(group))) {
6435 printk(KERN_CONT "\n");
6436 printk(KERN_ERR "ERROR: empty group\n");
6440 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6441 printk(KERN_CONT "\n");
6442 printk(KERN_ERR "ERROR: repeated CPUs\n");
6446 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6448 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6450 printk(KERN_CONT " %s", str);
6451 if (group->cpu_power != SCHED_LOAD_SCALE) {
6452 printk(KERN_CONT " (cpu_power = %d)",
6456 group = group->next;
6457 } while (group != sd->groups);
6458 printk(KERN_CONT "\n");
6460 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6461 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6464 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6465 printk(KERN_ERR "ERROR: parent span is not a superset "
6466 "of domain->span\n");
6470 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6472 cpumask_var_t groupmask;
6475 if (!sched_domain_debug_enabled)
6479 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6483 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6485 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6486 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6491 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6498 free_cpumask_var(groupmask);
6500 #else /* !CONFIG_SCHED_DEBUG */
6501 # define sched_domain_debug(sd, cpu) do { } while (0)
6502 #endif /* CONFIG_SCHED_DEBUG */
6504 static int sd_degenerate(struct sched_domain *sd)
6506 if (cpumask_weight(sched_domain_span(sd)) == 1)
6509 /* Following flags need at least 2 groups */
6510 if (sd->flags & (SD_LOAD_BALANCE |
6511 SD_BALANCE_NEWIDLE |
6515 SD_SHARE_PKG_RESOURCES)) {
6516 if (sd->groups != sd->groups->next)
6520 /* Following flags don't use groups */
6521 if (sd->flags & (SD_WAKE_AFFINE))
6528 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6530 unsigned long cflags = sd->flags, pflags = parent->flags;
6532 if (sd_degenerate(parent))
6535 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6538 /* Flags needing groups don't count if only 1 group in parent */
6539 if (parent->groups == parent->groups->next) {
6540 pflags &= ~(SD_LOAD_BALANCE |
6541 SD_BALANCE_NEWIDLE |
6545 SD_SHARE_PKG_RESOURCES);
6546 if (nr_node_ids == 1)
6547 pflags &= ~SD_SERIALIZE;
6549 if (~cflags & pflags)
6555 static void free_rootdomain(struct root_domain *rd)
6557 synchronize_sched();
6559 cpupri_cleanup(&rd->cpupri);
6561 free_cpumask_var(rd->rto_mask);
6562 free_cpumask_var(rd->online);
6563 free_cpumask_var(rd->span);
6567 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6569 struct root_domain *old_rd = NULL;
6570 unsigned long flags;
6572 raw_spin_lock_irqsave(&rq->lock, flags);
6577 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6580 cpumask_clear_cpu(rq->cpu, old_rd->span);
6583 * If we dont want to free the old_rt yet then
6584 * set old_rd to NULL to skip the freeing later
6587 if (!atomic_dec_and_test(&old_rd->refcount))
6591 atomic_inc(&rd->refcount);
6594 cpumask_set_cpu(rq->cpu, rd->span);
6595 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6598 raw_spin_unlock_irqrestore(&rq->lock, flags);
6601 free_rootdomain(old_rd);
6604 static int init_rootdomain(struct root_domain *rd)
6606 memset(rd, 0, sizeof(*rd));
6608 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
6610 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
6612 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6615 if (cpupri_init(&rd->cpupri) != 0)
6620 free_cpumask_var(rd->rto_mask);
6622 free_cpumask_var(rd->online);
6624 free_cpumask_var(rd->span);
6629 static void init_defrootdomain(void)
6631 init_rootdomain(&def_root_domain);
6633 atomic_set(&def_root_domain.refcount, 1);
6636 static struct root_domain *alloc_rootdomain(void)
6638 struct root_domain *rd;
6640 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6644 if (init_rootdomain(rd) != 0) {
6653 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6654 * hold the hotplug lock.
6657 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6659 struct rq *rq = cpu_rq(cpu);
6660 struct sched_domain *tmp;
6662 for (tmp = sd; tmp; tmp = tmp->parent)
6663 tmp->span_weight = cpumask_weight(sched_domain_span(tmp));
6665 /* Remove the sched domains which do not contribute to scheduling. */
6666 for (tmp = sd; tmp; ) {
6667 struct sched_domain *parent = tmp->parent;
6671 if (sd_parent_degenerate(tmp, parent)) {
6672 tmp->parent = parent->parent;
6674 parent->parent->child = tmp;
6679 if (sd && sd_degenerate(sd)) {
6685 sched_domain_debug(sd, cpu);
6687 rq_attach_root(rq, rd);
6688 rcu_assign_pointer(rq->sd, sd);
6691 /* cpus with isolated domains */
6692 static cpumask_var_t cpu_isolated_map;
6694 /* Setup the mask of cpus configured for isolated domains */
6695 static int __init isolated_cpu_setup(char *str)
6697 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6698 cpulist_parse(str, cpu_isolated_map);
6702 __setup("isolcpus=", isolated_cpu_setup);
6705 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6706 * to a function which identifies what group(along with sched group) a CPU
6707 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6708 * (due to the fact that we keep track of groups covered with a struct cpumask).
6710 * init_sched_build_groups will build a circular linked list of the groups
6711 * covered by the given span, and will set each group's ->cpumask correctly,
6712 * and ->cpu_power to 0.
6715 init_sched_build_groups(const struct cpumask *span,
6716 const struct cpumask *cpu_map,
6717 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
6718 struct sched_group **sg,
6719 struct cpumask *tmpmask),
6720 struct cpumask *covered, struct cpumask *tmpmask)
6722 struct sched_group *first = NULL, *last = NULL;
6725 cpumask_clear(covered);
6727 for_each_cpu(i, span) {
6728 struct sched_group *sg;
6729 int group = group_fn(i, cpu_map, &sg, tmpmask);
6732 if (cpumask_test_cpu(i, covered))
6735 cpumask_clear(sched_group_cpus(sg));
6738 for_each_cpu(j, span) {
6739 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6742 cpumask_set_cpu(j, covered);
6743 cpumask_set_cpu(j, sched_group_cpus(sg));
6754 #define SD_NODES_PER_DOMAIN 16
6759 * find_next_best_node - find the next node to include in a sched_domain
6760 * @node: node whose sched_domain we're building
6761 * @used_nodes: nodes already in the sched_domain
6763 * Find the next node to include in a given scheduling domain. Simply
6764 * finds the closest node not already in the @used_nodes map.
6766 * Should use nodemask_t.
6768 static int find_next_best_node(int node, nodemask_t *used_nodes)
6770 int i, n, val, min_val, best_node = 0;
6774 for (i = 0; i < nr_node_ids; i++) {
6775 /* Start at @node */
6776 n = (node + i) % nr_node_ids;
6778 if (!nr_cpus_node(n))
6781 /* Skip already used nodes */
6782 if (node_isset(n, *used_nodes))
6785 /* Simple min distance search */
6786 val = node_distance(node, n);
6788 if (val < min_val) {
6794 node_set(best_node, *used_nodes);
6799 * sched_domain_node_span - get a cpumask for a node's sched_domain
6800 * @node: node whose cpumask we're constructing
6801 * @span: resulting cpumask
6803 * Given a node, construct a good cpumask for its sched_domain to span. It
6804 * should be one that prevents unnecessary balancing, but also spreads tasks
6807 static void sched_domain_node_span(int node, struct cpumask *span)
6809 nodemask_t used_nodes;
6812 cpumask_clear(span);
6813 nodes_clear(used_nodes);
6815 cpumask_or(span, span, cpumask_of_node(node));
6816 node_set(node, used_nodes);
6818 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6819 int next_node = find_next_best_node(node, &used_nodes);
6821 cpumask_or(span, span, cpumask_of_node(next_node));
6824 #endif /* CONFIG_NUMA */
6826 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6829 * The cpus mask in sched_group and sched_domain hangs off the end.
6831 * ( See the the comments in include/linux/sched.h:struct sched_group
6832 * and struct sched_domain. )
6834 struct static_sched_group {
6835 struct sched_group sg;
6836 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
6839 struct static_sched_domain {
6840 struct sched_domain sd;
6841 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
6847 cpumask_var_t domainspan;
6848 cpumask_var_t covered;
6849 cpumask_var_t notcovered;
6851 cpumask_var_t nodemask;
6852 cpumask_var_t this_sibling_map;
6853 cpumask_var_t this_core_map;
6854 cpumask_var_t this_book_map;
6855 cpumask_var_t send_covered;
6856 cpumask_var_t tmpmask;
6857 struct sched_group **sched_group_nodes;
6858 struct root_domain *rd;
6862 sa_sched_groups = 0,
6868 sa_this_sibling_map,
6870 sa_sched_group_nodes,
6880 * SMT sched-domains:
6882 #ifdef CONFIG_SCHED_SMT
6883 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
6884 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
6887 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
6888 struct sched_group **sg, struct cpumask *unused)
6891 *sg = &per_cpu(sched_groups, cpu).sg;
6894 #endif /* CONFIG_SCHED_SMT */
6897 * multi-core sched-domains:
6899 #ifdef CONFIG_SCHED_MC
6900 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
6901 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
6904 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6905 struct sched_group **sg, struct cpumask *mask)
6908 #ifdef CONFIG_SCHED_SMT
6909 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6910 group = cpumask_first(mask);
6915 *sg = &per_cpu(sched_group_core, group).sg;
6918 #endif /* CONFIG_SCHED_MC */
6921 * book sched-domains:
6923 #ifdef CONFIG_SCHED_BOOK
6924 static DEFINE_PER_CPU(struct static_sched_domain, book_domains);
6925 static DEFINE_PER_CPU(struct static_sched_group, sched_group_book);
6928 cpu_to_book_group(int cpu, const struct cpumask *cpu_map,
6929 struct sched_group **sg, struct cpumask *mask)
6932 #ifdef CONFIG_SCHED_MC
6933 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6934 group = cpumask_first(mask);
6935 #elif defined(CONFIG_SCHED_SMT)
6936 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6937 group = cpumask_first(mask);
6940 *sg = &per_cpu(sched_group_book, group).sg;
6943 #endif /* CONFIG_SCHED_BOOK */
6945 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
6946 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
6949 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
6950 struct sched_group **sg, struct cpumask *mask)
6953 #ifdef CONFIG_SCHED_BOOK
6954 cpumask_and(mask, cpu_book_mask(cpu), cpu_map);
6955 group = cpumask_first(mask);
6956 #elif defined(CONFIG_SCHED_MC)
6957 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6958 group = cpumask_first(mask);
6959 #elif defined(CONFIG_SCHED_SMT)
6960 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6961 group = cpumask_first(mask);
6966 *sg = &per_cpu(sched_group_phys, group).sg;
6972 * The init_sched_build_groups can't handle what we want to do with node
6973 * groups, so roll our own. Now each node has its own list of groups which
6974 * gets dynamically allocated.
6976 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
6977 static struct sched_group ***sched_group_nodes_bycpu;
6979 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
6980 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
6982 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
6983 struct sched_group **sg,
6984 struct cpumask *nodemask)
6988 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
6989 group = cpumask_first(nodemask);
6992 *sg = &per_cpu(sched_group_allnodes, group).sg;
6996 static void init_numa_sched_groups_power(struct sched_group *group_head)
6998 struct sched_group *sg = group_head;
7004 for_each_cpu(j, sched_group_cpus(sg)) {
7005 struct sched_domain *sd;
7007 sd = &per_cpu(phys_domains, j).sd;
7008 if (j != group_first_cpu(sd->groups)) {
7010 * Only add "power" once for each
7016 sg->cpu_power += sd->groups->cpu_power;
7019 } while (sg != group_head);
7022 static int build_numa_sched_groups(struct s_data *d,
7023 const struct cpumask *cpu_map, int num)
7025 struct sched_domain *sd;
7026 struct sched_group *sg, *prev;
7029 cpumask_clear(d->covered);
7030 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
7031 if (cpumask_empty(d->nodemask)) {
7032 d->sched_group_nodes[num] = NULL;
7036 sched_domain_node_span(num, d->domainspan);
7037 cpumask_and(d->domainspan, d->domainspan, cpu_map);
7039 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
7042 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
7046 d->sched_group_nodes[num] = sg;
7048 for_each_cpu(j, d->nodemask) {
7049 sd = &per_cpu(node_domains, j).sd;
7054 cpumask_copy(sched_group_cpus(sg), d->nodemask);
7056 cpumask_or(d->covered, d->covered, d->nodemask);
7059 for (j = 0; j < nr_node_ids; j++) {
7060 n = (num + j) % nr_node_ids;
7061 cpumask_complement(d->notcovered, d->covered);
7062 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
7063 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
7064 if (cpumask_empty(d->tmpmask))
7066 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
7067 if (cpumask_empty(d->tmpmask))
7069 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
7073 "Can not alloc domain group for node %d\n", j);
7077 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
7078 sg->next = prev->next;
7079 cpumask_or(d->covered, d->covered, d->tmpmask);
7086 #endif /* CONFIG_NUMA */
7089 /* Free memory allocated for various sched_group structures */
7090 static void free_sched_groups(const struct cpumask *cpu_map,
7091 struct cpumask *nodemask)
7095 for_each_cpu(cpu, cpu_map) {
7096 struct sched_group **sched_group_nodes
7097 = sched_group_nodes_bycpu[cpu];
7099 if (!sched_group_nodes)
7102 for (i = 0; i < nr_node_ids; i++) {
7103 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7105 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
7106 if (cpumask_empty(nodemask))
7116 if (oldsg != sched_group_nodes[i])
7119 kfree(sched_group_nodes);
7120 sched_group_nodes_bycpu[cpu] = NULL;
7123 #else /* !CONFIG_NUMA */
7124 static void free_sched_groups(const struct cpumask *cpu_map,
7125 struct cpumask *nodemask)
7128 #endif /* CONFIG_NUMA */
7131 * Initialize sched groups cpu_power.
7133 * cpu_power indicates the capacity of sched group, which is used while
7134 * distributing the load between different sched groups in a sched domain.
7135 * Typically cpu_power for all the groups in a sched domain will be same unless
7136 * there are asymmetries in the topology. If there are asymmetries, group
7137 * having more cpu_power will pickup more load compared to the group having
7140 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7142 struct sched_domain *child;
7143 struct sched_group *group;
7147 WARN_ON(!sd || !sd->groups);
7149 if (cpu != group_first_cpu(sd->groups))
7152 sd->groups->group_weight = cpumask_weight(sched_group_cpus(sd->groups));
7156 sd->groups->cpu_power = 0;
7159 power = SCHED_LOAD_SCALE;
7160 weight = cpumask_weight(sched_domain_span(sd));
7162 * SMT siblings share the power of a single core.
7163 * Usually multiple threads get a better yield out of
7164 * that one core than a single thread would have,
7165 * reflect that in sd->smt_gain.
7167 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
7168 power *= sd->smt_gain;
7170 power >>= SCHED_LOAD_SHIFT;
7172 sd->groups->cpu_power += power;
7177 * Add cpu_power of each child group to this groups cpu_power.
7179 group = child->groups;
7181 sd->groups->cpu_power += group->cpu_power;
7182 group = group->next;
7183 } while (group != child->groups);
7187 * Initializers for schedule domains
7188 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7191 #ifdef CONFIG_SCHED_DEBUG
7192 # define SD_INIT_NAME(sd, type) sd->name = #type
7194 # define SD_INIT_NAME(sd, type) do { } while (0)
7197 #define SD_INIT(sd, type) sd_init_##type(sd)
7199 #define SD_INIT_FUNC(type) \
7200 static noinline void sd_init_##type(struct sched_domain *sd) \
7202 memset(sd, 0, sizeof(*sd)); \
7203 *sd = SD_##type##_INIT; \
7204 sd->level = SD_LV_##type; \
7205 SD_INIT_NAME(sd, type); \
7210 SD_INIT_FUNC(ALLNODES)
7213 #ifdef CONFIG_SCHED_SMT
7214 SD_INIT_FUNC(SIBLING)
7216 #ifdef CONFIG_SCHED_MC
7219 #ifdef CONFIG_SCHED_BOOK
7223 static int default_relax_domain_level = -1;
7225 static int __init setup_relax_domain_level(char *str)
7229 val = simple_strtoul(str, NULL, 0);
7230 if (val < SD_LV_MAX)
7231 default_relax_domain_level = val;
7235 __setup("relax_domain_level=", setup_relax_domain_level);
7237 static void set_domain_attribute(struct sched_domain *sd,
7238 struct sched_domain_attr *attr)
7242 if (!attr || attr->relax_domain_level < 0) {
7243 if (default_relax_domain_level < 0)
7246 request = default_relax_domain_level;
7248 request = attr->relax_domain_level;
7249 if (request < sd->level) {
7250 /* turn off idle balance on this domain */
7251 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7253 /* turn on idle balance on this domain */
7254 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7258 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
7259 const struct cpumask *cpu_map)
7262 case sa_sched_groups:
7263 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
7264 d->sched_group_nodes = NULL;
7266 free_rootdomain(d->rd); /* fall through */
7268 free_cpumask_var(d->tmpmask); /* fall through */
7269 case sa_send_covered:
7270 free_cpumask_var(d->send_covered); /* fall through */
7271 case sa_this_book_map:
7272 free_cpumask_var(d->this_book_map); /* fall through */
7273 case sa_this_core_map:
7274 free_cpumask_var(d->this_core_map); /* fall through */
7275 case sa_this_sibling_map:
7276 free_cpumask_var(d->this_sibling_map); /* fall through */
7278 free_cpumask_var(d->nodemask); /* fall through */
7279 case sa_sched_group_nodes:
7281 kfree(d->sched_group_nodes); /* fall through */
7283 free_cpumask_var(d->notcovered); /* fall through */
7285 free_cpumask_var(d->covered); /* fall through */
7287 free_cpumask_var(d->domainspan); /* fall through */
7294 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
7295 const struct cpumask *cpu_map)
7298 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
7300 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
7301 return sa_domainspan;
7302 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
7304 /* Allocate the per-node list of sched groups */
7305 d->sched_group_nodes = kcalloc(nr_node_ids,
7306 sizeof(struct sched_group *), GFP_KERNEL);
7307 if (!d->sched_group_nodes) {
7308 printk(KERN_WARNING "Can not alloc sched group node list\n");
7309 return sa_notcovered;
7311 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
7313 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
7314 return sa_sched_group_nodes;
7315 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
7317 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
7318 return sa_this_sibling_map;
7319 if (!alloc_cpumask_var(&d->this_book_map, GFP_KERNEL))
7320 return sa_this_core_map;
7321 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
7322 return sa_this_book_map;
7323 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
7324 return sa_send_covered;
7325 d->rd = alloc_rootdomain();
7327 printk(KERN_WARNING "Cannot alloc root domain\n");
7330 return sa_rootdomain;
7333 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
7334 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
7336 struct sched_domain *sd = NULL;
7338 struct sched_domain *parent;
7341 if (cpumask_weight(cpu_map) >
7342 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
7343 sd = &per_cpu(allnodes_domains, i).sd;
7344 SD_INIT(sd, ALLNODES);
7345 set_domain_attribute(sd, attr);
7346 cpumask_copy(sched_domain_span(sd), cpu_map);
7347 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
7352 sd = &per_cpu(node_domains, i).sd;
7354 set_domain_attribute(sd, attr);
7355 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
7356 sd->parent = parent;
7359 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
7364 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
7365 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7366 struct sched_domain *parent, int i)
7368 struct sched_domain *sd;
7369 sd = &per_cpu(phys_domains, i).sd;
7371 set_domain_attribute(sd, attr);
7372 cpumask_copy(sched_domain_span(sd), d->nodemask);
7373 sd->parent = parent;
7376 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
7380 static struct sched_domain *__build_book_sched_domain(struct s_data *d,
7381 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7382 struct sched_domain *parent, int i)
7384 struct sched_domain *sd = parent;
7385 #ifdef CONFIG_SCHED_BOOK
7386 sd = &per_cpu(book_domains, i).sd;
7388 set_domain_attribute(sd, attr);
7389 cpumask_and(sched_domain_span(sd), cpu_map, cpu_book_mask(i));
7390 sd->parent = parent;
7392 cpu_to_book_group(i, cpu_map, &sd->groups, d->tmpmask);
7397 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
7398 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7399 struct sched_domain *parent, int i)
7401 struct sched_domain *sd = parent;
7402 #ifdef CONFIG_SCHED_MC
7403 sd = &per_cpu(core_domains, i).sd;
7405 set_domain_attribute(sd, attr);
7406 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
7407 sd->parent = parent;
7409 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
7414 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
7415 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7416 struct sched_domain *parent, int i)
7418 struct sched_domain *sd = parent;
7419 #ifdef CONFIG_SCHED_SMT
7420 sd = &per_cpu(cpu_domains, i).sd;
7421 SD_INIT(sd, SIBLING);
7422 set_domain_attribute(sd, attr);
7423 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
7424 sd->parent = parent;
7426 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
7431 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
7432 const struct cpumask *cpu_map, int cpu)
7435 #ifdef CONFIG_SCHED_SMT
7436 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
7437 cpumask_and(d->this_sibling_map, cpu_map,
7438 topology_thread_cpumask(cpu));
7439 if (cpu == cpumask_first(d->this_sibling_map))
7440 init_sched_build_groups(d->this_sibling_map, cpu_map,
7442 d->send_covered, d->tmpmask);
7445 #ifdef CONFIG_SCHED_MC
7446 case SD_LV_MC: /* set up multi-core groups */
7447 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
7448 if (cpu == cpumask_first(d->this_core_map))
7449 init_sched_build_groups(d->this_core_map, cpu_map,
7451 d->send_covered, d->tmpmask);
7454 #ifdef CONFIG_SCHED_BOOK
7455 case SD_LV_BOOK: /* set up book groups */
7456 cpumask_and(d->this_book_map, cpu_map, cpu_book_mask(cpu));
7457 if (cpu == cpumask_first(d->this_book_map))
7458 init_sched_build_groups(d->this_book_map, cpu_map,
7460 d->send_covered, d->tmpmask);
7463 case SD_LV_CPU: /* set up physical groups */
7464 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
7465 if (!cpumask_empty(d->nodemask))
7466 init_sched_build_groups(d->nodemask, cpu_map,
7468 d->send_covered, d->tmpmask);
7471 case SD_LV_ALLNODES:
7472 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
7473 d->send_covered, d->tmpmask);
7482 * Build sched domains for a given set of cpus and attach the sched domains
7483 * to the individual cpus
7485 static int __build_sched_domains(const struct cpumask *cpu_map,
7486 struct sched_domain_attr *attr)
7488 enum s_alloc alloc_state = sa_none;
7490 struct sched_domain *sd;
7496 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7497 if (alloc_state != sa_rootdomain)
7499 alloc_state = sa_sched_groups;
7502 * Set up domains for cpus specified by the cpu_map.
7504 for_each_cpu(i, cpu_map) {
7505 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
7508 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
7509 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
7510 sd = __build_book_sched_domain(&d, cpu_map, attr, sd, i);
7511 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
7512 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
7515 for_each_cpu(i, cpu_map) {
7516 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
7517 build_sched_groups(&d, SD_LV_BOOK, cpu_map, i);
7518 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
7521 /* Set up physical groups */
7522 for (i = 0; i < nr_node_ids; i++)
7523 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
7526 /* Set up node groups */
7528 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
7530 for (i = 0; i < nr_node_ids; i++)
7531 if (build_numa_sched_groups(&d, cpu_map, i))
7535 /* Calculate CPU power for physical packages and nodes */
7536 #ifdef CONFIG_SCHED_SMT
7537 for_each_cpu(i, cpu_map) {
7538 sd = &per_cpu(cpu_domains, i).sd;
7539 init_sched_groups_power(i, sd);
7542 #ifdef CONFIG_SCHED_MC
7543 for_each_cpu(i, cpu_map) {
7544 sd = &per_cpu(core_domains, i).sd;
7545 init_sched_groups_power(i, sd);
7548 #ifdef CONFIG_SCHED_BOOK
7549 for_each_cpu(i, cpu_map) {
7550 sd = &per_cpu(book_domains, i).sd;
7551 init_sched_groups_power(i, sd);
7555 for_each_cpu(i, cpu_map) {
7556 sd = &per_cpu(phys_domains, i).sd;
7557 init_sched_groups_power(i, sd);
7561 for (i = 0; i < nr_node_ids; i++)
7562 init_numa_sched_groups_power(d.sched_group_nodes[i]);
7564 if (d.sd_allnodes) {
7565 struct sched_group *sg;
7567 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7569 init_numa_sched_groups_power(sg);
7573 /* Attach the domains */
7574 for_each_cpu(i, cpu_map) {
7575 #ifdef CONFIG_SCHED_SMT
7576 sd = &per_cpu(cpu_domains, i).sd;
7577 #elif defined(CONFIG_SCHED_MC)
7578 sd = &per_cpu(core_domains, i).sd;
7579 #elif defined(CONFIG_SCHED_BOOK)
7580 sd = &per_cpu(book_domains, i).sd;
7582 sd = &per_cpu(phys_domains, i).sd;
7584 cpu_attach_domain(sd, d.rd, i);
7587 d.sched_group_nodes = NULL; /* don't free this we still need it */
7588 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
7592 __free_domain_allocs(&d, alloc_state, cpu_map);
7596 static int build_sched_domains(const struct cpumask *cpu_map)
7598 return __build_sched_domains(cpu_map, NULL);
7601 static cpumask_var_t *doms_cur; /* current sched domains */
7602 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7603 static struct sched_domain_attr *dattr_cur;
7604 /* attribues of custom domains in 'doms_cur' */
7607 * Special case: If a kmalloc of a doms_cur partition (array of
7608 * cpumask) fails, then fallback to a single sched domain,
7609 * as determined by the single cpumask fallback_doms.
7611 static cpumask_var_t fallback_doms;
7614 * arch_update_cpu_topology lets virtualized architectures update the
7615 * cpu core maps. It is supposed to return 1 if the topology changed
7616 * or 0 if it stayed the same.
7618 int __attribute__((weak)) arch_update_cpu_topology(void)
7623 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7626 cpumask_var_t *doms;
7628 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7631 for (i = 0; i < ndoms; i++) {
7632 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7633 free_sched_domains(doms, i);
7640 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7643 for (i = 0; i < ndoms; i++)
7644 free_cpumask_var(doms[i]);
7649 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7650 * For now this just excludes isolated cpus, but could be used to
7651 * exclude other special cases in the future.
7653 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7657 arch_update_cpu_topology();
7659 doms_cur = alloc_sched_domains(ndoms_cur);
7661 doms_cur = &fallback_doms;
7662 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7664 err = build_sched_domains(doms_cur[0]);
7665 register_sched_domain_sysctl();
7670 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7671 struct cpumask *tmpmask)
7673 free_sched_groups(cpu_map, tmpmask);
7677 * Detach sched domains from a group of cpus specified in cpu_map
7678 * These cpus will now be attached to the NULL domain
7680 static void detach_destroy_domains(const struct cpumask *cpu_map)
7682 /* Save because hotplug lock held. */
7683 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7686 for_each_cpu(i, cpu_map)
7687 cpu_attach_domain(NULL, &def_root_domain, i);
7688 synchronize_sched();
7689 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7692 /* handle null as "default" */
7693 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7694 struct sched_domain_attr *new, int idx_new)
7696 struct sched_domain_attr tmp;
7703 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7704 new ? (new + idx_new) : &tmp,
7705 sizeof(struct sched_domain_attr));
7709 * Partition sched domains as specified by the 'ndoms_new'
7710 * cpumasks in the array doms_new[] of cpumasks. This compares
7711 * doms_new[] to the current sched domain partitioning, doms_cur[].
7712 * It destroys each deleted domain and builds each new domain.
7714 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7715 * The masks don't intersect (don't overlap.) We should setup one
7716 * sched domain for each mask. CPUs not in any of the cpumasks will
7717 * not be load balanced. If the same cpumask appears both in the
7718 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7721 * The passed in 'doms_new' should be allocated using
7722 * alloc_sched_domains. This routine takes ownership of it and will
7723 * free_sched_domains it when done with it. If the caller failed the
7724 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7725 * and partition_sched_domains() will fallback to the single partition
7726 * 'fallback_doms', it also forces the domains to be rebuilt.
7728 * If doms_new == NULL it will be replaced with cpu_online_mask.
7729 * ndoms_new == 0 is a special case for destroying existing domains,
7730 * and it will not create the default domain.
7732 * Call with hotplug lock held
7734 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7735 struct sched_domain_attr *dattr_new)
7740 mutex_lock(&sched_domains_mutex);
7742 /* always unregister in case we don't destroy any domains */
7743 unregister_sched_domain_sysctl();
7745 /* Let architecture update cpu core mappings. */
7746 new_topology = arch_update_cpu_topology();
7748 n = doms_new ? ndoms_new : 0;
7750 /* Destroy deleted domains */
7751 for (i = 0; i < ndoms_cur; i++) {
7752 for (j = 0; j < n && !new_topology; j++) {
7753 if (cpumask_equal(doms_cur[i], doms_new[j])
7754 && dattrs_equal(dattr_cur, i, dattr_new, j))
7757 /* no match - a current sched domain not in new doms_new[] */
7758 detach_destroy_domains(doms_cur[i]);
7763 if (doms_new == NULL) {
7765 doms_new = &fallback_doms;
7766 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7767 WARN_ON_ONCE(dattr_new);
7770 /* Build new domains */
7771 for (i = 0; i < ndoms_new; i++) {
7772 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7773 if (cpumask_equal(doms_new[i], doms_cur[j])
7774 && dattrs_equal(dattr_new, i, dattr_cur, j))
7777 /* no match - add a new doms_new */
7778 __build_sched_domains(doms_new[i],
7779 dattr_new ? dattr_new + i : NULL);
7784 /* Remember the new sched domains */
7785 if (doms_cur != &fallback_doms)
7786 free_sched_domains(doms_cur, ndoms_cur);
7787 kfree(dattr_cur); /* kfree(NULL) is safe */
7788 doms_cur = doms_new;
7789 dattr_cur = dattr_new;
7790 ndoms_cur = ndoms_new;
7792 register_sched_domain_sysctl();
7794 mutex_unlock(&sched_domains_mutex);
7797 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7798 static void arch_reinit_sched_domains(void)
7802 /* Destroy domains first to force the rebuild */
7803 partition_sched_domains(0, NULL, NULL);
7805 rebuild_sched_domains();
7809 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7811 unsigned int level = 0;
7813 if (sscanf(buf, "%u", &level) != 1)
7817 * level is always be positive so don't check for
7818 * level < POWERSAVINGS_BALANCE_NONE which is 0
7819 * What happens on 0 or 1 byte write,
7820 * need to check for count as well?
7823 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7827 sched_smt_power_savings = level;
7829 sched_mc_power_savings = level;
7831 arch_reinit_sched_domains();
7836 #ifdef CONFIG_SCHED_MC
7837 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7838 struct sysdev_class_attribute *attr,
7841 return sprintf(page, "%u\n", sched_mc_power_savings);
7843 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7844 struct sysdev_class_attribute *attr,
7845 const char *buf, size_t count)
7847 return sched_power_savings_store(buf, count, 0);
7849 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7850 sched_mc_power_savings_show,
7851 sched_mc_power_savings_store);
7854 #ifdef CONFIG_SCHED_SMT
7855 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7856 struct sysdev_class_attribute *attr,
7859 return sprintf(page, "%u\n", sched_smt_power_savings);
7861 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7862 struct sysdev_class_attribute *attr,
7863 const char *buf, size_t count)
7865 return sched_power_savings_store(buf, count, 1);
7867 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7868 sched_smt_power_savings_show,
7869 sched_smt_power_savings_store);
7872 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7876 #ifdef CONFIG_SCHED_SMT
7878 err = sysfs_create_file(&cls->kset.kobj,
7879 &attr_sched_smt_power_savings.attr);
7881 #ifdef CONFIG_SCHED_MC
7882 if (!err && mc_capable())
7883 err = sysfs_create_file(&cls->kset.kobj,
7884 &attr_sched_mc_power_savings.attr);
7888 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7891 * Update cpusets according to cpu_active mask. If cpusets are
7892 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7893 * around partition_sched_domains().
7895 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7898 switch (action & ~CPU_TASKS_FROZEN) {
7900 case CPU_DOWN_FAILED:
7901 cpuset_update_active_cpus();
7908 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7911 switch (action & ~CPU_TASKS_FROZEN) {
7912 case CPU_DOWN_PREPARE:
7913 cpuset_update_active_cpus();
7920 static int update_runtime(struct notifier_block *nfb,
7921 unsigned long action, void *hcpu)
7923 int cpu = (int)(long)hcpu;
7926 case CPU_DOWN_PREPARE:
7927 case CPU_DOWN_PREPARE_FROZEN:
7928 disable_runtime(cpu_rq(cpu));
7931 case CPU_DOWN_FAILED:
7932 case CPU_DOWN_FAILED_FROZEN:
7934 case CPU_ONLINE_FROZEN:
7935 enable_runtime(cpu_rq(cpu));
7943 void __init sched_init_smp(void)
7945 cpumask_var_t non_isolated_cpus;
7947 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7948 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7950 #if defined(CONFIG_NUMA)
7951 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7953 BUG_ON(sched_group_nodes_bycpu == NULL);
7956 mutex_lock(&sched_domains_mutex);
7957 arch_init_sched_domains(cpu_active_mask);
7958 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7959 if (cpumask_empty(non_isolated_cpus))
7960 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7961 mutex_unlock(&sched_domains_mutex);
7964 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7965 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7967 /* RT runtime code needs to handle some hotplug events */
7968 hotcpu_notifier(update_runtime, 0);
7972 /* Move init over to a non-isolated CPU */
7973 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7975 sched_init_granularity();
7976 free_cpumask_var(non_isolated_cpus);
7978 init_sched_rt_class();
7981 void __init sched_init_smp(void)
7983 sched_init_granularity();
7985 #endif /* CONFIG_SMP */
7987 const_debug unsigned int sysctl_timer_migration = 1;
7989 int in_sched_functions(unsigned long addr)
7991 return in_lock_functions(addr) ||
7992 (addr >= (unsigned long)__sched_text_start
7993 && addr < (unsigned long)__sched_text_end);
7996 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7998 cfs_rq->tasks_timeline = RB_ROOT;
7999 INIT_LIST_HEAD(&cfs_rq->tasks);
8000 #ifdef CONFIG_FAIR_GROUP_SCHED
8002 /* allow initial update_cfs_load() to truncate */
8004 cfs_rq->load_stamp = 1;
8007 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8010 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
8012 struct rt_prio_array *array;
8015 array = &rt_rq->active;
8016 for (i = 0; i < MAX_RT_PRIO; i++) {
8017 INIT_LIST_HEAD(array->queue + i);
8018 __clear_bit(i, array->bitmap);
8020 /* delimiter for bitsearch: */
8021 __set_bit(MAX_RT_PRIO, array->bitmap);
8023 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8024 rt_rq->highest_prio.curr = MAX_RT_PRIO;
8026 rt_rq->highest_prio.next = MAX_RT_PRIO;
8030 rt_rq->rt_nr_migratory = 0;
8031 rt_rq->overloaded = 0;
8032 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
8036 rt_rq->rt_throttled = 0;
8037 rt_rq->rt_runtime = 0;
8038 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
8040 #ifdef CONFIG_RT_GROUP_SCHED
8041 rt_rq->rt_nr_boosted = 0;
8046 #ifdef CONFIG_FAIR_GROUP_SCHED
8047 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8048 struct sched_entity *se, int cpu,
8049 struct sched_entity *parent)
8051 struct rq *rq = cpu_rq(cpu);
8052 tg->cfs_rq[cpu] = cfs_rq;
8053 init_cfs_rq(cfs_rq, rq);
8057 /* se could be NULL for root_task_group */
8062 se->cfs_rq = &rq->cfs;
8064 se->cfs_rq = parent->my_q;
8067 update_load_set(&se->load, 0);
8068 se->parent = parent;
8072 #ifdef CONFIG_RT_GROUP_SCHED
8073 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8074 struct sched_rt_entity *rt_se, int cpu,
8075 struct sched_rt_entity *parent)
8077 struct rq *rq = cpu_rq(cpu);
8079 tg->rt_rq[cpu] = rt_rq;
8080 init_rt_rq(rt_rq, rq);
8082 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8084 tg->rt_se[cpu] = rt_se;
8089 rt_se->rt_rq = &rq->rt;
8091 rt_se->rt_rq = parent->my_q;
8093 rt_se->my_q = rt_rq;
8094 rt_se->parent = parent;
8095 INIT_LIST_HEAD(&rt_se->run_list);
8099 void __init sched_init(void)
8102 unsigned long alloc_size = 0, ptr;
8104 #ifdef CONFIG_FAIR_GROUP_SCHED
8105 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8107 #ifdef CONFIG_RT_GROUP_SCHED
8108 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8110 #ifdef CONFIG_CPUMASK_OFFSTACK
8111 alloc_size += num_possible_cpus() * cpumask_size();
8114 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
8116 #ifdef CONFIG_FAIR_GROUP_SCHED
8117 root_task_group.se = (struct sched_entity **)ptr;
8118 ptr += nr_cpu_ids * sizeof(void **);
8120 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8121 ptr += nr_cpu_ids * sizeof(void **);
8123 #endif /* CONFIG_FAIR_GROUP_SCHED */
8124 #ifdef CONFIG_RT_GROUP_SCHED
8125 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8126 ptr += nr_cpu_ids * sizeof(void **);
8128 root_task_group.rt_rq = (struct rt_rq **)ptr;
8129 ptr += nr_cpu_ids * sizeof(void **);
8131 #endif /* CONFIG_RT_GROUP_SCHED */
8132 #ifdef CONFIG_CPUMASK_OFFSTACK
8133 for_each_possible_cpu(i) {
8134 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
8135 ptr += cpumask_size();
8137 #endif /* CONFIG_CPUMASK_OFFSTACK */
8141 init_defrootdomain();
8144 init_rt_bandwidth(&def_rt_bandwidth,
8145 global_rt_period(), global_rt_runtime());
8147 #ifdef CONFIG_RT_GROUP_SCHED
8148 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8149 global_rt_period(), global_rt_runtime());
8150 #endif /* CONFIG_RT_GROUP_SCHED */
8152 #ifdef CONFIG_CGROUP_SCHED
8153 list_add(&root_task_group.list, &task_groups);
8154 INIT_LIST_HEAD(&root_task_group.children);
8155 autogroup_init(&init_task);
8156 #endif /* CONFIG_CGROUP_SCHED */
8158 for_each_possible_cpu(i) {
8162 raw_spin_lock_init(&rq->lock);
8164 rq->calc_load_active = 0;
8165 rq->calc_load_update = jiffies + LOAD_FREQ;
8166 init_cfs_rq(&rq->cfs, rq);
8167 init_rt_rq(&rq->rt, rq);
8168 #ifdef CONFIG_FAIR_GROUP_SCHED
8169 root_task_group.shares = root_task_group_load;
8170 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8172 * How much cpu bandwidth does root_task_group get?
8174 * In case of task-groups formed thr' the cgroup filesystem, it
8175 * gets 100% of the cpu resources in the system. This overall
8176 * system cpu resource is divided among the tasks of
8177 * root_task_group and its child task-groups in a fair manner,
8178 * based on each entity's (task or task-group's) weight
8179 * (se->load.weight).
8181 * In other words, if root_task_group has 10 tasks of weight
8182 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8183 * then A0's share of the cpu resource is:
8185 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8187 * We achieve this by letting root_task_group's tasks sit
8188 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
8190 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
8191 #endif /* CONFIG_FAIR_GROUP_SCHED */
8193 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8194 #ifdef CONFIG_RT_GROUP_SCHED
8195 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8196 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
8199 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8200 rq->cpu_load[j] = 0;
8202 rq->last_load_update_tick = jiffies;
8207 rq->cpu_power = SCHED_LOAD_SCALE;
8208 rq->post_schedule = 0;
8209 rq->active_balance = 0;
8210 rq->next_balance = jiffies;
8215 rq->avg_idle = 2*sysctl_sched_migration_cost;
8216 rq_attach_root(rq, &def_root_domain);
8218 rq->nohz_balance_kick = 0;
8219 init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb, i));
8223 atomic_set(&rq->nr_iowait, 0);
8226 set_load_weight(&init_task);
8228 #ifdef CONFIG_PREEMPT_NOTIFIERS
8229 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8233 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8236 #ifdef CONFIG_RT_MUTEXES
8237 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
8241 * The boot idle thread does lazy MMU switching as well:
8243 atomic_inc(&init_mm.mm_count);
8244 enter_lazy_tlb(&init_mm, current);
8247 * Make us the idle thread. Technically, schedule() should not be
8248 * called from this thread, however somewhere below it might be,
8249 * but because we are the idle thread, we just pick up running again
8250 * when this runqueue becomes "idle".
8252 init_idle(current, smp_processor_id());
8254 calc_load_update = jiffies + LOAD_FREQ;
8257 * During early bootup we pretend to be a normal task:
8259 current->sched_class = &fair_sched_class;
8261 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8262 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
8265 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8266 alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT);
8267 atomic_set(&nohz.load_balancer, nr_cpu_ids);
8268 atomic_set(&nohz.first_pick_cpu, nr_cpu_ids);
8269 atomic_set(&nohz.second_pick_cpu, nr_cpu_ids);
8271 /* May be allocated at isolcpus cmdline parse time */
8272 if (cpu_isolated_map == NULL)
8273 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
8276 scheduler_running = 1;
8279 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8280 static inline int preempt_count_equals(int preempt_offset)
8282 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
8284 return (nested == preempt_offset);
8287 void __might_sleep(const char *file, int line, int preempt_offset)
8290 static unsigned long prev_jiffy; /* ratelimiting */
8292 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
8293 system_state != SYSTEM_RUNNING || oops_in_progress)
8295 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8297 prev_jiffy = jiffies;
8300 "BUG: sleeping function called from invalid context at %s:%d\n",
8303 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8304 in_atomic(), irqs_disabled(),
8305 current->pid, current->comm);
8307 debug_show_held_locks(current);
8308 if (irqs_disabled())
8309 print_irqtrace_events(current);
8313 EXPORT_SYMBOL(__might_sleep);
8316 #ifdef CONFIG_MAGIC_SYSRQ
8317 static void normalize_task(struct rq *rq, struct task_struct *p)
8319 const struct sched_class *prev_class = p->sched_class;
8320 int old_prio = p->prio;
8325 deactivate_task(rq, p, 0);
8326 __setscheduler(rq, p, SCHED_NORMAL, 0);
8328 activate_task(rq, p, 0);
8329 resched_task(rq->curr);
8332 check_class_changed(rq, p, prev_class, old_prio);
8335 void normalize_rt_tasks(void)
8337 struct task_struct *g, *p;
8338 unsigned long flags;
8341 read_lock_irqsave(&tasklist_lock, flags);
8342 do_each_thread(g, p) {
8344 * Only normalize user tasks:
8349 p->se.exec_start = 0;
8350 #ifdef CONFIG_SCHEDSTATS
8351 p->se.statistics.wait_start = 0;
8352 p->se.statistics.sleep_start = 0;
8353 p->se.statistics.block_start = 0;
8358 * Renice negative nice level userspace
8361 if (TASK_NICE(p) < 0 && p->mm)
8362 set_user_nice(p, 0);
8366 raw_spin_lock(&p->pi_lock);
8367 rq = __task_rq_lock(p);
8369 normalize_task(rq, p);
8371 __task_rq_unlock(rq);
8372 raw_spin_unlock(&p->pi_lock);
8373 } while_each_thread(g, p);
8375 read_unlock_irqrestore(&tasklist_lock, flags);
8378 #endif /* CONFIG_MAGIC_SYSRQ */
8380 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8382 * These functions are only useful for the IA64 MCA handling, or kdb.
8384 * They can only be called when the whole system has been
8385 * stopped - every CPU needs to be quiescent, and no scheduling
8386 * activity can take place. Using them for anything else would
8387 * be a serious bug, and as a result, they aren't even visible
8388 * under any other configuration.
8392 * curr_task - return the current task for a given cpu.
8393 * @cpu: the processor in question.
8395 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8397 struct task_struct *curr_task(int cpu)
8399 return cpu_curr(cpu);
8402 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8406 * set_curr_task - set the current task for a given cpu.
8407 * @cpu: the processor in question.
8408 * @p: the task pointer to set.
8410 * Description: This function must only be used when non-maskable interrupts
8411 * are serviced on a separate stack. It allows the architecture to switch the
8412 * notion of the current task on a cpu in a non-blocking manner. This function
8413 * must be called with all CPU's synchronized, and interrupts disabled, the
8414 * and caller must save the original value of the current task (see
8415 * curr_task() above) and restore that value before reenabling interrupts and
8416 * re-starting the system.
8418 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8420 void set_curr_task(int cpu, struct task_struct *p)
8427 #ifdef CONFIG_FAIR_GROUP_SCHED
8428 static void free_fair_sched_group(struct task_group *tg)
8432 for_each_possible_cpu(i) {
8434 kfree(tg->cfs_rq[i]);
8444 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8446 struct cfs_rq *cfs_rq;
8447 struct sched_entity *se;
8450 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8453 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8457 tg->shares = NICE_0_LOAD;
8459 for_each_possible_cpu(i) {
8460 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8461 GFP_KERNEL, cpu_to_node(i));
8465 se = kzalloc_node(sizeof(struct sched_entity),
8466 GFP_KERNEL, cpu_to_node(i));
8470 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8481 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8483 struct rq *rq = cpu_rq(cpu);
8484 unsigned long flags;
8487 * Only empty task groups can be destroyed; so we can speculatively
8488 * check on_list without danger of it being re-added.
8490 if (!tg->cfs_rq[cpu]->on_list)
8493 raw_spin_lock_irqsave(&rq->lock, flags);
8494 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8495 raw_spin_unlock_irqrestore(&rq->lock, flags);
8497 #else /* !CONFG_FAIR_GROUP_SCHED */
8498 static inline void free_fair_sched_group(struct task_group *tg)
8503 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8508 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8511 #endif /* CONFIG_FAIR_GROUP_SCHED */
8513 #ifdef CONFIG_RT_GROUP_SCHED
8514 static void free_rt_sched_group(struct task_group *tg)
8518 destroy_rt_bandwidth(&tg->rt_bandwidth);
8520 for_each_possible_cpu(i) {
8522 kfree(tg->rt_rq[i]);
8524 kfree(tg->rt_se[i]);
8532 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8534 struct rt_rq *rt_rq;
8535 struct sched_rt_entity *rt_se;
8539 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8542 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8546 init_rt_bandwidth(&tg->rt_bandwidth,
8547 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8549 for_each_possible_cpu(i) {
8552 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8553 GFP_KERNEL, cpu_to_node(i));
8557 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8558 GFP_KERNEL, cpu_to_node(i));
8562 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
8572 #else /* !CONFIG_RT_GROUP_SCHED */
8573 static inline void free_rt_sched_group(struct task_group *tg)
8578 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8582 #endif /* CONFIG_RT_GROUP_SCHED */
8584 #ifdef CONFIG_CGROUP_SCHED
8585 static void free_sched_group(struct task_group *tg)
8587 free_fair_sched_group(tg);
8588 free_rt_sched_group(tg);
8593 /* allocate runqueue etc for a new task group */
8594 struct task_group *sched_create_group(struct task_group *parent)
8596 struct task_group *tg;
8597 unsigned long flags;
8599 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8601 return ERR_PTR(-ENOMEM);
8603 if (!alloc_fair_sched_group(tg, parent))
8606 if (!alloc_rt_sched_group(tg, parent))
8609 spin_lock_irqsave(&task_group_lock, flags);
8610 list_add_rcu(&tg->list, &task_groups);
8612 WARN_ON(!parent); /* root should already exist */
8614 tg->parent = parent;
8615 INIT_LIST_HEAD(&tg->children);
8616 list_add_rcu(&tg->siblings, &parent->children);
8617 spin_unlock_irqrestore(&task_group_lock, flags);
8622 free_sched_group(tg);
8623 return ERR_PTR(-ENOMEM);
8626 /* rcu callback to free various structures associated with a task group */
8627 static void free_sched_group_rcu(struct rcu_head *rhp)
8629 /* now it should be safe to free those cfs_rqs */
8630 free_sched_group(container_of(rhp, struct task_group, rcu));
8633 /* Destroy runqueue etc associated with a task group */
8634 void sched_destroy_group(struct task_group *tg)
8636 unsigned long flags;
8639 /* end participation in shares distribution */
8640 for_each_possible_cpu(i)
8641 unregister_fair_sched_group(tg, i);
8643 spin_lock_irqsave(&task_group_lock, flags);
8644 list_del_rcu(&tg->list);
8645 list_del_rcu(&tg->siblings);
8646 spin_unlock_irqrestore(&task_group_lock, flags);
8648 /* wait for possible concurrent references to cfs_rqs complete */
8649 call_rcu(&tg->rcu, free_sched_group_rcu);
8652 /* change task's runqueue when it moves between groups.
8653 * The caller of this function should have put the task in its new group
8654 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8655 * reflect its new group.
8657 void sched_move_task(struct task_struct *tsk)
8660 unsigned long flags;
8663 rq = task_rq_lock(tsk, &flags);
8665 running = task_current(rq, tsk);
8669 dequeue_task(rq, tsk, 0);
8670 if (unlikely(running))
8671 tsk->sched_class->put_prev_task(rq, tsk);
8673 #ifdef CONFIG_FAIR_GROUP_SCHED
8674 if (tsk->sched_class->task_move_group)
8675 tsk->sched_class->task_move_group(tsk, on_rq);
8678 set_task_rq(tsk, task_cpu(tsk));
8680 if (unlikely(running))
8681 tsk->sched_class->set_curr_task(rq);
8683 enqueue_task(rq, tsk, 0);
8685 task_rq_unlock(rq, tsk, &flags);
8687 #endif /* CONFIG_CGROUP_SCHED */
8689 #ifdef CONFIG_FAIR_GROUP_SCHED
8690 static DEFINE_MUTEX(shares_mutex);
8692 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8695 unsigned long flags;
8698 * We can't change the weight of the root cgroup.
8703 if (shares < MIN_SHARES)
8704 shares = MIN_SHARES;
8705 else if (shares > MAX_SHARES)
8706 shares = MAX_SHARES;
8708 mutex_lock(&shares_mutex);
8709 if (tg->shares == shares)
8712 tg->shares = shares;
8713 for_each_possible_cpu(i) {
8714 struct rq *rq = cpu_rq(i);
8715 struct sched_entity *se;
8718 /* Propagate contribution to hierarchy */
8719 raw_spin_lock_irqsave(&rq->lock, flags);
8720 for_each_sched_entity(se)
8721 update_cfs_shares(group_cfs_rq(se));
8722 raw_spin_unlock_irqrestore(&rq->lock, flags);
8726 mutex_unlock(&shares_mutex);
8730 unsigned long sched_group_shares(struct task_group *tg)
8736 #ifdef CONFIG_RT_GROUP_SCHED
8738 * Ensure that the real time constraints are schedulable.
8740 static DEFINE_MUTEX(rt_constraints_mutex);
8742 static unsigned long to_ratio(u64 period, u64 runtime)
8744 if (runtime == RUNTIME_INF)
8747 return div64_u64(runtime << 20, period);
8750 /* Must be called with tasklist_lock held */
8751 static inline int tg_has_rt_tasks(struct task_group *tg)
8753 struct task_struct *g, *p;
8755 do_each_thread(g, p) {
8756 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8758 } while_each_thread(g, p);
8763 struct rt_schedulable_data {
8764 struct task_group *tg;
8769 static int tg_schedulable(struct task_group *tg, void *data)
8771 struct rt_schedulable_data *d = data;
8772 struct task_group *child;
8773 unsigned long total, sum = 0;
8774 u64 period, runtime;
8776 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8777 runtime = tg->rt_bandwidth.rt_runtime;
8780 period = d->rt_period;
8781 runtime = d->rt_runtime;
8785 * Cannot have more runtime than the period.
8787 if (runtime > period && runtime != RUNTIME_INF)
8791 * Ensure we don't starve existing RT tasks.
8793 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8796 total = to_ratio(period, runtime);
8799 * Nobody can have more than the global setting allows.
8801 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8805 * The sum of our children's runtime should not exceed our own.
8807 list_for_each_entry_rcu(child, &tg->children, siblings) {
8808 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8809 runtime = child->rt_bandwidth.rt_runtime;
8811 if (child == d->tg) {
8812 period = d->rt_period;
8813 runtime = d->rt_runtime;
8816 sum += to_ratio(period, runtime);
8825 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8827 struct rt_schedulable_data data = {
8829 .rt_period = period,
8830 .rt_runtime = runtime,
8833 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8836 static int tg_set_bandwidth(struct task_group *tg,
8837 u64 rt_period, u64 rt_runtime)
8841 mutex_lock(&rt_constraints_mutex);
8842 read_lock(&tasklist_lock);
8843 err = __rt_schedulable(tg, rt_period, rt_runtime);
8847 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8848 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8849 tg->rt_bandwidth.rt_runtime = rt_runtime;
8851 for_each_possible_cpu(i) {
8852 struct rt_rq *rt_rq = tg->rt_rq[i];
8854 raw_spin_lock(&rt_rq->rt_runtime_lock);
8855 rt_rq->rt_runtime = rt_runtime;
8856 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8858 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8860 read_unlock(&tasklist_lock);
8861 mutex_unlock(&rt_constraints_mutex);
8866 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8868 u64 rt_runtime, rt_period;
8870 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8871 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8872 if (rt_runtime_us < 0)
8873 rt_runtime = RUNTIME_INF;
8875 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8878 long sched_group_rt_runtime(struct task_group *tg)
8882 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8885 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8886 do_div(rt_runtime_us, NSEC_PER_USEC);
8887 return rt_runtime_us;
8890 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8892 u64 rt_runtime, rt_period;
8894 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8895 rt_runtime = tg->rt_bandwidth.rt_runtime;
8900 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8903 long sched_group_rt_period(struct task_group *tg)
8907 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8908 do_div(rt_period_us, NSEC_PER_USEC);
8909 return rt_period_us;
8912 static int sched_rt_global_constraints(void)
8914 u64 runtime, period;
8917 if (sysctl_sched_rt_period <= 0)
8920 runtime = global_rt_runtime();
8921 period = global_rt_period();
8924 * Sanity check on the sysctl variables.
8926 if (runtime > period && runtime != RUNTIME_INF)
8929 mutex_lock(&rt_constraints_mutex);
8930 read_lock(&tasklist_lock);
8931 ret = __rt_schedulable(NULL, 0, 0);
8932 read_unlock(&tasklist_lock);
8933 mutex_unlock(&rt_constraints_mutex);
8938 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8940 /* Don't accept realtime tasks when there is no way for them to run */
8941 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8947 #else /* !CONFIG_RT_GROUP_SCHED */
8948 static int sched_rt_global_constraints(void)
8950 unsigned long flags;
8953 if (sysctl_sched_rt_period <= 0)
8957 * There's always some RT tasks in the root group
8958 * -- migration, kstopmachine etc..
8960 if (sysctl_sched_rt_runtime == 0)
8963 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8964 for_each_possible_cpu(i) {
8965 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8967 raw_spin_lock(&rt_rq->rt_runtime_lock);
8968 rt_rq->rt_runtime = global_rt_runtime();
8969 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8971 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8975 #endif /* CONFIG_RT_GROUP_SCHED */
8977 int sched_rt_handler(struct ctl_table *table, int write,
8978 void __user *buffer, size_t *lenp,
8982 int old_period, old_runtime;
8983 static DEFINE_MUTEX(mutex);
8986 old_period = sysctl_sched_rt_period;
8987 old_runtime = sysctl_sched_rt_runtime;
8989 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8991 if (!ret && write) {
8992 ret = sched_rt_global_constraints();
8994 sysctl_sched_rt_period = old_period;
8995 sysctl_sched_rt_runtime = old_runtime;
8997 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8998 def_rt_bandwidth.rt_period =
8999 ns_to_ktime(global_rt_period());
9002 mutex_unlock(&mutex);
9007 #ifdef CONFIG_CGROUP_SCHED
9009 /* return corresponding task_group object of a cgroup */
9010 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9012 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9013 struct task_group, css);
9016 static struct cgroup_subsys_state *
9017 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9019 struct task_group *tg, *parent;
9021 if (!cgrp->parent) {
9022 /* This is early initialization for the top cgroup */
9023 return &root_task_group.css;
9026 parent = cgroup_tg(cgrp->parent);
9027 tg = sched_create_group(parent);
9029 return ERR_PTR(-ENOMEM);
9035 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9037 struct task_group *tg = cgroup_tg(cgrp);
9039 sched_destroy_group(tg);
9043 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
9045 #ifdef CONFIG_RT_GROUP_SCHED
9046 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
9049 /* We don't support RT-tasks being in separate groups */
9050 if (tsk->sched_class != &fair_sched_class)
9057 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9058 struct task_struct *tsk, bool threadgroup)
9060 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
9064 struct task_struct *c;
9066 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
9067 retval = cpu_cgroup_can_attach_task(cgrp, c);
9079 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9080 struct cgroup *old_cont, struct task_struct *tsk,
9083 sched_move_task(tsk);
9085 struct task_struct *c;
9087 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
9095 cpu_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
9096 struct cgroup *old_cgrp, struct task_struct *task)
9099 * cgroup_exit() is called in the copy_process() failure path.
9100 * Ignore this case since the task hasn't ran yet, this avoids
9101 * trying to poke a half freed task state from generic code.
9103 if (!(task->flags & PF_EXITING))
9106 sched_move_task(task);
9109 #ifdef CONFIG_FAIR_GROUP_SCHED
9110 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9113 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
9116 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9118 struct task_group *tg = cgroup_tg(cgrp);
9120 return (u64) tg->shares;
9122 #endif /* CONFIG_FAIR_GROUP_SCHED */
9124 #ifdef CONFIG_RT_GROUP_SCHED
9125 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9128 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9131 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9133 return sched_group_rt_runtime(cgroup_tg(cgrp));
9136 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9139 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9142 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9144 return sched_group_rt_period(cgroup_tg(cgrp));
9146 #endif /* CONFIG_RT_GROUP_SCHED */
9148 static struct cftype cpu_files[] = {
9149 #ifdef CONFIG_FAIR_GROUP_SCHED
9152 .read_u64 = cpu_shares_read_u64,
9153 .write_u64 = cpu_shares_write_u64,
9156 #ifdef CONFIG_RT_GROUP_SCHED
9158 .name = "rt_runtime_us",
9159 .read_s64 = cpu_rt_runtime_read,
9160 .write_s64 = cpu_rt_runtime_write,
9163 .name = "rt_period_us",
9164 .read_u64 = cpu_rt_period_read_uint,
9165 .write_u64 = cpu_rt_period_write_uint,
9170 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9172 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9175 struct cgroup_subsys cpu_cgroup_subsys = {
9177 .create = cpu_cgroup_create,
9178 .destroy = cpu_cgroup_destroy,
9179 .can_attach = cpu_cgroup_can_attach,
9180 .attach = cpu_cgroup_attach,
9181 .exit = cpu_cgroup_exit,
9182 .populate = cpu_cgroup_populate,
9183 .subsys_id = cpu_cgroup_subsys_id,
9187 #endif /* CONFIG_CGROUP_SCHED */
9189 #ifdef CONFIG_CGROUP_CPUACCT
9192 * CPU accounting code for task groups.
9194 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9195 * (balbir@in.ibm.com).
9198 /* track cpu usage of a group of tasks and its child groups */
9200 struct cgroup_subsys_state css;
9201 /* cpuusage holds pointer to a u64-type object on every cpu */
9202 u64 __percpu *cpuusage;
9203 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
9204 struct cpuacct *parent;
9207 struct cgroup_subsys cpuacct_subsys;
9209 /* return cpu accounting group corresponding to this container */
9210 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9212 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9213 struct cpuacct, css);
9216 /* return cpu accounting group to which this task belongs */
9217 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9219 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9220 struct cpuacct, css);
9223 /* create a new cpu accounting group */
9224 static struct cgroup_subsys_state *cpuacct_create(
9225 struct cgroup_subsys *ss, struct cgroup *cgrp)
9227 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9233 ca->cpuusage = alloc_percpu(u64);
9237 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9238 if (percpu_counter_init(&ca->cpustat[i], 0))
9239 goto out_free_counters;
9242 ca->parent = cgroup_ca(cgrp->parent);
9248 percpu_counter_destroy(&ca->cpustat[i]);
9249 free_percpu(ca->cpuusage);
9253 return ERR_PTR(-ENOMEM);
9256 /* destroy an existing cpu accounting group */
9258 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9260 struct cpuacct *ca = cgroup_ca(cgrp);
9263 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9264 percpu_counter_destroy(&ca->cpustat[i]);
9265 free_percpu(ca->cpuusage);
9269 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
9271 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9274 #ifndef CONFIG_64BIT
9276 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9278 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9280 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9288 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9290 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9292 #ifndef CONFIG_64BIT
9294 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9296 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9298 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9304 /* return total cpu usage (in nanoseconds) of a group */
9305 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9307 struct cpuacct *ca = cgroup_ca(cgrp);
9308 u64 totalcpuusage = 0;
9311 for_each_present_cpu(i)
9312 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9314 return totalcpuusage;
9317 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9320 struct cpuacct *ca = cgroup_ca(cgrp);
9329 for_each_present_cpu(i)
9330 cpuacct_cpuusage_write(ca, i, 0);
9336 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9339 struct cpuacct *ca = cgroup_ca(cgroup);
9343 for_each_present_cpu(i) {
9344 percpu = cpuacct_cpuusage_read(ca, i);
9345 seq_printf(m, "%llu ", (unsigned long long) percpu);
9347 seq_printf(m, "\n");
9351 static const char *cpuacct_stat_desc[] = {
9352 [CPUACCT_STAT_USER] = "user",
9353 [CPUACCT_STAT_SYSTEM] = "system",
9356 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
9357 struct cgroup_map_cb *cb)
9359 struct cpuacct *ca = cgroup_ca(cgrp);
9362 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
9363 s64 val = percpu_counter_read(&ca->cpustat[i]);
9364 val = cputime64_to_clock_t(val);
9365 cb->fill(cb, cpuacct_stat_desc[i], val);
9370 static struct cftype files[] = {
9373 .read_u64 = cpuusage_read,
9374 .write_u64 = cpuusage_write,
9377 .name = "usage_percpu",
9378 .read_seq_string = cpuacct_percpu_seq_read,
9382 .read_map = cpuacct_stats_show,
9386 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9388 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9392 * charge this task's execution time to its accounting group.
9394 * called with rq->lock held.
9396 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9401 if (unlikely(!cpuacct_subsys.active))
9404 cpu = task_cpu(tsk);
9410 for (; ca; ca = ca->parent) {
9411 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9412 *cpuusage += cputime;
9419 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9420 * in cputime_t units. As a result, cpuacct_update_stats calls
9421 * percpu_counter_add with values large enough to always overflow the
9422 * per cpu batch limit causing bad SMP scalability.
9424 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9425 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9426 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9429 #define CPUACCT_BATCH \
9430 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9432 #define CPUACCT_BATCH 0
9436 * Charge the system/user time to the task's accounting group.
9438 static void cpuacct_update_stats(struct task_struct *tsk,
9439 enum cpuacct_stat_index idx, cputime_t val)
9442 int batch = CPUACCT_BATCH;
9444 if (unlikely(!cpuacct_subsys.active))
9451 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9457 struct cgroup_subsys cpuacct_subsys = {
9459 .create = cpuacct_create,
9460 .destroy = cpuacct_destroy,
9461 .populate = cpuacct_populate,
9462 .subsys_id = cpuacct_subsys_id,
9464 #endif /* CONFIG_CGROUP_CPUACCT */