4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/perf_event.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/proc_fs.h>
59 #include <linux/seq_file.h>
60 #include <linux/stop_machine.h>
61 #include <linux/sysctl.h>
62 #include <linux/syscalls.h>
63 #include <linux/times.h>
64 #include <linux/tsacct_kern.h>
65 #include <linux/kprobes.h>
66 #include <linux/delayacct.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
73 #include <linux/ftrace.h>
74 #include <linux/slab.h>
75 #include <linux/cpuacct.h>
78 #include <asm/irq_regs.h>
79 #include <asm/mutex.h>
81 #include "sched_cpupri.h"
82 #include "workqueue_sched.h"
83 #include "sched_autogroup.h"
85 #define CREATE_TRACE_POINTS
86 #include <trace/events/sched.h>
89 * Convert user-nice values [ -20 ... 0 ... 19 ]
90 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
93 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
94 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
95 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
98 * 'User priority' is the nice value converted to something we
99 * can work with better when scaling various scheduler parameters,
100 * it's a [ 0 ... 39 ] range.
102 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
103 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
104 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
107 * Helpers for converting nanosecond timing to jiffy resolution
109 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
111 #define NICE_0_LOAD SCHED_LOAD_SCALE
112 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
115 * These are the 'tuning knobs' of the scheduler:
117 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
118 * Timeslices get refilled after they expire.
120 #define DEF_TIMESLICE (100 * HZ / 1000)
123 * single value that denotes runtime == period, ie unlimited time.
125 #define RUNTIME_INF ((u64)~0ULL)
127 static inline int rt_policy(int policy)
129 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
134 static inline int task_has_rt_policy(struct task_struct *p)
136 return rt_policy(p->policy);
140 * This is the priority-queue data structure of the RT scheduling class:
142 struct rt_prio_array {
143 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
144 struct list_head queue[MAX_RT_PRIO];
147 struct rt_bandwidth {
148 /* nests inside the rq lock: */
149 raw_spinlock_t rt_runtime_lock;
152 struct hrtimer rt_period_timer;
155 static struct rt_bandwidth def_rt_bandwidth;
157 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
159 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
161 struct rt_bandwidth *rt_b =
162 container_of(timer, struct rt_bandwidth, rt_period_timer);
168 now = hrtimer_cb_get_time(timer);
169 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
174 idle = do_sched_rt_period_timer(rt_b, overrun);
177 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
181 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
183 rt_b->rt_period = ns_to_ktime(period);
184 rt_b->rt_runtime = runtime;
186 raw_spin_lock_init(&rt_b->rt_runtime_lock);
188 hrtimer_init(&rt_b->rt_period_timer,
189 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
190 rt_b->rt_period_timer.function = sched_rt_period_timer;
193 static inline int rt_bandwidth_enabled(void)
195 return sysctl_sched_rt_runtime >= 0;
198 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
202 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
205 if (hrtimer_active(&rt_b->rt_period_timer))
208 raw_spin_lock(&rt_b->rt_runtime_lock);
213 if (hrtimer_active(&rt_b->rt_period_timer))
216 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
217 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
219 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
220 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
221 delta = ktime_to_ns(ktime_sub(hard, soft));
222 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
223 HRTIMER_MODE_ABS_PINNED, 0);
225 raw_spin_unlock(&rt_b->rt_runtime_lock);
228 #ifdef CONFIG_RT_GROUP_SCHED
229 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
231 hrtimer_cancel(&rt_b->rt_period_timer);
236 * sched_domains_mutex serializes calls to arch_init_sched_domains,
237 * detach_destroy_domains and partition_sched_domains.
239 static DEFINE_MUTEX(sched_domains_mutex);
241 #ifdef CONFIG_CGROUP_SCHED
243 #include <linux/cgroup.h>
247 static LIST_HEAD(task_groups);
249 /* task group related information */
251 struct cgroup_subsys_state css;
253 #ifdef CONFIG_FAIR_GROUP_SCHED
254 /* schedulable entities of this group on each cpu */
255 struct sched_entity **se;
256 /* runqueue "owned" by this group on each cpu */
257 struct cfs_rq **cfs_rq;
258 unsigned long shares;
260 atomic_t load_weight;
263 #ifdef CONFIG_RT_GROUP_SCHED
264 struct sched_rt_entity **rt_se;
265 struct rt_rq **rt_rq;
267 struct rt_bandwidth rt_bandwidth;
271 struct list_head list;
273 struct task_group *parent;
274 struct list_head siblings;
275 struct list_head children;
277 #ifdef CONFIG_SCHED_AUTOGROUP
278 struct autogroup *autogroup;
282 /* task_group_lock serializes the addition/removal of task groups */
283 static DEFINE_SPINLOCK(task_group_lock);
285 #ifdef CONFIG_FAIR_GROUP_SCHED
287 # define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
290 * A weight of 0 or 1 can cause arithmetics problems.
291 * A weight of a cfs_rq is the sum of weights of which entities
292 * are queued on this cfs_rq, so a weight of a entity should not be
293 * too large, so as the shares value of a task group.
294 * (The default weight is 1024 - so there's no practical
295 * limitation from this.)
298 #define MAX_SHARES (1UL << 18)
300 static int root_task_group_load = ROOT_TASK_GROUP_LOAD;
303 /* Default task group.
304 * Every task in system belong to this group at bootup.
306 struct task_group root_task_group;
308 #endif /* CONFIG_CGROUP_SCHED */
310 /* CFS-related fields in a runqueue */
312 struct load_weight load;
313 unsigned long nr_running;
318 struct rb_root tasks_timeline;
319 struct rb_node *rb_leftmost;
321 struct list_head tasks;
322 struct list_head *balance_iterator;
325 * 'curr' points to currently running entity on this cfs_rq.
326 * It is set to NULL otherwise (i.e when none are currently running).
328 struct sched_entity *curr, *next, *last;
330 unsigned int nr_spread_over;
332 #ifdef CONFIG_FAIR_GROUP_SCHED
333 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
336 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
337 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
338 * (like users, containers etc.)
340 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
341 * list is used during load balance.
344 struct list_head leaf_cfs_rq_list;
345 struct task_group *tg; /* group that "owns" this runqueue */
349 * the part of load.weight contributed by tasks
351 unsigned long task_weight;
354 * h_load = weight * f(tg)
356 * Where f(tg) is the recursive weight fraction assigned to
359 unsigned long h_load;
362 * Maintaining per-cpu shares distribution for group scheduling
364 * load_stamp is the last time we updated the load average
365 * load_last is the last time we updated the load average and saw load
366 * load_unacc_exec_time is currently unaccounted execution time
370 u64 load_stamp, load_last, load_unacc_exec_time;
372 unsigned long load_contribution;
377 /* Real-Time classes' related field in a runqueue: */
379 struct rt_prio_array active;
380 unsigned long rt_nr_running;
381 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
383 int curr; /* highest queued rt task prio */
385 int next; /* next highest */
390 unsigned long rt_nr_migratory;
391 unsigned long rt_nr_total;
393 struct plist_head pushable_tasks;
398 /* Nests inside the rq lock: */
399 raw_spinlock_t rt_runtime_lock;
401 #ifdef CONFIG_RT_GROUP_SCHED
402 unsigned long rt_nr_boosted;
405 struct list_head leaf_rt_rq_list;
406 struct task_group *tg;
413 * We add the notion of a root-domain which will be used to define per-domain
414 * variables. Each exclusive cpuset essentially defines an island domain by
415 * fully partitioning the member cpus from any other cpuset. Whenever a new
416 * exclusive cpuset is created, we also create and attach a new root-domain
423 cpumask_var_t online;
426 * The "RT overload" flag: it gets set if a CPU has more than
427 * one runnable RT task.
429 cpumask_var_t rto_mask;
431 struct cpupri cpupri;
435 * By default the system creates a single root-domain with all cpus as
436 * members (mimicking the global state we have today).
438 static struct root_domain def_root_domain;
440 #endif /* CONFIG_SMP */
443 * This is the main, per-CPU runqueue data structure.
445 * Locking rule: those places that want to lock multiple runqueues
446 * (such as the load balancing or the thread migration code), lock
447 * acquire operations must be ordered by ascending &runqueue.
454 * nr_running and cpu_load should be in the same cacheline because
455 * remote CPUs use both these fields when doing load calculation.
457 unsigned long nr_running;
458 #define CPU_LOAD_IDX_MAX 5
459 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
460 unsigned long last_load_update_tick;
463 unsigned char nohz_balance_kick;
465 unsigned int skip_clock_update;
467 /* capture load from *all* tasks on this cpu: */
468 struct load_weight load;
469 unsigned long nr_load_updates;
475 #ifdef CONFIG_FAIR_GROUP_SCHED
476 /* list of leaf cfs_rq on this cpu: */
477 struct list_head leaf_cfs_rq_list;
479 #ifdef CONFIG_RT_GROUP_SCHED
480 struct list_head leaf_rt_rq_list;
484 * This is part of a global counter where only the total sum
485 * over all CPUs matters. A task can increase this counter on
486 * one CPU and if it got migrated afterwards it may decrease
487 * it on another CPU. Always updated under the runqueue lock:
489 unsigned long nr_uninterruptible;
491 struct task_struct *curr, *idle, *stop;
492 unsigned long next_balance;
493 struct mm_struct *prev_mm;
501 struct root_domain *rd;
502 struct sched_domain *sd;
504 unsigned long cpu_power;
506 unsigned char idle_at_tick;
507 /* For active balancing */
511 struct cpu_stop_work active_balance_work;
512 /* cpu of this runqueue: */
516 unsigned long avg_load_per_task;
524 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
528 /* calc_load related fields */
529 unsigned long calc_load_update;
530 long calc_load_active;
532 #ifdef CONFIG_SCHED_HRTICK
534 int hrtick_csd_pending;
535 struct call_single_data hrtick_csd;
537 struct hrtimer hrtick_timer;
540 #ifdef CONFIG_SCHEDSTATS
542 struct sched_info rq_sched_info;
543 unsigned long long rq_cpu_time;
544 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
546 /* sys_sched_yield() stats */
547 unsigned int yld_count;
549 /* schedule() stats */
550 unsigned int sched_switch;
551 unsigned int sched_count;
552 unsigned int sched_goidle;
554 /* try_to_wake_up() stats */
555 unsigned int ttwu_count;
556 unsigned int ttwu_local;
560 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
563 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags);
565 static inline int cpu_of(struct rq *rq)
574 #define rcu_dereference_check_sched_domain(p) \
575 rcu_dereference_check((p), \
576 rcu_read_lock_sched_held() || \
577 lockdep_is_held(&sched_domains_mutex))
580 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
581 * See detach_destroy_domains: synchronize_sched for details.
583 * The domain tree of any CPU may only be accessed from within
584 * preempt-disabled sections.
586 #define for_each_domain(cpu, __sd) \
587 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
589 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
590 #define this_rq() (&__get_cpu_var(runqueues))
591 #define task_rq(p) cpu_rq(task_cpu(p))
592 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
593 #define raw_rq() (&__raw_get_cpu_var(runqueues))
595 #ifdef CONFIG_CGROUP_SCHED
598 * Return the group to which this tasks belongs.
600 * We use task_subsys_state_check() and extend the RCU verification
601 * with lockdep_is_held(&task_rq(p)->lock) because cpu_cgroup_attach()
602 * holds that lock for each task it moves into the cgroup. Therefore
603 * by holding that lock, we pin the task to the current cgroup.
605 static inline struct task_group *task_group(struct task_struct *p)
607 struct task_group *tg;
608 struct cgroup_subsys_state *css;
610 if (p->flags & PF_EXITING)
611 return &root_task_group;
613 css = task_subsys_state_check(p, cpu_cgroup_subsys_id,
614 lockdep_is_held(&task_rq(p)->lock));
615 tg = container_of(css, struct task_group, css);
617 return autogroup_task_group(p, tg);
620 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
621 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
623 #ifdef CONFIG_FAIR_GROUP_SCHED
624 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
625 p->se.parent = task_group(p)->se[cpu];
628 #ifdef CONFIG_RT_GROUP_SCHED
629 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
630 p->rt.parent = task_group(p)->rt_se[cpu];
634 #else /* CONFIG_CGROUP_SCHED */
636 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
637 static inline struct task_group *task_group(struct task_struct *p)
642 #endif /* CONFIG_CGROUP_SCHED */
644 static void update_rq_clock_task(struct rq *rq, s64 delta);
646 static void update_rq_clock(struct rq *rq)
650 if (rq->skip_clock_update)
653 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
655 update_rq_clock_task(rq, delta);
659 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
661 #ifdef CONFIG_SCHED_DEBUG
662 # define const_debug __read_mostly
664 # define const_debug static const
669 * @cpu: the processor in question.
671 * Returns true if the current cpu runqueue is locked.
672 * This interface allows printk to be called with the runqueue lock
673 * held and know whether or not it is OK to wake up the klogd.
675 int runqueue_is_locked(int cpu)
677 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
681 * Debugging: various feature bits
684 #define SCHED_FEAT(name, enabled) \
685 __SCHED_FEAT_##name ,
688 #include "sched_features.h"
693 #define SCHED_FEAT(name, enabled) \
694 (1UL << __SCHED_FEAT_##name) * enabled |
696 const_debug unsigned int sysctl_sched_features =
697 #include "sched_features.h"
702 #ifdef CONFIG_SCHED_DEBUG
703 #define SCHED_FEAT(name, enabled) \
706 static __read_mostly char *sched_feat_names[] = {
707 #include "sched_features.h"
713 static int sched_feat_show(struct seq_file *m, void *v)
717 for (i = 0; sched_feat_names[i]; i++) {
718 if (!(sysctl_sched_features & (1UL << i)))
720 seq_printf(m, "%s ", sched_feat_names[i]);
728 sched_feat_write(struct file *filp, const char __user *ubuf,
729 size_t cnt, loff_t *ppos)
739 if (copy_from_user(&buf, ubuf, cnt))
745 if (strncmp(cmp, "NO_", 3) == 0) {
750 for (i = 0; sched_feat_names[i]; i++) {
751 if (strcmp(cmp, sched_feat_names[i]) == 0) {
753 sysctl_sched_features &= ~(1UL << i);
755 sysctl_sched_features |= (1UL << i);
760 if (!sched_feat_names[i])
768 static int sched_feat_open(struct inode *inode, struct file *filp)
770 return single_open(filp, sched_feat_show, NULL);
773 static const struct file_operations sched_feat_fops = {
774 .open = sched_feat_open,
775 .write = sched_feat_write,
778 .release = single_release,
781 static __init int sched_init_debug(void)
783 debugfs_create_file("sched_features", 0644, NULL, NULL,
788 late_initcall(sched_init_debug);
792 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
795 * Number of tasks to iterate in a single balance run.
796 * Limited because this is done with IRQs disabled.
798 const_debug unsigned int sysctl_sched_nr_migrate = 32;
801 * period over which we average the RT time consumption, measured
806 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
809 * period over which we measure -rt task cpu usage in us.
812 unsigned int sysctl_sched_rt_period = 1000000;
814 static __read_mostly int scheduler_running;
817 * part of the period that we allow rt tasks to run in us.
820 int sysctl_sched_rt_runtime = 950000;
822 static inline u64 global_rt_period(void)
824 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
827 static inline u64 global_rt_runtime(void)
829 if (sysctl_sched_rt_runtime < 0)
832 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
835 #ifndef prepare_arch_switch
836 # define prepare_arch_switch(next) do { } while (0)
838 #ifndef finish_arch_switch
839 # define finish_arch_switch(prev) do { } while (0)
842 static inline int task_current(struct rq *rq, struct task_struct *p)
844 return rq->curr == p;
847 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
848 static inline int task_running(struct rq *rq, struct task_struct *p)
850 return task_current(rq, p);
853 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
857 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
859 #ifdef CONFIG_DEBUG_SPINLOCK
860 /* this is a valid case when another task releases the spinlock */
861 rq->lock.owner = current;
864 * If we are tracking spinlock dependencies then we have to
865 * fix up the runqueue lock - which gets 'carried over' from
868 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
870 raw_spin_unlock_irq(&rq->lock);
873 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
874 static inline int task_running(struct rq *rq, struct task_struct *p)
879 return task_current(rq, p);
883 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
887 * We can optimise this out completely for !SMP, because the
888 * SMP rebalancing from interrupt is the only thing that cares
893 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
894 raw_spin_unlock_irq(&rq->lock);
896 raw_spin_unlock(&rq->lock);
900 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
904 * After ->oncpu is cleared, the task can be moved to a different CPU.
905 * We must ensure this doesn't happen until the switch is completely
911 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
915 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
918 * Check whether the task is waking, we use this to synchronize ->cpus_allowed
921 static inline int task_is_waking(struct task_struct *p)
923 return unlikely(p->state == TASK_WAKING);
927 * __task_rq_lock - lock the runqueue a given task resides on.
928 * Must be called interrupts disabled.
930 static inline struct rq *__task_rq_lock(struct task_struct *p)
937 raw_spin_lock(&rq->lock);
938 if (likely(rq == task_rq(p)))
940 raw_spin_unlock(&rq->lock);
945 * task_rq_lock - lock the runqueue a given task resides on and disable
946 * interrupts. Note the ordering: we can safely lookup the task_rq without
947 * explicitly disabling preemption.
949 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
955 local_irq_save(*flags);
957 raw_spin_lock(&rq->lock);
958 if (likely(rq == task_rq(p)))
960 raw_spin_unlock_irqrestore(&rq->lock, *flags);
964 static void __task_rq_unlock(struct rq *rq)
967 raw_spin_unlock(&rq->lock);
970 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
973 raw_spin_unlock_irqrestore(&rq->lock, *flags);
977 * this_rq_lock - lock this runqueue and disable interrupts.
979 static struct rq *this_rq_lock(void)
986 raw_spin_lock(&rq->lock);
991 #ifdef CONFIG_SCHED_HRTICK
993 * Use HR-timers to deliver accurate preemption points.
995 * Its all a bit involved since we cannot program an hrt while holding the
996 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
999 * When we get rescheduled we reprogram the hrtick_timer outside of the
1005 * - enabled by features
1006 * - hrtimer is actually high res
1008 static inline int hrtick_enabled(struct rq *rq)
1010 if (!sched_feat(HRTICK))
1012 if (!cpu_active(cpu_of(rq)))
1014 return hrtimer_is_hres_active(&rq->hrtick_timer);
1017 static void hrtick_clear(struct rq *rq)
1019 if (hrtimer_active(&rq->hrtick_timer))
1020 hrtimer_cancel(&rq->hrtick_timer);
1024 * High-resolution timer tick.
1025 * Runs from hardirq context with interrupts disabled.
1027 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1029 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1031 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1033 raw_spin_lock(&rq->lock);
1034 update_rq_clock(rq);
1035 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1036 raw_spin_unlock(&rq->lock);
1038 return HRTIMER_NORESTART;
1043 * called from hardirq (IPI) context
1045 static void __hrtick_start(void *arg)
1047 struct rq *rq = arg;
1049 raw_spin_lock(&rq->lock);
1050 hrtimer_restart(&rq->hrtick_timer);
1051 rq->hrtick_csd_pending = 0;
1052 raw_spin_unlock(&rq->lock);
1056 * Called to set the hrtick timer state.
1058 * called with rq->lock held and irqs disabled
1060 static void hrtick_start(struct rq *rq, u64 delay)
1062 struct hrtimer *timer = &rq->hrtick_timer;
1063 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1065 hrtimer_set_expires(timer, time);
1067 if (rq == this_rq()) {
1068 hrtimer_restart(timer);
1069 } else if (!rq->hrtick_csd_pending) {
1070 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1071 rq->hrtick_csd_pending = 1;
1076 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1078 int cpu = (int)(long)hcpu;
1081 case CPU_UP_CANCELED:
1082 case CPU_UP_CANCELED_FROZEN:
1083 case CPU_DOWN_PREPARE:
1084 case CPU_DOWN_PREPARE_FROZEN:
1086 case CPU_DEAD_FROZEN:
1087 hrtick_clear(cpu_rq(cpu));
1094 static __init void init_hrtick(void)
1096 hotcpu_notifier(hotplug_hrtick, 0);
1100 * Called to set the hrtick timer state.
1102 * called with rq->lock held and irqs disabled
1104 static void hrtick_start(struct rq *rq, u64 delay)
1106 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1107 HRTIMER_MODE_REL_PINNED, 0);
1110 static inline void init_hrtick(void)
1113 #endif /* CONFIG_SMP */
1115 static void init_rq_hrtick(struct rq *rq)
1118 rq->hrtick_csd_pending = 0;
1120 rq->hrtick_csd.flags = 0;
1121 rq->hrtick_csd.func = __hrtick_start;
1122 rq->hrtick_csd.info = rq;
1125 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1126 rq->hrtick_timer.function = hrtick;
1128 #else /* CONFIG_SCHED_HRTICK */
1129 static inline void hrtick_clear(struct rq *rq)
1133 static inline void init_rq_hrtick(struct rq *rq)
1137 static inline void init_hrtick(void)
1140 #endif /* CONFIG_SCHED_HRTICK */
1143 * resched_task - mark a task 'to be rescheduled now'.
1145 * On UP this means the setting of the need_resched flag, on SMP it
1146 * might also involve a cross-CPU call to trigger the scheduler on
1151 #ifndef tsk_is_polling
1152 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1155 static void resched_task(struct task_struct *p)
1159 assert_raw_spin_locked(&task_rq(p)->lock);
1161 if (test_tsk_need_resched(p))
1164 set_tsk_need_resched(p);
1167 if (cpu == smp_processor_id())
1170 /* NEED_RESCHED must be visible before we test polling */
1172 if (!tsk_is_polling(p))
1173 smp_send_reschedule(cpu);
1176 static void resched_cpu(int cpu)
1178 struct rq *rq = cpu_rq(cpu);
1179 unsigned long flags;
1181 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1183 resched_task(cpu_curr(cpu));
1184 raw_spin_unlock_irqrestore(&rq->lock, flags);
1189 * In the semi idle case, use the nearest busy cpu for migrating timers
1190 * from an idle cpu. This is good for power-savings.
1192 * We don't do similar optimization for completely idle system, as
1193 * selecting an idle cpu will add more delays to the timers than intended
1194 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1196 int get_nohz_timer_target(void)
1198 int cpu = smp_processor_id();
1200 struct sched_domain *sd;
1202 for_each_domain(cpu, sd) {
1203 for_each_cpu(i, sched_domain_span(sd))
1210 * When add_timer_on() enqueues a timer into the timer wheel of an
1211 * idle CPU then this timer might expire before the next timer event
1212 * which is scheduled to wake up that CPU. In case of a completely
1213 * idle system the next event might even be infinite time into the
1214 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1215 * leaves the inner idle loop so the newly added timer is taken into
1216 * account when the CPU goes back to idle and evaluates the timer
1217 * wheel for the next timer event.
1219 void wake_up_idle_cpu(int cpu)
1221 struct rq *rq = cpu_rq(cpu);
1223 if (cpu == smp_processor_id())
1227 * This is safe, as this function is called with the timer
1228 * wheel base lock of (cpu) held. When the CPU is on the way
1229 * to idle and has not yet set rq->curr to idle then it will
1230 * be serialized on the timer wheel base lock and take the new
1231 * timer into account automatically.
1233 if (rq->curr != rq->idle)
1237 * We can set TIF_RESCHED on the idle task of the other CPU
1238 * lockless. The worst case is that the other CPU runs the
1239 * idle task through an additional NOOP schedule()
1241 set_tsk_need_resched(rq->idle);
1243 /* NEED_RESCHED must be visible before we test polling */
1245 if (!tsk_is_polling(rq->idle))
1246 smp_send_reschedule(cpu);
1249 #endif /* CONFIG_NO_HZ */
1251 static u64 sched_avg_period(void)
1253 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1256 static void sched_avg_update(struct rq *rq)
1258 s64 period = sched_avg_period();
1260 while ((s64)(rq->clock - rq->age_stamp) > period) {
1262 * Inline assembly required to prevent the compiler
1263 * optimising this loop into a divmod call.
1264 * See __iter_div_u64_rem() for another example of this.
1266 asm("" : "+rm" (rq->age_stamp));
1267 rq->age_stamp += period;
1272 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1274 rq->rt_avg += rt_delta;
1275 sched_avg_update(rq);
1278 #else /* !CONFIG_SMP */
1279 static void resched_task(struct task_struct *p)
1281 assert_raw_spin_locked(&task_rq(p)->lock);
1282 set_tsk_need_resched(p);
1285 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1289 static void sched_avg_update(struct rq *rq)
1292 #endif /* CONFIG_SMP */
1294 #if BITS_PER_LONG == 32
1295 # define WMULT_CONST (~0UL)
1297 # define WMULT_CONST (1UL << 32)
1300 #define WMULT_SHIFT 32
1303 * Shift right and round:
1305 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1308 * delta *= weight / lw
1310 static unsigned long
1311 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1312 struct load_weight *lw)
1316 if (!lw->inv_weight) {
1317 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1320 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1324 tmp = (u64)delta_exec * weight;
1326 * Check whether we'd overflow the 64-bit multiplication:
1328 if (unlikely(tmp > WMULT_CONST))
1329 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1332 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1334 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1337 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1343 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1349 static inline void update_load_set(struct load_weight *lw, unsigned long w)
1356 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1357 * of tasks with abnormal "nice" values across CPUs the contribution that
1358 * each task makes to its run queue's load is weighted according to its
1359 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1360 * scaled version of the new time slice allocation that they receive on time
1364 #define WEIGHT_IDLEPRIO 3
1365 #define WMULT_IDLEPRIO 1431655765
1368 * Nice levels are multiplicative, with a gentle 10% change for every
1369 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1370 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1371 * that remained on nice 0.
1373 * The "10% effect" is relative and cumulative: from _any_ nice level,
1374 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1375 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1376 * If a task goes up by ~10% and another task goes down by ~10% then
1377 * the relative distance between them is ~25%.)
1379 static const int prio_to_weight[40] = {
1380 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1381 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1382 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1383 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1384 /* 0 */ 1024, 820, 655, 526, 423,
1385 /* 5 */ 335, 272, 215, 172, 137,
1386 /* 10 */ 110, 87, 70, 56, 45,
1387 /* 15 */ 36, 29, 23, 18, 15,
1391 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1393 * In cases where the weight does not change often, we can use the
1394 * precalculated inverse to speed up arithmetics by turning divisions
1395 * into multiplications:
1397 static const u32 prio_to_wmult[40] = {
1398 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1399 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1400 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1401 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1402 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1403 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1404 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1405 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1408 /* Time spent by the tasks of the cpu accounting group executing in ... */
1409 enum cpuacct_stat_index {
1410 CPUACCT_STAT_USER, /* ... user mode */
1411 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1413 CPUACCT_STAT_NSTATS,
1416 #ifdef CONFIG_CGROUP_CPUACCT
1417 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1418 static void cpuacct_update_stats(struct task_struct *tsk,
1419 enum cpuacct_stat_index idx, cputime_t val);
1421 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1422 static inline void cpuacct_update_stats(struct task_struct *tsk,
1423 enum cpuacct_stat_index idx, cputime_t val) {}
1426 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1428 update_load_add(&rq->load, load);
1431 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1433 update_load_sub(&rq->load, load);
1436 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1437 typedef int (*tg_visitor)(struct task_group *, void *);
1440 * Iterate the full tree, calling @down when first entering a node and @up when
1441 * leaving it for the final time.
1443 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1445 struct task_group *parent, *child;
1449 parent = &root_task_group;
1451 ret = (*down)(parent, data);
1454 list_for_each_entry_rcu(child, &parent->children, siblings) {
1461 ret = (*up)(parent, data);
1466 parent = parent->parent;
1475 static int tg_nop(struct task_group *tg, void *data)
1482 /* Used instead of source_load when we know the type == 0 */
1483 static unsigned long weighted_cpuload(const int cpu)
1485 return cpu_rq(cpu)->load.weight;
1489 * Return a low guess at the load of a migration-source cpu weighted
1490 * according to the scheduling class and "nice" value.
1492 * We want to under-estimate the load of migration sources, to
1493 * balance conservatively.
1495 static unsigned long source_load(int cpu, int type)
1497 struct rq *rq = cpu_rq(cpu);
1498 unsigned long total = weighted_cpuload(cpu);
1500 if (type == 0 || !sched_feat(LB_BIAS))
1503 return min(rq->cpu_load[type-1], total);
1507 * Return a high guess at the load of a migration-target cpu weighted
1508 * according to the scheduling class and "nice" value.
1510 static unsigned long target_load(int cpu, int type)
1512 struct rq *rq = cpu_rq(cpu);
1513 unsigned long total = weighted_cpuload(cpu);
1515 if (type == 0 || !sched_feat(LB_BIAS))
1518 return max(rq->cpu_load[type-1], total);
1521 static unsigned long power_of(int cpu)
1523 return cpu_rq(cpu)->cpu_power;
1526 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1528 static unsigned long cpu_avg_load_per_task(int cpu)
1530 struct rq *rq = cpu_rq(cpu);
1531 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1534 rq->avg_load_per_task = rq->load.weight / nr_running;
1536 rq->avg_load_per_task = 0;
1538 return rq->avg_load_per_task;
1541 #ifdef CONFIG_FAIR_GROUP_SCHED
1544 * Compute the cpu's hierarchical load factor for each task group.
1545 * This needs to be done in a top-down fashion because the load of a child
1546 * group is a fraction of its parents load.
1548 static int tg_load_down(struct task_group *tg, void *data)
1551 long cpu = (long)data;
1554 load = cpu_rq(cpu)->load.weight;
1556 load = tg->parent->cfs_rq[cpu]->h_load;
1557 load *= tg->se[cpu]->load.weight;
1558 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1561 tg->cfs_rq[cpu]->h_load = load;
1566 static void update_h_load(long cpu)
1568 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1573 #ifdef CONFIG_PREEMPT
1575 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1578 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1579 * way at the expense of forcing extra atomic operations in all
1580 * invocations. This assures that the double_lock is acquired using the
1581 * same underlying policy as the spinlock_t on this architecture, which
1582 * reduces latency compared to the unfair variant below. However, it
1583 * also adds more overhead and therefore may reduce throughput.
1585 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1586 __releases(this_rq->lock)
1587 __acquires(busiest->lock)
1588 __acquires(this_rq->lock)
1590 raw_spin_unlock(&this_rq->lock);
1591 double_rq_lock(this_rq, busiest);
1598 * Unfair double_lock_balance: Optimizes throughput at the expense of
1599 * latency by eliminating extra atomic operations when the locks are
1600 * already in proper order on entry. This favors lower cpu-ids and will
1601 * grant the double lock to lower cpus over higher ids under contention,
1602 * regardless of entry order into the function.
1604 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1605 __releases(this_rq->lock)
1606 __acquires(busiest->lock)
1607 __acquires(this_rq->lock)
1611 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1612 if (busiest < this_rq) {
1613 raw_spin_unlock(&this_rq->lock);
1614 raw_spin_lock(&busiest->lock);
1615 raw_spin_lock_nested(&this_rq->lock,
1616 SINGLE_DEPTH_NESTING);
1619 raw_spin_lock_nested(&busiest->lock,
1620 SINGLE_DEPTH_NESTING);
1625 #endif /* CONFIG_PREEMPT */
1628 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1630 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1632 if (unlikely(!irqs_disabled())) {
1633 /* printk() doesn't work good under rq->lock */
1634 raw_spin_unlock(&this_rq->lock);
1638 return _double_lock_balance(this_rq, busiest);
1641 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1642 __releases(busiest->lock)
1644 raw_spin_unlock(&busiest->lock);
1645 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1649 * double_rq_lock - safely lock two runqueues
1651 * Note this does not disable interrupts like task_rq_lock,
1652 * you need to do so manually before calling.
1654 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1655 __acquires(rq1->lock)
1656 __acquires(rq2->lock)
1658 BUG_ON(!irqs_disabled());
1660 raw_spin_lock(&rq1->lock);
1661 __acquire(rq2->lock); /* Fake it out ;) */
1664 raw_spin_lock(&rq1->lock);
1665 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1667 raw_spin_lock(&rq2->lock);
1668 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1674 * double_rq_unlock - safely unlock two runqueues
1676 * Note this does not restore interrupts like task_rq_unlock,
1677 * you need to do so manually after calling.
1679 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1680 __releases(rq1->lock)
1681 __releases(rq2->lock)
1683 raw_spin_unlock(&rq1->lock);
1685 raw_spin_unlock(&rq2->lock);
1687 __release(rq2->lock);
1692 static void calc_load_account_idle(struct rq *this_rq);
1693 static void update_sysctl(void);
1694 static int get_update_sysctl_factor(void);
1695 static void update_cpu_load(struct rq *this_rq);
1697 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1699 set_task_rq(p, cpu);
1702 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1703 * successfuly executed on another CPU. We must ensure that updates of
1704 * per-task data have been completed by this moment.
1707 task_thread_info(p)->cpu = cpu;
1711 static const struct sched_class rt_sched_class;
1713 #define sched_class_highest (&stop_sched_class)
1714 #define for_each_class(class) \
1715 for (class = sched_class_highest; class; class = class->next)
1717 #include "sched_stats.h"
1719 static void inc_nr_running(struct rq *rq)
1724 static void dec_nr_running(struct rq *rq)
1729 static void set_load_weight(struct task_struct *p)
1732 * SCHED_IDLE tasks get minimal weight:
1734 if (p->policy == SCHED_IDLE) {
1735 p->se.load.weight = WEIGHT_IDLEPRIO;
1736 p->se.load.inv_weight = WMULT_IDLEPRIO;
1740 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1741 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1744 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1746 update_rq_clock(rq);
1747 sched_info_queued(p);
1748 p->sched_class->enqueue_task(rq, p, flags);
1752 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1754 update_rq_clock(rq);
1755 sched_info_dequeued(p);
1756 p->sched_class->dequeue_task(rq, p, flags);
1761 * activate_task - move a task to the runqueue.
1763 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1765 if (task_contributes_to_load(p))
1766 rq->nr_uninterruptible--;
1768 enqueue_task(rq, p, flags);
1773 * deactivate_task - remove a task from the runqueue.
1775 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1777 if (task_contributes_to_load(p))
1778 rq->nr_uninterruptible++;
1780 dequeue_task(rq, p, flags);
1784 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1787 * There are no locks covering percpu hardirq/softirq time.
1788 * They are only modified in account_system_vtime, on corresponding CPU
1789 * with interrupts disabled. So, writes are safe.
1790 * They are read and saved off onto struct rq in update_rq_clock().
1791 * This may result in other CPU reading this CPU's irq time and can
1792 * race with irq/account_system_vtime on this CPU. We would either get old
1793 * or new value with a side effect of accounting a slice of irq time to wrong
1794 * task when irq is in progress while we read rq->clock. That is a worthy
1795 * compromise in place of having locks on each irq in account_system_time.
1797 static DEFINE_PER_CPU(u64, cpu_hardirq_time);
1798 static DEFINE_PER_CPU(u64, cpu_softirq_time);
1800 static DEFINE_PER_CPU(u64, irq_start_time);
1801 static int sched_clock_irqtime;
1803 void enable_sched_clock_irqtime(void)
1805 sched_clock_irqtime = 1;
1808 void disable_sched_clock_irqtime(void)
1810 sched_clock_irqtime = 0;
1813 #ifndef CONFIG_64BIT
1814 static DEFINE_PER_CPU(seqcount_t, irq_time_seq);
1816 static inline void irq_time_write_begin(void)
1818 __this_cpu_inc(irq_time_seq.sequence);
1822 static inline void irq_time_write_end(void)
1825 __this_cpu_inc(irq_time_seq.sequence);
1828 static inline u64 irq_time_read(int cpu)
1834 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
1835 irq_time = per_cpu(cpu_softirq_time, cpu) +
1836 per_cpu(cpu_hardirq_time, cpu);
1837 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
1841 #else /* CONFIG_64BIT */
1842 static inline void irq_time_write_begin(void)
1846 static inline void irq_time_write_end(void)
1850 static inline u64 irq_time_read(int cpu)
1852 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
1854 #endif /* CONFIG_64BIT */
1857 * Called before incrementing preempt_count on {soft,}irq_enter
1858 * and before decrementing preempt_count on {soft,}irq_exit.
1860 void account_system_vtime(struct task_struct *curr)
1862 unsigned long flags;
1866 if (!sched_clock_irqtime)
1869 local_irq_save(flags);
1871 cpu = smp_processor_id();
1872 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
1873 __this_cpu_add(irq_start_time, delta);
1875 irq_time_write_begin();
1877 * We do not account for softirq time from ksoftirqd here.
1878 * We want to continue accounting softirq time to ksoftirqd thread
1879 * in that case, so as not to confuse scheduler with a special task
1880 * that do not consume any time, but still wants to run.
1882 if (hardirq_count())
1883 __this_cpu_add(cpu_hardirq_time, delta);
1884 else if (in_serving_softirq() && !(curr->flags & PF_KSOFTIRQD))
1885 __this_cpu_add(cpu_softirq_time, delta);
1887 irq_time_write_end();
1888 local_irq_restore(flags);
1890 EXPORT_SYMBOL_GPL(account_system_vtime);
1892 static void update_rq_clock_task(struct rq *rq, s64 delta)
1896 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
1899 * Since irq_time is only updated on {soft,}irq_exit, we might run into
1900 * this case when a previous update_rq_clock() happened inside a
1901 * {soft,}irq region.
1903 * When this happens, we stop ->clock_task and only update the
1904 * prev_irq_time stamp to account for the part that fit, so that a next
1905 * update will consume the rest. This ensures ->clock_task is
1908 * It does however cause some slight miss-attribution of {soft,}irq
1909 * time, a more accurate solution would be to update the irq_time using
1910 * the current rq->clock timestamp, except that would require using
1913 if (irq_delta > delta)
1916 rq->prev_irq_time += irq_delta;
1918 rq->clock_task += delta;
1920 if (irq_delta && sched_feat(NONIRQ_POWER))
1921 sched_rt_avg_update(rq, irq_delta);
1924 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
1926 static void update_rq_clock_task(struct rq *rq, s64 delta)
1928 rq->clock_task += delta;
1931 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
1933 #include "sched_idletask.c"
1934 #include "sched_fair.c"
1935 #include "sched_rt.c"
1936 #include "sched_autogroup.c"
1937 #include "sched_stoptask.c"
1938 #ifdef CONFIG_SCHED_DEBUG
1939 # include "sched_debug.c"
1942 void sched_set_stop_task(int cpu, struct task_struct *stop)
1944 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
1945 struct task_struct *old_stop = cpu_rq(cpu)->stop;
1949 * Make it appear like a SCHED_FIFO task, its something
1950 * userspace knows about and won't get confused about.
1952 * Also, it will make PI more or less work without too
1953 * much confusion -- but then, stop work should not
1954 * rely on PI working anyway.
1956 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
1958 stop->sched_class = &stop_sched_class;
1961 cpu_rq(cpu)->stop = stop;
1965 * Reset it back to a normal scheduling class so that
1966 * it can die in pieces.
1968 old_stop->sched_class = &rt_sched_class;
1973 * __normal_prio - return the priority that is based on the static prio
1975 static inline int __normal_prio(struct task_struct *p)
1977 return p->static_prio;
1981 * Calculate the expected normal priority: i.e. priority
1982 * without taking RT-inheritance into account. Might be
1983 * boosted by interactivity modifiers. Changes upon fork,
1984 * setprio syscalls, and whenever the interactivity
1985 * estimator recalculates.
1987 static inline int normal_prio(struct task_struct *p)
1991 if (task_has_rt_policy(p))
1992 prio = MAX_RT_PRIO-1 - p->rt_priority;
1994 prio = __normal_prio(p);
1999 * Calculate the current priority, i.e. the priority
2000 * taken into account by the scheduler. This value might
2001 * be boosted by RT tasks, or might be boosted by
2002 * interactivity modifiers. Will be RT if the task got
2003 * RT-boosted. If not then it returns p->normal_prio.
2005 static int effective_prio(struct task_struct *p)
2007 p->normal_prio = normal_prio(p);
2009 * If we are RT tasks or we were boosted to RT priority,
2010 * keep the priority unchanged. Otherwise, update priority
2011 * to the normal priority:
2013 if (!rt_prio(p->prio))
2014 return p->normal_prio;
2019 * task_curr - is this task currently executing on a CPU?
2020 * @p: the task in question.
2022 inline int task_curr(const struct task_struct *p)
2024 return cpu_curr(task_cpu(p)) == p;
2027 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2028 const struct sched_class *prev_class,
2029 int oldprio, int running)
2031 if (prev_class != p->sched_class) {
2032 if (prev_class->switched_from)
2033 prev_class->switched_from(rq, p, running);
2034 p->sched_class->switched_to(rq, p, running);
2036 p->sched_class->prio_changed(rq, p, oldprio, running);
2039 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2041 const struct sched_class *class;
2043 if (p->sched_class == rq->curr->sched_class) {
2044 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2046 for_each_class(class) {
2047 if (class == rq->curr->sched_class)
2049 if (class == p->sched_class) {
2050 resched_task(rq->curr);
2057 * A queue event has occurred, and we're going to schedule. In
2058 * this case, we can save a useless back to back clock update.
2060 if (rq->curr->se.on_rq && test_tsk_need_resched(rq->curr))
2061 rq->skip_clock_update = 1;
2066 * Is this task likely cache-hot:
2069 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2073 if (p->sched_class != &fair_sched_class)
2076 if (unlikely(p->policy == SCHED_IDLE))
2080 * Buddy candidates are cache hot:
2082 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2083 (&p->se == cfs_rq_of(&p->se)->next ||
2084 &p->se == cfs_rq_of(&p->se)->last))
2087 if (sysctl_sched_migration_cost == -1)
2089 if (sysctl_sched_migration_cost == 0)
2092 delta = now - p->se.exec_start;
2094 return delta < (s64)sysctl_sched_migration_cost;
2097 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2099 #ifdef CONFIG_SCHED_DEBUG
2101 * We should never call set_task_cpu() on a blocked task,
2102 * ttwu() will sort out the placement.
2104 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2105 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2108 trace_sched_migrate_task(p, new_cpu);
2110 if (task_cpu(p) != new_cpu) {
2111 p->se.nr_migrations++;
2112 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2115 __set_task_cpu(p, new_cpu);
2118 struct migration_arg {
2119 struct task_struct *task;
2123 static int migration_cpu_stop(void *data);
2126 * The task's runqueue lock must be held.
2127 * Returns true if you have to wait for migration thread.
2129 static bool migrate_task(struct task_struct *p, struct rq *rq)
2132 * If the task is not on a runqueue (and not running), then
2133 * the next wake-up will properly place the task.
2135 return p->se.on_rq || task_running(rq, p);
2139 * wait_task_inactive - wait for a thread to unschedule.
2141 * If @match_state is nonzero, it's the @p->state value just checked and
2142 * not expected to change. If it changes, i.e. @p might have woken up,
2143 * then return zero. When we succeed in waiting for @p to be off its CPU,
2144 * we return a positive number (its total switch count). If a second call
2145 * a short while later returns the same number, the caller can be sure that
2146 * @p has remained unscheduled the whole time.
2148 * The caller must ensure that the task *will* unschedule sometime soon,
2149 * else this function might spin for a *long* time. This function can't
2150 * be called with interrupts off, or it may introduce deadlock with
2151 * smp_call_function() if an IPI is sent by the same process we are
2152 * waiting to become inactive.
2154 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2156 unsigned long flags;
2163 * We do the initial early heuristics without holding
2164 * any task-queue locks at all. We'll only try to get
2165 * the runqueue lock when things look like they will
2171 * If the task is actively running on another CPU
2172 * still, just relax and busy-wait without holding
2175 * NOTE! Since we don't hold any locks, it's not
2176 * even sure that "rq" stays as the right runqueue!
2177 * But we don't care, since "task_running()" will
2178 * return false if the runqueue has changed and p
2179 * is actually now running somewhere else!
2181 while (task_running(rq, p)) {
2182 if (match_state && unlikely(p->state != match_state))
2188 * Ok, time to look more closely! We need the rq
2189 * lock now, to be *sure*. If we're wrong, we'll
2190 * just go back and repeat.
2192 rq = task_rq_lock(p, &flags);
2193 trace_sched_wait_task(p);
2194 running = task_running(rq, p);
2195 on_rq = p->se.on_rq;
2197 if (!match_state || p->state == match_state)
2198 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2199 task_rq_unlock(rq, &flags);
2202 * If it changed from the expected state, bail out now.
2204 if (unlikely(!ncsw))
2208 * Was it really running after all now that we
2209 * checked with the proper locks actually held?
2211 * Oops. Go back and try again..
2213 if (unlikely(running)) {
2219 * It's not enough that it's not actively running,
2220 * it must be off the runqueue _entirely_, and not
2223 * So if it was still runnable (but just not actively
2224 * running right now), it's preempted, and we should
2225 * yield - it could be a while.
2227 if (unlikely(on_rq)) {
2228 schedule_timeout_uninterruptible(1);
2233 * Ahh, all good. It wasn't running, and it wasn't
2234 * runnable, which means that it will never become
2235 * running in the future either. We're all done!
2244 * kick_process - kick a running thread to enter/exit the kernel
2245 * @p: the to-be-kicked thread
2247 * Cause a process which is running on another CPU to enter
2248 * kernel-mode, without any delay. (to get signals handled.)
2250 * NOTE: this function doesnt have to take the runqueue lock,
2251 * because all it wants to ensure is that the remote task enters
2252 * the kernel. If the IPI races and the task has been migrated
2253 * to another CPU then no harm is done and the purpose has been
2256 void kick_process(struct task_struct *p)
2262 if ((cpu != smp_processor_id()) && task_curr(p))
2263 smp_send_reschedule(cpu);
2266 EXPORT_SYMBOL_GPL(kick_process);
2267 #endif /* CONFIG_SMP */
2270 * task_oncpu_function_call - call a function on the cpu on which a task runs
2271 * @p: the task to evaluate
2272 * @func: the function to be called
2273 * @info: the function call argument
2275 * Calls the function @func when the task is currently running. This might
2276 * be on the current CPU, which just calls the function directly
2278 void task_oncpu_function_call(struct task_struct *p,
2279 void (*func) (void *info), void *info)
2286 smp_call_function_single(cpu, func, info, 1);
2292 * ->cpus_allowed is protected by either TASK_WAKING or rq->lock held.
2294 static int select_fallback_rq(int cpu, struct task_struct *p)
2297 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2299 /* Look for allowed, online CPU in same node. */
2300 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2301 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2304 /* Any allowed, online CPU? */
2305 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2306 if (dest_cpu < nr_cpu_ids)
2309 /* No more Mr. Nice Guy. */
2310 dest_cpu = cpuset_cpus_allowed_fallback(p);
2312 * Don't tell them about moving exiting tasks or
2313 * kernel threads (both mm NULL), since they never
2316 if (p->mm && printk_ratelimit()) {
2317 printk(KERN_INFO "process %d (%s) no longer affine to cpu%d\n",
2318 task_pid_nr(p), p->comm, cpu);
2325 * The caller (fork, wakeup) owns TASK_WAKING, ->cpus_allowed is stable.
2328 int select_task_rq(struct rq *rq, struct task_struct *p, int sd_flags, int wake_flags)
2330 int cpu = p->sched_class->select_task_rq(rq, p, sd_flags, wake_flags);
2333 * In order not to call set_task_cpu() on a blocking task we need
2334 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2337 * Since this is common to all placement strategies, this lives here.
2339 * [ this allows ->select_task() to simply return task_cpu(p) and
2340 * not worry about this generic constraint ]
2342 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2344 cpu = select_fallback_rq(task_cpu(p), p);
2349 static void update_avg(u64 *avg, u64 sample)
2351 s64 diff = sample - *avg;
2356 static inline void ttwu_activate(struct task_struct *p, struct rq *rq,
2357 bool is_sync, bool is_migrate, bool is_local,
2358 unsigned long en_flags)
2360 schedstat_inc(p, se.statistics.nr_wakeups);
2362 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2364 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2366 schedstat_inc(p, se.statistics.nr_wakeups_local);
2368 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2370 activate_task(rq, p, en_flags);
2373 static inline void ttwu_post_activation(struct task_struct *p, struct rq *rq,
2374 int wake_flags, bool success)
2376 trace_sched_wakeup(p, success);
2377 check_preempt_curr(rq, p, wake_flags);
2379 p->state = TASK_RUNNING;
2381 if (p->sched_class->task_woken)
2382 p->sched_class->task_woken(rq, p);
2384 if (unlikely(rq->idle_stamp)) {
2385 u64 delta = rq->clock - rq->idle_stamp;
2386 u64 max = 2*sysctl_sched_migration_cost;
2391 update_avg(&rq->avg_idle, delta);
2395 /* if a worker is waking up, notify workqueue */
2396 if ((p->flags & PF_WQ_WORKER) && success)
2397 wq_worker_waking_up(p, cpu_of(rq));
2401 * try_to_wake_up - wake up a thread
2402 * @p: the thread to be awakened
2403 * @state: the mask of task states that can be woken
2404 * @wake_flags: wake modifier flags (WF_*)
2406 * Put it on the run-queue if it's not already there. The "current"
2407 * thread is always on the run-queue (except when the actual
2408 * re-schedule is in progress), and as such you're allowed to do
2409 * the simpler "current->state = TASK_RUNNING" to mark yourself
2410 * runnable without the overhead of this.
2412 * Returns %true if @p was woken up, %false if it was already running
2413 * or @state didn't match @p's state.
2415 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2418 int cpu, orig_cpu, this_cpu, success = 0;
2419 unsigned long flags;
2420 unsigned long en_flags = ENQUEUE_WAKEUP;
2423 this_cpu = get_cpu();
2426 rq = task_rq_lock(p, &flags);
2427 if (!(p->state & state))
2437 if (unlikely(task_running(rq, p)))
2441 * In order to handle concurrent wakeups and release the rq->lock
2442 * we put the task in TASK_WAKING state.
2444 * First fix up the nr_uninterruptible count:
2446 if (task_contributes_to_load(p)) {
2447 if (likely(cpu_online(orig_cpu)))
2448 rq->nr_uninterruptible--;
2450 this_rq()->nr_uninterruptible--;
2452 p->state = TASK_WAKING;
2454 if (p->sched_class->task_waking) {
2455 p->sched_class->task_waking(rq, p);
2456 en_flags |= ENQUEUE_WAKING;
2459 cpu = select_task_rq(rq, p, SD_BALANCE_WAKE, wake_flags);
2460 if (cpu != orig_cpu)
2461 set_task_cpu(p, cpu);
2462 __task_rq_unlock(rq);
2465 raw_spin_lock(&rq->lock);
2468 * We migrated the task without holding either rq->lock, however
2469 * since the task is not on the task list itself, nobody else
2470 * will try and migrate the task, hence the rq should match the
2471 * cpu we just moved it to.
2473 WARN_ON(task_cpu(p) != cpu);
2474 WARN_ON(p->state != TASK_WAKING);
2476 #ifdef CONFIG_SCHEDSTATS
2477 schedstat_inc(rq, ttwu_count);
2478 if (cpu == this_cpu)
2479 schedstat_inc(rq, ttwu_local);
2481 struct sched_domain *sd;
2482 for_each_domain(this_cpu, sd) {
2483 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2484 schedstat_inc(sd, ttwu_wake_remote);
2489 #endif /* CONFIG_SCHEDSTATS */
2492 #endif /* CONFIG_SMP */
2493 ttwu_activate(p, rq, wake_flags & WF_SYNC, orig_cpu != cpu,
2494 cpu == this_cpu, en_flags);
2497 ttwu_post_activation(p, rq, wake_flags, success);
2499 task_rq_unlock(rq, &flags);
2506 * try_to_wake_up_local - try to wake up a local task with rq lock held
2507 * @p: the thread to be awakened
2509 * Put @p on the run-queue if it's not already there. The caller must
2510 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2511 * the current task. this_rq() stays locked over invocation.
2513 static void try_to_wake_up_local(struct task_struct *p)
2515 struct rq *rq = task_rq(p);
2516 bool success = false;
2518 BUG_ON(rq != this_rq());
2519 BUG_ON(p == current);
2520 lockdep_assert_held(&rq->lock);
2522 if (!(p->state & TASK_NORMAL))
2526 if (likely(!task_running(rq, p))) {
2527 schedstat_inc(rq, ttwu_count);
2528 schedstat_inc(rq, ttwu_local);
2530 ttwu_activate(p, rq, false, false, true, ENQUEUE_WAKEUP);
2533 ttwu_post_activation(p, rq, 0, success);
2537 * wake_up_process - Wake up a specific process
2538 * @p: The process to be woken up.
2540 * Attempt to wake up the nominated process and move it to the set of runnable
2541 * processes. Returns 1 if the process was woken up, 0 if it was already
2544 * It may be assumed that this function implies a write memory barrier before
2545 * changing the task state if and only if any tasks are woken up.
2547 int wake_up_process(struct task_struct *p)
2549 return try_to_wake_up(p, TASK_ALL, 0);
2551 EXPORT_SYMBOL(wake_up_process);
2553 int wake_up_state(struct task_struct *p, unsigned int state)
2555 return try_to_wake_up(p, state, 0);
2559 * Perform scheduler related setup for a newly forked process p.
2560 * p is forked by current.
2562 * __sched_fork() is basic setup used by init_idle() too:
2564 static void __sched_fork(struct task_struct *p)
2566 p->se.exec_start = 0;
2567 p->se.sum_exec_runtime = 0;
2568 p->se.prev_sum_exec_runtime = 0;
2569 p->se.nr_migrations = 0;
2571 #ifdef CONFIG_SCHEDSTATS
2572 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2575 INIT_LIST_HEAD(&p->rt.run_list);
2577 INIT_LIST_HEAD(&p->se.group_node);
2579 #ifdef CONFIG_PREEMPT_NOTIFIERS
2580 INIT_HLIST_HEAD(&p->preempt_notifiers);
2585 * fork()/clone()-time setup:
2587 void sched_fork(struct task_struct *p, int clone_flags)
2589 int cpu = get_cpu();
2593 * We mark the process as running here. This guarantees that
2594 * nobody will actually run it, and a signal or other external
2595 * event cannot wake it up and insert it on the runqueue either.
2597 p->state = TASK_RUNNING;
2600 * Revert to default priority/policy on fork if requested.
2602 if (unlikely(p->sched_reset_on_fork)) {
2603 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2604 p->policy = SCHED_NORMAL;
2605 p->normal_prio = p->static_prio;
2608 if (PRIO_TO_NICE(p->static_prio) < 0) {
2609 p->static_prio = NICE_TO_PRIO(0);
2610 p->normal_prio = p->static_prio;
2615 * We don't need the reset flag anymore after the fork. It has
2616 * fulfilled its duty:
2618 p->sched_reset_on_fork = 0;
2622 * Make sure we do not leak PI boosting priority to the child.
2624 p->prio = current->normal_prio;
2626 if (!rt_prio(p->prio))
2627 p->sched_class = &fair_sched_class;
2629 if (p->sched_class->task_fork)
2630 p->sched_class->task_fork(p);
2633 * The child is not yet in the pid-hash so no cgroup attach races,
2634 * and the cgroup is pinned to this child due to cgroup_fork()
2635 * is ran before sched_fork().
2637 * Silence PROVE_RCU.
2640 set_task_cpu(p, cpu);
2643 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2644 if (likely(sched_info_on()))
2645 memset(&p->sched_info, 0, sizeof(p->sched_info));
2647 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2650 #ifdef CONFIG_PREEMPT
2651 /* Want to start with kernel preemption disabled. */
2652 task_thread_info(p)->preempt_count = 1;
2655 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2662 * wake_up_new_task - wake up a newly created task for the first time.
2664 * This function will do some initial scheduler statistics housekeeping
2665 * that must be done for every newly created context, then puts the task
2666 * on the runqueue and wakes it.
2668 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2670 unsigned long flags;
2672 int cpu __maybe_unused = get_cpu();
2675 rq = task_rq_lock(p, &flags);
2676 p->state = TASK_WAKING;
2679 * Fork balancing, do it here and not earlier because:
2680 * - cpus_allowed can change in the fork path
2681 * - any previously selected cpu might disappear through hotplug
2683 * We set TASK_WAKING so that select_task_rq() can drop rq->lock
2684 * without people poking at ->cpus_allowed.
2686 cpu = select_task_rq(rq, p, SD_BALANCE_FORK, 0);
2687 set_task_cpu(p, cpu);
2689 p->state = TASK_RUNNING;
2690 task_rq_unlock(rq, &flags);
2693 rq = task_rq_lock(p, &flags);
2694 activate_task(rq, p, 0);
2695 trace_sched_wakeup_new(p, 1);
2696 check_preempt_curr(rq, p, WF_FORK);
2698 if (p->sched_class->task_woken)
2699 p->sched_class->task_woken(rq, p);
2701 task_rq_unlock(rq, &flags);
2705 #ifdef CONFIG_PREEMPT_NOTIFIERS
2708 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2709 * @notifier: notifier struct to register
2711 void preempt_notifier_register(struct preempt_notifier *notifier)
2713 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2715 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2718 * preempt_notifier_unregister - no longer interested in preemption notifications
2719 * @notifier: notifier struct to unregister
2721 * This is safe to call from within a preemption notifier.
2723 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2725 hlist_del(¬ifier->link);
2727 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2729 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2731 struct preempt_notifier *notifier;
2732 struct hlist_node *node;
2734 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2735 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2739 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2740 struct task_struct *next)
2742 struct preempt_notifier *notifier;
2743 struct hlist_node *node;
2745 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2746 notifier->ops->sched_out(notifier, next);
2749 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2751 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2756 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2757 struct task_struct *next)
2761 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2764 * prepare_task_switch - prepare to switch tasks
2765 * @rq: the runqueue preparing to switch
2766 * @prev: the current task that is being switched out
2767 * @next: the task we are going to switch to.
2769 * This is called with the rq lock held and interrupts off. It must
2770 * be paired with a subsequent finish_task_switch after the context
2773 * prepare_task_switch sets up locking and calls architecture specific
2777 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2778 struct task_struct *next)
2780 fire_sched_out_preempt_notifiers(prev, next);
2781 prepare_lock_switch(rq, next);
2782 prepare_arch_switch(next);
2786 * finish_task_switch - clean up after a task-switch
2787 * @rq: runqueue associated with task-switch
2788 * @prev: the thread we just switched away from.
2790 * finish_task_switch must be called after the context switch, paired
2791 * with a prepare_task_switch call before the context switch.
2792 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2793 * and do any other architecture-specific cleanup actions.
2795 * Note that we may have delayed dropping an mm in context_switch(). If
2796 * so, we finish that here outside of the runqueue lock. (Doing it
2797 * with the lock held can cause deadlocks; see schedule() for
2800 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2801 __releases(rq->lock)
2803 struct mm_struct *mm = rq->prev_mm;
2809 * A task struct has one reference for the use as "current".
2810 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2811 * schedule one last time. The schedule call will never return, and
2812 * the scheduled task must drop that reference.
2813 * The test for TASK_DEAD must occur while the runqueue locks are
2814 * still held, otherwise prev could be scheduled on another cpu, die
2815 * there before we look at prev->state, and then the reference would
2817 * Manfred Spraul <manfred@colorfullife.com>
2819 prev_state = prev->state;
2820 finish_arch_switch(prev);
2821 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2822 local_irq_disable();
2823 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2824 perf_event_task_sched_in(current);
2825 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2827 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2828 finish_lock_switch(rq, prev);
2830 fire_sched_in_preempt_notifiers(current);
2833 if (unlikely(prev_state == TASK_DEAD)) {
2835 * Remove function-return probe instances associated with this
2836 * task and put them back on the free list.
2838 kprobe_flush_task(prev);
2839 put_task_struct(prev);
2845 /* assumes rq->lock is held */
2846 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2848 if (prev->sched_class->pre_schedule)
2849 prev->sched_class->pre_schedule(rq, prev);
2852 /* rq->lock is NOT held, but preemption is disabled */
2853 static inline void post_schedule(struct rq *rq)
2855 if (rq->post_schedule) {
2856 unsigned long flags;
2858 raw_spin_lock_irqsave(&rq->lock, flags);
2859 if (rq->curr->sched_class->post_schedule)
2860 rq->curr->sched_class->post_schedule(rq);
2861 raw_spin_unlock_irqrestore(&rq->lock, flags);
2863 rq->post_schedule = 0;
2869 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2873 static inline void post_schedule(struct rq *rq)
2880 * schedule_tail - first thing a freshly forked thread must call.
2881 * @prev: the thread we just switched away from.
2883 asmlinkage void schedule_tail(struct task_struct *prev)
2884 __releases(rq->lock)
2886 struct rq *rq = this_rq();
2888 finish_task_switch(rq, prev);
2891 * FIXME: do we need to worry about rq being invalidated by the
2896 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2897 /* In this case, finish_task_switch does not reenable preemption */
2900 if (current->set_child_tid)
2901 put_user(task_pid_vnr(current), current->set_child_tid);
2905 * context_switch - switch to the new MM and the new
2906 * thread's register state.
2909 context_switch(struct rq *rq, struct task_struct *prev,
2910 struct task_struct *next)
2912 struct mm_struct *mm, *oldmm;
2914 prepare_task_switch(rq, prev, next);
2915 trace_sched_switch(prev, next);
2917 oldmm = prev->active_mm;
2919 * For paravirt, this is coupled with an exit in switch_to to
2920 * combine the page table reload and the switch backend into
2923 arch_start_context_switch(prev);
2926 next->active_mm = oldmm;
2927 atomic_inc(&oldmm->mm_count);
2928 enter_lazy_tlb(oldmm, next);
2930 switch_mm(oldmm, mm, next);
2933 prev->active_mm = NULL;
2934 rq->prev_mm = oldmm;
2937 * Since the runqueue lock will be released by the next
2938 * task (which is an invalid locking op but in the case
2939 * of the scheduler it's an obvious special-case), so we
2940 * do an early lockdep release here:
2942 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2943 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2946 /* Here we just switch the register state and the stack. */
2947 switch_to(prev, next, prev);
2951 * this_rq must be evaluated again because prev may have moved
2952 * CPUs since it called schedule(), thus the 'rq' on its stack
2953 * frame will be invalid.
2955 finish_task_switch(this_rq(), prev);
2959 * nr_running, nr_uninterruptible and nr_context_switches:
2961 * externally visible scheduler statistics: current number of runnable
2962 * threads, current number of uninterruptible-sleeping threads, total
2963 * number of context switches performed since bootup.
2965 unsigned long nr_running(void)
2967 unsigned long i, sum = 0;
2969 for_each_online_cpu(i)
2970 sum += cpu_rq(i)->nr_running;
2975 unsigned long nr_uninterruptible(void)
2977 unsigned long i, sum = 0;
2979 for_each_possible_cpu(i)
2980 sum += cpu_rq(i)->nr_uninterruptible;
2983 * Since we read the counters lockless, it might be slightly
2984 * inaccurate. Do not allow it to go below zero though:
2986 if (unlikely((long)sum < 0))
2992 unsigned long long nr_context_switches(void)
2995 unsigned long long sum = 0;
2997 for_each_possible_cpu(i)
2998 sum += cpu_rq(i)->nr_switches;
3003 unsigned long nr_iowait(void)
3005 unsigned long i, sum = 0;
3007 for_each_possible_cpu(i)
3008 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3013 unsigned long nr_iowait_cpu(int cpu)
3015 struct rq *this = cpu_rq(cpu);
3016 return atomic_read(&this->nr_iowait);
3019 unsigned long this_cpu_load(void)
3021 struct rq *this = this_rq();
3022 return this->cpu_load[0];
3026 /* Variables and functions for calc_load */
3027 static atomic_long_t calc_load_tasks;
3028 static unsigned long calc_load_update;
3029 unsigned long avenrun[3];
3030 EXPORT_SYMBOL(avenrun);
3032 static long calc_load_fold_active(struct rq *this_rq)
3034 long nr_active, delta = 0;
3036 nr_active = this_rq->nr_running;
3037 nr_active += (long) this_rq->nr_uninterruptible;
3039 if (nr_active != this_rq->calc_load_active) {
3040 delta = nr_active - this_rq->calc_load_active;
3041 this_rq->calc_load_active = nr_active;
3047 static unsigned long
3048 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3051 load += active * (FIXED_1 - exp);
3052 load += 1UL << (FSHIFT - 1);
3053 return load >> FSHIFT;
3058 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
3060 * When making the ILB scale, we should try to pull this in as well.
3062 static atomic_long_t calc_load_tasks_idle;
3064 static void calc_load_account_idle(struct rq *this_rq)
3068 delta = calc_load_fold_active(this_rq);
3070 atomic_long_add(delta, &calc_load_tasks_idle);
3073 static long calc_load_fold_idle(void)
3078 * Its got a race, we don't care...
3080 if (atomic_long_read(&calc_load_tasks_idle))
3081 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
3087 * fixed_power_int - compute: x^n, in O(log n) time
3089 * @x: base of the power
3090 * @frac_bits: fractional bits of @x
3091 * @n: power to raise @x to.
3093 * By exploiting the relation between the definition of the natural power
3094 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
3095 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
3096 * (where: n_i \elem {0, 1}, the binary vector representing n),
3097 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
3098 * of course trivially computable in O(log_2 n), the length of our binary
3101 static unsigned long
3102 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
3104 unsigned long result = 1UL << frac_bits;
3109 result += 1UL << (frac_bits - 1);
3110 result >>= frac_bits;
3116 x += 1UL << (frac_bits - 1);
3124 * a1 = a0 * e + a * (1 - e)
3126 * a2 = a1 * e + a * (1 - e)
3127 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
3128 * = a0 * e^2 + a * (1 - e) * (1 + e)
3130 * a3 = a2 * e + a * (1 - e)
3131 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
3132 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
3136 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
3137 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
3138 * = a0 * e^n + a * (1 - e^n)
3140 * [1] application of the geometric series:
3143 * S_n := \Sum x^i = -------------
3146 static unsigned long
3147 calc_load_n(unsigned long load, unsigned long exp,
3148 unsigned long active, unsigned int n)
3151 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
3155 * NO_HZ can leave us missing all per-cpu ticks calling
3156 * calc_load_account_active(), but since an idle CPU folds its delta into
3157 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
3158 * in the pending idle delta if our idle period crossed a load cycle boundary.
3160 * Once we've updated the global active value, we need to apply the exponential
3161 * weights adjusted to the number of cycles missed.
3163 static void calc_global_nohz(unsigned long ticks)
3165 long delta, active, n;
3167 if (time_before(jiffies, calc_load_update))
3171 * If we crossed a calc_load_update boundary, make sure to fold
3172 * any pending idle changes, the respective CPUs might have
3173 * missed the tick driven calc_load_account_active() update
3176 delta = calc_load_fold_idle();
3178 atomic_long_add(delta, &calc_load_tasks);
3181 * If we were idle for multiple load cycles, apply them.
3183 if (ticks >= LOAD_FREQ) {
3184 n = ticks / LOAD_FREQ;
3186 active = atomic_long_read(&calc_load_tasks);
3187 active = active > 0 ? active * FIXED_1 : 0;
3189 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
3190 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
3191 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
3193 calc_load_update += n * LOAD_FREQ;
3197 * Its possible the remainder of the above division also crosses
3198 * a LOAD_FREQ period, the regular check in calc_global_load()
3199 * which comes after this will take care of that.
3201 * Consider us being 11 ticks before a cycle completion, and us
3202 * sleeping for 4*LOAD_FREQ + 22 ticks, then the above code will
3203 * age us 4 cycles, and the test in calc_global_load() will
3204 * pick up the final one.
3208 static void calc_load_account_idle(struct rq *this_rq)
3212 static inline long calc_load_fold_idle(void)
3217 static void calc_global_nohz(unsigned long ticks)
3223 * get_avenrun - get the load average array
3224 * @loads: pointer to dest load array
3225 * @offset: offset to add
3226 * @shift: shift count to shift the result left
3228 * These values are estimates at best, so no need for locking.
3230 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3232 loads[0] = (avenrun[0] + offset) << shift;
3233 loads[1] = (avenrun[1] + offset) << shift;
3234 loads[2] = (avenrun[2] + offset) << shift;
3238 * calc_load - update the avenrun load estimates 10 ticks after the
3239 * CPUs have updated calc_load_tasks.
3241 void calc_global_load(unsigned long ticks)
3245 calc_global_nohz(ticks);
3247 if (time_before(jiffies, calc_load_update + 10))
3250 active = atomic_long_read(&calc_load_tasks);
3251 active = active > 0 ? active * FIXED_1 : 0;
3253 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3254 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3255 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3257 calc_load_update += LOAD_FREQ;
3261 * Called from update_cpu_load() to periodically update this CPU's
3264 static void calc_load_account_active(struct rq *this_rq)
3268 if (time_before(jiffies, this_rq->calc_load_update))
3271 delta = calc_load_fold_active(this_rq);
3272 delta += calc_load_fold_idle();
3274 atomic_long_add(delta, &calc_load_tasks);
3276 this_rq->calc_load_update += LOAD_FREQ;
3280 * The exact cpuload at various idx values, calculated at every tick would be
3281 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3283 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3284 * on nth tick when cpu may be busy, then we have:
3285 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3286 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3288 * decay_load_missed() below does efficient calculation of
3289 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3290 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3292 * The calculation is approximated on a 128 point scale.
3293 * degrade_zero_ticks is the number of ticks after which load at any
3294 * particular idx is approximated to be zero.
3295 * degrade_factor is a precomputed table, a row for each load idx.
3296 * Each column corresponds to degradation factor for a power of two ticks,
3297 * based on 128 point scale.
3299 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3300 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3302 * With this power of 2 load factors, we can degrade the load n times
3303 * by looking at 1 bits in n and doing as many mult/shift instead of
3304 * n mult/shifts needed by the exact degradation.
3306 #define DEGRADE_SHIFT 7
3307 static const unsigned char
3308 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3309 static const unsigned char
3310 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3311 {0, 0, 0, 0, 0, 0, 0, 0},
3312 {64, 32, 8, 0, 0, 0, 0, 0},
3313 {96, 72, 40, 12, 1, 0, 0},
3314 {112, 98, 75, 43, 15, 1, 0},
3315 {120, 112, 98, 76, 45, 16, 2} };
3318 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3319 * would be when CPU is idle and so we just decay the old load without
3320 * adding any new load.
3322 static unsigned long
3323 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3327 if (!missed_updates)
3330 if (missed_updates >= degrade_zero_ticks[idx])
3334 return load >> missed_updates;
3336 while (missed_updates) {
3337 if (missed_updates % 2)
3338 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3340 missed_updates >>= 1;
3347 * Update rq->cpu_load[] statistics. This function is usually called every
3348 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3349 * every tick. We fix it up based on jiffies.
3351 static void update_cpu_load(struct rq *this_rq)
3353 unsigned long this_load = this_rq->load.weight;
3354 unsigned long curr_jiffies = jiffies;
3355 unsigned long pending_updates;
3358 this_rq->nr_load_updates++;
3360 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3361 if (curr_jiffies == this_rq->last_load_update_tick)
3364 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3365 this_rq->last_load_update_tick = curr_jiffies;
3367 /* Update our load: */
3368 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3369 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3370 unsigned long old_load, new_load;
3372 /* scale is effectively 1 << i now, and >> i divides by scale */
3374 old_load = this_rq->cpu_load[i];
3375 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3376 new_load = this_load;
3378 * Round up the averaging division if load is increasing. This
3379 * prevents us from getting stuck on 9 if the load is 10, for
3382 if (new_load > old_load)
3383 new_load += scale - 1;
3385 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3388 sched_avg_update(this_rq);
3391 static void update_cpu_load_active(struct rq *this_rq)
3393 update_cpu_load(this_rq);
3395 calc_load_account_active(this_rq);
3401 * sched_exec - execve() is a valuable balancing opportunity, because at
3402 * this point the task has the smallest effective memory and cache footprint.
3404 void sched_exec(void)
3406 struct task_struct *p = current;
3407 unsigned long flags;
3411 rq = task_rq_lock(p, &flags);
3412 dest_cpu = p->sched_class->select_task_rq(rq, p, SD_BALANCE_EXEC, 0);
3413 if (dest_cpu == smp_processor_id())
3417 * select_task_rq() can race against ->cpus_allowed
3419 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed) &&
3420 likely(cpu_active(dest_cpu)) && migrate_task(p, rq)) {
3421 struct migration_arg arg = { p, dest_cpu };
3423 task_rq_unlock(rq, &flags);
3424 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
3428 task_rq_unlock(rq, &flags);
3433 DEFINE_PER_CPU(struct kernel_stat, kstat);
3435 EXPORT_PER_CPU_SYMBOL(kstat);
3438 * Return any ns on the sched_clock that have not yet been accounted in
3439 * @p in case that task is currently running.
3441 * Called with task_rq_lock() held on @rq.
3443 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3447 if (task_current(rq, p)) {
3448 update_rq_clock(rq);
3449 ns = rq->clock_task - p->se.exec_start;
3457 unsigned long long task_delta_exec(struct task_struct *p)
3459 unsigned long flags;
3463 rq = task_rq_lock(p, &flags);
3464 ns = do_task_delta_exec(p, rq);
3465 task_rq_unlock(rq, &flags);
3471 * Return accounted runtime for the task.
3472 * In case the task is currently running, return the runtime plus current's
3473 * pending runtime that have not been accounted yet.
3475 unsigned long long task_sched_runtime(struct task_struct *p)
3477 unsigned long flags;
3481 rq = task_rq_lock(p, &flags);
3482 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3483 task_rq_unlock(rq, &flags);
3489 * Return sum_exec_runtime for the thread group.
3490 * In case the task is currently running, return the sum plus current's
3491 * pending runtime that have not been accounted yet.
3493 * Note that the thread group might have other running tasks as well,
3494 * so the return value not includes other pending runtime that other
3495 * running tasks might have.
3497 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3499 struct task_cputime totals;
3500 unsigned long flags;
3504 rq = task_rq_lock(p, &flags);
3505 thread_group_cputime(p, &totals);
3506 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3507 task_rq_unlock(rq, &flags);
3513 * Account user cpu time to a process.
3514 * @p: the process that the cpu time gets accounted to
3515 * @cputime: the cpu time spent in user space since the last update
3516 * @cputime_scaled: cputime scaled by cpu frequency
3518 void account_user_time(struct task_struct *p, cputime_t cputime,
3519 cputime_t cputime_scaled)
3521 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3524 /* Add user time to process. */
3525 p->utime = cputime_add(p->utime, cputime);
3526 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3527 account_group_user_time(p, cputime);
3529 /* Add user time to cpustat. */
3530 tmp = cputime_to_cputime64(cputime);
3531 if (TASK_NICE(p) > 0)
3532 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3534 cpustat->user = cputime64_add(cpustat->user, tmp);
3536 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3537 /* Account for user time used */
3538 acct_update_integrals(p);
3542 * Account guest cpu time to a process.
3543 * @p: the process that the cpu time gets accounted to
3544 * @cputime: the cpu time spent in virtual machine since the last update
3545 * @cputime_scaled: cputime scaled by cpu frequency
3547 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3548 cputime_t cputime_scaled)
3551 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3553 tmp = cputime_to_cputime64(cputime);
3555 /* Add guest time to process. */
3556 p->utime = cputime_add(p->utime, cputime);
3557 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3558 account_group_user_time(p, cputime);
3559 p->gtime = cputime_add(p->gtime, cputime);
3561 /* Add guest time to cpustat. */
3562 if (TASK_NICE(p) > 0) {
3563 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3564 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3566 cpustat->user = cputime64_add(cpustat->user, tmp);
3567 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3572 * Account system cpu time to a process.
3573 * @p: the process that the cpu time gets accounted to
3574 * @hardirq_offset: the offset to subtract from hardirq_count()
3575 * @cputime: the cpu time spent in kernel space since the last update
3576 * @cputime_scaled: cputime scaled by cpu frequency
3578 void account_system_time(struct task_struct *p, int hardirq_offset,
3579 cputime_t cputime, cputime_t cputime_scaled)
3581 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3584 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3585 account_guest_time(p, cputime, cputime_scaled);
3589 /* Add system time to process. */
3590 p->stime = cputime_add(p->stime, cputime);
3591 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3592 account_group_system_time(p, cputime);
3594 /* Add system time to cpustat. */
3595 tmp = cputime_to_cputime64(cputime);
3596 if (hardirq_count() - hardirq_offset)
3597 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3598 else if (in_serving_softirq())
3599 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3601 cpustat->system = cputime64_add(cpustat->system, tmp);
3603 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3605 /* Account for system time used */
3606 acct_update_integrals(p);
3610 * Account for involuntary wait time.
3611 * @steal: the cpu time spent in involuntary wait
3613 void account_steal_time(cputime_t cputime)
3615 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3616 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3618 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3622 * Account for idle time.
3623 * @cputime: the cpu time spent in idle wait
3625 void account_idle_time(cputime_t cputime)
3627 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3628 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3629 struct rq *rq = this_rq();
3631 if (atomic_read(&rq->nr_iowait) > 0)
3632 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3634 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3637 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3640 * Account a single tick of cpu time.
3641 * @p: the process that the cpu time gets accounted to
3642 * @user_tick: indicates if the tick is a user or a system tick
3644 void account_process_tick(struct task_struct *p, int user_tick)
3646 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3647 struct rq *rq = this_rq();
3650 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3651 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3652 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3655 account_idle_time(cputime_one_jiffy);
3659 * Account multiple ticks of steal time.
3660 * @p: the process from which the cpu time has been stolen
3661 * @ticks: number of stolen ticks
3663 void account_steal_ticks(unsigned long ticks)
3665 account_steal_time(jiffies_to_cputime(ticks));
3669 * Account multiple ticks of idle time.
3670 * @ticks: number of stolen ticks
3672 void account_idle_ticks(unsigned long ticks)
3674 account_idle_time(jiffies_to_cputime(ticks));
3680 * Use precise platform statistics if available:
3682 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3683 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3689 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3691 struct task_cputime cputime;
3693 thread_group_cputime(p, &cputime);
3695 *ut = cputime.utime;
3696 *st = cputime.stime;
3700 #ifndef nsecs_to_cputime
3701 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3704 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3706 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
3709 * Use CFS's precise accounting:
3711 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3717 do_div(temp, total);
3718 utime = (cputime_t)temp;
3723 * Compare with previous values, to keep monotonicity:
3725 p->prev_utime = max(p->prev_utime, utime);
3726 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
3728 *ut = p->prev_utime;
3729 *st = p->prev_stime;
3733 * Must be called with siglock held.
3735 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3737 struct signal_struct *sig = p->signal;
3738 struct task_cputime cputime;
3739 cputime_t rtime, utime, total;
3741 thread_group_cputime(p, &cputime);
3743 total = cputime_add(cputime.utime, cputime.stime);
3744 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3749 temp *= cputime.utime;
3750 do_div(temp, total);
3751 utime = (cputime_t)temp;
3755 sig->prev_utime = max(sig->prev_utime, utime);
3756 sig->prev_stime = max(sig->prev_stime,
3757 cputime_sub(rtime, sig->prev_utime));
3759 *ut = sig->prev_utime;
3760 *st = sig->prev_stime;
3765 * This function gets called by the timer code, with HZ frequency.
3766 * We call it with interrupts disabled.
3768 * It also gets called by the fork code, when changing the parent's
3771 void scheduler_tick(void)
3773 int cpu = smp_processor_id();
3774 struct rq *rq = cpu_rq(cpu);
3775 struct task_struct *curr = rq->curr;
3779 raw_spin_lock(&rq->lock);
3780 update_rq_clock(rq);
3781 update_cpu_load_active(rq);
3782 curr->sched_class->task_tick(rq, curr, 0);
3783 raw_spin_unlock(&rq->lock);
3785 perf_event_task_tick();
3788 rq->idle_at_tick = idle_cpu(cpu);
3789 trigger_load_balance(rq, cpu);
3793 notrace unsigned long get_parent_ip(unsigned long addr)
3795 if (in_lock_functions(addr)) {
3796 addr = CALLER_ADDR2;
3797 if (in_lock_functions(addr))
3798 addr = CALLER_ADDR3;
3803 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3804 defined(CONFIG_PREEMPT_TRACER))
3806 void __kprobes add_preempt_count(int val)
3808 #ifdef CONFIG_DEBUG_PREEMPT
3812 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3815 preempt_count() += val;
3816 #ifdef CONFIG_DEBUG_PREEMPT
3818 * Spinlock count overflowing soon?
3820 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3823 if (preempt_count() == val)
3824 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3826 EXPORT_SYMBOL(add_preempt_count);
3828 void __kprobes sub_preempt_count(int val)
3830 #ifdef CONFIG_DEBUG_PREEMPT
3834 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3837 * Is the spinlock portion underflowing?
3839 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3840 !(preempt_count() & PREEMPT_MASK)))
3844 if (preempt_count() == val)
3845 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3846 preempt_count() -= val;
3848 EXPORT_SYMBOL(sub_preempt_count);
3853 * Print scheduling while atomic bug:
3855 static noinline void __schedule_bug(struct task_struct *prev)
3857 struct pt_regs *regs = get_irq_regs();
3859 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3860 prev->comm, prev->pid, preempt_count());
3862 debug_show_held_locks(prev);
3864 if (irqs_disabled())
3865 print_irqtrace_events(prev);
3874 * Various schedule()-time debugging checks and statistics:
3876 static inline void schedule_debug(struct task_struct *prev)
3879 * Test if we are atomic. Since do_exit() needs to call into
3880 * schedule() atomically, we ignore that path for now.
3881 * Otherwise, whine if we are scheduling when we should not be.
3883 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3884 __schedule_bug(prev);
3886 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3888 schedstat_inc(this_rq(), sched_count);
3889 #ifdef CONFIG_SCHEDSTATS
3890 if (unlikely(prev->lock_depth >= 0)) {
3891 schedstat_inc(this_rq(), rq_sched_info.bkl_count);
3892 schedstat_inc(prev, sched_info.bkl_count);
3897 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3900 update_rq_clock(rq);
3901 prev->sched_class->put_prev_task(rq, prev);
3905 * Pick up the highest-prio task:
3907 static inline struct task_struct *
3908 pick_next_task(struct rq *rq)
3910 const struct sched_class *class;
3911 struct task_struct *p;
3914 * Optimization: we know that if all tasks are in
3915 * the fair class we can call that function directly:
3917 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3918 p = fair_sched_class.pick_next_task(rq);
3923 for_each_class(class) {
3924 p = class->pick_next_task(rq);
3929 BUG(); /* the idle class will always have a runnable task */
3933 * schedule() is the main scheduler function.
3935 asmlinkage void __sched schedule(void)
3937 struct task_struct *prev, *next;
3938 unsigned long *switch_count;
3944 cpu = smp_processor_id();
3946 rcu_note_context_switch(cpu);
3949 release_kernel_lock(prev);
3950 need_resched_nonpreemptible:
3952 schedule_debug(prev);
3954 if (sched_feat(HRTICK))
3957 raw_spin_lock_irq(&rq->lock);
3959 switch_count = &prev->nivcsw;
3960 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3961 if (unlikely(signal_pending_state(prev->state, prev))) {
3962 prev->state = TASK_RUNNING;
3965 * If a worker is going to sleep, notify and
3966 * ask workqueue whether it wants to wake up a
3967 * task to maintain concurrency. If so, wake
3970 if (prev->flags & PF_WQ_WORKER) {
3971 struct task_struct *to_wakeup;
3973 to_wakeup = wq_worker_sleeping(prev, cpu);
3975 try_to_wake_up_local(to_wakeup);
3977 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3979 switch_count = &prev->nvcsw;
3982 pre_schedule(rq, prev);
3984 if (unlikely(!rq->nr_running))
3985 idle_balance(cpu, rq);
3987 put_prev_task(rq, prev);
3988 next = pick_next_task(rq);
3989 clear_tsk_need_resched(prev);
3990 rq->skip_clock_update = 0;
3992 if (likely(prev != next)) {
3993 sched_info_switch(prev, next);
3994 perf_event_task_sched_out(prev, next);
4000 context_switch(rq, prev, next); /* unlocks the rq */
4002 * The context switch have flipped the stack from under us
4003 * and restored the local variables which were saved when
4004 * this task called schedule() in the past. prev == current
4005 * is still correct, but it can be moved to another cpu/rq.
4007 cpu = smp_processor_id();
4010 raw_spin_unlock_irq(&rq->lock);
4014 if (unlikely(reacquire_kernel_lock(prev)))
4015 goto need_resched_nonpreemptible;
4017 preempt_enable_no_resched();
4021 EXPORT_SYMBOL(schedule);
4023 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
4025 * Look out! "owner" is an entirely speculative pointer
4026 * access and not reliable.
4028 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
4033 if (!sched_feat(OWNER_SPIN))
4036 #ifdef CONFIG_DEBUG_PAGEALLOC
4038 * Need to access the cpu field knowing that
4039 * DEBUG_PAGEALLOC could have unmapped it if
4040 * the mutex owner just released it and exited.
4042 if (probe_kernel_address(&owner->cpu, cpu))
4049 * Even if the access succeeded (likely case),
4050 * the cpu field may no longer be valid.
4052 if (cpu >= nr_cpumask_bits)
4056 * We need to validate that we can do a
4057 * get_cpu() and that we have the percpu area.
4059 if (!cpu_online(cpu))
4066 * Owner changed, break to re-assess state.
4068 if (lock->owner != owner) {
4070 * If the lock has switched to a different owner,
4071 * we likely have heavy contention. Return 0 to quit
4072 * optimistic spinning and not contend further:
4080 * Is that owner really running on that cpu?
4082 if (task_thread_info(rq->curr) != owner || need_resched())
4085 arch_mutex_cpu_relax();
4092 #ifdef CONFIG_PREEMPT
4094 * this is the entry point to schedule() from in-kernel preemption
4095 * off of preempt_enable. Kernel preemptions off return from interrupt
4096 * occur there and call schedule directly.
4098 asmlinkage void __sched notrace preempt_schedule(void)
4100 struct thread_info *ti = current_thread_info();
4103 * If there is a non-zero preempt_count or interrupts are disabled,
4104 * we do not want to preempt the current task. Just return..
4106 if (likely(ti->preempt_count || irqs_disabled()))
4110 add_preempt_count_notrace(PREEMPT_ACTIVE);
4112 sub_preempt_count_notrace(PREEMPT_ACTIVE);
4115 * Check again in case we missed a preemption opportunity
4116 * between schedule and now.
4119 } while (need_resched());
4121 EXPORT_SYMBOL(preempt_schedule);
4124 * this is the entry point to schedule() from kernel preemption
4125 * off of irq context.
4126 * Note, that this is called and return with irqs disabled. This will
4127 * protect us against recursive calling from irq.
4129 asmlinkage void __sched preempt_schedule_irq(void)
4131 struct thread_info *ti = current_thread_info();
4133 /* Catch callers which need to be fixed */
4134 BUG_ON(ti->preempt_count || !irqs_disabled());
4137 add_preempt_count(PREEMPT_ACTIVE);
4140 local_irq_disable();
4141 sub_preempt_count(PREEMPT_ACTIVE);
4144 * Check again in case we missed a preemption opportunity
4145 * between schedule and now.
4148 } while (need_resched());
4151 #endif /* CONFIG_PREEMPT */
4153 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
4156 return try_to_wake_up(curr->private, mode, wake_flags);
4158 EXPORT_SYMBOL(default_wake_function);
4161 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4162 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4163 * number) then we wake all the non-exclusive tasks and one exclusive task.
4165 * There are circumstances in which we can try to wake a task which has already
4166 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4167 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4169 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4170 int nr_exclusive, int wake_flags, void *key)
4172 wait_queue_t *curr, *next;
4174 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4175 unsigned flags = curr->flags;
4177 if (curr->func(curr, mode, wake_flags, key) &&
4178 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4184 * __wake_up - wake up threads blocked on a waitqueue.
4186 * @mode: which threads
4187 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4188 * @key: is directly passed to the wakeup function
4190 * It may be assumed that this function implies a write memory barrier before
4191 * changing the task state if and only if any tasks are woken up.
4193 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4194 int nr_exclusive, void *key)
4196 unsigned long flags;
4198 spin_lock_irqsave(&q->lock, flags);
4199 __wake_up_common(q, mode, nr_exclusive, 0, key);
4200 spin_unlock_irqrestore(&q->lock, flags);
4202 EXPORT_SYMBOL(__wake_up);
4205 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4207 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4209 __wake_up_common(q, mode, 1, 0, NULL);
4211 EXPORT_SYMBOL_GPL(__wake_up_locked);
4213 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
4215 __wake_up_common(q, mode, 1, 0, key);
4217 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
4220 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4222 * @mode: which threads
4223 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4224 * @key: opaque value to be passed to wakeup targets
4226 * The sync wakeup differs that the waker knows that it will schedule
4227 * away soon, so while the target thread will be woken up, it will not
4228 * be migrated to another CPU - ie. the two threads are 'synchronized'
4229 * with each other. This can prevent needless bouncing between CPUs.
4231 * On UP it can prevent extra preemption.
4233 * It may be assumed that this function implies a write memory barrier before
4234 * changing the task state if and only if any tasks are woken up.
4236 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4237 int nr_exclusive, void *key)
4239 unsigned long flags;
4240 int wake_flags = WF_SYNC;
4245 if (unlikely(!nr_exclusive))
4248 spin_lock_irqsave(&q->lock, flags);
4249 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4250 spin_unlock_irqrestore(&q->lock, flags);
4252 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4255 * __wake_up_sync - see __wake_up_sync_key()
4257 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4259 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4261 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4264 * complete: - signals a single thread waiting on this completion
4265 * @x: holds the state of this particular completion
4267 * This will wake up a single thread waiting on this completion. Threads will be
4268 * awakened in the same order in which they were queued.
4270 * See also complete_all(), wait_for_completion() and related routines.
4272 * It may be assumed that this function implies a write memory barrier before
4273 * changing the task state if and only if any tasks are woken up.
4275 void complete(struct completion *x)
4277 unsigned long flags;
4279 spin_lock_irqsave(&x->wait.lock, flags);
4281 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4282 spin_unlock_irqrestore(&x->wait.lock, flags);
4284 EXPORT_SYMBOL(complete);
4287 * complete_all: - signals all threads waiting on this completion
4288 * @x: holds the state of this particular completion
4290 * This will wake up all threads waiting on this particular completion event.
4292 * It may be assumed that this function implies a write memory barrier before
4293 * changing the task state if and only if any tasks are woken up.
4295 void complete_all(struct completion *x)
4297 unsigned long flags;
4299 spin_lock_irqsave(&x->wait.lock, flags);
4300 x->done += UINT_MAX/2;
4301 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4302 spin_unlock_irqrestore(&x->wait.lock, flags);
4304 EXPORT_SYMBOL(complete_all);
4306 static inline long __sched
4307 do_wait_for_common(struct completion *x, long timeout, int state)
4310 DECLARE_WAITQUEUE(wait, current);
4312 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4314 if (signal_pending_state(state, current)) {
4315 timeout = -ERESTARTSYS;
4318 __set_current_state(state);
4319 spin_unlock_irq(&x->wait.lock);
4320 timeout = schedule_timeout(timeout);
4321 spin_lock_irq(&x->wait.lock);
4322 } while (!x->done && timeout);
4323 __remove_wait_queue(&x->wait, &wait);
4328 return timeout ?: 1;
4332 wait_for_common(struct completion *x, long timeout, int state)
4336 spin_lock_irq(&x->wait.lock);
4337 timeout = do_wait_for_common(x, timeout, state);
4338 spin_unlock_irq(&x->wait.lock);
4343 * wait_for_completion: - waits for completion of a task
4344 * @x: holds the state of this particular completion
4346 * This waits to be signaled for completion of a specific task. It is NOT
4347 * interruptible and there is no timeout.
4349 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4350 * and interrupt capability. Also see complete().
4352 void __sched wait_for_completion(struct completion *x)
4354 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4356 EXPORT_SYMBOL(wait_for_completion);
4359 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4360 * @x: holds the state of this particular completion
4361 * @timeout: timeout value in jiffies
4363 * This waits for either a completion of a specific task to be signaled or for a
4364 * specified timeout to expire. The timeout is in jiffies. It is not
4367 unsigned long __sched
4368 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4370 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4372 EXPORT_SYMBOL(wait_for_completion_timeout);
4375 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4376 * @x: holds the state of this particular completion
4378 * This waits for completion of a specific task to be signaled. It is
4381 int __sched wait_for_completion_interruptible(struct completion *x)
4383 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4384 if (t == -ERESTARTSYS)
4388 EXPORT_SYMBOL(wait_for_completion_interruptible);
4391 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4392 * @x: holds the state of this particular completion
4393 * @timeout: timeout value in jiffies
4395 * This waits for either a completion of a specific task to be signaled or for a
4396 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4399 wait_for_completion_interruptible_timeout(struct completion *x,
4400 unsigned long timeout)
4402 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4404 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4407 * wait_for_completion_killable: - waits for completion of a task (killable)
4408 * @x: holds the state of this particular completion
4410 * This waits to be signaled for completion of a specific task. It can be
4411 * interrupted by a kill signal.
4413 int __sched wait_for_completion_killable(struct completion *x)
4415 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4416 if (t == -ERESTARTSYS)
4420 EXPORT_SYMBOL(wait_for_completion_killable);
4423 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4424 * @x: holds the state of this particular completion
4425 * @timeout: timeout value in jiffies
4427 * This waits for either a completion of a specific task to be
4428 * signaled or for a specified timeout to expire. It can be
4429 * interrupted by a kill signal. The timeout is in jiffies.
4432 wait_for_completion_killable_timeout(struct completion *x,
4433 unsigned long timeout)
4435 return wait_for_common(x, timeout, TASK_KILLABLE);
4437 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
4440 * try_wait_for_completion - try to decrement a completion without blocking
4441 * @x: completion structure
4443 * Returns: 0 if a decrement cannot be done without blocking
4444 * 1 if a decrement succeeded.
4446 * If a completion is being used as a counting completion,
4447 * attempt to decrement the counter without blocking. This
4448 * enables us to avoid waiting if the resource the completion
4449 * is protecting is not available.
4451 bool try_wait_for_completion(struct completion *x)
4453 unsigned long flags;
4456 spin_lock_irqsave(&x->wait.lock, flags);
4461 spin_unlock_irqrestore(&x->wait.lock, flags);
4464 EXPORT_SYMBOL(try_wait_for_completion);
4467 * completion_done - Test to see if a completion has any waiters
4468 * @x: completion structure
4470 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4471 * 1 if there are no waiters.
4474 bool completion_done(struct completion *x)
4476 unsigned long flags;
4479 spin_lock_irqsave(&x->wait.lock, flags);
4482 spin_unlock_irqrestore(&x->wait.lock, flags);
4485 EXPORT_SYMBOL(completion_done);
4488 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4490 unsigned long flags;
4493 init_waitqueue_entry(&wait, current);
4495 __set_current_state(state);
4497 spin_lock_irqsave(&q->lock, flags);
4498 __add_wait_queue(q, &wait);
4499 spin_unlock(&q->lock);
4500 timeout = schedule_timeout(timeout);
4501 spin_lock_irq(&q->lock);
4502 __remove_wait_queue(q, &wait);
4503 spin_unlock_irqrestore(&q->lock, flags);
4508 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4510 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4512 EXPORT_SYMBOL(interruptible_sleep_on);
4515 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4517 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4519 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4521 void __sched sleep_on(wait_queue_head_t *q)
4523 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4525 EXPORT_SYMBOL(sleep_on);
4527 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4529 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4531 EXPORT_SYMBOL(sleep_on_timeout);
4533 #ifdef CONFIG_RT_MUTEXES
4536 * rt_mutex_setprio - set the current priority of a task
4538 * @prio: prio value (kernel-internal form)
4540 * This function changes the 'effective' priority of a task. It does
4541 * not touch ->normal_prio like __setscheduler().
4543 * Used by the rt_mutex code to implement priority inheritance logic.
4545 void rt_mutex_setprio(struct task_struct *p, int prio)
4547 unsigned long flags;
4548 int oldprio, on_rq, running;
4550 const struct sched_class *prev_class;
4552 BUG_ON(prio < 0 || prio > MAX_PRIO);
4554 rq = task_rq_lock(p, &flags);
4556 trace_sched_pi_setprio(p, prio);
4558 prev_class = p->sched_class;
4559 on_rq = p->se.on_rq;
4560 running = task_current(rq, p);
4562 dequeue_task(rq, p, 0);
4564 p->sched_class->put_prev_task(rq, p);
4567 p->sched_class = &rt_sched_class;
4569 p->sched_class = &fair_sched_class;
4574 p->sched_class->set_curr_task(rq);
4576 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4578 check_class_changed(rq, p, prev_class, oldprio, running);
4580 task_rq_unlock(rq, &flags);
4585 void set_user_nice(struct task_struct *p, long nice)
4587 int old_prio, delta, on_rq;
4588 unsigned long flags;
4591 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4594 * We have to be careful, if called from sys_setpriority(),
4595 * the task might be in the middle of scheduling on another CPU.
4597 rq = task_rq_lock(p, &flags);
4599 * The RT priorities are set via sched_setscheduler(), but we still
4600 * allow the 'normal' nice value to be set - but as expected
4601 * it wont have any effect on scheduling until the task is
4602 * SCHED_FIFO/SCHED_RR:
4604 if (task_has_rt_policy(p)) {
4605 p->static_prio = NICE_TO_PRIO(nice);
4608 on_rq = p->se.on_rq;
4610 dequeue_task(rq, p, 0);
4612 p->static_prio = NICE_TO_PRIO(nice);
4615 p->prio = effective_prio(p);
4616 delta = p->prio - old_prio;
4619 enqueue_task(rq, p, 0);
4621 * If the task increased its priority or is running and
4622 * lowered its priority, then reschedule its CPU:
4624 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4625 resched_task(rq->curr);
4628 task_rq_unlock(rq, &flags);
4630 EXPORT_SYMBOL(set_user_nice);
4633 * can_nice - check if a task can reduce its nice value
4637 int can_nice(const struct task_struct *p, const int nice)
4639 /* convert nice value [19,-20] to rlimit style value [1,40] */
4640 int nice_rlim = 20 - nice;
4642 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4643 capable(CAP_SYS_NICE));
4646 #ifdef __ARCH_WANT_SYS_NICE
4649 * sys_nice - change the priority of the current process.
4650 * @increment: priority increment
4652 * sys_setpriority is a more generic, but much slower function that
4653 * does similar things.
4655 SYSCALL_DEFINE1(nice, int, increment)
4660 * Setpriority might change our priority at the same moment.
4661 * We don't have to worry. Conceptually one call occurs first
4662 * and we have a single winner.
4664 if (increment < -40)
4669 nice = TASK_NICE(current) + increment;
4675 if (increment < 0 && !can_nice(current, nice))
4678 retval = security_task_setnice(current, nice);
4682 set_user_nice(current, nice);
4689 * task_prio - return the priority value of a given task.
4690 * @p: the task in question.
4692 * This is the priority value as seen by users in /proc.
4693 * RT tasks are offset by -200. Normal tasks are centered
4694 * around 0, value goes from -16 to +15.
4696 int task_prio(const struct task_struct *p)
4698 return p->prio - MAX_RT_PRIO;
4702 * task_nice - return the nice value of a given task.
4703 * @p: the task in question.
4705 int task_nice(const struct task_struct *p)
4707 return TASK_NICE(p);
4709 EXPORT_SYMBOL(task_nice);
4712 * idle_cpu - is a given cpu idle currently?
4713 * @cpu: the processor in question.
4715 int idle_cpu(int cpu)
4717 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4721 * idle_task - return the idle task for a given cpu.
4722 * @cpu: the processor in question.
4724 struct task_struct *idle_task(int cpu)
4726 return cpu_rq(cpu)->idle;
4730 * find_process_by_pid - find a process with a matching PID value.
4731 * @pid: the pid in question.
4733 static struct task_struct *find_process_by_pid(pid_t pid)
4735 return pid ? find_task_by_vpid(pid) : current;
4738 /* Actually do priority change: must hold rq lock. */
4740 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4742 BUG_ON(p->se.on_rq);
4745 p->rt_priority = prio;
4746 p->normal_prio = normal_prio(p);
4747 /* we are holding p->pi_lock already */
4748 p->prio = rt_mutex_getprio(p);
4749 if (rt_prio(p->prio))
4750 p->sched_class = &rt_sched_class;
4752 p->sched_class = &fair_sched_class;
4757 * check the target process has a UID that matches the current process's
4759 static bool check_same_owner(struct task_struct *p)
4761 const struct cred *cred = current_cred(), *pcred;
4765 pcred = __task_cred(p);
4766 match = (cred->euid == pcred->euid ||
4767 cred->euid == pcred->uid);
4772 static int __sched_setscheduler(struct task_struct *p, int policy,
4773 const struct sched_param *param, bool user)
4775 int retval, oldprio, oldpolicy = -1, on_rq, running;
4776 unsigned long flags;
4777 const struct sched_class *prev_class;
4781 /* may grab non-irq protected spin_locks */
4782 BUG_ON(in_interrupt());
4784 /* double check policy once rq lock held */
4786 reset_on_fork = p->sched_reset_on_fork;
4787 policy = oldpolicy = p->policy;
4789 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4790 policy &= ~SCHED_RESET_ON_FORK;
4792 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4793 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4794 policy != SCHED_IDLE)
4799 * Valid priorities for SCHED_FIFO and SCHED_RR are
4800 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4801 * SCHED_BATCH and SCHED_IDLE is 0.
4803 if (param->sched_priority < 0 ||
4804 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4805 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4807 if (rt_policy(policy) != (param->sched_priority != 0))
4811 * Allow unprivileged RT tasks to decrease priority:
4813 if (user && !capable(CAP_SYS_NICE)) {
4814 if (rt_policy(policy)) {
4815 unsigned long rlim_rtprio =
4816 task_rlimit(p, RLIMIT_RTPRIO);
4818 /* can't set/change the rt policy */
4819 if (policy != p->policy && !rlim_rtprio)
4822 /* can't increase priority */
4823 if (param->sched_priority > p->rt_priority &&
4824 param->sched_priority > rlim_rtprio)
4828 * Like positive nice levels, dont allow tasks to
4829 * move out of SCHED_IDLE either:
4831 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4834 /* can't change other user's priorities */
4835 if (!check_same_owner(p))
4838 /* Normal users shall not reset the sched_reset_on_fork flag */
4839 if (p->sched_reset_on_fork && !reset_on_fork)
4844 retval = security_task_setscheduler(p);
4850 * make sure no PI-waiters arrive (or leave) while we are
4851 * changing the priority of the task:
4853 raw_spin_lock_irqsave(&p->pi_lock, flags);
4855 * To be able to change p->policy safely, the apropriate
4856 * runqueue lock must be held.
4858 rq = __task_rq_lock(p);
4861 * Changing the policy of the stop threads its a very bad idea
4863 if (p == rq->stop) {
4864 __task_rq_unlock(rq);
4865 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4869 #ifdef CONFIG_RT_GROUP_SCHED
4872 * Do not allow realtime tasks into groups that have no runtime
4875 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4876 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4877 !task_group_is_autogroup(task_group(p))) {
4878 __task_rq_unlock(rq);
4879 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4885 /* recheck policy now with rq lock held */
4886 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4887 policy = oldpolicy = -1;
4888 __task_rq_unlock(rq);
4889 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4892 on_rq = p->se.on_rq;
4893 running = task_current(rq, p);
4895 deactivate_task(rq, p, 0);
4897 p->sched_class->put_prev_task(rq, p);
4899 p->sched_reset_on_fork = reset_on_fork;
4902 prev_class = p->sched_class;
4903 __setscheduler(rq, p, policy, param->sched_priority);
4906 p->sched_class->set_curr_task(rq);
4908 activate_task(rq, p, 0);
4910 check_class_changed(rq, p, prev_class, oldprio, running);
4912 __task_rq_unlock(rq);
4913 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4915 rt_mutex_adjust_pi(p);
4921 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4922 * @p: the task in question.
4923 * @policy: new policy.
4924 * @param: structure containing the new RT priority.
4926 * NOTE that the task may be already dead.
4928 int sched_setscheduler(struct task_struct *p, int policy,
4929 const struct sched_param *param)
4931 return __sched_setscheduler(p, policy, param, true);
4933 EXPORT_SYMBOL_GPL(sched_setscheduler);
4936 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4937 * @p: the task in question.
4938 * @policy: new policy.
4939 * @param: structure containing the new RT priority.
4941 * Just like sched_setscheduler, only don't bother checking if the
4942 * current context has permission. For example, this is needed in
4943 * stop_machine(): we create temporary high priority worker threads,
4944 * but our caller might not have that capability.
4946 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4947 const struct sched_param *param)
4949 return __sched_setscheduler(p, policy, param, false);
4953 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4955 struct sched_param lparam;
4956 struct task_struct *p;
4959 if (!param || pid < 0)
4961 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4966 p = find_process_by_pid(pid);
4968 retval = sched_setscheduler(p, policy, &lparam);
4975 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4976 * @pid: the pid in question.
4977 * @policy: new policy.
4978 * @param: structure containing the new RT priority.
4980 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4981 struct sched_param __user *, param)
4983 /* negative values for policy are not valid */
4987 return do_sched_setscheduler(pid, policy, param);
4991 * sys_sched_setparam - set/change the RT priority of a thread
4992 * @pid: the pid in question.
4993 * @param: structure containing the new RT priority.
4995 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4997 return do_sched_setscheduler(pid, -1, param);
5001 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5002 * @pid: the pid in question.
5004 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5006 struct task_struct *p;
5014 p = find_process_by_pid(pid);
5016 retval = security_task_getscheduler(p);
5019 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
5026 * sys_sched_getparam - get the RT priority of a thread
5027 * @pid: the pid in question.
5028 * @param: structure containing the RT priority.
5030 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5032 struct sched_param lp;
5033 struct task_struct *p;
5036 if (!param || pid < 0)
5040 p = find_process_by_pid(pid);
5045 retval = security_task_getscheduler(p);
5049 lp.sched_priority = p->rt_priority;
5053 * This one might sleep, we cannot do it with a spinlock held ...
5055 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5064 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5066 cpumask_var_t cpus_allowed, new_mask;
5067 struct task_struct *p;
5073 p = find_process_by_pid(pid);
5080 /* Prevent p going away */
5084 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5088 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5090 goto out_free_cpus_allowed;
5093 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
5096 retval = security_task_setscheduler(p);
5100 cpuset_cpus_allowed(p, cpus_allowed);
5101 cpumask_and(new_mask, in_mask, cpus_allowed);
5103 retval = set_cpus_allowed_ptr(p, new_mask);
5106 cpuset_cpus_allowed(p, cpus_allowed);
5107 if (!cpumask_subset(new_mask, cpus_allowed)) {
5109 * We must have raced with a concurrent cpuset
5110 * update. Just reset the cpus_allowed to the
5111 * cpuset's cpus_allowed
5113 cpumask_copy(new_mask, cpus_allowed);
5118 free_cpumask_var(new_mask);
5119 out_free_cpus_allowed:
5120 free_cpumask_var(cpus_allowed);
5127 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5128 struct cpumask *new_mask)
5130 if (len < cpumask_size())
5131 cpumask_clear(new_mask);
5132 else if (len > cpumask_size())
5133 len = cpumask_size();
5135 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5139 * sys_sched_setaffinity - set the cpu affinity of a process
5140 * @pid: pid of the process
5141 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5142 * @user_mask_ptr: user-space pointer to the new cpu mask
5144 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5145 unsigned long __user *, user_mask_ptr)
5147 cpumask_var_t new_mask;
5150 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5153 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5155 retval = sched_setaffinity(pid, new_mask);
5156 free_cpumask_var(new_mask);
5160 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5162 struct task_struct *p;
5163 unsigned long flags;
5171 p = find_process_by_pid(pid);
5175 retval = security_task_getscheduler(p);
5179 rq = task_rq_lock(p, &flags);
5180 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5181 task_rq_unlock(rq, &flags);
5191 * sys_sched_getaffinity - get the cpu affinity of a process
5192 * @pid: pid of the process
5193 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5194 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5196 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5197 unsigned long __user *, user_mask_ptr)
5202 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5204 if (len & (sizeof(unsigned long)-1))
5207 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5210 ret = sched_getaffinity(pid, mask);
5212 size_t retlen = min_t(size_t, len, cpumask_size());
5214 if (copy_to_user(user_mask_ptr, mask, retlen))
5219 free_cpumask_var(mask);
5225 * sys_sched_yield - yield the current processor to other threads.
5227 * This function yields the current CPU to other tasks. If there are no
5228 * other threads running on this CPU then this function will return.
5230 SYSCALL_DEFINE0(sched_yield)
5232 struct rq *rq = this_rq_lock();
5234 schedstat_inc(rq, yld_count);
5235 current->sched_class->yield_task(rq);
5238 * Since we are going to call schedule() anyway, there's
5239 * no need to preempt or enable interrupts:
5241 __release(rq->lock);
5242 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5243 do_raw_spin_unlock(&rq->lock);
5244 preempt_enable_no_resched();
5251 static inline int should_resched(void)
5253 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5256 static void __cond_resched(void)
5258 add_preempt_count(PREEMPT_ACTIVE);
5260 sub_preempt_count(PREEMPT_ACTIVE);
5263 int __sched _cond_resched(void)
5265 if (should_resched()) {
5271 EXPORT_SYMBOL(_cond_resched);
5274 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5275 * call schedule, and on return reacquire the lock.
5277 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5278 * operations here to prevent schedule() from being called twice (once via
5279 * spin_unlock(), once by hand).
5281 int __cond_resched_lock(spinlock_t *lock)
5283 int resched = should_resched();
5286 lockdep_assert_held(lock);
5288 if (spin_needbreak(lock) || resched) {
5299 EXPORT_SYMBOL(__cond_resched_lock);
5301 int __sched __cond_resched_softirq(void)
5303 BUG_ON(!in_softirq());
5305 if (should_resched()) {
5313 EXPORT_SYMBOL(__cond_resched_softirq);
5316 * yield - yield the current processor to other threads.
5318 * This is a shortcut for kernel-space yielding - it marks the
5319 * thread runnable and calls sys_sched_yield().
5321 void __sched yield(void)
5323 set_current_state(TASK_RUNNING);
5326 EXPORT_SYMBOL(yield);
5329 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5330 * that process accounting knows that this is a task in IO wait state.
5332 void __sched io_schedule(void)
5334 struct rq *rq = raw_rq();
5336 delayacct_blkio_start();
5337 atomic_inc(&rq->nr_iowait);
5338 current->in_iowait = 1;
5340 current->in_iowait = 0;
5341 atomic_dec(&rq->nr_iowait);
5342 delayacct_blkio_end();
5344 EXPORT_SYMBOL(io_schedule);
5346 long __sched io_schedule_timeout(long timeout)
5348 struct rq *rq = raw_rq();
5351 delayacct_blkio_start();
5352 atomic_inc(&rq->nr_iowait);
5353 current->in_iowait = 1;
5354 ret = schedule_timeout(timeout);
5355 current->in_iowait = 0;
5356 atomic_dec(&rq->nr_iowait);
5357 delayacct_blkio_end();
5362 * sys_sched_get_priority_max - return maximum RT priority.
5363 * @policy: scheduling class.
5365 * this syscall returns the maximum rt_priority that can be used
5366 * by a given scheduling class.
5368 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5375 ret = MAX_USER_RT_PRIO-1;
5387 * sys_sched_get_priority_min - return minimum RT priority.
5388 * @policy: scheduling class.
5390 * this syscall returns the minimum rt_priority that can be used
5391 * by a given scheduling class.
5393 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5411 * sys_sched_rr_get_interval - return the default timeslice of a process.
5412 * @pid: pid of the process.
5413 * @interval: userspace pointer to the timeslice value.
5415 * this syscall writes the default timeslice value of a given process
5416 * into the user-space timespec buffer. A value of '0' means infinity.
5418 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5419 struct timespec __user *, interval)
5421 struct task_struct *p;
5422 unsigned int time_slice;
5423 unsigned long flags;
5433 p = find_process_by_pid(pid);
5437 retval = security_task_getscheduler(p);
5441 rq = task_rq_lock(p, &flags);
5442 time_slice = p->sched_class->get_rr_interval(rq, p);
5443 task_rq_unlock(rq, &flags);
5446 jiffies_to_timespec(time_slice, &t);
5447 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5455 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5457 void sched_show_task(struct task_struct *p)
5459 unsigned long free = 0;
5462 state = p->state ? __ffs(p->state) + 1 : 0;
5463 printk(KERN_INFO "%-15.15s %c", p->comm,
5464 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5465 #if BITS_PER_LONG == 32
5466 if (state == TASK_RUNNING)
5467 printk(KERN_CONT " running ");
5469 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5471 if (state == TASK_RUNNING)
5472 printk(KERN_CONT " running task ");
5474 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5476 #ifdef CONFIG_DEBUG_STACK_USAGE
5477 free = stack_not_used(p);
5479 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5480 task_pid_nr(p), task_pid_nr(p->real_parent),
5481 (unsigned long)task_thread_info(p)->flags);
5483 show_stack(p, NULL);
5486 void show_state_filter(unsigned long state_filter)
5488 struct task_struct *g, *p;
5490 #if BITS_PER_LONG == 32
5492 " task PC stack pid father\n");
5495 " task PC stack pid father\n");
5497 read_lock(&tasklist_lock);
5498 do_each_thread(g, p) {
5500 * reset the NMI-timeout, listing all files on a slow
5501 * console might take alot of time:
5503 touch_nmi_watchdog();
5504 if (!state_filter || (p->state & state_filter))
5506 } while_each_thread(g, p);
5508 touch_all_softlockup_watchdogs();
5510 #ifdef CONFIG_SCHED_DEBUG
5511 sysrq_sched_debug_show();
5513 read_unlock(&tasklist_lock);
5515 * Only show locks if all tasks are dumped:
5518 debug_show_all_locks();
5521 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5523 idle->sched_class = &idle_sched_class;
5527 * init_idle - set up an idle thread for a given CPU
5528 * @idle: task in question
5529 * @cpu: cpu the idle task belongs to
5531 * NOTE: this function does not set the idle thread's NEED_RESCHED
5532 * flag, to make booting more robust.
5534 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5536 struct rq *rq = cpu_rq(cpu);
5537 unsigned long flags;
5539 raw_spin_lock_irqsave(&rq->lock, flags);
5542 idle->state = TASK_RUNNING;
5543 idle->se.exec_start = sched_clock();
5545 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
5547 * We're having a chicken and egg problem, even though we are
5548 * holding rq->lock, the cpu isn't yet set to this cpu so the
5549 * lockdep check in task_group() will fail.
5551 * Similar case to sched_fork(). / Alternatively we could
5552 * use task_rq_lock() here and obtain the other rq->lock.
5557 __set_task_cpu(idle, cpu);
5560 rq->curr = rq->idle = idle;
5561 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5564 raw_spin_unlock_irqrestore(&rq->lock, flags);
5566 /* Set the preempt count _outside_ the spinlocks! */
5567 #if defined(CONFIG_PREEMPT)
5568 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5570 task_thread_info(idle)->preempt_count = 0;
5573 * The idle tasks have their own, simple scheduling class:
5575 idle->sched_class = &idle_sched_class;
5576 ftrace_graph_init_task(idle);
5580 * In a system that switches off the HZ timer nohz_cpu_mask
5581 * indicates which cpus entered this state. This is used
5582 * in the rcu update to wait only for active cpus. For system
5583 * which do not switch off the HZ timer nohz_cpu_mask should
5584 * always be CPU_BITS_NONE.
5586 cpumask_var_t nohz_cpu_mask;
5589 * Increase the granularity value when there are more CPUs,
5590 * because with more CPUs the 'effective latency' as visible
5591 * to users decreases. But the relationship is not linear,
5592 * so pick a second-best guess by going with the log2 of the
5595 * This idea comes from the SD scheduler of Con Kolivas:
5597 static int get_update_sysctl_factor(void)
5599 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5600 unsigned int factor;
5602 switch (sysctl_sched_tunable_scaling) {
5603 case SCHED_TUNABLESCALING_NONE:
5606 case SCHED_TUNABLESCALING_LINEAR:
5609 case SCHED_TUNABLESCALING_LOG:
5611 factor = 1 + ilog2(cpus);
5618 static void update_sysctl(void)
5620 unsigned int factor = get_update_sysctl_factor();
5622 #define SET_SYSCTL(name) \
5623 (sysctl_##name = (factor) * normalized_sysctl_##name)
5624 SET_SYSCTL(sched_min_granularity);
5625 SET_SYSCTL(sched_latency);
5626 SET_SYSCTL(sched_wakeup_granularity);
5630 static inline void sched_init_granularity(void)
5637 * This is how migration works:
5639 * 1) we invoke migration_cpu_stop() on the target CPU using
5641 * 2) stopper starts to run (implicitly forcing the migrated thread
5643 * 3) it checks whether the migrated task is still in the wrong runqueue.
5644 * 4) if it's in the wrong runqueue then the migration thread removes
5645 * it and puts it into the right queue.
5646 * 5) stopper completes and stop_one_cpu() returns and the migration
5651 * Change a given task's CPU affinity. Migrate the thread to a
5652 * proper CPU and schedule it away if the CPU it's executing on
5653 * is removed from the allowed bitmask.
5655 * NOTE: the caller must have a valid reference to the task, the
5656 * task must not exit() & deallocate itself prematurely. The
5657 * call is not atomic; no spinlocks may be held.
5659 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5661 unsigned long flags;
5663 unsigned int dest_cpu;
5667 * Serialize against TASK_WAKING so that ttwu() and wunt() can
5668 * drop the rq->lock and still rely on ->cpus_allowed.
5671 while (task_is_waking(p))
5673 rq = task_rq_lock(p, &flags);
5674 if (task_is_waking(p)) {
5675 task_rq_unlock(rq, &flags);
5679 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5684 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5685 !cpumask_equal(&p->cpus_allowed, new_mask))) {
5690 if (p->sched_class->set_cpus_allowed)
5691 p->sched_class->set_cpus_allowed(p, new_mask);
5693 cpumask_copy(&p->cpus_allowed, new_mask);
5694 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5697 /* Can the task run on the task's current CPU? If so, we're done */
5698 if (cpumask_test_cpu(task_cpu(p), new_mask))
5701 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
5702 if (migrate_task(p, rq)) {
5703 struct migration_arg arg = { p, dest_cpu };
5704 /* Need help from migration thread: drop lock and wait. */
5705 task_rq_unlock(rq, &flags);
5706 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
5707 tlb_migrate_finish(p->mm);
5711 task_rq_unlock(rq, &flags);
5715 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5718 * Move (not current) task off this cpu, onto dest cpu. We're doing
5719 * this because either it can't run here any more (set_cpus_allowed()
5720 * away from this CPU, or CPU going down), or because we're
5721 * attempting to rebalance this task on exec (sched_exec).
5723 * So we race with normal scheduler movements, but that's OK, as long
5724 * as the task is no longer on this CPU.
5726 * Returns non-zero if task was successfully migrated.
5728 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5730 struct rq *rq_dest, *rq_src;
5733 if (unlikely(!cpu_active(dest_cpu)))
5736 rq_src = cpu_rq(src_cpu);
5737 rq_dest = cpu_rq(dest_cpu);
5739 double_rq_lock(rq_src, rq_dest);
5740 /* Already moved. */
5741 if (task_cpu(p) != src_cpu)
5743 /* Affinity changed (again). */
5744 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
5748 * If we're not on a rq, the next wake-up will ensure we're
5752 deactivate_task(rq_src, p, 0);
5753 set_task_cpu(p, dest_cpu);
5754 activate_task(rq_dest, p, 0);
5755 check_preempt_curr(rq_dest, p, 0);
5760 double_rq_unlock(rq_src, rq_dest);
5765 * migration_cpu_stop - this will be executed by a highprio stopper thread
5766 * and performs thread migration by bumping thread off CPU then
5767 * 'pushing' onto another runqueue.
5769 static int migration_cpu_stop(void *data)
5771 struct migration_arg *arg = data;
5774 * The original target cpu might have gone down and we might
5775 * be on another cpu but it doesn't matter.
5777 local_irq_disable();
5778 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
5783 #ifdef CONFIG_HOTPLUG_CPU
5786 * Ensures that the idle task is using init_mm right before its cpu goes
5789 void idle_task_exit(void)
5791 struct mm_struct *mm = current->active_mm;
5793 BUG_ON(cpu_online(smp_processor_id()));
5796 switch_mm(mm, &init_mm, current);
5801 * While a dead CPU has no uninterruptible tasks queued at this point,
5802 * it might still have a nonzero ->nr_uninterruptible counter, because
5803 * for performance reasons the counter is not stricly tracking tasks to
5804 * their home CPUs. So we just add the counter to another CPU's counter,
5805 * to keep the global sum constant after CPU-down:
5807 static void migrate_nr_uninterruptible(struct rq *rq_src)
5809 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5811 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5812 rq_src->nr_uninterruptible = 0;
5816 * remove the tasks which were accounted by rq from calc_load_tasks.
5818 static void calc_global_load_remove(struct rq *rq)
5820 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5821 rq->calc_load_active = 0;
5825 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5826 * try_to_wake_up()->select_task_rq().
5828 * Called with rq->lock held even though we'er in stop_machine() and
5829 * there's no concurrency possible, we hold the required locks anyway
5830 * because of lock validation efforts.
5832 static void migrate_tasks(unsigned int dead_cpu)
5834 struct rq *rq = cpu_rq(dead_cpu);
5835 struct task_struct *next, *stop = rq->stop;
5839 * Fudge the rq selection such that the below task selection loop
5840 * doesn't get stuck on the currently eligible stop task.
5842 * We're currently inside stop_machine() and the rq is either stuck
5843 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5844 * either way we should never end up calling schedule() until we're
5851 * There's this thread running, bail when that's the only
5854 if (rq->nr_running == 1)
5857 next = pick_next_task(rq);
5859 next->sched_class->put_prev_task(rq, next);
5861 /* Find suitable destination for @next, with force if needed. */
5862 dest_cpu = select_fallback_rq(dead_cpu, next);
5863 raw_spin_unlock(&rq->lock);
5865 __migrate_task(next, dead_cpu, dest_cpu);
5867 raw_spin_lock(&rq->lock);
5873 #endif /* CONFIG_HOTPLUG_CPU */
5875 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5877 static struct ctl_table sd_ctl_dir[] = {
5879 .procname = "sched_domain",
5885 static struct ctl_table sd_ctl_root[] = {
5887 .procname = "kernel",
5889 .child = sd_ctl_dir,
5894 static struct ctl_table *sd_alloc_ctl_entry(int n)
5896 struct ctl_table *entry =
5897 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5902 static void sd_free_ctl_entry(struct ctl_table **tablep)
5904 struct ctl_table *entry;
5907 * In the intermediate directories, both the child directory and
5908 * procname are dynamically allocated and could fail but the mode
5909 * will always be set. In the lowest directory the names are
5910 * static strings and all have proc handlers.
5912 for (entry = *tablep; entry->mode; entry++) {
5914 sd_free_ctl_entry(&entry->child);
5915 if (entry->proc_handler == NULL)
5916 kfree(entry->procname);
5924 set_table_entry(struct ctl_table *entry,
5925 const char *procname, void *data, int maxlen,
5926 mode_t mode, proc_handler *proc_handler)
5928 entry->procname = procname;
5930 entry->maxlen = maxlen;
5932 entry->proc_handler = proc_handler;
5935 static struct ctl_table *
5936 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5938 struct ctl_table *table = sd_alloc_ctl_entry(13);
5943 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5944 sizeof(long), 0644, proc_doulongvec_minmax);
5945 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5946 sizeof(long), 0644, proc_doulongvec_minmax);
5947 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5948 sizeof(int), 0644, proc_dointvec_minmax);
5949 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5950 sizeof(int), 0644, proc_dointvec_minmax);
5951 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5952 sizeof(int), 0644, proc_dointvec_minmax);
5953 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5954 sizeof(int), 0644, proc_dointvec_minmax);
5955 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5956 sizeof(int), 0644, proc_dointvec_minmax);
5957 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5958 sizeof(int), 0644, proc_dointvec_minmax);
5959 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5960 sizeof(int), 0644, proc_dointvec_minmax);
5961 set_table_entry(&table[9], "cache_nice_tries",
5962 &sd->cache_nice_tries,
5963 sizeof(int), 0644, proc_dointvec_minmax);
5964 set_table_entry(&table[10], "flags", &sd->flags,
5965 sizeof(int), 0644, proc_dointvec_minmax);
5966 set_table_entry(&table[11], "name", sd->name,
5967 CORENAME_MAX_SIZE, 0444, proc_dostring);
5968 /* &table[12] is terminator */
5973 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5975 struct ctl_table *entry, *table;
5976 struct sched_domain *sd;
5977 int domain_num = 0, i;
5980 for_each_domain(cpu, sd)
5982 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5987 for_each_domain(cpu, sd) {
5988 snprintf(buf, 32, "domain%d", i);
5989 entry->procname = kstrdup(buf, GFP_KERNEL);
5991 entry->child = sd_alloc_ctl_domain_table(sd);
5998 static struct ctl_table_header *sd_sysctl_header;
5999 static void register_sched_domain_sysctl(void)
6001 int i, cpu_num = num_possible_cpus();
6002 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6005 WARN_ON(sd_ctl_dir[0].child);
6006 sd_ctl_dir[0].child = entry;
6011 for_each_possible_cpu(i) {
6012 snprintf(buf, 32, "cpu%d", i);
6013 entry->procname = kstrdup(buf, GFP_KERNEL);
6015 entry->child = sd_alloc_ctl_cpu_table(i);
6019 WARN_ON(sd_sysctl_header);
6020 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6023 /* may be called multiple times per register */
6024 static void unregister_sched_domain_sysctl(void)
6026 if (sd_sysctl_header)
6027 unregister_sysctl_table(sd_sysctl_header);
6028 sd_sysctl_header = NULL;
6029 if (sd_ctl_dir[0].child)
6030 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6033 static void register_sched_domain_sysctl(void)
6036 static void unregister_sched_domain_sysctl(void)
6041 static void set_rq_online(struct rq *rq)
6044 const struct sched_class *class;
6046 cpumask_set_cpu(rq->cpu, rq->rd->online);
6049 for_each_class(class) {
6050 if (class->rq_online)
6051 class->rq_online(rq);
6056 static void set_rq_offline(struct rq *rq)
6059 const struct sched_class *class;
6061 for_each_class(class) {
6062 if (class->rq_offline)
6063 class->rq_offline(rq);
6066 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6072 * migration_call - callback that gets triggered when a CPU is added.
6073 * Here we can start up the necessary migration thread for the new CPU.
6075 static int __cpuinit
6076 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6078 int cpu = (long)hcpu;
6079 unsigned long flags;
6080 struct rq *rq = cpu_rq(cpu);
6082 switch (action & ~CPU_TASKS_FROZEN) {
6084 case CPU_UP_PREPARE:
6085 rq->calc_load_update = calc_load_update;
6089 /* Update our root-domain */
6090 raw_spin_lock_irqsave(&rq->lock, flags);
6092 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6096 raw_spin_unlock_irqrestore(&rq->lock, flags);
6099 #ifdef CONFIG_HOTPLUG_CPU
6101 /* Update our root-domain */
6102 raw_spin_lock_irqsave(&rq->lock, flags);
6104 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6108 BUG_ON(rq->nr_running != 1); /* the migration thread */
6109 raw_spin_unlock_irqrestore(&rq->lock, flags);
6111 migrate_nr_uninterruptible(rq);
6112 calc_global_load_remove(rq);
6120 * Register at high priority so that task migration (migrate_all_tasks)
6121 * happens before everything else. This has to be lower priority than
6122 * the notifier in the perf_event subsystem, though.
6124 static struct notifier_block __cpuinitdata migration_notifier = {
6125 .notifier_call = migration_call,
6126 .priority = CPU_PRI_MIGRATION,
6129 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
6130 unsigned long action, void *hcpu)
6132 switch (action & ~CPU_TASKS_FROZEN) {
6134 case CPU_DOWN_FAILED:
6135 set_cpu_active((long)hcpu, true);
6142 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
6143 unsigned long action, void *hcpu)
6145 switch (action & ~CPU_TASKS_FROZEN) {
6146 case CPU_DOWN_PREPARE:
6147 set_cpu_active((long)hcpu, false);
6154 static int __init migration_init(void)
6156 void *cpu = (void *)(long)smp_processor_id();
6159 /* Initialize migration for the boot CPU */
6160 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6161 BUG_ON(err == NOTIFY_BAD);
6162 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6163 register_cpu_notifier(&migration_notifier);
6165 /* Register cpu active notifiers */
6166 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
6167 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
6171 early_initcall(migration_init);
6176 #ifdef CONFIG_SCHED_DEBUG
6178 static __read_mostly int sched_domain_debug_enabled;
6180 static int __init sched_domain_debug_setup(char *str)
6182 sched_domain_debug_enabled = 1;
6186 early_param("sched_debug", sched_domain_debug_setup);
6188 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6189 struct cpumask *groupmask)
6191 struct sched_group *group = sd->groups;
6194 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6195 cpumask_clear(groupmask);
6197 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6199 if (!(sd->flags & SD_LOAD_BALANCE)) {
6200 printk("does not load-balance\n");
6202 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6207 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6209 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6210 printk(KERN_ERR "ERROR: domain->span does not contain "
6213 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6214 printk(KERN_ERR "ERROR: domain->groups does not contain"
6218 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6222 printk(KERN_ERR "ERROR: group is NULL\n");
6226 if (!group->cpu_power) {
6227 printk(KERN_CONT "\n");
6228 printk(KERN_ERR "ERROR: domain->cpu_power not "
6233 if (!cpumask_weight(sched_group_cpus(group))) {
6234 printk(KERN_CONT "\n");
6235 printk(KERN_ERR "ERROR: empty group\n");
6239 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6240 printk(KERN_CONT "\n");
6241 printk(KERN_ERR "ERROR: repeated CPUs\n");
6245 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6247 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6249 printk(KERN_CONT " %s", str);
6250 if (group->cpu_power != SCHED_LOAD_SCALE) {
6251 printk(KERN_CONT " (cpu_power = %d)",
6255 group = group->next;
6256 } while (group != sd->groups);
6257 printk(KERN_CONT "\n");
6259 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6260 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6263 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6264 printk(KERN_ERR "ERROR: parent span is not a superset "
6265 "of domain->span\n");
6269 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6271 cpumask_var_t groupmask;
6274 if (!sched_domain_debug_enabled)
6278 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6282 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6284 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6285 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6290 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6297 free_cpumask_var(groupmask);
6299 #else /* !CONFIG_SCHED_DEBUG */
6300 # define sched_domain_debug(sd, cpu) do { } while (0)
6301 #endif /* CONFIG_SCHED_DEBUG */
6303 static int sd_degenerate(struct sched_domain *sd)
6305 if (cpumask_weight(sched_domain_span(sd)) == 1)
6308 /* Following flags need at least 2 groups */
6309 if (sd->flags & (SD_LOAD_BALANCE |
6310 SD_BALANCE_NEWIDLE |
6314 SD_SHARE_PKG_RESOURCES)) {
6315 if (sd->groups != sd->groups->next)
6319 /* Following flags don't use groups */
6320 if (sd->flags & (SD_WAKE_AFFINE))
6327 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6329 unsigned long cflags = sd->flags, pflags = parent->flags;
6331 if (sd_degenerate(parent))
6334 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6337 /* Flags needing groups don't count if only 1 group in parent */
6338 if (parent->groups == parent->groups->next) {
6339 pflags &= ~(SD_LOAD_BALANCE |
6340 SD_BALANCE_NEWIDLE |
6344 SD_SHARE_PKG_RESOURCES);
6345 if (nr_node_ids == 1)
6346 pflags &= ~SD_SERIALIZE;
6348 if (~cflags & pflags)
6354 static void free_rootdomain(struct root_domain *rd)
6356 synchronize_sched();
6358 cpupri_cleanup(&rd->cpupri);
6360 free_cpumask_var(rd->rto_mask);
6361 free_cpumask_var(rd->online);
6362 free_cpumask_var(rd->span);
6366 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6368 struct root_domain *old_rd = NULL;
6369 unsigned long flags;
6371 raw_spin_lock_irqsave(&rq->lock, flags);
6376 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6379 cpumask_clear_cpu(rq->cpu, old_rd->span);
6382 * If we dont want to free the old_rt yet then
6383 * set old_rd to NULL to skip the freeing later
6386 if (!atomic_dec_and_test(&old_rd->refcount))
6390 atomic_inc(&rd->refcount);
6393 cpumask_set_cpu(rq->cpu, rd->span);
6394 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6397 raw_spin_unlock_irqrestore(&rq->lock, flags);
6400 free_rootdomain(old_rd);
6403 static int init_rootdomain(struct root_domain *rd)
6405 memset(rd, 0, sizeof(*rd));
6407 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
6409 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
6411 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6414 if (cpupri_init(&rd->cpupri) != 0)
6419 free_cpumask_var(rd->rto_mask);
6421 free_cpumask_var(rd->online);
6423 free_cpumask_var(rd->span);
6428 static void init_defrootdomain(void)
6430 init_rootdomain(&def_root_domain);
6432 atomic_set(&def_root_domain.refcount, 1);
6435 static struct root_domain *alloc_rootdomain(void)
6437 struct root_domain *rd;
6439 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6443 if (init_rootdomain(rd) != 0) {
6452 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6453 * hold the hotplug lock.
6456 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6458 struct rq *rq = cpu_rq(cpu);
6459 struct sched_domain *tmp;
6461 for (tmp = sd; tmp; tmp = tmp->parent)
6462 tmp->span_weight = cpumask_weight(sched_domain_span(tmp));
6464 /* Remove the sched domains which do not contribute to scheduling. */
6465 for (tmp = sd; tmp; ) {
6466 struct sched_domain *parent = tmp->parent;
6470 if (sd_parent_degenerate(tmp, parent)) {
6471 tmp->parent = parent->parent;
6473 parent->parent->child = tmp;
6478 if (sd && sd_degenerate(sd)) {
6484 sched_domain_debug(sd, cpu);
6486 rq_attach_root(rq, rd);
6487 rcu_assign_pointer(rq->sd, sd);
6490 /* cpus with isolated domains */
6491 static cpumask_var_t cpu_isolated_map;
6493 /* Setup the mask of cpus configured for isolated domains */
6494 static int __init isolated_cpu_setup(char *str)
6496 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6497 cpulist_parse(str, cpu_isolated_map);
6501 __setup("isolcpus=", isolated_cpu_setup);
6504 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6505 * to a function which identifies what group(along with sched group) a CPU
6506 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6507 * (due to the fact that we keep track of groups covered with a struct cpumask).
6509 * init_sched_build_groups will build a circular linked list of the groups
6510 * covered by the given span, and will set each group's ->cpumask correctly,
6511 * and ->cpu_power to 0.
6514 init_sched_build_groups(const struct cpumask *span,
6515 const struct cpumask *cpu_map,
6516 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
6517 struct sched_group **sg,
6518 struct cpumask *tmpmask),
6519 struct cpumask *covered, struct cpumask *tmpmask)
6521 struct sched_group *first = NULL, *last = NULL;
6524 cpumask_clear(covered);
6526 for_each_cpu(i, span) {
6527 struct sched_group *sg;
6528 int group = group_fn(i, cpu_map, &sg, tmpmask);
6531 if (cpumask_test_cpu(i, covered))
6534 cpumask_clear(sched_group_cpus(sg));
6537 for_each_cpu(j, span) {
6538 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6541 cpumask_set_cpu(j, covered);
6542 cpumask_set_cpu(j, sched_group_cpus(sg));
6553 #define SD_NODES_PER_DOMAIN 16
6558 * find_next_best_node - find the next node to include in a sched_domain
6559 * @node: node whose sched_domain we're building
6560 * @used_nodes: nodes already in the sched_domain
6562 * Find the next node to include in a given scheduling domain. Simply
6563 * finds the closest node not already in the @used_nodes map.
6565 * Should use nodemask_t.
6567 static int find_next_best_node(int node, nodemask_t *used_nodes)
6569 int i, n, val, min_val, best_node = 0;
6573 for (i = 0; i < nr_node_ids; i++) {
6574 /* Start at @node */
6575 n = (node + i) % nr_node_ids;
6577 if (!nr_cpus_node(n))
6580 /* Skip already used nodes */
6581 if (node_isset(n, *used_nodes))
6584 /* Simple min distance search */
6585 val = node_distance(node, n);
6587 if (val < min_val) {
6593 node_set(best_node, *used_nodes);
6598 * sched_domain_node_span - get a cpumask for a node's sched_domain
6599 * @node: node whose cpumask we're constructing
6600 * @span: resulting cpumask
6602 * Given a node, construct a good cpumask for its sched_domain to span. It
6603 * should be one that prevents unnecessary balancing, but also spreads tasks
6606 static void sched_domain_node_span(int node, struct cpumask *span)
6608 nodemask_t used_nodes;
6611 cpumask_clear(span);
6612 nodes_clear(used_nodes);
6614 cpumask_or(span, span, cpumask_of_node(node));
6615 node_set(node, used_nodes);
6617 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6618 int next_node = find_next_best_node(node, &used_nodes);
6620 cpumask_or(span, span, cpumask_of_node(next_node));
6623 #endif /* CONFIG_NUMA */
6625 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6628 * The cpus mask in sched_group and sched_domain hangs off the end.
6630 * ( See the the comments in include/linux/sched.h:struct sched_group
6631 * and struct sched_domain. )
6633 struct static_sched_group {
6634 struct sched_group sg;
6635 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
6638 struct static_sched_domain {
6639 struct sched_domain sd;
6640 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
6646 cpumask_var_t domainspan;
6647 cpumask_var_t covered;
6648 cpumask_var_t notcovered;
6650 cpumask_var_t nodemask;
6651 cpumask_var_t this_sibling_map;
6652 cpumask_var_t this_core_map;
6653 cpumask_var_t this_book_map;
6654 cpumask_var_t send_covered;
6655 cpumask_var_t tmpmask;
6656 struct sched_group **sched_group_nodes;
6657 struct root_domain *rd;
6661 sa_sched_groups = 0,
6667 sa_this_sibling_map,
6669 sa_sched_group_nodes,
6679 * SMT sched-domains:
6681 #ifdef CONFIG_SCHED_SMT
6682 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
6683 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
6686 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
6687 struct sched_group **sg, struct cpumask *unused)
6690 *sg = &per_cpu(sched_groups, cpu).sg;
6693 #endif /* CONFIG_SCHED_SMT */
6696 * multi-core sched-domains:
6698 #ifdef CONFIG_SCHED_MC
6699 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
6700 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
6703 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6704 struct sched_group **sg, struct cpumask *mask)
6707 #ifdef CONFIG_SCHED_SMT
6708 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6709 group = cpumask_first(mask);
6714 *sg = &per_cpu(sched_group_core, group).sg;
6717 #endif /* CONFIG_SCHED_MC */
6720 * book sched-domains:
6722 #ifdef CONFIG_SCHED_BOOK
6723 static DEFINE_PER_CPU(struct static_sched_domain, book_domains);
6724 static DEFINE_PER_CPU(struct static_sched_group, sched_group_book);
6727 cpu_to_book_group(int cpu, const struct cpumask *cpu_map,
6728 struct sched_group **sg, struct cpumask *mask)
6731 #ifdef CONFIG_SCHED_MC
6732 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6733 group = cpumask_first(mask);
6734 #elif defined(CONFIG_SCHED_SMT)
6735 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6736 group = cpumask_first(mask);
6739 *sg = &per_cpu(sched_group_book, group).sg;
6742 #endif /* CONFIG_SCHED_BOOK */
6744 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
6745 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
6748 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
6749 struct sched_group **sg, struct cpumask *mask)
6752 #ifdef CONFIG_SCHED_BOOK
6753 cpumask_and(mask, cpu_book_mask(cpu), cpu_map);
6754 group = cpumask_first(mask);
6755 #elif defined(CONFIG_SCHED_MC)
6756 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6757 group = cpumask_first(mask);
6758 #elif defined(CONFIG_SCHED_SMT)
6759 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6760 group = cpumask_first(mask);
6765 *sg = &per_cpu(sched_group_phys, group).sg;
6771 * The init_sched_build_groups can't handle what we want to do with node
6772 * groups, so roll our own. Now each node has its own list of groups which
6773 * gets dynamically allocated.
6775 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
6776 static struct sched_group ***sched_group_nodes_bycpu;
6778 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
6779 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
6781 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
6782 struct sched_group **sg,
6783 struct cpumask *nodemask)
6787 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
6788 group = cpumask_first(nodemask);
6791 *sg = &per_cpu(sched_group_allnodes, group).sg;
6795 static void init_numa_sched_groups_power(struct sched_group *group_head)
6797 struct sched_group *sg = group_head;
6803 for_each_cpu(j, sched_group_cpus(sg)) {
6804 struct sched_domain *sd;
6806 sd = &per_cpu(phys_domains, j).sd;
6807 if (j != group_first_cpu(sd->groups)) {
6809 * Only add "power" once for each
6815 sg->cpu_power += sd->groups->cpu_power;
6818 } while (sg != group_head);
6821 static int build_numa_sched_groups(struct s_data *d,
6822 const struct cpumask *cpu_map, int num)
6824 struct sched_domain *sd;
6825 struct sched_group *sg, *prev;
6828 cpumask_clear(d->covered);
6829 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
6830 if (cpumask_empty(d->nodemask)) {
6831 d->sched_group_nodes[num] = NULL;
6835 sched_domain_node_span(num, d->domainspan);
6836 cpumask_and(d->domainspan, d->domainspan, cpu_map);
6838 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6841 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
6845 d->sched_group_nodes[num] = sg;
6847 for_each_cpu(j, d->nodemask) {
6848 sd = &per_cpu(node_domains, j).sd;
6853 cpumask_copy(sched_group_cpus(sg), d->nodemask);
6855 cpumask_or(d->covered, d->covered, d->nodemask);
6858 for (j = 0; j < nr_node_ids; j++) {
6859 n = (num + j) % nr_node_ids;
6860 cpumask_complement(d->notcovered, d->covered);
6861 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
6862 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
6863 if (cpumask_empty(d->tmpmask))
6865 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
6866 if (cpumask_empty(d->tmpmask))
6868 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6872 "Can not alloc domain group for node %d\n", j);
6876 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
6877 sg->next = prev->next;
6878 cpumask_or(d->covered, d->covered, d->tmpmask);
6885 #endif /* CONFIG_NUMA */
6888 /* Free memory allocated for various sched_group structures */
6889 static void free_sched_groups(const struct cpumask *cpu_map,
6890 struct cpumask *nodemask)
6894 for_each_cpu(cpu, cpu_map) {
6895 struct sched_group **sched_group_nodes
6896 = sched_group_nodes_bycpu[cpu];
6898 if (!sched_group_nodes)
6901 for (i = 0; i < nr_node_ids; i++) {
6902 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6904 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
6905 if (cpumask_empty(nodemask))
6915 if (oldsg != sched_group_nodes[i])
6918 kfree(sched_group_nodes);
6919 sched_group_nodes_bycpu[cpu] = NULL;
6922 #else /* !CONFIG_NUMA */
6923 static void free_sched_groups(const struct cpumask *cpu_map,
6924 struct cpumask *nodemask)
6927 #endif /* CONFIG_NUMA */
6930 * Initialize sched groups cpu_power.
6932 * cpu_power indicates the capacity of sched group, which is used while
6933 * distributing the load between different sched groups in a sched domain.
6934 * Typically cpu_power for all the groups in a sched domain will be same unless
6935 * there are asymmetries in the topology. If there are asymmetries, group
6936 * having more cpu_power will pickup more load compared to the group having
6939 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6941 struct sched_domain *child;
6942 struct sched_group *group;
6946 WARN_ON(!sd || !sd->groups);
6948 if (cpu != group_first_cpu(sd->groups))
6951 sd->groups->group_weight = cpumask_weight(sched_group_cpus(sd->groups));
6955 sd->groups->cpu_power = 0;
6958 power = SCHED_LOAD_SCALE;
6959 weight = cpumask_weight(sched_domain_span(sd));
6961 * SMT siblings share the power of a single core.
6962 * Usually multiple threads get a better yield out of
6963 * that one core than a single thread would have,
6964 * reflect that in sd->smt_gain.
6966 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
6967 power *= sd->smt_gain;
6969 power >>= SCHED_LOAD_SHIFT;
6971 sd->groups->cpu_power += power;
6976 * Add cpu_power of each child group to this groups cpu_power.
6978 group = child->groups;
6980 sd->groups->cpu_power += group->cpu_power;
6981 group = group->next;
6982 } while (group != child->groups);
6986 * Initializers for schedule domains
6987 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6990 #ifdef CONFIG_SCHED_DEBUG
6991 # define SD_INIT_NAME(sd, type) sd->name = #type
6993 # define SD_INIT_NAME(sd, type) do { } while (0)
6996 #define SD_INIT(sd, type) sd_init_##type(sd)
6998 #define SD_INIT_FUNC(type) \
6999 static noinline void sd_init_##type(struct sched_domain *sd) \
7001 memset(sd, 0, sizeof(*sd)); \
7002 *sd = SD_##type##_INIT; \
7003 sd->level = SD_LV_##type; \
7004 SD_INIT_NAME(sd, type); \
7009 SD_INIT_FUNC(ALLNODES)
7012 #ifdef CONFIG_SCHED_SMT
7013 SD_INIT_FUNC(SIBLING)
7015 #ifdef CONFIG_SCHED_MC
7018 #ifdef CONFIG_SCHED_BOOK
7022 static int default_relax_domain_level = -1;
7024 static int __init setup_relax_domain_level(char *str)
7028 val = simple_strtoul(str, NULL, 0);
7029 if (val < SD_LV_MAX)
7030 default_relax_domain_level = val;
7034 __setup("relax_domain_level=", setup_relax_domain_level);
7036 static void set_domain_attribute(struct sched_domain *sd,
7037 struct sched_domain_attr *attr)
7041 if (!attr || attr->relax_domain_level < 0) {
7042 if (default_relax_domain_level < 0)
7045 request = default_relax_domain_level;
7047 request = attr->relax_domain_level;
7048 if (request < sd->level) {
7049 /* turn off idle balance on this domain */
7050 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7052 /* turn on idle balance on this domain */
7053 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7057 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
7058 const struct cpumask *cpu_map)
7061 case sa_sched_groups:
7062 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
7063 d->sched_group_nodes = NULL;
7065 free_rootdomain(d->rd); /* fall through */
7067 free_cpumask_var(d->tmpmask); /* fall through */
7068 case sa_send_covered:
7069 free_cpumask_var(d->send_covered); /* fall through */
7070 case sa_this_book_map:
7071 free_cpumask_var(d->this_book_map); /* fall through */
7072 case sa_this_core_map:
7073 free_cpumask_var(d->this_core_map); /* fall through */
7074 case sa_this_sibling_map:
7075 free_cpumask_var(d->this_sibling_map); /* fall through */
7077 free_cpumask_var(d->nodemask); /* fall through */
7078 case sa_sched_group_nodes:
7080 kfree(d->sched_group_nodes); /* fall through */
7082 free_cpumask_var(d->notcovered); /* fall through */
7084 free_cpumask_var(d->covered); /* fall through */
7086 free_cpumask_var(d->domainspan); /* fall through */
7093 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
7094 const struct cpumask *cpu_map)
7097 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
7099 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
7100 return sa_domainspan;
7101 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
7103 /* Allocate the per-node list of sched groups */
7104 d->sched_group_nodes = kcalloc(nr_node_ids,
7105 sizeof(struct sched_group *), GFP_KERNEL);
7106 if (!d->sched_group_nodes) {
7107 printk(KERN_WARNING "Can not alloc sched group node list\n");
7108 return sa_notcovered;
7110 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
7112 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
7113 return sa_sched_group_nodes;
7114 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
7116 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
7117 return sa_this_sibling_map;
7118 if (!alloc_cpumask_var(&d->this_book_map, GFP_KERNEL))
7119 return sa_this_core_map;
7120 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
7121 return sa_this_book_map;
7122 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
7123 return sa_send_covered;
7124 d->rd = alloc_rootdomain();
7126 printk(KERN_WARNING "Cannot alloc root domain\n");
7129 return sa_rootdomain;
7132 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
7133 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
7135 struct sched_domain *sd = NULL;
7137 struct sched_domain *parent;
7140 if (cpumask_weight(cpu_map) >
7141 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
7142 sd = &per_cpu(allnodes_domains, i).sd;
7143 SD_INIT(sd, ALLNODES);
7144 set_domain_attribute(sd, attr);
7145 cpumask_copy(sched_domain_span(sd), cpu_map);
7146 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
7151 sd = &per_cpu(node_domains, i).sd;
7153 set_domain_attribute(sd, attr);
7154 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
7155 sd->parent = parent;
7158 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
7163 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
7164 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7165 struct sched_domain *parent, int i)
7167 struct sched_domain *sd;
7168 sd = &per_cpu(phys_domains, i).sd;
7170 set_domain_attribute(sd, attr);
7171 cpumask_copy(sched_domain_span(sd), d->nodemask);
7172 sd->parent = parent;
7175 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
7179 static struct sched_domain *__build_book_sched_domain(struct s_data *d,
7180 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7181 struct sched_domain *parent, int i)
7183 struct sched_domain *sd = parent;
7184 #ifdef CONFIG_SCHED_BOOK
7185 sd = &per_cpu(book_domains, i).sd;
7187 set_domain_attribute(sd, attr);
7188 cpumask_and(sched_domain_span(sd), cpu_map, cpu_book_mask(i));
7189 sd->parent = parent;
7191 cpu_to_book_group(i, cpu_map, &sd->groups, d->tmpmask);
7196 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
7197 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7198 struct sched_domain *parent, int i)
7200 struct sched_domain *sd = parent;
7201 #ifdef CONFIG_SCHED_MC
7202 sd = &per_cpu(core_domains, i).sd;
7204 set_domain_attribute(sd, attr);
7205 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
7206 sd->parent = parent;
7208 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
7213 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
7214 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7215 struct sched_domain *parent, int i)
7217 struct sched_domain *sd = parent;
7218 #ifdef CONFIG_SCHED_SMT
7219 sd = &per_cpu(cpu_domains, i).sd;
7220 SD_INIT(sd, SIBLING);
7221 set_domain_attribute(sd, attr);
7222 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
7223 sd->parent = parent;
7225 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
7230 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
7231 const struct cpumask *cpu_map, int cpu)
7234 #ifdef CONFIG_SCHED_SMT
7235 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
7236 cpumask_and(d->this_sibling_map, cpu_map,
7237 topology_thread_cpumask(cpu));
7238 if (cpu == cpumask_first(d->this_sibling_map))
7239 init_sched_build_groups(d->this_sibling_map, cpu_map,
7241 d->send_covered, d->tmpmask);
7244 #ifdef CONFIG_SCHED_MC
7245 case SD_LV_MC: /* set up multi-core groups */
7246 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
7247 if (cpu == cpumask_first(d->this_core_map))
7248 init_sched_build_groups(d->this_core_map, cpu_map,
7250 d->send_covered, d->tmpmask);
7253 #ifdef CONFIG_SCHED_BOOK
7254 case SD_LV_BOOK: /* set up book groups */
7255 cpumask_and(d->this_book_map, cpu_map, cpu_book_mask(cpu));
7256 if (cpu == cpumask_first(d->this_book_map))
7257 init_sched_build_groups(d->this_book_map, cpu_map,
7259 d->send_covered, d->tmpmask);
7262 case SD_LV_CPU: /* set up physical groups */
7263 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
7264 if (!cpumask_empty(d->nodemask))
7265 init_sched_build_groups(d->nodemask, cpu_map,
7267 d->send_covered, d->tmpmask);
7270 case SD_LV_ALLNODES:
7271 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
7272 d->send_covered, d->tmpmask);
7281 * Build sched domains for a given set of cpus and attach the sched domains
7282 * to the individual cpus
7284 static int __build_sched_domains(const struct cpumask *cpu_map,
7285 struct sched_domain_attr *attr)
7287 enum s_alloc alloc_state = sa_none;
7289 struct sched_domain *sd;
7295 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7296 if (alloc_state != sa_rootdomain)
7298 alloc_state = sa_sched_groups;
7301 * Set up domains for cpus specified by the cpu_map.
7303 for_each_cpu(i, cpu_map) {
7304 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
7307 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
7308 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
7309 sd = __build_book_sched_domain(&d, cpu_map, attr, sd, i);
7310 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
7311 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
7314 for_each_cpu(i, cpu_map) {
7315 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
7316 build_sched_groups(&d, SD_LV_BOOK, cpu_map, i);
7317 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
7320 /* Set up physical groups */
7321 for (i = 0; i < nr_node_ids; i++)
7322 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
7325 /* Set up node groups */
7327 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
7329 for (i = 0; i < nr_node_ids; i++)
7330 if (build_numa_sched_groups(&d, cpu_map, i))
7334 /* Calculate CPU power for physical packages and nodes */
7335 #ifdef CONFIG_SCHED_SMT
7336 for_each_cpu(i, cpu_map) {
7337 sd = &per_cpu(cpu_domains, i).sd;
7338 init_sched_groups_power(i, sd);
7341 #ifdef CONFIG_SCHED_MC
7342 for_each_cpu(i, cpu_map) {
7343 sd = &per_cpu(core_domains, i).sd;
7344 init_sched_groups_power(i, sd);
7347 #ifdef CONFIG_SCHED_BOOK
7348 for_each_cpu(i, cpu_map) {
7349 sd = &per_cpu(book_domains, i).sd;
7350 init_sched_groups_power(i, sd);
7354 for_each_cpu(i, cpu_map) {
7355 sd = &per_cpu(phys_domains, i).sd;
7356 init_sched_groups_power(i, sd);
7360 for (i = 0; i < nr_node_ids; i++)
7361 init_numa_sched_groups_power(d.sched_group_nodes[i]);
7363 if (d.sd_allnodes) {
7364 struct sched_group *sg;
7366 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7368 init_numa_sched_groups_power(sg);
7372 /* Attach the domains */
7373 for_each_cpu(i, cpu_map) {
7374 #ifdef CONFIG_SCHED_SMT
7375 sd = &per_cpu(cpu_domains, i).sd;
7376 #elif defined(CONFIG_SCHED_MC)
7377 sd = &per_cpu(core_domains, i).sd;
7378 #elif defined(CONFIG_SCHED_BOOK)
7379 sd = &per_cpu(book_domains, i).sd;
7381 sd = &per_cpu(phys_domains, i).sd;
7383 cpu_attach_domain(sd, d.rd, i);
7386 d.sched_group_nodes = NULL; /* don't free this we still need it */
7387 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
7391 __free_domain_allocs(&d, alloc_state, cpu_map);
7395 static int build_sched_domains(const struct cpumask *cpu_map)
7397 return __build_sched_domains(cpu_map, NULL);
7400 static cpumask_var_t *doms_cur; /* current sched domains */
7401 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7402 static struct sched_domain_attr *dattr_cur;
7403 /* attribues of custom domains in 'doms_cur' */
7406 * Special case: If a kmalloc of a doms_cur partition (array of
7407 * cpumask) fails, then fallback to a single sched domain,
7408 * as determined by the single cpumask fallback_doms.
7410 static cpumask_var_t fallback_doms;
7413 * arch_update_cpu_topology lets virtualized architectures update the
7414 * cpu core maps. It is supposed to return 1 if the topology changed
7415 * or 0 if it stayed the same.
7417 int __attribute__((weak)) arch_update_cpu_topology(void)
7422 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7425 cpumask_var_t *doms;
7427 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7430 for (i = 0; i < ndoms; i++) {
7431 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7432 free_sched_domains(doms, i);
7439 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7442 for (i = 0; i < ndoms; i++)
7443 free_cpumask_var(doms[i]);
7448 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7449 * For now this just excludes isolated cpus, but could be used to
7450 * exclude other special cases in the future.
7452 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7456 arch_update_cpu_topology();
7458 doms_cur = alloc_sched_domains(ndoms_cur);
7460 doms_cur = &fallback_doms;
7461 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7463 err = build_sched_domains(doms_cur[0]);
7464 register_sched_domain_sysctl();
7469 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7470 struct cpumask *tmpmask)
7472 free_sched_groups(cpu_map, tmpmask);
7476 * Detach sched domains from a group of cpus specified in cpu_map
7477 * These cpus will now be attached to the NULL domain
7479 static void detach_destroy_domains(const struct cpumask *cpu_map)
7481 /* Save because hotplug lock held. */
7482 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7485 for_each_cpu(i, cpu_map)
7486 cpu_attach_domain(NULL, &def_root_domain, i);
7487 synchronize_sched();
7488 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7491 /* handle null as "default" */
7492 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7493 struct sched_domain_attr *new, int idx_new)
7495 struct sched_domain_attr tmp;
7502 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7503 new ? (new + idx_new) : &tmp,
7504 sizeof(struct sched_domain_attr));
7508 * Partition sched domains as specified by the 'ndoms_new'
7509 * cpumasks in the array doms_new[] of cpumasks. This compares
7510 * doms_new[] to the current sched domain partitioning, doms_cur[].
7511 * It destroys each deleted domain and builds each new domain.
7513 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7514 * The masks don't intersect (don't overlap.) We should setup one
7515 * sched domain for each mask. CPUs not in any of the cpumasks will
7516 * not be load balanced. If the same cpumask appears both in the
7517 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7520 * The passed in 'doms_new' should be allocated using
7521 * alloc_sched_domains. This routine takes ownership of it and will
7522 * free_sched_domains it when done with it. If the caller failed the
7523 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7524 * and partition_sched_domains() will fallback to the single partition
7525 * 'fallback_doms', it also forces the domains to be rebuilt.
7527 * If doms_new == NULL it will be replaced with cpu_online_mask.
7528 * ndoms_new == 0 is a special case for destroying existing domains,
7529 * and it will not create the default domain.
7531 * Call with hotplug lock held
7533 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7534 struct sched_domain_attr *dattr_new)
7539 mutex_lock(&sched_domains_mutex);
7541 /* always unregister in case we don't destroy any domains */
7542 unregister_sched_domain_sysctl();
7544 /* Let architecture update cpu core mappings. */
7545 new_topology = arch_update_cpu_topology();
7547 n = doms_new ? ndoms_new : 0;
7549 /* Destroy deleted domains */
7550 for (i = 0; i < ndoms_cur; i++) {
7551 for (j = 0; j < n && !new_topology; j++) {
7552 if (cpumask_equal(doms_cur[i], doms_new[j])
7553 && dattrs_equal(dattr_cur, i, dattr_new, j))
7556 /* no match - a current sched domain not in new doms_new[] */
7557 detach_destroy_domains(doms_cur[i]);
7562 if (doms_new == NULL) {
7564 doms_new = &fallback_doms;
7565 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7566 WARN_ON_ONCE(dattr_new);
7569 /* Build new domains */
7570 for (i = 0; i < ndoms_new; i++) {
7571 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7572 if (cpumask_equal(doms_new[i], doms_cur[j])
7573 && dattrs_equal(dattr_new, i, dattr_cur, j))
7576 /* no match - add a new doms_new */
7577 __build_sched_domains(doms_new[i],
7578 dattr_new ? dattr_new + i : NULL);
7583 /* Remember the new sched domains */
7584 if (doms_cur != &fallback_doms)
7585 free_sched_domains(doms_cur, ndoms_cur);
7586 kfree(dattr_cur); /* kfree(NULL) is safe */
7587 doms_cur = doms_new;
7588 dattr_cur = dattr_new;
7589 ndoms_cur = ndoms_new;
7591 register_sched_domain_sysctl();
7593 mutex_unlock(&sched_domains_mutex);
7596 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7597 static void arch_reinit_sched_domains(void)
7601 /* Destroy domains first to force the rebuild */
7602 partition_sched_domains(0, NULL, NULL);
7604 rebuild_sched_domains();
7608 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7610 unsigned int level = 0;
7612 if (sscanf(buf, "%u", &level) != 1)
7616 * level is always be positive so don't check for
7617 * level < POWERSAVINGS_BALANCE_NONE which is 0
7618 * What happens on 0 or 1 byte write,
7619 * need to check for count as well?
7622 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7626 sched_smt_power_savings = level;
7628 sched_mc_power_savings = level;
7630 arch_reinit_sched_domains();
7635 #ifdef CONFIG_SCHED_MC
7636 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7637 struct sysdev_class_attribute *attr,
7640 return sprintf(page, "%u\n", sched_mc_power_savings);
7642 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7643 struct sysdev_class_attribute *attr,
7644 const char *buf, size_t count)
7646 return sched_power_savings_store(buf, count, 0);
7648 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7649 sched_mc_power_savings_show,
7650 sched_mc_power_savings_store);
7653 #ifdef CONFIG_SCHED_SMT
7654 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7655 struct sysdev_class_attribute *attr,
7658 return sprintf(page, "%u\n", sched_smt_power_savings);
7660 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7661 struct sysdev_class_attribute *attr,
7662 const char *buf, size_t count)
7664 return sched_power_savings_store(buf, count, 1);
7666 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7667 sched_smt_power_savings_show,
7668 sched_smt_power_savings_store);
7671 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7675 #ifdef CONFIG_SCHED_SMT
7677 err = sysfs_create_file(&cls->kset.kobj,
7678 &attr_sched_smt_power_savings.attr);
7680 #ifdef CONFIG_SCHED_MC
7681 if (!err && mc_capable())
7682 err = sysfs_create_file(&cls->kset.kobj,
7683 &attr_sched_mc_power_savings.attr);
7687 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7690 * Update cpusets according to cpu_active mask. If cpusets are
7691 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7692 * around partition_sched_domains().
7694 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7697 switch (action & ~CPU_TASKS_FROZEN) {
7699 case CPU_DOWN_FAILED:
7700 cpuset_update_active_cpus();
7707 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7710 switch (action & ~CPU_TASKS_FROZEN) {
7711 case CPU_DOWN_PREPARE:
7712 cpuset_update_active_cpus();
7719 static int update_runtime(struct notifier_block *nfb,
7720 unsigned long action, void *hcpu)
7722 int cpu = (int)(long)hcpu;
7725 case CPU_DOWN_PREPARE:
7726 case CPU_DOWN_PREPARE_FROZEN:
7727 disable_runtime(cpu_rq(cpu));
7730 case CPU_DOWN_FAILED:
7731 case CPU_DOWN_FAILED_FROZEN:
7733 case CPU_ONLINE_FROZEN:
7734 enable_runtime(cpu_rq(cpu));
7742 void __init sched_init_smp(void)
7744 cpumask_var_t non_isolated_cpus;
7746 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7747 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7749 #if defined(CONFIG_NUMA)
7750 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7752 BUG_ON(sched_group_nodes_bycpu == NULL);
7755 mutex_lock(&sched_domains_mutex);
7756 arch_init_sched_domains(cpu_active_mask);
7757 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7758 if (cpumask_empty(non_isolated_cpus))
7759 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7760 mutex_unlock(&sched_domains_mutex);
7763 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7764 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7766 /* RT runtime code needs to handle some hotplug events */
7767 hotcpu_notifier(update_runtime, 0);
7771 /* Move init over to a non-isolated CPU */
7772 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7774 sched_init_granularity();
7775 free_cpumask_var(non_isolated_cpus);
7777 init_sched_rt_class();
7780 void __init sched_init_smp(void)
7782 sched_init_granularity();
7784 #endif /* CONFIG_SMP */
7786 const_debug unsigned int sysctl_timer_migration = 1;
7788 int in_sched_functions(unsigned long addr)
7790 return in_lock_functions(addr) ||
7791 (addr >= (unsigned long)__sched_text_start
7792 && addr < (unsigned long)__sched_text_end);
7795 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7797 cfs_rq->tasks_timeline = RB_ROOT;
7798 INIT_LIST_HEAD(&cfs_rq->tasks);
7799 #ifdef CONFIG_FAIR_GROUP_SCHED
7802 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7805 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7807 struct rt_prio_array *array;
7810 array = &rt_rq->active;
7811 for (i = 0; i < MAX_RT_PRIO; i++) {
7812 INIT_LIST_HEAD(array->queue + i);
7813 __clear_bit(i, array->bitmap);
7815 /* delimiter for bitsearch: */
7816 __set_bit(MAX_RT_PRIO, array->bitmap);
7818 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7819 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7821 rt_rq->highest_prio.next = MAX_RT_PRIO;
7825 rt_rq->rt_nr_migratory = 0;
7826 rt_rq->overloaded = 0;
7827 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
7831 rt_rq->rt_throttled = 0;
7832 rt_rq->rt_runtime = 0;
7833 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7835 #ifdef CONFIG_RT_GROUP_SCHED
7836 rt_rq->rt_nr_boosted = 0;
7841 #ifdef CONFIG_FAIR_GROUP_SCHED
7842 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7843 struct sched_entity *se, int cpu,
7844 struct sched_entity *parent)
7846 struct rq *rq = cpu_rq(cpu);
7847 tg->cfs_rq[cpu] = cfs_rq;
7848 init_cfs_rq(cfs_rq, rq);
7852 /* se could be NULL for root_task_group */
7857 se->cfs_rq = &rq->cfs;
7859 se->cfs_rq = parent->my_q;
7862 update_load_set(&se->load, 0);
7863 se->parent = parent;
7867 #ifdef CONFIG_RT_GROUP_SCHED
7868 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7869 struct sched_rt_entity *rt_se, int cpu,
7870 struct sched_rt_entity *parent)
7872 struct rq *rq = cpu_rq(cpu);
7874 tg->rt_rq[cpu] = rt_rq;
7875 init_rt_rq(rt_rq, rq);
7877 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7879 tg->rt_se[cpu] = rt_se;
7884 rt_se->rt_rq = &rq->rt;
7886 rt_se->rt_rq = parent->my_q;
7888 rt_se->my_q = rt_rq;
7889 rt_se->parent = parent;
7890 INIT_LIST_HEAD(&rt_se->run_list);
7894 void __init sched_init(void)
7897 unsigned long alloc_size = 0, ptr;
7899 #ifdef CONFIG_FAIR_GROUP_SCHED
7900 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7902 #ifdef CONFIG_RT_GROUP_SCHED
7903 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7905 #ifdef CONFIG_CPUMASK_OFFSTACK
7906 alloc_size += num_possible_cpus() * cpumask_size();
7909 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7911 #ifdef CONFIG_FAIR_GROUP_SCHED
7912 root_task_group.se = (struct sched_entity **)ptr;
7913 ptr += nr_cpu_ids * sizeof(void **);
7915 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7916 ptr += nr_cpu_ids * sizeof(void **);
7918 #endif /* CONFIG_FAIR_GROUP_SCHED */
7919 #ifdef CONFIG_RT_GROUP_SCHED
7920 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7921 ptr += nr_cpu_ids * sizeof(void **);
7923 root_task_group.rt_rq = (struct rt_rq **)ptr;
7924 ptr += nr_cpu_ids * sizeof(void **);
7926 #endif /* CONFIG_RT_GROUP_SCHED */
7927 #ifdef CONFIG_CPUMASK_OFFSTACK
7928 for_each_possible_cpu(i) {
7929 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
7930 ptr += cpumask_size();
7932 #endif /* CONFIG_CPUMASK_OFFSTACK */
7936 init_defrootdomain();
7939 init_rt_bandwidth(&def_rt_bandwidth,
7940 global_rt_period(), global_rt_runtime());
7942 #ifdef CONFIG_RT_GROUP_SCHED
7943 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7944 global_rt_period(), global_rt_runtime());
7945 #endif /* CONFIG_RT_GROUP_SCHED */
7947 #ifdef CONFIG_CGROUP_SCHED
7948 list_add(&root_task_group.list, &task_groups);
7949 INIT_LIST_HEAD(&root_task_group.children);
7950 autogroup_init(&init_task);
7951 #endif /* CONFIG_CGROUP_SCHED */
7953 for_each_possible_cpu(i) {
7957 raw_spin_lock_init(&rq->lock);
7959 rq->calc_load_active = 0;
7960 rq->calc_load_update = jiffies + LOAD_FREQ;
7961 init_cfs_rq(&rq->cfs, rq);
7962 init_rt_rq(&rq->rt, rq);
7963 #ifdef CONFIG_FAIR_GROUP_SCHED
7964 root_task_group.shares = root_task_group_load;
7965 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7967 * How much cpu bandwidth does root_task_group get?
7969 * In case of task-groups formed thr' the cgroup filesystem, it
7970 * gets 100% of the cpu resources in the system. This overall
7971 * system cpu resource is divided among the tasks of
7972 * root_task_group and its child task-groups in a fair manner,
7973 * based on each entity's (task or task-group's) weight
7974 * (se->load.weight).
7976 * In other words, if root_task_group has 10 tasks of weight
7977 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7978 * then A0's share of the cpu resource is:
7980 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7982 * We achieve this by letting root_task_group's tasks sit
7983 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7985 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7986 #endif /* CONFIG_FAIR_GROUP_SCHED */
7988 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7989 #ifdef CONFIG_RT_GROUP_SCHED
7990 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7991 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7994 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7995 rq->cpu_load[j] = 0;
7997 rq->last_load_update_tick = jiffies;
8002 rq->cpu_power = SCHED_LOAD_SCALE;
8003 rq->post_schedule = 0;
8004 rq->active_balance = 0;
8005 rq->next_balance = jiffies;
8010 rq->avg_idle = 2*sysctl_sched_migration_cost;
8011 rq_attach_root(rq, &def_root_domain);
8013 rq->nohz_balance_kick = 0;
8014 init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb, i));
8018 atomic_set(&rq->nr_iowait, 0);
8021 set_load_weight(&init_task);
8023 #ifdef CONFIG_PREEMPT_NOTIFIERS
8024 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8028 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8031 #ifdef CONFIG_RT_MUTEXES
8032 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
8036 * The boot idle thread does lazy MMU switching as well:
8038 atomic_inc(&init_mm.mm_count);
8039 enter_lazy_tlb(&init_mm, current);
8042 * Make us the idle thread. Technically, schedule() should not be
8043 * called from this thread, however somewhere below it might be,
8044 * but because we are the idle thread, we just pick up running again
8045 * when this runqueue becomes "idle".
8047 init_idle(current, smp_processor_id());
8049 calc_load_update = jiffies + LOAD_FREQ;
8052 * During early bootup we pretend to be a normal task:
8054 current->sched_class = &fair_sched_class;
8056 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8057 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
8060 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8061 alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT);
8062 atomic_set(&nohz.load_balancer, nr_cpu_ids);
8063 atomic_set(&nohz.first_pick_cpu, nr_cpu_ids);
8064 atomic_set(&nohz.second_pick_cpu, nr_cpu_ids);
8066 /* May be allocated at isolcpus cmdline parse time */
8067 if (cpu_isolated_map == NULL)
8068 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
8071 scheduler_running = 1;
8074 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8075 static inline int preempt_count_equals(int preempt_offset)
8077 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
8079 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
8082 static int __might_sleep_init_called;
8083 int __init __might_sleep_init(void)
8085 __might_sleep_init_called = 1;
8088 early_initcall(__might_sleep_init);
8090 void __might_sleep(const char *file, int line, int preempt_offset)
8093 static unsigned long prev_jiffy; /* ratelimiting */
8095 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
8098 if (system_state != SYSTEM_RUNNING &&
8099 (!__might_sleep_init_called || system_state != SYSTEM_BOOTING))
8101 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8103 prev_jiffy = jiffies;
8106 "BUG: sleeping function called from invalid context at %s:%d\n",
8109 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8110 in_atomic(), irqs_disabled(),
8111 current->pid, current->comm);
8113 debug_show_held_locks(current);
8114 if (irqs_disabled())
8115 print_irqtrace_events(current);
8119 EXPORT_SYMBOL(__might_sleep);
8122 #ifdef CONFIG_MAGIC_SYSRQ
8123 static void normalize_task(struct rq *rq, struct task_struct *p)
8127 on_rq = p->se.on_rq;
8129 deactivate_task(rq, p, 0);
8130 __setscheduler(rq, p, SCHED_NORMAL, 0);
8132 activate_task(rq, p, 0);
8133 resched_task(rq->curr);
8137 void normalize_rt_tasks(void)
8139 struct task_struct *g, *p;
8140 unsigned long flags;
8143 read_lock_irqsave(&tasklist_lock, flags);
8144 do_each_thread(g, p) {
8146 * Only normalize user tasks:
8151 p->se.exec_start = 0;
8152 #ifdef CONFIG_SCHEDSTATS
8153 p->se.statistics.wait_start = 0;
8154 p->se.statistics.sleep_start = 0;
8155 p->se.statistics.block_start = 0;
8160 * Renice negative nice level userspace
8163 if (TASK_NICE(p) < 0 && p->mm)
8164 set_user_nice(p, 0);
8168 raw_spin_lock(&p->pi_lock);
8169 rq = __task_rq_lock(p);
8171 normalize_task(rq, p);
8173 __task_rq_unlock(rq);
8174 raw_spin_unlock(&p->pi_lock);
8175 } while_each_thread(g, p);
8177 read_unlock_irqrestore(&tasklist_lock, flags);
8180 #endif /* CONFIG_MAGIC_SYSRQ */
8182 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8184 * These functions are only useful for the IA64 MCA handling, or kdb.
8186 * They can only be called when the whole system has been
8187 * stopped - every CPU needs to be quiescent, and no scheduling
8188 * activity can take place. Using them for anything else would
8189 * be a serious bug, and as a result, they aren't even visible
8190 * under any other configuration.
8194 * curr_task - return the current task for a given cpu.
8195 * @cpu: the processor in question.
8197 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8199 struct task_struct *curr_task(int cpu)
8201 return cpu_curr(cpu);
8204 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8208 * set_curr_task - set the current task for a given cpu.
8209 * @cpu: the processor in question.
8210 * @p: the task pointer to set.
8212 * Description: This function must only be used when non-maskable interrupts
8213 * are serviced on a separate stack. It allows the architecture to switch the
8214 * notion of the current task on a cpu in a non-blocking manner. This function
8215 * must be called with all CPU's synchronized, and interrupts disabled, the
8216 * and caller must save the original value of the current task (see
8217 * curr_task() above) and restore that value before reenabling interrupts and
8218 * re-starting the system.
8220 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8222 void set_curr_task(int cpu, struct task_struct *p)
8229 #ifdef CONFIG_FAIR_GROUP_SCHED
8230 static void free_fair_sched_group(struct task_group *tg)
8234 for_each_possible_cpu(i) {
8236 kfree(tg->cfs_rq[i]);
8246 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8248 struct cfs_rq *cfs_rq;
8249 struct sched_entity *se;
8253 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8256 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8260 tg->shares = NICE_0_LOAD;
8262 for_each_possible_cpu(i) {
8265 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8266 GFP_KERNEL, cpu_to_node(i));
8270 se = kzalloc_node(sizeof(struct sched_entity),
8271 GFP_KERNEL, cpu_to_node(i));
8275 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8286 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8288 struct rq *rq = cpu_rq(cpu);
8289 unsigned long flags;
8292 * Only empty task groups can be destroyed; so we can speculatively
8293 * check on_list without danger of it being re-added.
8295 if (!tg->cfs_rq[cpu]->on_list)
8298 raw_spin_lock_irqsave(&rq->lock, flags);
8299 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8300 raw_spin_unlock_irqrestore(&rq->lock, flags);
8302 #else /* !CONFG_FAIR_GROUP_SCHED */
8303 static inline void free_fair_sched_group(struct task_group *tg)
8308 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8313 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8316 #endif /* CONFIG_FAIR_GROUP_SCHED */
8318 #ifdef CONFIG_RT_GROUP_SCHED
8319 static void free_rt_sched_group(struct task_group *tg)
8323 destroy_rt_bandwidth(&tg->rt_bandwidth);
8325 for_each_possible_cpu(i) {
8327 kfree(tg->rt_rq[i]);
8329 kfree(tg->rt_se[i]);
8337 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8339 struct rt_rq *rt_rq;
8340 struct sched_rt_entity *rt_se;
8344 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8347 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8351 init_rt_bandwidth(&tg->rt_bandwidth,
8352 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8354 for_each_possible_cpu(i) {
8357 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8358 GFP_KERNEL, cpu_to_node(i));
8362 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8363 GFP_KERNEL, cpu_to_node(i));
8367 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
8377 #else /* !CONFIG_RT_GROUP_SCHED */
8378 static inline void free_rt_sched_group(struct task_group *tg)
8383 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8387 #endif /* CONFIG_RT_GROUP_SCHED */
8389 #ifdef CONFIG_CGROUP_SCHED
8390 static void free_sched_group(struct task_group *tg)
8392 free_fair_sched_group(tg);
8393 free_rt_sched_group(tg);
8398 /* allocate runqueue etc for a new task group */
8399 struct task_group *sched_create_group(struct task_group *parent)
8401 struct task_group *tg;
8402 unsigned long flags;
8404 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8406 return ERR_PTR(-ENOMEM);
8408 if (!alloc_fair_sched_group(tg, parent))
8411 if (!alloc_rt_sched_group(tg, parent))
8414 spin_lock_irqsave(&task_group_lock, flags);
8415 list_add_rcu(&tg->list, &task_groups);
8417 WARN_ON(!parent); /* root should already exist */
8419 tg->parent = parent;
8420 INIT_LIST_HEAD(&tg->children);
8421 list_add_rcu(&tg->siblings, &parent->children);
8422 spin_unlock_irqrestore(&task_group_lock, flags);
8427 free_sched_group(tg);
8428 return ERR_PTR(-ENOMEM);
8431 /* rcu callback to free various structures associated with a task group */
8432 static void free_sched_group_rcu(struct rcu_head *rhp)
8434 /* now it should be safe to free those cfs_rqs */
8435 free_sched_group(container_of(rhp, struct task_group, rcu));
8438 /* Destroy runqueue etc associated with a task group */
8439 void sched_destroy_group(struct task_group *tg)
8441 unsigned long flags;
8444 /* end participation in shares distribution */
8445 for_each_possible_cpu(i)
8446 unregister_fair_sched_group(tg, i);
8448 spin_lock_irqsave(&task_group_lock, flags);
8449 list_del_rcu(&tg->list);
8450 list_del_rcu(&tg->siblings);
8451 spin_unlock_irqrestore(&task_group_lock, flags);
8453 /* wait for possible concurrent references to cfs_rqs complete */
8454 call_rcu(&tg->rcu, free_sched_group_rcu);
8457 /* change task's runqueue when it moves between groups.
8458 * The caller of this function should have put the task in its new group
8459 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8460 * reflect its new group.
8462 void sched_move_task(struct task_struct *tsk)
8465 unsigned long flags;
8468 rq = task_rq_lock(tsk, &flags);
8470 running = task_current(rq, tsk);
8471 on_rq = tsk->se.on_rq;
8474 dequeue_task(rq, tsk, 0);
8475 if (unlikely(running))
8476 tsk->sched_class->put_prev_task(rq, tsk);
8478 #ifdef CONFIG_FAIR_GROUP_SCHED
8479 if (tsk->sched_class->task_move_group)
8480 tsk->sched_class->task_move_group(tsk, on_rq);
8483 set_task_rq(tsk, task_cpu(tsk));
8485 if (unlikely(running))
8486 tsk->sched_class->set_curr_task(rq);
8488 enqueue_task(rq, tsk, 0);
8490 task_rq_unlock(rq, &flags);
8492 #endif /* CONFIG_CGROUP_SCHED */
8494 #ifdef CONFIG_FAIR_GROUP_SCHED
8495 static DEFINE_MUTEX(shares_mutex);
8497 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8500 unsigned long flags;
8503 * We can't change the weight of the root cgroup.
8508 if (shares < MIN_SHARES)
8509 shares = MIN_SHARES;
8510 else if (shares > MAX_SHARES)
8511 shares = MAX_SHARES;
8513 mutex_lock(&shares_mutex);
8514 if (tg->shares == shares)
8517 tg->shares = shares;
8518 for_each_possible_cpu(i) {
8519 struct rq *rq = cpu_rq(i);
8520 struct sched_entity *se;
8523 /* Propagate contribution to hierarchy */
8524 raw_spin_lock_irqsave(&rq->lock, flags);
8525 for_each_sched_entity(se)
8526 update_cfs_shares(group_cfs_rq(se), 0);
8527 raw_spin_unlock_irqrestore(&rq->lock, flags);
8531 mutex_unlock(&shares_mutex);
8535 unsigned long sched_group_shares(struct task_group *tg)
8541 #ifdef CONFIG_RT_GROUP_SCHED
8543 * Ensure that the real time constraints are schedulable.
8545 static DEFINE_MUTEX(rt_constraints_mutex);
8547 static unsigned long to_ratio(u64 period, u64 runtime)
8549 if (runtime == RUNTIME_INF)
8552 return div64_u64(runtime << 20, period);
8555 /* Must be called with tasklist_lock held */
8556 static inline int tg_has_rt_tasks(struct task_group *tg)
8558 struct task_struct *g, *p;
8560 do_each_thread(g, p) {
8561 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8563 } while_each_thread(g, p);
8568 struct rt_schedulable_data {
8569 struct task_group *tg;
8574 static int tg_schedulable(struct task_group *tg, void *data)
8576 struct rt_schedulable_data *d = data;
8577 struct task_group *child;
8578 unsigned long total, sum = 0;
8579 u64 period, runtime;
8581 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8582 runtime = tg->rt_bandwidth.rt_runtime;
8585 period = d->rt_period;
8586 runtime = d->rt_runtime;
8590 * Cannot have more runtime than the period.
8592 if (runtime > period && runtime != RUNTIME_INF)
8596 * Ensure we don't starve existing RT tasks.
8598 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8601 total = to_ratio(period, runtime);
8604 * Nobody can have more than the global setting allows.
8606 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8610 * The sum of our children's runtime should not exceed our own.
8612 list_for_each_entry_rcu(child, &tg->children, siblings) {
8613 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8614 runtime = child->rt_bandwidth.rt_runtime;
8616 if (child == d->tg) {
8617 period = d->rt_period;
8618 runtime = d->rt_runtime;
8621 sum += to_ratio(period, runtime);
8630 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8632 struct rt_schedulable_data data = {
8634 .rt_period = period,
8635 .rt_runtime = runtime,
8638 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8641 static int tg_set_bandwidth(struct task_group *tg,
8642 u64 rt_period, u64 rt_runtime)
8646 mutex_lock(&rt_constraints_mutex);
8647 read_lock(&tasklist_lock);
8648 err = __rt_schedulable(tg, rt_period, rt_runtime);
8652 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8653 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8654 tg->rt_bandwidth.rt_runtime = rt_runtime;
8656 for_each_possible_cpu(i) {
8657 struct rt_rq *rt_rq = tg->rt_rq[i];
8659 raw_spin_lock(&rt_rq->rt_runtime_lock);
8660 rt_rq->rt_runtime = rt_runtime;
8661 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8663 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8665 read_unlock(&tasklist_lock);
8666 mutex_unlock(&rt_constraints_mutex);
8671 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8673 u64 rt_runtime, rt_period;
8675 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8676 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8677 if (rt_runtime_us < 0)
8678 rt_runtime = RUNTIME_INF;
8680 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8683 long sched_group_rt_runtime(struct task_group *tg)
8687 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8690 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8691 do_div(rt_runtime_us, NSEC_PER_USEC);
8692 return rt_runtime_us;
8695 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8697 u64 rt_runtime, rt_period;
8699 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8700 rt_runtime = tg->rt_bandwidth.rt_runtime;
8705 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8708 long sched_group_rt_period(struct task_group *tg)
8712 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8713 do_div(rt_period_us, NSEC_PER_USEC);
8714 return rt_period_us;
8717 static int sched_rt_global_constraints(void)
8719 u64 runtime, period;
8722 if (sysctl_sched_rt_period <= 0)
8725 runtime = global_rt_runtime();
8726 period = global_rt_period();
8729 * Sanity check on the sysctl variables.
8731 if (runtime > period && runtime != RUNTIME_INF)
8734 mutex_lock(&rt_constraints_mutex);
8735 read_lock(&tasklist_lock);
8736 ret = __rt_schedulable(NULL, 0, 0);
8737 read_unlock(&tasklist_lock);
8738 mutex_unlock(&rt_constraints_mutex);
8743 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8745 /* Don't accept realtime tasks when there is no way for them to run */
8746 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8752 #else /* !CONFIG_RT_GROUP_SCHED */
8753 static int sched_rt_global_constraints(void)
8755 unsigned long flags;
8758 if (sysctl_sched_rt_period <= 0)
8762 * There's always some RT tasks in the root group
8763 * -- migration, kstopmachine etc..
8765 if (sysctl_sched_rt_runtime == 0)
8768 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8769 for_each_possible_cpu(i) {
8770 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8772 raw_spin_lock(&rt_rq->rt_runtime_lock);
8773 rt_rq->rt_runtime = global_rt_runtime();
8774 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8776 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8780 #endif /* CONFIG_RT_GROUP_SCHED */
8782 int sched_rt_handler(struct ctl_table *table, int write,
8783 void __user *buffer, size_t *lenp,
8787 int old_period, old_runtime;
8788 static DEFINE_MUTEX(mutex);
8791 old_period = sysctl_sched_rt_period;
8792 old_runtime = sysctl_sched_rt_runtime;
8794 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8796 if (!ret && write) {
8797 ret = sched_rt_global_constraints();
8799 sysctl_sched_rt_period = old_period;
8800 sysctl_sched_rt_runtime = old_runtime;
8802 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8803 def_rt_bandwidth.rt_period =
8804 ns_to_ktime(global_rt_period());
8807 mutex_unlock(&mutex);
8812 #ifdef CONFIG_CGROUP_SCHED
8814 /* return corresponding task_group object of a cgroup */
8815 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8817 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8818 struct task_group, css);
8821 static struct cgroup_subsys_state *
8822 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8824 struct task_group *tg, *parent;
8826 if (!cgrp->parent) {
8827 /* This is early initialization for the top cgroup */
8828 return &root_task_group.css;
8831 parent = cgroup_tg(cgrp->parent);
8832 tg = sched_create_group(parent);
8834 return ERR_PTR(-ENOMEM);
8840 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8842 struct task_group *tg = cgroup_tg(cgrp);
8844 sched_destroy_group(tg);
8848 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8850 if ((current != tsk) && (!capable(CAP_SYS_NICE))) {
8851 const struct cred *cred = current_cred(), *tcred;
8853 tcred = __task_cred(tsk);
8855 if (cred->euid != tcred->uid && cred->euid != tcred->suid)
8859 #ifdef CONFIG_RT_GROUP_SCHED
8860 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
8863 /* We don't support RT-tasks being in separate groups */
8864 if (tsk->sched_class != &fair_sched_class)
8871 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8872 struct task_struct *tsk, bool threadgroup)
8874 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
8878 struct task_struct *c;
8880 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8881 retval = cpu_cgroup_can_attach_task(cgrp, c);
8893 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8894 struct cgroup *old_cont, struct task_struct *tsk,
8897 sched_move_task(tsk);
8899 struct task_struct *c;
8901 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8909 cpu_cgroup_exit(struct cgroup_subsys *ss, struct task_struct *task)
8912 * cgroup_exit() is called in the copy_process() failure path.
8913 * Ignore this case since the task hasn't ran yet, this avoids
8914 * trying to poke a half freed task state from generic code.
8916 if (!(task->flags & PF_EXITING))
8919 sched_move_task(task);
8922 #ifdef CONFIG_FAIR_GROUP_SCHED
8923 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8926 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8929 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8931 struct task_group *tg = cgroup_tg(cgrp);
8933 return (u64) tg->shares;
8935 #endif /* CONFIG_FAIR_GROUP_SCHED */
8937 #ifdef CONFIG_RT_GROUP_SCHED
8938 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8941 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8944 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8946 return sched_group_rt_runtime(cgroup_tg(cgrp));
8949 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8952 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8955 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8957 return sched_group_rt_period(cgroup_tg(cgrp));
8959 #endif /* CONFIG_RT_GROUP_SCHED */
8961 static struct cftype cpu_files[] = {
8962 #ifdef CONFIG_FAIR_GROUP_SCHED
8965 .read_u64 = cpu_shares_read_u64,
8966 .write_u64 = cpu_shares_write_u64,
8969 #ifdef CONFIG_RT_GROUP_SCHED
8971 .name = "rt_runtime_us",
8972 .read_s64 = cpu_rt_runtime_read,
8973 .write_s64 = cpu_rt_runtime_write,
8976 .name = "rt_period_us",
8977 .read_u64 = cpu_rt_period_read_uint,
8978 .write_u64 = cpu_rt_period_write_uint,
8983 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8985 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8988 struct cgroup_subsys cpu_cgroup_subsys = {
8990 .create = cpu_cgroup_create,
8991 .destroy = cpu_cgroup_destroy,
8992 .can_attach = cpu_cgroup_can_attach,
8993 .attach = cpu_cgroup_attach,
8994 .exit = cpu_cgroup_exit,
8995 .populate = cpu_cgroup_populate,
8996 .subsys_id = cpu_cgroup_subsys_id,
9000 #endif /* CONFIG_CGROUP_SCHED */
9002 #ifdef CONFIG_CGROUP_CPUACCT
9005 * CPU accounting code for task groups.
9007 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9008 * (balbir@in.ibm.com).
9011 /* track cpu usage of a group of tasks and its child groups */
9013 struct cgroup_subsys_state css;
9014 /* cpuusage holds pointer to a u64-type object on every cpu */
9015 u64 __percpu *cpuusage;
9016 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
9017 struct cpuacct *parent;
9018 struct cpuacct_charge_calls *cpufreq_fn;
9022 static struct cpuacct *cpuacct_root;
9024 /* Default calls for cpufreq accounting */
9025 static struct cpuacct_charge_calls *cpuacct_cpufreq;
9026 int cpuacct_register_cpufreq(struct cpuacct_charge_calls *fn)
9028 cpuacct_cpufreq = fn;
9031 * Root node is created before platform can register callbacks,
9034 if (cpuacct_root && fn) {
9035 cpuacct_root->cpufreq_fn = fn;
9037 fn->init(&cpuacct_root->cpuacct_data);
9042 struct cgroup_subsys cpuacct_subsys;
9044 /* return cpu accounting group corresponding to this container */
9045 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9047 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9048 struct cpuacct, css);
9051 /* return cpu accounting group to which this task belongs */
9052 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9054 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9055 struct cpuacct, css);
9058 /* create a new cpu accounting group */
9059 static struct cgroup_subsys_state *cpuacct_create(
9060 struct cgroup_subsys *ss, struct cgroup *cgrp)
9062 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9068 ca->cpuusage = alloc_percpu(u64);
9072 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9073 if (percpu_counter_init(&ca->cpustat[i], 0))
9074 goto out_free_counters;
9076 ca->cpufreq_fn = cpuacct_cpufreq;
9078 /* If available, have platform code initalize cpu frequency table */
9079 if (ca->cpufreq_fn && ca->cpufreq_fn->init)
9080 ca->cpufreq_fn->init(&ca->cpuacct_data);
9083 ca->parent = cgroup_ca(cgrp->parent);
9091 percpu_counter_destroy(&ca->cpustat[i]);
9092 free_percpu(ca->cpuusage);
9096 return ERR_PTR(-ENOMEM);
9099 /* destroy an existing cpu accounting group */
9101 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9103 struct cpuacct *ca = cgroup_ca(cgrp);
9106 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9107 percpu_counter_destroy(&ca->cpustat[i]);
9108 free_percpu(ca->cpuusage);
9112 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
9114 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9117 #ifndef CONFIG_64BIT
9119 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9121 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9123 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9131 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9133 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9135 #ifndef CONFIG_64BIT
9137 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9139 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9141 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9147 /* return total cpu usage (in nanoseconds) of a group */
9148 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9150 struct cpuacct *ca = cgroup_ca(cgrp);
9151 u64 totalcpuusage = 0;
9154 for_each_present_cpu(i)
9155 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9157 return totalcpuusage;
9160 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9163 struct cpuacct *ca = cgroup_ca(cgrp);
9172 for_each_present_cpu(i)
9173 cpuacct_cpuusage_write(ca, i, 0);
9179 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9182 struct cpuacct *ca = cgroup_ca(cgroup);
9186 for_each_present_cpu(i) {
9187 percpu = cpuacct_cpuusage_read(ca, i);
9188 seq_printf(m, "%llu ", (unsigned long long) percpu);
9190 seq_printf(m, "\n");
9194 static const char *cpuacct_stat_desc[] = {
9195 [CPUACCT_STAT_USER] = "user",
9196 [CPUACCT_STAT_SYSTEM] = "system",
9199 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
9200 struct cgroup_map_cb *cb)
9202 struct cpuacct *ca = cgroup_ca(cgrp);
9205 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
9206 s64 val = percpu_counter_read(&ca->cpustat[i]);
9207 val = cputime64_to_clock_t(val);
9208 cb->fill(cb, cpuacct_stat_desc[i], val);
9213 static int cpuacct_cpufreq_show(struct cgroup *cgrp, struct cftype *cft,
9214 struct cgroup_map_cb *cb)
9216 struct cpuacct *ca = cgroup_ca(cgrp);
9217 if (ca->cpufreq_fn && ca->cpufreq_fn->cpufreq_show)
9218 ca->cpufreq_fn->cpufreq_show(ca->cpuacct_data, cb);
9223 /* return total cpu power usage (milliWatt second) of a group */
9224 static u64 cpuacct_powerusage_read(struct cgroup *cgrp, struct cftype *cft)
9227 struct cpuacct *ca = cgroup_ca(cgrp);
9230 if (ca->cpufreq_fn && ca->cpufreq_fn->power_usage)
9231 for_each_present_cpu(i) {
9232 totalpower += ca->cpufreq_fn->power_usage(
9239 static struct cftype files[] = {
9242 .read_u64 = cpuusage_read,
9243 .write_u64 = cpuusage_write,
9246 .name = "usage_percpu",
9247 .read_seq_string = cpuacct_percpu_seq_read,
9251 .read_map = cpuacct_stats_show,
9255 .read_map = cpuacct_cpufreq_show,
9259 .read_u64 = cpuacct_powerusage_read
9263 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9265 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9269 * charge this task's execution time to its accounting group.
9271 * called with rq->lock held.
9273 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9278 if (unlikely(!cpuacct_subsys.active))
9281 cpu = task_cpu(tsk);
9287 for (; ca; ca = ca->parent) {
9288 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9289 *cpuusage += cputime;
9291 /* Call back into platform code to account for CPU speeds */
9292 if (ca->cpufreq_fn && ca->cpufreq_fn->charge)
9293 ca->cpufreq_fn->charge(ca->cpuacct_data, cputime, cpu);
9300 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9301 * in cputime_t units. As a result, cpuacct_update_stats calls
9302 * percpu_counter_add with values large enough to always overflow the
9303 * per cpu batch limit causing bad SMP scalability.
9305 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9306 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9307 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9310 #define CPUACCT_BATCH \
9311 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9313 #define CPUACCT_BATCH 0
9317 * Charge the system/user time to the task's accounting group.
9319 static void cpuacct_update_stats(struct task_struct *tsk,
9320 enum cpuacct_stat_index idx, cputime_t val)
9323 int batch = CPUACCT_BATCH;
9325 if (unlikely(!cpuacct_subsys.active))
9332 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9338 struct cgroup_subsys cpuacct_subsys = {
9340 .create = cpuacct_create,
9341 .destroy = cpuacct_destroy,
9342 .populate = cpuacct_populate,
9343 .subsys_id = cpuacct_subsys_id,
9345 #endif /* CONFIG_CGROUP_CPUACCT */