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>
77 #include <asm/irq_regs.h>
79 #include "sched_cpupri.h"
80 #include "workqueue_sched.h"
82 #define CREATE_TRACE_POINTS
83 #include <trace/events/sched.h>
86 * Convert user-nice values [ -20 ... 0 ... 19 ]
87 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
90 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
91 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
92 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
95 * 'User priority' is the nice value converted to something we
96 * can work with better when scaling various scheduler parameters,
97 * it's a [ 0 ... 39 ] range.
99 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
100 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
101 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
104 * Helpers for converting nanosecond timing to jiffy resolution
106 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
108 #define NICE_0_LOAD SCHED_LOAD_SCALE
109 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
112 * These are the 'tuning knobs' of the scheduler:
114 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
115 * Timeslices get refilled after they expire.
117 #define DEF_TIMESLICE (100 * HZ / 1000)
120 * single value that denotes runtime == period, ie unlimited time.
122 #define RUNTIME_INF ((u64)~0ULL)
124 static inline int rt_policy(int policy)
126 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
131 static inline int task_has_rt_policy(struct task_struct *p)
133 return rt_policy(p->policy);
137 * This is the priority-queue data structure of the RT scheduling class:
139 struct rt_prio_array {
140 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
141 struct list_head queue[MAX_RT_PRIO];
144 struct rt_bandwidth {
145 /* nests inside the rq lock: */
146 raw_spinlock_t rt_runtime_lock;
149 struct hrtimer rt_period_timer;
152 static struct rt_bandwidth def_rt_bandwidth;
154 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
156 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
158 struct rt_bandwidth *rt_b =
159 container_of(timer, struct rt_bandwidth, rt_period_timer);
165 now = hrtimer_cb_get_time(timer);
166 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
171 idle = do_sched_rt_period_timer(rt_b, overrun);
174 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
178 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
180 rt_b->rt_period = ns_to_ktime(period);
181 rt_b->rt_runtime = runtime;
183 raw_spin_lock_init(&rt_b->rt_runtime_lock);
185 hrtimer_init(&rt_b->rt_period_timer,
186 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
187 rt_b->rt_period_timer.function = sched_rt_period_timer;
190 static inline int rt_bandwidth_enabled(void)
192 return sysctl_sched_rt_runtime >= 0;
195 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
199 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
202 if (hrtimer_active(&rt_b->rt_period_timer))
205 raw_spin_lock(&rt_b->rt_runtime_lock);
210 if (hrtimer_active(&rt_b->rt_period_timer))
213 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
214 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
216 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
217 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
218 delta = ktime_to_ns(ktime_sub(hard, soft));
219 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
220 HRTIMER_MODE_ABS_PINNED, 0);
222 raw_spin_unlock(&rt_b->rt_runtime_lock);
225 #ifdef CONFIG_RT_GROUP_SCHED
226 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
228 hrtimer_cancel(&rt_b->rt_period_timer);
233 * sched_domains_mutex serializes calls to arch_init_sched_domains,
234 * detach_destroy_domains and partition_sched_domains.
236 static DEFINE_MUTEX(sched_domains_mutex);
238 #ifdef CONFIG_CGROUP_SCHED
240 #include <linux/cgroup.h>
244 static LIST_HEAD(task_groups);
246 /* task group related information */
248 struct cgroup_subsys_state css;
250 #ifdef CONFIG_FAIR_GROUP_SCHED
251 /* schedulable entities of this group on each cpu */
252 struct sched_entity **se;
253 /* runqueue "owned" by this group on each cpu */
254 struct cfs_rq **cfs_rq;
255 unsigned long shares;
258 #ifdef CONFIG_RT_GROUP_SCHED
259 struct sched_rt_entity **rt_se;
260 struct rt_rq **rt_rq;
262 struct rt_bandwidth rt_bandwidth;
266 struct list_head list;
268 struct task_group *parent;
269 struct list_head siblings;
270 struct list_head children;
273 #define root_task_group init_task_group
275 /* task_group_lock serializes add/remove of task groups and also changes to
276 * a task group's cpu shares.
278 static DEFINE_SPINLOCK(task_group_lock);
280 #ifdef CONFIG_FAIR_GROUP_SCHED
283 static int root_task_group_empty(void)
285 return list_empty(&root_task_group.children);
289 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
292 * A weight of 0 or 1 can cause arithmetics problems.
293 * A weight of a cfs_rq is the sum of weights of which entities
294 * are queued on this cfs_rq, so a weight of a entity should not be
295 * too large, so as the shares value of a task group.
296 * (The default weight is 1024 - so there's no practical
297 * limitation from this.)
300 #define MAX_SHARES (1UL << 18)
302 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
305 /* Default task group.
306 * Every task in system belong to this group at bootup.
308 struct task_group init_task_group;
310 #endif /* CONFIG_CGROUP_SCHED */
312 /* CFS-related fields in a runqueue */
314 struct load_weight load;
315 unsigned long nr_running;
320 struct rb_root tasks_timeline;
321 struct rb_node *rb_leftmost;
323 struct list_head tasks;
324 struct list_head *balance_iterator;
327 * 'curr' points to currently running entity on this cfs_rq.
328 * It is set to NULL otherwise (i.e when none are currently running).
330 struct sched_entity *curr, *next, *last;
332 unsigned int nr_spread_over;
334 #ifdef CONFIG_FAIR_GROUP_SCHED
335 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
338 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
339 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
340 * (like users, containers etc.)
342 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
343 * list is used during load balance.
345 struct list_head leaf_cfs_rq_list;
346 struct task_group *tg; /* group that "owns" this runqueue */
350 * the part of load.weight contributed by tasks
352 unsigned long task_weight;
355 * h_load = weight * f(tg)
357 * Where f(tg) is the recursive weight fraction assigned to
360 unsigned long h_load;
363 * this cpu's part of tg->shares
365 unsigned long shares;
368 * load.weight at the time we set shares
370 unsigned long rq_weight;
375 /* Real-Time classes' related field in a runqueue: */
377 struct rt_prio_array active;
378 unsigned long rt_nr_running;
379 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
381 int curr; /* highest queued rt task prio */
383 int next; /* next highest */
388 unsigned long rt_nr_migratory;
389 unsigned long rt_nr_total;
391 struct plist_head pushable_tasks;
396 /* Nests inside the rq lock: */
397 raw_spinlock_t rt_runtime_lock;
399 #ifdef CONFIG_RT_GROUP_SCHED
400 unsigned long rt_nr_boosted;
403 struct list_head leaf_rt_rq_list;
404 struct task_group *tg;
411 * We add the notion of a root-domain which will be used to define per-domain
412 * variables. Each exclusive cpuset essentially defines an island domain by
413 * fully partitioning the member cpus from any other cpuset. Whenever a new
414 * exclusive cpuset is created, we also create and attach a new root-domain
421 cpumask_var_t online;
424 * The "RT overload" flag: it gets set if a CPU has more than
425 * one runnable RT task.
427 cpumask_var_t rto_mask;
430 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;
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;
492 unsigned long next_balance;
493 struct mm_struct *prev_mm;
500 struct root_domain *rd;
501 struct sched_domain *sd;
503 unsigned long cpu_power;
505 unsigned char idle_at_tick;
506 /* For active balancing */
510 struct cpu_stop_work active_balance_work;
511 /* cpu of this runqueue: */
515 unsigned long avg_load_per_task;
523 /* calc_load related fields */
524 unsigned long calc_load_update;
525 long calc_load_active;
527 #ifdef CONFIG_SCHED_HRTICK
529 int hrtick_csd_pending;
530 struct call_single_data hrtick_csd;
532 struct hrtimer hrtick_timer;
535 #ifdef CONFIG_SCHEDSTATS
537 struct sched_info rq_sched_info;
538 unsigned long long rq_cpu_time;
539 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
541 /* sys_sched_yield() stats */
542 unsigned int yld_count;
544 /* schedule() stats */
545 unsigned int sched_switch;
546 unsigned int sched_count;
547 unsigned int sched_goidle;
549 /* try_to_wake_up() stats */
550 unsigned int ttwu_count;
551 unsigned int ttwu_local;
554 unsigned int bkl_count;
558 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
561 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
563 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
566 * A queue event has occurred, and we're going to schedule. In
567 * this case, we can save a useless back to back clock update.
569 if (test_tsk_need_resched(p))
570 rq->skip_clock_update = 1;
573 static inline int cpu_of(struct rq *rq)
582 #define rcu_dereference_check_sched_domain(p) \
583 rcu_dereference_check((p), \
584 rcu_read_lock_sched_held() || \
585 lockdep_is_held(&sched_domains_mutex))
588 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
589 * See detach_destroy_domains: synchronize_sched for details.
591 * The domain tree of any CPU may only be accessed from within
592 * preempt-disabled sections.
594 #define for_each_domain(cpu, __sd) \
595 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
597 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
598 #define this_rq() (&__get_cpu_var(runqueues))
599 #define task_rq(p) cpu_rq(task_cpu(p))
600 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
601 #define raw_rq() (&__raw_get_cpu_var(runqueues))
603 #ifdef CONFIG_CGROUP_SCHED
606 * Return the group to which this tasks belongs.
608 * We use task_subsys_state_check() and extend the RCU verification
609 * with lockdep_is_held(&task_rq(p)->lock) because cpu_cgroup_attach()
610 * holds that lock for each task it moves into the cgroup. Therefore
611 * by holding that lock, we pin the task to the current cgroup.
613 static inline struct task_group *task_group(struct task_struct *p)
615 struct cgroup_subsys_state *css;
617 css = task_subsys_state_check(p, cpu_cgroup_subsys_id,
618 lockdep_is_held(&task_rq(p)->lock));
619 return container_of(css, struct task_group, css);
622 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
623 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
625 #ifdef CONFIG_FAIR_GROUP_SCHED
626 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
627 p->se.parent = task_group(p)->se[cpu];
630 #ifdef CONFIG_RT_GROUP_SCHED
631 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
632 p->rt.parent = task_group(p)->rt_se[cpu];
636 #else /* CONFIG_CGROUP_SCHED */
638 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
639 static inline struct task_group *task_group(struct task_struct *p)
644 #endif /* CONFIG_CGROUP_SCHED */
646 inline void update_rq_clock(struct rq *rq)
648 if (!rq->skip_clock_update)
649 rq->clock = sched_clock_cpu(cpu_of(rq));
653 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
655 #ifdef CONFIG_SCHED_DEBUG
656 # define const_debug __read_mostly
658 # define const_debug static const
663 * @cpu: the processor in question.
665 * Returns true if the current cpu runqueue is locked.
666 * This interface allows printk to be called with the runqueue lock
667 * held and know whether or not it is OK to wake up the klogd.
669 int runqueue_is_locked(int cpu)
671 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
675 * Debugging: various feature bits
678 #define SCHED_FEAT(name, enabled) \
679 __SCHED_FEAT_##name ,
682 #include "sched_features.h"
687 #define SCHED_FEAT(name, enabled) \
688 (1UL << __SCHED_FEAT_##name) * enabled |
690 const_debug unsigned int sysctl_sched_features =
691 #include "sched_features.h"
696 #ifdef CONFIG_SCHED_DEBUG
697 #define SCHED_FEAT(name, enabled) \
700 static __read_mostly char *sched_feat_names[] = {
701 #include "sched_features.h"
707 static int sched_feat_show(struct seq_file *m, void *v)
711 for (i = 0; sched_feat_names[i]; i++) {
712 if (!(sysctl_sched_features & (1UL << i)))
714 seq_printf(m, "%s ", sched_feat_names[i]);
722 sched_feat_write(struct file *filp, const char __user *ubuf,
723 size_t cnt, loff_t *ppos)
733 if (copy_from_user(&buf, ubuf, cnt))
738 if (strncmp(buf, "NO_", 3) == 0) {
743 for (i = 0; sched_feat_names[i]; i++) {
744 int len = strlen(sched_feat_names[i]);
746 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
748 sysctl_sched_features &= ~(1UL << i);
750 sysctl_sched_features |= (1UL << i);
755 if (!sched_feat_names[i])
763 static int sched_feat_open(struct inode *inode, struct file *filp)
765 return single_open(filp, sched_feat_show, NULL);
768 static const struct file_operations sched_feat_fops = {
769 .open = sched_feat_open,
770 .write = sched_feat_write,
773 .release = single_release,
776 static __init int sched_init_debug(void)
778 debugfs_create_file("sched_features", 0644, NULL, NULL,
783 late_initcall(sched_init_debug);
787 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
790 * Number of tasks to iterate in a single balance run.
791 * Limited because this is done with IRQs disabled.
793 const_debug unsigned int sysctl_sched_nr_migrate = 32;
796 * ratelimit for updating the group shares.
799 unsigned int sysctl_sched_shares_ratelimit = 250000;
800 unsigned int normalized_sysctl_sched_shares_ratelimit = 250000;
803 * Inject some fuzzyness into changing the per-cpu group shares
804 * this avoids remote rq-locks at the expense of fairness.
807 unsigned int sysctl_sched_shares_thresh = 4;
810 * period over which we average the RT time consumption, measured
815 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
818 * period over which we measure -rt task cpu usage in us.
821 unsigned int sysctl_sched_rt_period = 1000000;
823 static __read_mostly int scheduler_running;
826 * part of the period that we allow rt tasks to run in us.
829 int sysctl_sched_rt_runtime = 950000;
831 static inline u64 global_rt_period(void)
833 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
836 static inline u64 global_rt_runtime(void)
838 if (sysctl_sched_rt_runtime < 0)
841 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
844 #ifndef prepare_arch_switch
845 # define prepare_arch_switch(next) do { } while (0)
847 #ifndef finish_arch_switch
848 # define finish_arch_switch(prev) do { } while (0)
851 static inline int task_current(struct rq *rq, struct task_struct *p)
853 return rq->curr == p;
856 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
857 static inline int task_running(struct rq *rq, struct task_struct *p)
859 return task_current(rq, p);
862 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
866 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
868 #ifdef CONFIG_DEBUG_SPINLOCK
869 /* this is a valid case when another task releases the spinlock */
870 rq->lock.owner = current;
873 * If we are tracking spinlock dependencies then we have to
874 * fix up the runqueue lock - which gets 'carried over' from
877 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
879 raw_spin_unlock_irq(&rq->lock);
882 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
883 static inline int task_running(struct rq *rq, struct task_struct *p)
888 return task_current(rq, p);
892 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
896 * We can optimise this out completely for !SMP, because the
897 * SMP rebalancing from interrupt is the only thing that cares
902 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
903 raw_spin_unlock_irq(&rq->lock);
905 raw_spin_unlock(&rq->lock);
909 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
913 * After ->oncpu is cleared, the task can be moved to a different CPU.
914 * We must ensure this doesn't happen until the switch is completely
920 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
924 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
927 * Check whether the task is waking, we use this to synchronize ->cpus_allowed
930 static inline int task_is_waking(struct task_struct *p)
932 return unlikely(p->state == TASK_WAKING);
936 * __task_rq_lock - lock the runqueue a given task resides on.
937 * Must be called interrupts disabled.
939 static inline struct rq *__task_rq_lock(struct task_struct *p)
946 raw_spin_lock(&rq->lock);
947 if (likely(rq == task_rq(p)))
949 raw_spin_unlock(&rq->lock);
954 * task_rq_lock - lock the runqueue a given task resides on and disable
955 * interrupts. Note the ordering: we can safely lookup the task_rq without
956 * explicitly disabling preemption.
958 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
964 local_irq_save(*flags);
966 raw_spin_lock(&rq->lock);
967 if (likely(rq == task_rq(p)))
969 raw_spin_unlock_irqrestore(&rq->lock, *flags);
973 static void __task_rq_unlock(struct rq *rq)
976 raw_spin_unlock(&rq->lock);
979 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
982 raw_spin_unlock_irqrestore(&rq->lock, *flags);
986 * this_rq_lock - lock this runqueue and disable interrupts.
988 static struct rq *this_rq_lock(void)
995 raw_spin_lock(&rq->lock);
1000 #ifdef CONFIG_SCHED_HRTICK
1002 * Use HR-timers to deliver accurate preemption points.
1004 * Its all a bit involved since we cannot program an hrt while holding the
1005 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1008 * When we get rescheduled we reprogram the hrtick_timer outside of the
1014 * - enabled by features
1015 * - hrtimer is actually high res
1017 static inline int hrtick_enabled(struct rq *rq)
1019 if (!sched_feat(HRTICK))
1021 if (!cpu_active(cpu_of(rq)))
1023 return hrtimer_is_hres_active(&rq->hrtick_timer);
1026 static void hrtick_clear(struct rq *rq)
1028 if (hrtimer_active(&rq->hrtick_timer))
1029 hrtimer_cancel(&rq->hrtick_timer);
1033 * High-resolution timer tick.
1034 * Runs from hardirq context with interrupts disabled.
1036 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1038 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1040 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1042 raw_spin_lock(&rq->lock);
1043 update_rq_clock(rq);
1044 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1045 raw_spin_unlock(&rq->lock);
1047 return HRTIMER_NORESTART;
1052 * called from hardirq (IPI) context
1054 static void __hrtick_start(void *arg)
1056 struct rq *rq = arg;
1058 raw_spin_lock(&rq->lock);
1059 hrtimer_restart(&rq->hrtick_timer);
1060 rq->hrtick_csd_pending = 0;
1061 raw_spin_unlock(&rq->lock);
1065 * Called to set the hrtick timer state.
1067 * called with rq->lock held and irqs disabled
1069 static void hrtick_start(struct rq *rq, u64 delay)
1071 struct hrtimer *timer = &rq->hrtick_timer;
1072 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1074 hrtimer_set_expires(timer, time);
1076 if (rq == this_rq()) {
1077 hrtimer_restart(timer);
1078 } else if (!rq->hrtick_csd_pending) {
1079 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1080 rq->hrtick_csd_pending = 1;
1085 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1087 int cpu = (int)(long)hcpu;
1090 case CPU_UP_CANCELED:
1091 case CPU_UP_CANCELED_FROZEN:
1092 case CPU_DOWN_PREPARE:
1093 case CPU_DOWN_PREPARE_FROZEN:
1095 case CPU_DEAD_FROZEN:
1096 hrtick_clear(cpu_rq(cpu));
1103 static __init void init_hrtick(void)
1105 hotcpu_notifier(hotplug_hrtick, 0);
1109 * Called to set the hrtick timer state.
1111 * called with rq->lock held and irqs disabled
1113 static void hrtick_start(struct rq *rq, u64 delay)
1115 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1116 HRTIMER_MODE_REL_PINNED, 0);
1119 static inline void init_hrtick(void)
1122 #endif /* CONFIG_SMP */
1124 static void init_rq_hrtick(struct rq *rq)
1127 rq->hrtick_csd_pending = 0;
1129 rq->hrtick_csd.flags = 0;
1130 rq->hrtick_csd.func = __hrtick_start;
1131 rq->hrtick_csd.info = rq;
1134 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1135 rq->hrtick_timer.function = hrtick;
1137 #else /* CONFIG_SCHED_HRTICK */
1138 static inline void hrtick_clear(struct rq *rq)
1142 static inline void init_rq_hrtick(struct rq *rq)
1146 static inline void init_hrtick(void)
1149 #endif /* CONFIG_SCHED_HRTICK */
1152 * resched_task - mark a task 'to be rescheduled now'.
1154 * On UP this means the setting of the need_resched flag, on SMP it
1155 * might also involve a cross-CPU call to trigger the scheduler on
1160 #ifndef tsk_is_polling
1161 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1164 static void resched_task(struct task_struct *p)
1168 assert_raw_spin_locked(&task_rq(p)->lock);
1170 if (test_tsk_need_resched(p))
1173 set_tsk_need_resched(p);
1176 if (cpu == smp_processor_id())
1179 /* NEED_RESCHED must be visible before we test polling */
1181 if (!tsk_is_polling(p))
1182 smp_send_reschedule(cpu);
1185 static void resched_cpu(int cpu)
1187 struct rq *rq = cpu_rq(cpu);
1188 unsigned long flags;
1190 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1192 resched_task(cpu_curr(cpu));
1193 raw_spin_unlock_irqrestore(&rq->lock, flags);
1198 * In the semi idle case, use the nearest busy cpu for migrating timers
1199 * from an idle cpu. This is good for power-savings.
1201 * We don't do similar optimization for completely idle system, as
1202 * selecting an idle cpu will add more delays to the timers than intended
1203 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1205 int get_nohz_timer_target(void)
1207 int cpu = smp_processor_id();
1209 struct sched_domain *sd;
1211 for_each_domain(cpu, sd) {
1212 for_each_cpu(i, sched_domain_span(sd))
1219 * When add_timer_on() enqueues a timer into the timer wheel of an
1220 * idle CPU then this timer might expire before the next timer event
1221 * which is scheduled to wake up that CPU. In case of a completely
1222 * idle system the next event might even be infinite time into the
1223 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1224 * leaves the inner idle loop so the newly added timer is taken into
1225 * account when the CPU goes back to idle and evaluates the timer
1226 * wheel for the next timer event.
1228 void wake_up_idle_cpu(int cpu)
1230 struct rq *rq = cpu_rq(cpu);
1232 if (cpu == smp_processor_id())
1236 * This is safe, as this function is called with the timer
1237 * wheel base lock of (cpu) held. When the CPU is on the way
1238 * to idle and has not yet set rq->curr to idle then it will
1239 * be serialized on the timer wheel base lock and take the new
1240 * timer into account automatically.
1242 if (rq->curr != rq->idle)
1246 * We can set TIF_RESCHED on the idle task of the other CPU
1247 * lockless. The worst case is that the other CPU runs the
1248 * idle task through an additional NOOP schedule()
1250 set_tsk_need_resched(rq->idle);
1252 /* NEED_RESCHED must be visible before we test polling */
1254 if (!tsk_is_polling(rq->idle))
1255 smp_send_reschedule(cpu);
1258 int nohz_ratelimit(int cpu)
1260 struct rq *rq = cpu_rq(cpu);
1261 u64 diff = rq->clock - rq->nohz_stamp;
1263 rq->nohz_stamp = rq->clock;
1265 return diff < (NSEC_PER_SEC / HZ) >> 1;
1268 #endif /* CONFIG_NO_HZ */
1270 static u64 sched_avg_period(void)
1272 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1275 static void sched_avg_update(struct rq *rq)
1277 s64 period = sched_avg_period();
1279 while ((s64)(rq->clock - rq->age_stamp) > period) {
1280 rq->age_stamp += period;
1285 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1287 rq->rt_avg += rt_delta;
1288 sched_avg_update(rq);
1291 #else /* !CONFIG_SMP */
1292 static void resched_task(struct task_struct *p)
1294 assert_raw_spin_locked(&task_rq(p)->lock);
1295 set_tsk_need_resched(p);
1298 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1301 #endif /* CONFIG_SMP */
1303 #if BITS_PER_LONG == 32
1304 # define WMULT_CONST (~0UL)
1306 # define WMULT_CONST (1UL << 32)
1309 #define WMULT_SHIFT 32
1312 * Shift right and round:
1314 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1317 * delta *= weight / lw
1319 static unsigned long
1320 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1321 struct load_weight *lw)
1325 if (!lw->inv_weight) {
1326 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1329 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1333 tmp = (u64)delta_exec * weight;
1335 * Check whether we'd overflow the 64-bit multiplication:
1337 if (unlikely(tmp > WMULT_CONST))
1338 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1341 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1343 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1346 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1352 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1359 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1360 * of tasks with abnormal "nice" values across CPUs the contribution that
1361 * each task makes to its run queue's load is weighted according to its
1362 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1363 * scaled version of the new time slice allocation that they receive on time
1367 #define WEIGHT_IDLEPRIO 3
1368 #define WMULT_IDLEPRIO 1431655765
1371 * Nice levels are multiplicative, with a gentle 10% change for every
1372 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1373 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1374 * that remained on nice 0.
1376 * The "10% effect" is relative and cumulative: from _any_ nice level,
1377 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1378 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1379 * If a task goes up by ~10% and another task goes down by ~10% then
1380 * the relative distance between them is ~25%.)
1382 static const int prio_to_weight[40] = {
1383 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1384 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1385 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1386 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1387 /* 0 */ 1024, 820, 655, 526, 423,
1388 /* 5 */ 335, 272, 215, 172, 137,
1389 /* 10 */ 110, 87, 70, 56, 45,
1390 /* 15 */ 36, 29, 23, 18, 15,
1394 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1396 * In cases where the weight does not change often, we can use the
1397 * precalculated inverse to speed up arithmetics by turning divisions
1398 * into multiplications:
1400 static const u32 prio_to_wmult[40] = {
1401 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1402 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1403 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1404 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1405 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1406 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1407 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1408 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1411 /* Time spent by the tasks of the cpu accounting group executing in ... */
1412 enum cpuacct_stat_index {
1413 CPUACCT_STAT_USER, /* ... user mode */
1414 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1416 CPUACCT_STAT_NSTATS,
1419 #ifdef CONFIG_CGROUP_CPUACCT
1420 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1421 static void cpuacct_update_stats(struct task_struct *tsk,
1422 enum cpuacct_stat_index idx, cputime_t val);
1424 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1425 static inline void cpuacct_update_stats(struct task_struct *tsk,
1426 enum cpuacct_stat_index idx, cputime_t val) {}
1429 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1431 update_load_add(&rq->load, load);
1434 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1436 update_load_sub(&rq->load, load);
1439 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1440 typedef int (*tg_visitor)(struct task_group *, void *);
1443 * Iterate the full tree, calling @down when first entering a node and @up when
1444 * leaving it for the final time.
1446 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1448 struct task_group *parent, *child;
1452 parent = &root_task_group;
1454 ret = (*down)(parent, data);
1457 list_for_each_entry_rcu(child, &parent->children, siblings) {
1464 ret = (*up)(parent, data);
1469 parent = parent->parent;
1478 static int tg_nop(struct task_group *tg, void *data)
1485 /* Used instead of source_load when we know the type == 0 */
1486 static unsigned long weighted_cpuload(const int cpu)
1488 return cpu_rq(cpu)->load.weight;
1492 * Return a low guess at the load of a migration-source cpu weighted
1493 * according to the scheduling class and "nice" value.
1495 * We want to under-estimate the load of migration sources, to
1496 * balance conservatively.
1498 static unsigned long source_load(int cpu, int type)
1500 struct rq *rq = cpu_rq(cpu);
1501 unsigned long total = weighted_cpuload(cpu);
1503 if (type == 0 || !sched_feat(LB_BIAS))
1506 return min(rq->cpu_load[type-1], total);
1510 * Return a high guess at the load of a migration-target cpu weighted
1511 * according to the scheduling class and "nice" value.
1513 static unsigned long target_load(int cpu, int type)
1515 struct rq *rq = cpu_rq(cpu);
1516 unsigned long total = weighted_cpuload(cpu);
1518 if (type == 0 || !sched_feat(LB_BIAS))
1521 return max(rq->cpu_load[type-1], total);
1524 static unsigned long power_of(int cpu)
1526 return cpu_rq(cpu)->cpu_power;
1529 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1531 static unsigned long cpu_avg_load_per_task(int cpu)
1533 struct rq *rq = cpu_rq(cpu);
1534 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1537 rq->avg_load_per_task = rq->load.weight / nr_running;
1539 rq->avg_load_per_task = 0;
1541 return rq->avg_load_per_task;
1544 #ifdef CONFIG_FAIR_GROUP_SCHED
1546 static __read_mostly unsigned long __percpu *update_shares_data;
1548 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1551 * Calculate and set the cpu's group shares.
1553 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1554 unsigned long sd_shares,
1555 unsigned long sd_rq_weight,
1556 unsigned long *usd_rq_weight)
1558 unsigned long shares, rq_weight;
1561 rq_weight = usd_rq_weight[cpu];
1564 rq_weight = NICE_0_LOAD;
1568 * \Sum_j shares_j * rq_weight_i
1569 * shares_i = -----------------------------
1570 * \Sum_j rq_weight_j
1572 shares = (sd_shares * rq_weight) / sd_rq_weight;
1573 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1575 if (abs(shares - tg->se[cpu]->load.weight) >
1576 sysctl_sched_shares_thresh) {
1577 struct rq *rq = cpu_rq(cpu);
1578 unsigned long flags;
1580 raw_spin_lock_irqsave(&rq->lock, flags);
1581 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1582 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1583 __set_se_shares(tg->se[cpu], shares);
1584 raw_spin_unlock_irqrestore(&rq->lock, flags);
1589 * Re-compute the task group their per cpu shares over the given domain.
1590 * This needs to be done in a bottom-up fashion because the rq weight of a
1591 * parent group depends on the shares of its child groups.
1593 static int tg_shares_up(struct task_group *tg, void *data)
1595 unsigned long weight, rq_weight = 0, sum_weight = 0, shares = 0;
1596 unsigned long *usd_rq_weight;
1597 struct sched_domain *sd = data;
1598 unsigned long flags;
1604 local_irq_save(flags);
1605 usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id());
1607 for_each_cpu(i, sched_domain_span(sd)) {
1608 weight = tg->cfs_rq[i]->load.weight;
1609 usd_rq_weight[i] = weight;
1611 rq_weight += weight;
1613 * If there are currently no tasks on the cpu pretend there
1614 * is one of average load so that when a new task gets to
1615 * run here it will not get delayed by group starvation.
1618 weight = NICE_0_LOAD;
1620 sum_weight += weight;
1621 shares += tg->cfs_rq[i]->shares;
1625 rq_weight = sum_weight;
1627 if ((!shares && rq_weight) || shares > tg->shares)
1628 shares = tg->shares;
1630 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1631 shares = tg->shares;
1633 for_each_cpu(i, sched_domain_span(sd))
1634 update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight);
1636 local_irq_restore(flags);
1642 * Compute the cpu's hierarchical load factor for each task group.
1643 * This needs to be done in a top-down fashion because the load of a child
1644 * group is a fraction of its parents load.
1646 static int tg_load_down(struct task_group *tg, void *data)
1649 long cpu = (long)data;
1652 load = cpu_rq(cpu)->load.weight;
1654 load = tg->parent->cfs_rq[cpu]->h_load;
1655 load *= tg->cfs_rq[cpu]->shares;
1656 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1659 tg->cfs_rq[cpu]->h_load = load;
1664 static void update_shares(struct sched_domain *sd)
1669 if (root_task_group_empty())
1672 now = local_clock();
1673 elapsed = now - sd->last_update;
1675 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1676 sd->last_update = now;
1677 walk_tg_tree(tg_nop, tg_shares_up, sd);
1681 static void update_h_load(long cpu)
1683 if (root_task_group_empty())
1686 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1691 static inline void update_shares(struct sched_domain *sd)
1697 #ifdef CONFIG_PREEMPT
1699 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1702 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1703 * way at the expense of forcing extra atomic operations in all
1704 * invocations. This assures that the double_lock is acquired using the
1705 * same underlying policy as the spinlock_t on this architecture, which
1706 * reduces latency compared to the unfair variant below. However, it
1707 * also adds more overhead and therefore may reduce throughput.
1709 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1710 __releases(this_rq->lock)
1711 __acquires(busiest->lock)
1712 __acquires(this_rq->lock)
1714 raw_spin_unlock(&this_rq->lock);
1715 double_rq_lock(this_rq, busiest);
1722 * Unfair double_lock_balance: Optimizes throughput at the expense of
1723 * latency by eliminating extra atomic operations when the locks are
1724 * already in proper order on entry. This favors lower cpu-ids and will
1725 * grant the double lock to lower cpus over higher ids under contention,
1726 * regardless of entry order into the function.
1728 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1729 __releases(this_rq->lock)
1730 __acquires(busiest->lock)
1731 __acquires(this_rq->lock)
1735 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1736 if (busiest < this_rq) {
1737 raw_spin_unlock(&this_rq->lock);
1738 raw_spin_lock(&busiest->lock);
1739 raw_spin_lock_nested(&this_rq->lock,
1740 SINGLE_DEPTH_NESTING);
1743 raw_spin_lock_nested(&busiest->lock,
1744 SINGLE_DEPTH_NESTING);
1749 #endif /* CONFIG_PREEMPT */
1752 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1754 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1756 if (unlikely(!irqs_disabled())) {
1757 /* printk() doesn't work good under rq->lock */
1758 raw_spin_unlock(&this_rq->lock);
1762 return _double_lock_balance(this_rq, busiest);
1765 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1766 __releases(busiest->lock)
1768 raw_spin_unlock(&busiest->lock);
1769 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1773 * double_rq_lock - safely lock two runqueues
1775 * Note this does not disable interrupts like task_rq_lock,
1776 * you need to do so manually before calling.
1778 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1779 __acquires(rq1->lock)
1780 __acquires(rq2->lock)
1782 BUG_ON(!irqs_disabled());
1784 raw_spin_lock(&rq1->lock);
1785 __acquire(rq2->lock); /* Fake it out ;) */
1788 raw_spin_lock(&rq1->lock);
1789 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1791 raw_spin_lock(&rq2->lock);
1792 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1798 * double_rq_unlock - safely unlock two runqueues
1800 * Note this does not restore interrupts like task_rq_unlock,
1801 * you need to do so manually after calling.
1803 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1804 __releases(rq1->lock)
1805 __releases(rq2->lock)
1807 raw_spin_unlock(&rq1->lock);
1809 raw_spin_unlock(&rq2->lock);
1811 __release(rq2->lock);
1816 #ifdef CONFIG_FAIR_GROUP_SCHED
1817 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1820 cfs_rq->shares = shares;
1825 static void calc_load_account_idle(struct rq *this_rq);
1826 static void update_sysctl(void);
1827 static int get_update_sysctl_factor(void);
1828 static void update_cpu_load(struct rq *this_rq);
1830 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1832 set_task_rq(p, cpu);
1835 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1836 * successfuly executed on another CPU. We must ensure that updates of
1837 * per-task data have been completed by this moment.
1840 task_thread_info(p)->cpu = cpu;
1844 static const struct sched_class rt_sched_class;
1846 #define sched_class_highest (&rt_sched_class)
1847 #define for_each_class(class) \
1848 for (class = sched_class_highest; class; class = class->next)
1850 #include "sched_stats.h"
1852 static void inc_nr_running(struct rq *rq)
1857 static void dec_nr_running(struct rq *rq)
1862 static void set_load_weight(struct task_struct *p)
1864 if (task_has_rt_policy(p)) {
1865 p->se.load.weight = 0;
1866 p->se.load.inv_weight = WMULT_CONST;
1871 * SCHED_IDLE tasks get minimal weight:
1873 if (p->policy == SCHED_IDLE) {
1874 p->se.load.weight = WEIGHT_IDLEPRIO;
1875 p->se.load.inv_weight = WMULT_IDLEPRIO;
1879 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1880 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1883 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1885 update_rq_clock(rq);
1886 sched_info_queued(p);
1887 p->sched_class->enqueue_task(rq, p, flags);
1891 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1893 update_rq_clock(rq);
1894 sched_info_dequeued(p);
1895 p->sched_class->dequeue_task(rq, p, flags);
1900 * activate_task - move a task to the runqueue.
1902 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1904 if (task_contributes_to_load(p))
1905 rq->nr_uninterruptible--;
1907 enqueue_task(rq, p, flags);
1912 * deactivate_task - remove a task from the runqueue.
1914 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1916 if (task_contributes_to_load(p))
1917 rq->nr_uninterruptible++;
1919 dequeue_task(rq, p, flags);
1923 #include "sched_idletask.c"
1924 #include "sched_fair.c"
1925 #include "sched_rt.c"
1926 #ifdef CONFIG_SCHED_DEBUG
1927 # include "sched_debug.c"
1931 * __normal_prio - return the priority that is based on the static prio
1933 static inline int __normal_prio(struct task_struct *p)
1935 return p->static_prio;
1939 * Calculate the expected normal priority: i.e. priority
1940 * without taking RT-inheritance into account. Might be
1941 * boosted by interactivity modifiers. Changes upon fork,
1942 * setprio syscalls, and whenever the interactivity
1943 * estimator recalculates.
1945 static inline int normal_prio(struct task_struct *p)
1949 if (task_has_rt_policy(p))
1950 prio = MAX_RT_PRIO-1 - p->rt_priority;
1952 prio = __normal_prio(p);
1957 * Calculate the current priority, i.e. the priority
1958 * taken into account by the scheduler. This value might
1959 * be boosted by RT tasks, or might be boosted by
1960 * interactivity modifiers. Will be RT if the task got
1961 * RT-boosted. If not then it returns p->normal_prio.
1963 static int effective_prio(struct task_struct *p)
1965 p->normal_prio = normal_prio(p);
1967 * If we are RT tasks or we were boosted to RT priority,
1968 * keep the priority unchanged. Otherwise, update priority
1969 * to the normal priority:
1971 if (!rt_prio(p->prio))
1972 return p->normal_prio;
1977 * task_curr - is this task currently executing on a CPU?
1978 * @p: the task in question.
1980 inline int task_curr(const struct task_struct *p)
1982 return cpu_curr(task_cpu(p)) == p;
1985 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1986 const struct sched_class *prev_class,
1987 int oldprio, int running)
1989 if (prev_class != p->sched_class) {
1990 if (prev_class->switched_from)
1991 prev_class->switched_from(rq, p, running);
1992 p->sched_class->switched_to(rq, p, running);
1994 p->sched_class->prio_changed(rq, p, oldprio, running);
1999 * Is this task likely cache-hot:
2002 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2006 if (p->sched_class != &fair_sched_class)
2010 * Buddy candidates are cache hot:
2012 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2013 (&p->se == cfs_rq_of(&p->se)->next ||
2014 &p->se == cfs_rq_of(&p->se)->last))
2017 if (sysctl_sched_migration_cost == -1)
2019 if (sysctl_sched_migration_cost == 0)
2022 delta = now - p->se.exec_start;
2024 return delta < (s64)sysctl_sched_migration_cost;
2027 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2029 #ifdef CONFIG_SCHED_DEBUG
2031 * We should never call set_task_cpu() on a blocked task,
2032 * ttwu() will sort out the placement.
2034 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2035 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2038 trace_sched_migrate_task(p, new_cpu);
2040 if (task_cpu(p) != new_cpu) {
2041 p->se.nr_migrations++;
2042 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2045 __set_task_cpu(p, new_cpu);
2048 struct migration_arg {
2049 struct task_struct *task;
2053 static int migration_cpu_stop(void *data);
2056 * The task's runqueue lock must be held.
2057 * Returns true if you have to wait for migration thread.
2059 static bool migrate_task(struct task_struct *p, int dest_cpu)
2061 struct rq *rq = task_rq(p);
2064 * If the task is not on a runqueue (and not running), then
2065 * the next wake-up will properly place the task.
2067 return p->se.on_rq || task_running(rq, p);
2071 * wait_task_inactive - wait for a thread to unschedule.
2073 * If @match_state is nonzero, it's the @p->state value just checked and
2074 * not expected to change. If it changes, i.e. @p might have woken up,
2075 * then return zero. When we succeed in waiting for @p to be off its CPU,
2076 * we return a positive number (its total switch count). If a second call
2077 * a short while later returns the same number, the caller can be sure that
2078 * @p has remained unscheduled the whole time.
2080 * The caller must ensure that the task *will* unschedule sometime soon,
2081 * else this function might spin for a *long* time. This function can't
2082 * be called with interrupts off, or it may introduce deadlock with
2083 * smp_call_function() if an IPI is sent by the same process we are
2084 * waiting to become inactive.
2086 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2088 unsigned long flags;
2095 * We do the initial early heuristics without holding
2096 * any task-queue locks at all. We'll only try to get
2097 * the runqueue lock when things look like they will
2103 * If the task is actively running on another CPU
2104 * still, just relax and busy-wait without holding
2107 * NOTE! Since we don't hold any locks, it's not
2108 * even sure that "rq" stays as the right runqueue!
2109 * But we don't care, since "task_running()" will
2110 * return false if the runqueue has changed and p
2111 * is actually now running somewhere else!
2113 while (task_running(rq, p)) {
2114 if (match_state && unlikely(p->state != match_state))
2120 * Ok, time to look more closely! We need the rq
2121 * lock now, to be *sure*. If we're wrong, we'll
2122 * just go back and repeat.
2124 rq = task_rq_lock(p, &flags);
2125 trace_sched_wait_task(p);
2126 running = task_running(rq, p);
2127 on_rq = p->se.on_rq;
2129 if (!match_state || p->state == match_state)
2130 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2131 task_rq_unlock(rq, &flags);
2134 * If it changed from the expected state, bail out now.
2136 if (unlikely(!ncsw))
2140 * Was it really running after all now that we
2141 * checked with the proper locks actually held?
2143 * Oops. Go back and try again..
2145 if (unlikely(running)) {
2151 * It's not enough that it's not actively running,
2152 * it must be off the runqueue _entirely_, and not
2155 * So if it was still runnable (but just not actively
2156 * running right now), it's preempted, and we should
2157 * yield - it could be a while.
2159 if (unlikely(on_rq)) {
2160 schedule_timeout_uninterruptible(1);
2165 * Ahh, all good. It wasn't running, and it wasn't
2166 * runnable, which means that it will never become
2167 * running in the future either. We're all done!
2176 * kick_process - kick a running thread to enter/exit the kernel
2177 * @p: the to-be-kicked thread
2179 * Cause a process which is running on another CPU to enter
2180 * kernel-mode, without any delay. (to get signals handled.)
2182 * NOTE: this function doesnt have to take the runqueue lock,
2183 * because all it wants to ensure is that the remote task enters
2184 * the kernel. If the IPI races and the task has been migrated
2185 * to another CPU then no harm is done and the purpose has been
2188 void kick_process(struct task_struct *p)
2194 if ((cpu != smp_processor_id()) && task_curr(p))
2195 smp_send_reschedule(cpu);
2198 EXPORT_SYMBOL_GPL(kick_process);
2199 #endif /* CONFIG_SMP */
2202 * task_oncpu_function_call - call a function on the cpu on which a task runs
2203 * @p: the task to evaluate
2204 * @func: the function to be called
2205 * @info: the function call argument
2207 * Calls the function @func when the task is currently running. This might
2208 * be on the current CPU, which just calls the function directly
2210 void task_oncpu_function_call(struct task_struct *p,
2211 void (*func) (void *info), void *info)
2218 smp_call_function_single(cpu, func, info, 1);
2224 * ->cpus_allowed is protected by either TASK_WAKING or rq->lock held.
2226 static int select_fallback_rq(int cpu, struct task_struct *p)
2229 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2231 /* Look for allowed, online CPU in same node. */
2232 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2233 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2236 /* Any allowed, online CPU? */
2237 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2238 if (dest_cpu < nr_cpu_ids)
2241 /* No more Mr. Nice Guy. */
2242 if (unlikely(dest_cpu >= nr_cpu_ids)) {
2243 dest_cpu = cpuset_cpus_allowed_fallback(p);
2245 * Don't tell them about moving exiting tasks or
2246 * kernel threads (both mm NULL), since they never
2249 if (p->mm && printk_ratelimit()) {
2250 printk(KERN_INFO "process %d (%s) no "
2251 "longer affine to cpu%d\n",
2252 task_pid_nr(p), p->comm, cpu);
2260 * The caller (fork, wakeup) owns TASK_WAKING, ->cpus_allowed is stable.
2263 int select_task_rq(struct rq *rq, struct task_struct *p, int sd_flags, int wake_flags)
2265 int cpu = p->sched_class->select_task_rq(rq, p, sd_flags, wake_flags);
2268 * In order not to call set_task_cpu() on a blocking task we need
2269 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2272 * Since this is common to all placement strategies, this lives here.
2274 * [ this allows ->select_task() to simply return task_cpu(p) and
2275 * not worry about this generic constraint ]
2277 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2279 cpu = select_fallback_rq(task_cpu(p), p);
2284 static void update_avg(u64 *avg, u64 sample)
2286 s64 diff = sample - *avg;
2291 static inline void ttwu_activate(struct task_struct *p, struct rq *rq,
2292 bool is_sync, bool is_migrate, bool is_local,
2293 unsigned long en_flags)
2295 schedstat_inc(p, se.statistics.nr_wakeups);
2297 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2299 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2301 schedstat_inc(p, se.statistics.nr_wakeups_local);
2303 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2305 activate_task(rq, p, en_flags);
2308 static inline void ttwu_post_activation(struct task_struct *p, struct rq *rq,
2309 int wake_flags, bool success)
2311 trace_sched_wakeup(p, success);
2312 check_preempt_curr(rq, p, wake_flags);
2314 p->state = TASK_RUNNING;
2316 if (p->sched_class->task_woken)
2317 p->sched_class->task_woken(rq, p);
2319 if (unlikely(rq->idle_stamp)) {
2320 u64 delta = rq->clock - rq->idle_stamp;
2321 u64 max = 2*sysctl_sched_migration_cost;
2326 update_avg(&rq->avg_idle, delta);
2330 /* if a worker is waking up, notify workqueue */
2331 if ((p->flags & PF_WQ_WORKER) && success)
2332 wq_worker_waking_up(p, cpu_of(rq));
2336 * try_to_wake_up - wake up a thread
2337 * @p: the thread to be awakened
2338 * @state: the mask of task states that can be woken
2339 * @wake_flags: wake modifier flags (WF_*)
2341 * Put it on the run-queue if it's not already there. The "current"
2342 * thread is always on the run-queue (except when the actual
2343 * re-schedule is in progress), and as such you're allowed to do
2344 * the simpler "current->state = TASK_RUNNING" to mark yourself
2345 * runnable without the overhead of this.
2347 * Returns %true if @p was woken up, %false if it was already running
2348 * or @state didn't match @p's state.
2350 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2353 int cpu, orig_cpu, this_cpu, success = 0;
2354 unsigned long flags;
2355 unsigned long en_flags = ENQUEUE_WAKEUP;
2358 this_cpu = get_cpu();
2361 rq = task_rq_lock(p, &flags);
2362 if (!(p->state & state))
2372 if (unlikely(task_running(rq, p)))
2376 * In order to handle concurrent wakeups and release the rq->lock
2377 * we put the task in TASK_WAKING state.
2379 * First fix up the nr_uninterruptible count:
2381 if (task_contributes_to_load(p)) {
2382 if (likely(cpu_online(orig_cpu)))
2383 rq->nr_uninterruptible--;
2385 this_rq()->nr_uninterruptible--;
2387 p->state = TASK_WAKING;
2389 if (p->sched_class->task_waking) {
2390 p->sched_class->task_waking(rq, p);
2391 en_flags |= ENQUEUE_WAKING;
2394 cpu = select_task_rq(rq, p, SD_BALANCE_WAKE, wake_flags);
2395 if (cpu != orig_cpu)
2396 set_task_cpu(p, cpu);
2397 __task_rq_unlock(rq);
2400 raw_spin_lock(&rq->lock);
2403 * We migrated the task without holding either rq->lock, however
2404 * since the task is not on the task list itself, nobody else
2405 * will try and migrate the task, hence the rq should match the
2406 * cpu we just moved it to.
2408 WARN_ON(task_cpu(p) != cpu);
2409 WARN_ON(p->state != TASK_WAKING);
2411 #ifdef CONFIG_SCHEDSTATS
2412 schedstat_inc(rq, ttwu_count);
2413 if (cpu == this_cpu)
2414 schedstat_inc(rq, ttwu_local);
2416 struct sched_domain *sd;
2417 for_each_domain(this_cpu, sd) {
2418 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2419 schedstat_inc(sd, ttwu_wake_remote);
2424 #endif /* CONFIG_SCHEDSTATS */
2427 #endif /* CONFIG_SMP */
2428 ttwu_activate(p, rq, wake_flags & WF_SYNC, orig_cpu != cpu,
2429 cpu == this_cpu, en_flags);
2432 ttwu_post_activation(p, rq, wake_flags, success);
2434 task_rq_unlock(rq, &flags);
2441 * try_to_wake_up_local - try to wake up a local task with rq lock held
2442 * @p: the thread to be awakened
2444 * Put @p on the run-queue if it's not alredy there. The caller must
2445 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2446 * the current task. this_rq() stays locked over invocation.
2448 static void try_to_wake_up_local(struct task_struct *p)
2450 struct rq *rq = task_rq(p);
2451 bool success = false;
2453 BUG_ON(rq != this_rq());
2454 BUG_ON(p == current);
2455 lockdep_assert_held(&rq->lock);
2457 if (!(p->state & TASK_NORMAL))
2461 if (likely(!task_running(rq, p))) {
2462 schedstat_inc(rq, ttwu_count);
2463 schedstat_inc(rq, ttwu_local);
2465 ttwu_activate(p, rq, false, false, true, ENQUEUE_WAKEUP);
2468 ttwu_post_activation(p, rq, 0, success);
2472 * wake_up_process - Wake up a specific process
2473 * @p: The process to be woken up.
2475 * Attempt to wake up the nominated process and move it to the set of runnable
2476 * processes. Returns 1 if the process was woken up, 0 if it was already
2479 * It may be assumed that this function implies a write memory barrier before
2480 * changing the task state if and only if any tasks are woken up.
2482 int wake_up_process(struct task_struct *p)
2484 return try_to_wake_up(p, TASK_ALL, 0);
2486 EXPORT_SYMBOL(wake_up_process);
2488 int wake_up_state(struct task_struct *p, unsigned int state)
2490 return try_to_wake_up(p, state, 0);
2494 * Perform scheduler related setup for a newly forked process p.
2495 * p is forked by current.
2497 * __sched_fork() is basic setup used by init_idle() too:
2499 static void __sched_fork(struct task_struct *p)
2501 p->se.exec_start = 0;
2502 p->se.sum_exec_runtime = 0;
2503 p->se.prev_sum_exec_runtime = 0;
2504 p->se.nr_migrations = 0;
2506 #ifdef CONFIG_SCHEDSTATS
2507 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2510 INIT_LIST_HEAD(&p->rt.run_list);
2512 INIT_LIST_HEAD(&p->se.group_node);
2514 #ifdef CONFIG_PREEMPT_NOTIFIERS
2515 INIT_HLIST_HEAD(&p->preempt_notifiers);
2520 * fork()/clone()-time setup:
2522 void sched_fork(struct task_struct *p, int clone_flags)
2524 int cpu = get_cpu();
2528 * We mark the process as running here. This guarantees that
2529 * nobody will actually run it, and a signal or other external
2530 * event cannot wake it up and insert it on the runqueue either.
2532 p->state = TASK_RUNNING;
2535 * Revert to default priority/policy on fork if requested.
2537 if (unlikely(p->sched_reset_on_fork)) {
2538 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2539 p->policy = SCHED_NORMAL;
2540 p->normal_prio = p->static_prio;
2543 if (PRIO_TO_NICE(p->static_prio) < 0) {
2544 p->static_prio = NICE_TO_PRIO(0);
2545 p->normal_prio = p->static_prio;
2550 * We don't need the reset flag anymore after the fork. It has
2551 * fulfilled its duty:
2553 p->sched_reset_on_fork = 0;
2557 * Make sure we do not leak PI boosting priority to the child.
2559 p->prio = current->normal_prio;
2561 if (!rt_prio(p->prio))
2562 p->sched_class = &fair_sched_class;
2564 if (p->sched_class->task_fork)
2565 p->sched_class->task_fork(p);
2567 set_task_cpu(p, cpu);
2569 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2570 if (likely(sched_info_on()))
2571 memset(&p->sched_info, 0, sizeof(p->sched_info));
2573 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2576 #ifdef CONFIG_PREEMPT
2577 /* Want to start with kernel preemption disabled. */
2578 task_thread_info(p)->preempt_count = 1;
2580 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2586 * wake_up_new_task - wake up a newly created task for the first time.
2588 * This function will do some initial scheduler statistics housekeeping
2589 * that must be done for every newly created context, then puts the task
2590 * on the runqueue and wakes it.
2592 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2594 unsigned long flags;
2596 int cpu __maybe_unused = get_cpu();
2599 rq = task_rq_lock(p, &flags);
2600 p->state = TASK_WAKING;
2603 * Fork balancing, do it here and not earlier because:
2604 * - cpus_allowed can change in the fork path
2605 * - any previously selected cpu might disappear through hotplug
2607 * We set TASK_WAKING so that select_task_rq() can drop rq->lock
2608 * without people poking at ->cpus_allowed.
2610 cpu = select_task_rq(rq, p, SD_BALANCE_FORK, 0);
2611 set_task_cpu(p, cpu);
2613 p->state = TASK_RUNNING;
2614 task_rq_unlock(rq, &flags);
2617 rq = task_rq_lock(p, &flags);
2618 activate_task(rq, p, 0);
2619 trace_sched_wakeup_new(p, 1);
2620 check_preempt_curr(rq, p, WF_FORK);
2622 if (p->sched_class->task_woken)
2623 p->sched_class->task_woken(rq, p);
2625 task_rq_unlock(rq, &flags);
2629 #ifdef CONFIG_PREEMPT_NOTIFIERS
2632 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2633 * @notifier: notifier struct to register
2635 void preempt_notifier_register(struct preempt_notifier *notifier)
2637 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2639 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2642 * preempt_notifier_unregister - no longer interested in preemption notifications
2643 * @notifier: notifier struct to unregister
2645 * This is safe to call from within a preemption notifier.
2647 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2649 hlist_del(¬ifier->link);
2651 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2653 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2655 struct preempt_notifier *notifier;
2656 struct hlist_node *node;
2658 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2659 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2663 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2664 struct task_struct *next)
2666 struct preempt_notifier *notifier;
2667 struct hlist_node *node;
2669 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2670 notifier->ops->sched_out(notifier, next);
2673 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2675 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2680 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2681 struct task_struct *next)
2685 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2688 * prepare_task_switch - prepare to switch tasks
2689 * @rq: the runqueue preparing to switch
2690 * @prev: the current task that is being switched out
2691 * @next: the task we are going to switch to.
2693 * This is called with the rq lock held and interrupts off. It must
2694 * be paired with a subsequent finish_task_switch after the context
2697 * prepare_task_switch sets up locking and calls architecture specific
2701 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2702 struct task_struct *next)
2704 fire_sched_out_preempt_notifiers(prev, next);
2705 prepare_lock_switch(rq, next);
2706 prepare_arch_switch(next);
2710 * finish_task_switch - clean up after a task-switch
2711 * @rq: runqueue associated with task-switch
2712 * @prev: the thread we just switched away from.
2714 * finish_task_switch must be called after the context switch, paired
2715 * with a prepare_task_switch call before the context switch.
2716 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2717 * and do any other architecture-specific cleanup actions.
2719 * Note that we may have delayed dropping an mm in context_switch(). If
2720 * so, we finish that here outside of the runqueue lock. (Doing it
2721 * with the lock held can cause deadlocks; see schedule() for
2724 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2725 __releases(rq->lock)
2727 struct mm_struct *mm = rq->prev_mm;
2733 * A task struct has one reference for the use as "current".
2734 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2735 * schedule one last time. The schedule call will never return, and
2736 * the scheduled task must drop that reference.
2737 * The test for TASK_DEAD must occur while the runqueue locks are
2738 * still held, otherwise prev could be scheduled on another cpu, die
2739 * there before we look at prev->state, and then the reference would
2741 * Manfred Spraul <manfred@colorfullife.com>
2743 prev_state = prev->state;
2744 finish_arch_switch(prev);
2745 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2746 local_irq_disable();
2747 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2748 perf_event_task_sched_in(current);
2749 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2751 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2752 finish_lock_switch(rq, prev);
2754 fire_sched_in_preempt_notifiers(current);
2757 if (unlikely(prev_state == TASK_DEAD)) {
2759 * Remove function-return probe instances associated with this
2760 * task and put them back on the free list.
2762 kprobe_flush_task(prev);
2763 put_task_struct(prev);
2769 /* assumes rq->lock is held */
2770 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2772 if (prev->sched_class->pre_schedule)
2773 prev->sched_class->pre_schedule(rq, prev);
2776 /* rq->lock is NOT held, but preemption is disabled */
2777 static inline void post_schedule(struct rq *rq)
2779 if (rq->post_schedule) {
2780 unsigned long flags;
2782 raw_spin_lock_irqsave(&rq->lock, flags);
2783 if (rq->curr->sched_class->post_schedule)
2784 rq->curr->sched_class->post_schedule(rq);
2785 raw_spin_unlock_irqrestore(&rq->lock, flags);
2787 rq->post_schedule = 0;
2793 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2797 static inline void post_schedule(struct rq *rq)
2804 * schedule_tail - first thing a freshly forked thread must call.
2805 * @prev: the thread we just switched away from.
2807 asmlinkage void schedule_tail(struct task_struct *prev)
2808 __releases(rq->lock)
2810 struct rq *rq = this_rq();
2812 finish_task_switch(rq, prev);
2815 * FIXME: do we need to worry about rq being invalidated by the
2820 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2821 /* In this case, finish_task_switch does not reenable preemption */
2824 if (current->set_child_tid)
2825 put_user(task_pid_vnr(current), current->set_child_tid);
2829 * context_switch - switch to the new MM and the new
2830 * thread's register state.
2833 context_switch(struct rq *rq, struct task_struct *prev,
2834 struct task_struct *next)
2836 struct mm_struct *mm, *oldmm;
2838 prepare_task_switch(rq, prev, next);
2839 trace_sched_switch(prev, next);
2841 oldmm = prev->active_mm;
2843 * For paravirt, this is coupled with an exit in switch_to to
2844 * combine the page table reload and the switch backend into
2847 arch_start_context_switch(prev);
2850 next->active_mm = oldmm;
2851 atomic_inc(&oldmm->mm_count);
2852 enter_lazy_tlb(oldmm, next);
2854 switch_mm(oldmm, mm, next);
2856 if (likely(!prev->mm)) {
2857 prev->active_mm = NULL;
2858 rq->prev_mm = oldmm;
2861 * Since the runqueue lock will be released by the next
2862 * task (which is an invalid locking op but in the case
2863 * of the scheduler it's an obvious special-case), so we
2864 * do an early lockdep release here:
2866 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2867 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2870 /* Here we just switch the register state and the stack. */
2871 switch_to(prev, next, prev);
2875 * this_rq must be evaluated again because prev may have moved
2876 * CPUs since it called schedule(), thus the 'rq' on its stack
2877 * frame will be invalid.
2879 finish_task_switch(this_rq(), prev);
2883 * nr_running, nr_uninterruptible and nr_context_switches:
2885 * externally visible scheduler statistics: current number of runnable
2886 * threads, current number of uninterruptible-sleeping threads, total
2887 * number of context switches performed since bootup.
2889 unsigned long nr_running(void)
2891 unsigned long i, sum = 0;
2893 for_each_online_cpu(i)
2894 sum += cpu_rq(i)->nr_running;
2899 unsigned long nr_uninterruptible(void)
2901 unsigned long i, sum = 0;
2903 for_each_possible_cpu(i)
2904 sum += cpu_rq(i)->nr_uninterruptible;
2907 * Since we read the counters lockless, it might be slightly
2908 * inaccurate. Do not allow it to go below zero though:
2910 if (unlikely((long)sum < 0))
2916 unsigned long long nr_context_switches(void)
2919 unsigned long long sum = 0;
2921 for_each_possible_cpu(i)
2922 sum += cpu_rq(i)->nr_switches;
2927 unsigned long nr_iowait(void)
2929 unsigned long i, sum = 0;
2931 for_each_possible_cpu(i)
2932 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2937 unsigned long nr_iowait_cpu(void)
2939 struct rq *this = this_rq();
2940 return atomic_read(&this->nr_iowait);
2943 unsigned long this_cpu_load(void)
2945 struct rq *this = this_rq();
2946 return this->cpu_load[0];
2950 /* Variables and functions for calc_load */
2951 static atomic_long_t calc_load_tasks;
2952 static unsigned long calc_load_update;
2953 unsigned long avenrun[3];
2954 EXPORT_SYMBOL(avenrun);
2956 static long calc_load_fold_active(struct rq *this_rq)
2958 long nr_active, delta = 0;
2960 nr_active = this_rq->nr_running;
2961 nr_active += (long) this_rq->nr_uninterruptible;
2963 if (nr_active != this_rq->calc_load_active) {
2964 delta = nr_active - this_rq->calc_load_active;
2965 this_rq->calc_load_active = nr_active;
2973 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
2975 * When making the ILB scale, we should try to pull this in as well.
2977 static atomic_long_t calc_load_tasks_idle;
2979 static void calc_load_account_idle(struct rq *this_rq)
2983 delta = calc_load_fold_active(this_rq);
2985 atomic_long_add(delta, &calc_load_tasks_idle);
2988 static long calc_load_fold_idle(void)
2993 * Its got a race, we don't care...
2995 if (atomic_long_read(&calc_load_tasks_idle))
2996 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
3001 static void calc_load_account_idle(struct rq *this_rq)
3005 static inline long calc_load_fold_idle(void)
3012 * get_avenrun - get the load average array
3013 * @loads: pointer to dest load array
3014 * @offset: offset to add
3015 * @shift: shift count to shift the result left
3017 * These values are estimates at best, so no need for locking.
3019 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3021 loads[0] = (avenrun[0] + offset) << shift;
3022 loads[1] = (avenrun[1] + offset) << shift;
3023 loads[2] = (avenrun[2] + offset) << shift;
3026 static unsigned long
3027 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3030 load += active * (FIXED_1 - exp);
3031 return load >> FSHIFT;
3035 * calc_load - update the avenrun load estimates 10 ticks after the
3036 * CPUs have updated calc_load_tasks.
3038 void calc_global_load(void)
3040 unsigned long upd = calc_load_update + 10;
3043 if (time_before(jiffies, upd))
3046 active = atomic_long_read(&calc_load_tasks);
3047 active = active > 0 ? active * FIXED_1 : 0;
3049 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3050 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3051 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3053 calc_load_update += LOAD_FREQ;
3057 * Called from update_cpu_load() to periodically update this CPU's
3060 static void calc_load_account_active(struct rq *this_rq)
3064 if (time_before(jiffies, this_rq->calc_load_update))
3067 delta = calc_load_fold_active(this_rq);
3068 delta += calc_load_fold_idle();
3070 atomic_long_add(delta, &calc_load_tasks);
3072 this_rq->calc_load_update += LOAD_FREQ;
3076 * The exact cpuload at various idx values, calculated at every tick would be
3077 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3079 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3080 * on nth tick when cpu may be busy, then we have:
3081 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3082 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3084 * decay_load_missed() below does efficient calculation of
3085 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3086 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3088 * The calculation is approximated on a 128 point scale.
3089 * degrade_zero_ticks is the number of ticks after which load at any
3090 * particular idx is approximated to be zero.
3091 * degrade_factor is a precomputed table, a row for each load idx.
3092 * Each column corresponds to degradation factor for a power of two ticks,
3093 * based on 128 point scale.
3095 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3096 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3098 * With this power of 2 load factors, we can degrade the load n times
3099 * by looking at 1 bits in n and doing as many mult/shift instead of
3100 * n mult/shifts needed by the exact degradation.
3102 #define DEGRADE_SHIFT 7
3103 static const unsigned char
3104 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3105 static const unsigned char
3106 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3107 {0, 0, 0, 0, 0, 0, 0, 0},
3108 {64, 32, 8, 0, 0, 0, 0, 0},
3109 {96, 72, 40, 12, 1, 0, 0},
3110 {112, 98, 75, 43, 15, 1, 0},
3111 {120, 112, 98, 76, 45, 16, 2} };
3114 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3115 * would be when CPU is idle and so we just decay the old load without
3116 * adding any new load.
3118 static unsigned long
3119 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3123 if (!missed_updates)
3126 if (missed_updates >= degrade_zero_ticks[idx])
3130 return load >> missed_updates;
3132 while (missed_updates) {
3133 if (missed_updates % 2)
3134 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3136 missed_updates >>= 1;
3143 * Update rq->cpu_load[] statistics. This function is usually called every
3144 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3145 * every tick. We fix it up based on jiffies.
3147 static void update_cpu_load(struct rq *this_rq)
3149 unsigned long this_load = this_rq->load.weight;
3150 unsigned long curr_jiffies = jiffies;
3151 unsigned long pending_updates;
3154 this_rq->nr_load_updates++;
3156 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3157 if (curr_jiffies == this_rq->last_load_update_tick)
3160 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3161 this_rq->last_load_update_tick = curr_jiffies;
3163 /* Update our load: */
3164 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3165 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3166 unsigned long old_load, new_load;
3168 /* scale is effectively 1 << i now, and >> i divides by scale */
3170 old_load = this_rq->cpu_load[i];
3171 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3172 new_load = this_load;
3174 * Round up the averaging division if load is increasing. This
3175 * prevents us from getting stuck on 9 if the load is 10, for
3178 if (new_load > old_load)
3179 new_load += scale - 1;
3181 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3185 static void update_cpu_load_active(struct rq *this_rq)
3187 update_cpu_load(this_rq);
3189 calc_load_account_active(this_rq);
3195 * sched_exec - execve() is a valuable balancing opportunity, because at
3196 * this point the task has the smallest effective memory and cache footprint.
3198 void sched_exec(void)
3200 struct task_struct *p = current;
3201 unsigned long flags;
3205 rq = task_rq_lock(p, &flags);
3206 dest_cpu = p->sched_class->select_task_rq(rq, p, SD_BALANCE_EXEC, 0);
3207 if (dest_cpu == smp_processor_id())
3211 * select_task_rq() can race against ->cpus_allowed
3213 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed) &&
3214 likely(cpu_active(dest_cpu)) && migrate_task(p, dest_cpu)) {
3215 struct migration_arg arg = { p, dest_cpu };
3217 task_rq_unlock(rq, &flags);
3218 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
3222 task_rq_unlock(rq, &flags);
3227 DEFINE_PER_CPU(struct kernel_stat, kstat);
3229 EXPORT_PER_CPU_SYMBOL(kstat);
3232 * Return any ns on the sched_clock that have not yet been accounted in
3233 * @p in case that task is currently running.
3235 * Called with task_rq_lock() held on @rq.
3237 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3241 if (task_current(rq, p)) {
3242 update_rq_clock(rq);
3243 ns = rq->clock - p->se.exec_start;
3251 unsigned long long task_delta_exec(struct task_struct *p)
3253 unsigned long flags;
3257 rq = task_rq_lock(p, &flags);
3258 ns = do_task_delta_exec(p, rq);
3259 task_rq_unlock(rq, &flags);
3265 * Return accounted runtime for the task.
3266 * In case the task is currently running, return the runtime plus current's
3267 * pending runtime that have not been accounted yet.
3269 unsigned long long task_sched_runtime(struct task_struct *p)
3271 unsigned long flags;
3275 rq = task_rq_lock(p, &flags);
3276 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3277 task_rq_unlock(rq, &flags);
3283 * Return sum_exec_runtime for the thread group.
3284 * In case the task is currently running, return the sum plus current's
3285 * pending runtime that have not been accounted yet.
3287 * Note that the thread group might have other running tasks as well,
3288 * so the return value not includes other pending runtime that other
3289 * running tasks might have.
3291 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3293 struct task_cputime totals;
3294 unsigned long flags;
3298 rq = task_rq_lock(p, &flags);
3299 thread_group_cputime(p, &totals);
3300 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3301 task_rq_unlock(rq, &flags);
3307 * Account user cpu time to a process.
3308 * @p: the process that the cpu time gets accounted to
3309 * @cputime: the cpu time spent in user space since the last update
3310 * @cputime_scaled: cputime scaled by cpu frequency
3312 void account_user_time(struct task_struct *p, cputime_t cputime,
3313 cputime_t cputime_scaled)
3315 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3318 /* Add user time to process. */
3319 p->utime = cputime_add(p->utime, cputime);
3320 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3321 account_group_user_time(p, cputime);
3323 /* Add user time to cpustat. */
3324 tmp = cputime_to_cputime64(cputime);
3325 if (TASK_NICE(p) > 0)
3326 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3328 cpustat->user = cputime64_add(cpustat->user, tmp);
3330 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3331 /* Account for user time used */
3332 acct_update_integrals(p);
3336 * Account guest cpu time to a process.
3337 * @p: the process that the cpu time gets accounted to
3338 * @cputime: the cpu time spent in virtual machine since the last update
3339 * @cputime_scaled: cputime scaled by cpu frequency
3341 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3342 cputime_t cputime_scaled)
3345 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3347 tmp = cputime_to_cputime64(cputime);
3349 /* Add guest time to process. */
3350 p->utime = cputime_add(p->utime, cputime);
3351 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3352 account_group_user_time(p, cputime);
3353 p->gtime = cputime_add(p->gtime, cputime);
3355 /* Add guest time to cpustat. */
3356 if (TASK_NICE(p) > 0) {
3357 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3358 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3360 cpustat->user = cputime64_add(cpustat->user, tmp);
3361 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3366 * Account system cpu time to a process.
3367 * @p: the process that the cpu time gets accounted to
3368 * @hardirq_offset: the offset to subtract from hardirq_count()
3369 * @cputime: the cpu time spent in kernel space since the last update
3370 * @cputime_scaled: cputime scaled by cpu frequency
3372 void account_system_time(struct task_struct *p, int hardirq_offset,
3373 cputime_t cputime, cputime_t cputime_scaled)
3375 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3378 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3379 account_guest_time(p, cputime, cputime_scaled);
3383 /* Add system time to process. */
3384 p->stime = cputime_add(p->stime, cputime);
3385 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3386 account_group_system_time(p, cputime);
3388 /* Add system time to cpustat. */
3389 tmp = cputime_to_cputime64(cputime);
3390 if (hardirq_count() - hardirq_offset)
3391 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3392 else if (softirq_count())
3393 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3395 cpustat->system = cputime64_add(cpustat->system, tmp);
3397 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3399 /* Account for system time used */
3400 acct_update_integrals(p);
3404 * Account for involuntary wait time.
3405 * @steal: the cpu time spent in involuntary wait
3407 void account_steal_time(cputime_t cputime)
3409 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3410 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3412 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3416 * Account for idle time.
3417 * @cputime: the cpu time spent in idle wait
3419 void account_idle_time(cputime_t cputime)
3421 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3422 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3423 struct rq *rq = this_rq();
3425 if (atomic_read(&rq->nr_iowait) > 0)
3426 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3428 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3431 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3434 * Account a single tick of cpu time.
3435 * @p: the process that the cpu time gets accounted to
3436 * @user_tick: indicates if the tick is a user or a system tick
3438 void account_process_tick(struct task_struct *p, int user_tick)
3440 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3441 struct rq *rq = this_rq();
3444 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3445 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3446 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3449 account_idle_time(cputime_one_jiffy);
3453 * Account multiple ticks of steal time.
3454 * @p: the process from which the cpu time has been stolen
3455 * @ticks: number of stolen ticks
3457 void account_steal_ticks(unsigned long ticks)
3459 account_steal_time(jiffies_to_cputime(ticks));
3463 * Account multiple ticks of idle time.
3464 * @ticks: number of stolen ticks
3466 void account_idle_ticks(unsigned long ticks)
3468 account_idle_time(jiffies_to_cputime(ticks));
3474 * Use precise platform statistics if available:
3476 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3477 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3483 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3485 struct task_cputime cputime;
3487 thread_group_cputime(p, &cputime);
3489 *ut = cputime.utime;
3490 *st = cputime.stime;
3494 #ifndef nsecs_to_cputime
3495 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3498 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3500 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
3503 * Use CFS's precise accounting:
3505 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3510 temp = (u64)(rtime * utime);
3511 do_div(temp, total);
3512 utime = (cputime_t)temp;
3517 * Compare with previous values, to keep monotonicity:
3519 p->prev_utime = max(p->prev_utime, utime);
3520 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
3522 *ut = p->prev_utime;
3523 *st = p->prev_stime;
3527 * Must be called with siglock held.
3529 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3531 struct signal_struct *sig = p->signal;
3532 struct task_cputime cputime;
3533 cputime_t rtime, utime, total;
3535 thread_group_cputime(p, &cputime);
3537 total = cputime_add(cputime.utime, cputime.stime);
3538 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3543 temp = (u64)(rtime * cputime.utime);
3544 do_div(temp, total);
3545 utime = (cputime_t)temp;
3549 sig->prev_utime = max(sig->prev_utime, utime);
3550 sig->prev_stime = max(sig->prev_stime,
3551 cputime_sub(rtime, sig->prev_utime));
3553 *ut = sig->prev_utime;
3554 *st = sig->prev_stime;
3559 * This function gets called by the timer code, with HZ frequency.
3560 * We call it with interrupts disabled.
3562 * It also gets called by the fork code, when changing the parent's
3565 void scheduler_tick(void)
3567 int cpu = smp_processor_id();
3568 struct rq *rq = cpu_rq(cpu);
3569 struct task_struct *curr = rq->curr;
3573 raw_spin_lock(&rq->lock);
3574 update_rq_clock(rq);
3575 update_cpu_load_active(rq);
3576 curr->sched_class->task_tick(rq, curr, 0);
3577 raw_spin_unlock(&rq->lock);
3579 perf_event_task_tick(curr);
3582 rq->idle_at_tick = idle_cpu(cpu);
3583 trigger_load_balance(rq, cpu);
3587 notrace unsigned long get_parent_ip(unsigned long addr)
3589 if (in_lock_functions(addr)) {
3590 addr = CALLER_ADDR2;
3591 if (in_lock_functions(addr))
3592 addr = CALLER_ADDR3;
3597 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3598 defined(CONFIG_PREEMPT_TRACER))
3600 void __kprobes add_preempt_count(int val)
3602 #ifdef CONFIG_DEBUG_PREEMPT
3606 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3609 preempt_count() += val;
3610 #ifdef CONFIG_DEBUG_PREEMPT
3612 * Spinlock count overflowing soon?
3614 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3617 if (preempt_count() == val)
3618 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3620 EXPORT_SYMBOL(add_preempt_count);
3622 void __kprobes sub_preempt_count(int val)
3624 #ifdef CONFIG_DEBUG_PREEMPT
3628 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3631 * Is the spinlock portion underflowing?
3633 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3634 !(preempt_count() & PREEMPT_MASK)))
3638 if (preempt_count() == val)
3639 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3640 preempt_count() -= val;
3642 EXPORT_SYMBOL(sub_preempt_count);
3647 * Print scheduling while atomic bug:
3649 static noinline void __schedule_bug(struct task_struct *prev)
3651 struct pt_regs *regs = get_irq_regs();
3653 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3654 prev->comm, prev->pid, preempt_count());
3656 debug_show_held_locks(prev);
3658 if (irqs_disabled())
3659 print_irqtrace_events(prev);
3668 * Various schedule()-time debugging checks and statistics:
3670 static inline void schedule_debug(struct task_struct *prev)
3673 * Test if we are atomic. Since do_exit() needs to call into
3674 * schedule() atomically, we ignore that path for now.
3675 * Otherwise, whine if we are scheduling when we should not be.
3677 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3678 __schedule_bug(prev);
3680 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3682 schedstat_inc(this_rq(), sched_count);
3683 #ifdef CONFIG_SCHEDSTATS
3684 if (unlikely(prev->lock_depth >= 0)) {
3685 schedstat_inc(this_rq(), bkl_count);
3686 schedstat_inc(prev, sched_info.bkl_count);
3691 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3694 update_rq_clock(rq);
3695 rq->skip_clock_update = 0;
3696 prev->sched_class->put_prev_task(rq, prev);
3700 * Pick up the highest-prio task:
3702 static inline struct task_struct *
3703 pick_next_task(struct rq *rq)
3705 const struct sched_class *class;
3706 struct task_struct *p;
3709 * Optimization: we know that if all tasks are in
3710 * the fair class we can call that function directly:
3712 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3713 p = fair_sched_class.pick_next_task(rq);
3718 class = sched_class_highest;
3720 p = class->pick_next_task(rq);
3724 * Will never be NULL as the idle class always
3725 * returns a non-NULL p:
3727 class = class->next;
3732 * schedule() is the main scheduler function.
3734 asmlinkage void __sched schedule(void)
3736 struct task_struct *prev, *next;
3737 unsigned long *switch_count;
3743 cpu = smp_processor_id();
3745 rcu_note_context_switch(cpu);
3748 release_kernel_lock(prev);
3749 need_resched_nonpreemptible:
3751 schedule_debug(prev);
3753 if (sched_feat(HRTICK))
3756 raw_spin_lock_irq(&rq->lock);
3757 clear_tsk_need_resched(prev);
3759 switch_count = &prev->nivcsw;
3760 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3761 if (unlikely(signal_pending_state(prev->state, prev))) {
3762 prev->state = TASK_RUNNING;
3765 * If a worker is going to sleep, notify and
3766 * ask workqueue whether it wants to wake up a
3767 * task to maintain concurrency. If so, wake
3770 if (prev->flags & PF_WQ_WORKER) {
3771 struct task_struct *to_wakeup;
3773 to_wakeup = wq_worker_sleeping(prev, cpu);
3775 try_to_wake_up_local(to_wakeup);
3777 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3779 switch_count = &prev->nvcsw;
3782 pre_schedule(rq, prev);
3784 if (unlikely(!rq->nr_running))
3785 idle_balance(cpu, rq);
3787 put_prev_task(rq, prev);
3788 next = pick_next_task(rq);
3790 if (likely(prev != next)) {
3791 sched_info_switch(prev, next);
3792 perf_event_task_sched_out(prev, next);
3798 context_switch(rq, prev, next); /* unlocks the rq */
3800 * The context switch have flipped the stack from under us
3801 * and restored the local variables which were saved when
3802 * this task called schedule() in the past. prev == current
3803 * is still correct, but it can be moved to another cpu/rq.
3805 cpu = smp_processor_id();
3808 raw_spin_unlock_irq(&rq->lock);
3812 if (unlikely(reacquire_kernel_lock(prev)))
3813 goto need_resched_nonpreemptible;
3815 preempt_enable_no_resched();
3819 EXPORT_SYMBOL(schedule);
3821 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3823 * Look out! "owner" is an entirely speculative pointer
3824 * access and not reliable.
3826 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
3831 if (!sched_feat(OWNER_SPIN))
3834 #ifdef CONFIG_DEBUG_PAGEALLOC
3836 * Need to access the cpu field knowing that
3837 * DEBUG_PAGEALLOC could have unmapped it if
3838 * the mutex owner just released it and exited.
3840 if (probe_kernel_address(&owner->cpu, cpu))
3847 * Even if the access succeeded (likely case),
3848 * the cpu field may no longer be valid.
3850 if (cpu >= nr_cpumask_bits)
3854 * We need to validate that we can do a
3855 * get_cpu() and that we have the percpu area.
3857 if (!cpu_online(cpu))
3864 * Owner changed, break to re-assess state.
3866 if (lock->owner != owner)
3870 * Is that owner really running on that cpu?
3872 if (task_thread_info(rq->curr) != owner || need_resched())
3882 #ifdef CONFIG_PREEMPT
3884 * this is the entry point to schedule() from in-kernel preemption
3885 * off of preempt_enable. Kernel preemptions off return from interrupt
3886 * occur there and call schedule directly.
3888 asmlinkage void __sched preempt_schedule(void)
3890 struct thread_info *ti = current_thread_info();
3893 * If there is a non-zero preempt_count or interrupts are disabled,
3894 * we do not want to preempt the current task. Just return..
3896 if (likely(ti->preempt_count || irqs_disabled()))
3900 add_preempt_count(PREEMPT_ACTIVE);
3902 sub_preempt_count(PREEMPT_ACTIVE);
3905 * Check again in case we missed a preemption opportunity
3906 * between schedule and now.
3909 } while (need_resched());
3911 EXPORT_SYMBOL(preempt_schedule);
3914 * this is the entry point to schedule() from kernel preemption
3915 * off of irq context.
3916 * Note, that this is called and return with irqs disabled. This will
3917 * protect us against recursive calling from irq.
3919 asmlinkage void __sched preempt_schedule_irq(void)
3921 struct thread_info *ti = current_thread_info();
3923 /* Catch callers which need to be fixed */
3924 BUG_ON(ti->preempt_count || !irqs_disabled());
3927 add_preempt_count(PREEMPT_ACTIVE);
3930 local_irq_disable();
3931 sub_preempt_count(PREEMPT_ACTIVE);
3934 * Check again in case we missed a preemption opportunity
3935 * between schedule and now.
3938 } while (need_resched());
3941 #endif /* CONFIG_PREEMPT */
3943 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3946 return try_to_wake_up(curr->private, mode, wake_flags);
3948 EXPORT_SYMBOL(default_wake_function);
3951 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3952 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3953 * number) then we wake all the non-exclusive tasks and one exclusive task.
3955 * There are circumstances in which we can try to wake a task which has already
3956 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3957 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3959 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3960 int nr_exclusive, int wake_flags, void *key)
3962 wait_queue_t *curr, *next;
3964 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3965 unsigned flags = curr->flags;
3967 if (curr->func(curr, mode, wake_flags, key) &&
3968 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3974 * __wake_up - wake up threads blocked on a waitqueue.
3976 * @mode: which threads
3977 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3978 * @key: is directly passed to the wakeup function
3980 * It may be assumed that this function implies a write memory barrier before
3981 * changing the task state if and only if any tasks are woken up.
3983 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3984 int nr_exclusive, void *key)
3986 unsigned long flags;
3988 spin_lock_irqsave(&q->lock, flags);
3989 __wake_up_common(q, mode, nr_exclusive, 0, key);
3990 spin_unlock_irqrestore(&q->lock, flags);
3992 EXPORT_SYMBOL(__wake_up);
3995 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3997 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3999 __wake_up_common(q, mode, 1, 0, NULL);
4001 EXPORT_SYMBOL_GPL(__wake_up_locked);
4003 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
4005 __wake_up_common(q, mode, 1, 0, key);
4009 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4011 * @mode: which threads
4012 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4013 * @key: opaque value to be passed to wakeup targets
4015 * The sync wakeup differs that the waker knows that it will schedule
4016 * away soon, so while the target thread will be woken up, it will not
4017 * be migrated to another CPU - ie. the two threads are 'synchronized'
4018 * with each other. This can prevent needless bouncing between CPUs.
4020 * On UP it can prevent extra preemption.
4022 * It may be assumed that this function implies a write memory barrier before
4023 * changing the task state if and only if any tasks are woken up.
4025 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4026 int nr_exclusive, void *key)
4028 unsigned long flags;
4029 int wake_flags = WF_SYNC;
4034 if (unlikely(!nr_exclusive))
4037 spin_lock_irqsave(&q->lock, flags);
4038 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4039 spin_unlock_irqrestore(&q->lock, flags);
4041 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4044 * __wake_up_sync - see __wake_up_sync_key()
4046 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4048 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4050 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4053 * complete: - signals a single thread waiting on this completion
4054 * @x: holds the state of this particular completion
4056 * This will wake up a single thread waiting on this completion. Threads will be
4057 * awakened in the same order in which they were queued.
4059 * See also complete_all(), wait_for_completion() and related routines.
4061 * It may be assumed that this function implies a write memory barrier before
4062 * changing the task state if and only if any tasks are woken up.
4064 void complete(struct completion *x)
4066 unsigned long flags;
4068 spin_lock_irqsave(&x->wait.lock, flags);
4070 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4071 spin_unlock_irqrestore(&x->wait.lock, flags);
4073 EXPORT_SYMBOL(complete);
4076 * complete_all: - signals all threads waiting on this completion
4077 * @x: holds the state of this particular completion
4079 * This will wake up all threads waiting on this particular completion event.
4081 * It may be assumed that this function implies a write memory barrier before
4082 * changing the task state if and only if any tasks are woken up.
4084 void complete_all(struct completion *x)
4086 unsigned long flags;
4088 spin_lock_irqsave(&x->wait.lock, flags);
4089 x->done += UINT_MAX/2;
4090 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4091 spin_unlock_irqrestore(&x->wait.lock, flags);
4093 EXPORT_SYMBOL(complete_all);
4095 static inline long __sched
4096 do_wait_for_common(struct completion *x, long timeout, int state)
4099 DECLARE_WAITQUEUE(wait, current);
4101 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4103 if (signal_pending_state(state, current)) {
4104 timeout = -ERESTARTSYS;
4107 __set_current_state(state);
4108 spin_unlock_irq(&x->wait.lock);
4109 timeout = schedule_timeout(timeout);
4110 spin_lock_irq(&x->wait.lock);
4111 } while (!x->done && timeout);
4112 __remove_wait_queue(&x->wait, &wait);
4117 return timeout ?: 1;
4121 wait_for_common(struct completion *x, long timeout, int state)
4125 spin_lock_irq(&x->wait.lock);
4126 timeout = do_wait_for_common(x, timeout, state);
4127 spin_unlock_irq(&x->wait.lock);
4132 * wait_for_completion: - waits for completion of a task
4133 * @x: holds the state of this particular completion
4135 * This waits to be signaled for completion of a specific task. It is NOT
4136 * interruptible and there is no timeout.
4138 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4139 * and interrupt capability. Also see complete().
4141 void __sched wait_for_completion(struct completion *x)
4143 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4145 EXPORT_SYMBOL(wait_for_completion);
4148 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4149 * @x: holds the state of this particular completion
4150 * @timeout: timeout value in jiffies
4152 * This waits for either a completion of a specific task to be signaled or for a
4153 * specified timeout to expire. The timeout is in jiffies. It is not
4156 unsigned long __sched
4157 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4159 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4161 EXPORT_SYMBOL(wait_for_completion_timeout);
4164 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4165 * @x: holds the state of this particular completion
4167 * This waits for completion of a specific task to be signaled. It is
4170 int __sched wait_for_completion_interruptible(struct completion *x)
4172 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4173 if (t == -ERESTARTSYS)
4177 EXPORT_SYMBOL(wait_for_completion_interruptible);
4180 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4181 * @x: holds the state of this particular completion
4182 * @timeout: timeout value in jiffies
4184 * This waits for either a completion of a specific task to be signaled or for a
4185 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4187 unsigned long __sched
4188 wait_for_completion_interruptible_timeout(struct completion *x,
4189 unsigned long timeout)
4191 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4193 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4196 * wait_for_completion_killable: - waits for completion of a task (killable)
4197 * @x: holds the state of this particular completion
4199 * This waits to be signaled for completion of a specific task. It can be
4200 * interrupted by a kill signal.
4202 int __sched wait_for_completion_killable(struct completion *x)
4204 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4205 if (t == -ERESTARTSYS)
4209 EXPORT_SYMBOL(wait_for_completion_killable);
4212 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4213 * @x: holds the state of this particular completion
4214 * @timeout: timeout value in jiffies
4216 * This waits for either a completion of a specific task to be
4217 * signaled or for a specified timeout to expire. It can be
4218 * interrupted by a kill signal. The timeout is in jiffies.
4220 unsigned long __sched
4221 wait_for_completion_killable_timeout(struct completion *x,
4222 unsigned long timeout)
4224 return wait_for_common(x, timeout, TASK_KILLABLE);
4226 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
4229 * try_wait_for_completion - try to decrement a completion without blocking
4230 * @x: completion structure
4232 * Returns: 0 if a decrement cannot be done without blocking
4233 * 1 if a decrement succeeded.
4235 * If a completion is being used as a counting completion,
4236 * attempt to decrement the counter without blocking. This
4237 * enables us to avoid waiting if the resource the completion
4238 * is protecting is not available.
4240 bool try_wait_for_completion(struct completion *x)
4242 unsigned long flags;
4245 spin_lock_irqsave(&x->wait.lock, flags);
4250 spin_unlock_irqrestore(&x->wait.lock, flags);
4253 EXPORT_SYMBOL(try_wait_for_completion);
4256 * completion_done - Test to see if a completion has any waiters
4257 * @x: completion structure
4259 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4260 * 1 if there are no waiters.
4263 bool completion_done(struct completion *x)
4265 unsigned long flags;
4268 spin_lock_irqsave(&x->wait.lock, flags);
4271 spin_unlock_irqrestore(&x->wait.lock, flags);
4274 EXPORT_SYMBOL(completion_done);
4277 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4279 unsigned long flags;
4282 init_waitqueue_entry(&wait, current);
4284 __set_current_state(state);
4286 spin_lock_irqsave(&q->lock, flags);
4287 __add_wait_queue(q, &wait);
4288 spin_unlock(&q->lock);
4289 timeout = schedule_timeout(timeout);
4290 spin_lock_irq(&q->lock);
4291 __remove_wait_queue(q, &wait);
4292 spin_unlock_irqrestore(&q->lock, flags);
4297 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4299 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4301 EXPORT_SYMBOL(interruptible_sleep_on);
4304 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4306 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4308 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4310 void __sched sleep_on(wait_queue_head_t *q)
4312 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4314 EXPORT_SYMBOL(sleep_on);
4316 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4318 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4320 EXPORT_SYMBOL(sleep_on_timeout);
4322 #ifdef CONFIG_RT_MUTEXES
4325 * rt_mutex_setprio - set the current priority of a task
4327 * @prio: prio value (kernel-internal form)
4329 * This function changes the 'effective' priority of a task. It does
4330 * not touch ->normal_prio like __setscheduler().
4332 * Used by the rt_mutex code to implement priority inheritance logic.
4334 void rt_mutex_setprio(struct task_struct *p, int prio)
4336 unsigned long flags;
4337 int oldprio, on_rq, running;
4339 const struct sched_class *prev_class;
4341 BUG_ON(prio < 0 || prio > MAX_PRIO);
4343 rq = task_rq_lock(p, &flags);
4346 prev_class = p->sched_class;
4347 on_rq = p->se.on_rq;
4348 running = task_current(rq, p);
4350 dequeue_task(rq, p, 0);
4352 p->sched_class->put_prev_task(rq, p);
4355 p->sched_class = &rt_sched_class;
4357 p->sched_class = &fair_sched_class;
4362 p->sched_class->set_curr_task(rq);
4364 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4366 check_class_changed(rq, p, prev_class, oldprio, running);
4368 task_rq_unlock(rq, &flags);
4373 void set_user_nice(struct task_struct *p, long nice)
4375 int old_prio, delta, on_rq;
4376 unsigned long flags;
4379 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4382 * We have to be careful, if called from sys_setpriority(),
4383 * the task might be in the middle of scheduling on another CPU.
4385 rq = task_rq_lock(p, &flags);
4387 * The RT priorities are set via sched_setscheduler(), but we still
4388 * allow the 'normal' nice value to be set - but as expected
4389 * it wont have any effect on scheduling until the task is
4390 * SCHED_FIFO/SCHED_RR:
4392 if (task_has_rt_policy(p)) {
4393 p->static_prio = NICE_TO_PRIO(nice);
4396 on_rq = p->se.on_rq;
4398 dequeue_task(rq, p, 0);
4400 p->static_prio = NICE_TO_PRIO(nice);
4403 p->prio = effective_prio(p);
4404 delta = p->prio - old_prio;
4407 enqueue_task(rq, p, 0);
4409 * If the task increased its priority or is running and
4410 * lowered its priority, then reschedule its CPU:
4412 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4413 resched_task(rq->curr);
4416 task_rq_unlock(rq, &flags);
4418 EXPORT_SYMBOL(set_user_nice);
4421 * can_nice - check if a task can reduce its nice value
4425 int can_nice(const struct task_struct *p, const int nice)
4427 /* convert nice value [19,-20] to rlimit style value [1,40] */
4428 int nice_rlim = 20 - nice;
4430 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4431 capable(CAP_SYS_NICE));
4434 #ifdef __ARCH_WANT_SYS_NICE
4437 * sys_nice - change the priority of the current process.
4438 * @increment: priority increment
4440 * sys_setpriority is a more generic, but much slower function that
4441 * does similar things.
4443 SYSCALL_DEFINE1(nice, int, increment)
4448 * Setpriority might change our priority at the same moment.
4449 * We don't have to worry. Conceptually one call occurs first
4450 * and we have a single winner.
4452 if (increment < -40)
4457 nice = TASK_NICE(current) + increment;
4463 if (increment < 0 && !can_nice(current, nice))
4466 retval = security_task_setnice(current, nice);
4470 set_user_nice(current, nice);
4477 * task_prio - return the priority value of a given task.
4478 * @p: the task in question.
4480 * This is the priority value as seen by users in /proc.
4481 * RT tasks are offset by -200. Normal tasks are centered
4482 * around 0, value goes from -16 to +15.
4484 int task_prio(const struct task_struct *p)
4486 return p->prio - MAX_RT_PRIO;
4490 * task_nice - return the nice value of a given task.
4491 * @p: the task in question.
4493 int task_nice(const struct task_struct *p)
4495 return TASK_NICE(p);
4497 EXPORT_SYMBOL(task_nice);
4500 * idle_cpu - is a given cpu idle currently?
4501 * @cpu: the processor in question.
4503 int idle_cpu(int cpu)
4505 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4509 * idle_task - return the idle task for a given cpu.
4510 * @cpu: the processor in question.
4512 struct task_struct *idle_task(int cpu)
4514 return cpu_rq(cpu)->idle;
4518 * find_process_by_pid - find a process with a matching PID value.
4519 * @pid: the pid in question.
4521 static struct task_struct *find_process_by_pid(pid_t pid)
4523 return pid ? find_task_by_vpid(pid) : current;
4526 /* Actually do priority change: must hold rq lock. */
4528 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4530 BUG_ON(p->se.on_rq);
4533 p->rt_priority = prio;
4534 p->normal_prio = normal_prio(p);
4535 /* we are holding p->pi_lock already */
4536 p->prio = rt_mutex_getprio(p);
4537 if (rt_prio(p->prio))
4538 p->sched_class = &rt_sched_class;
4540 p->sched_class = &fair_sched_class;
4545 * check the target process has a UID that matches the current process's
4547 static bool check_same_owner(struct task_struct *p)
4549 const struct cred *cred = current_cred(), *pcred;
4553 pcred = __task_cred(p);
4554 match = (cred->euid == pcred->euid ||
4555 cred->euid == pcred->uid);
4560 static int __sched_setscheduler(struct task_struct *p, int policy,
4561 struct sched_param *param, bool user)
4563 int retval, oldprio, oldpolicy = -1, on_rq, running;
4564 unsigned long flags;
4565 const struct sched_class *prev_class;
4569 /* may grab non-irq protected spin_locks */
4570 BUG_ON(in_interrupt());
4572 /* double check policy once rq lock held */
4574 reset_on_fork = p->sched_reset_on_fork;
4575 policy = oldpolicy = p->policy;
4577 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4578 policy &= ~SCHED_RESET_ON_FORK;
4580 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4581 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4582 policy != SCHED_IDLE)
4587 * Valid priorities for SCHED_FIFO and SCHED_RR are
4588 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4589 * SCHED_BATCH and SCHED_IDLE is 0.
4591 if (param->sched_priority < 0 ||
4592 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4593 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4595 if (rt_policy(policy) != (param->sched_priority != 0))
4599 * Allow unprivileged RT tasks to decrease priority:
4601 if (user && !capable(CAP_SYS_NICE)) {
4602 if (rt_policy(policy)) {
4603 unsigned long rlim_rtprio;
4605 if (!lock_task_sighand(p, &flags))
4607 rlim_rtprio = task_rlimit(p, RLIMIT_RTPRIO);
4608 unlock_task_sighand(p, &flags);
4610 /* can't set/change the rt policy */
4611 if (policy != p->policy && !rlim_rtprio)
4614 /* can't increase priority */
4615 if (param->sched_priority > p->rt_priority &&
4616 param->sched_priority > rlim_rtprio)
4620 * Like positive nice levels, dont allow tasks to
4621 * move out of SCHED_IDLE either:
4623 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4626 /* can't change other user's priorities */
4627 if (!check_same_owner(p))
4630 /* Normal users shall not reset the sched_reset_on_fork flag */
4631 if (p->sched_reset_on_fork && !reset_on_fork)
4636 retval = security_task_setscheduler(p, policy, param);
4642 * make sure no PI-waiters arrive (or leave) while we are
4643 * changing the priority of the task:
4645 raw_spin_lock_irqsave(&p->pi_lock, flags);
4647 * To be able to change p->policy safely, the apropriate
4648 * runqueue lock must be held.
4650 rq = __task_rq_lock(p);
4652 #ifdef CONFIG_RT_GROUP_SCHED
4655 * Do not allow realtime tasks into groups that have no runtime
4658 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4659 task_group(p)->rt_bandwidth.rt_runtime == 0) {
4660 __task_rq_unlock(rq);
4661 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4667 /* recheck policy now with rq lock held */
4668 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4669 policy = oldpolicy = -1;
4670 __task_rq_unlock(rq);
4671 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4674 on_rq = p->se.on_rq;
4675 running = task_current(rq, p);
4677 deactivate_task(rq, p, 0);
4679 p->sched_class->put_prev_task(rq, p);
4681 p->sched_reset_on_fork = reset_on_fork;
4684 prev_class = p->sched_class;
4685 __setscheduler(rq, p, policy, param->sched_priority);
4688 p->sched_class->set_curr_task(rq);
4690 activate_task(rq, p, 0);
4692 check_class_changed(rq, p, prev_class, oldprio, running);
4694 __task_rq_unlock(rq);
4695 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4697 rt_mutex_adjust_pi(p);
4703 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4704 * @p: the task in question.
4705 * @policy: new policy.
4706 * @param: structure containing the new RT priority.
4708 * NOTE that the task may be already dead.
4710 int sched_setscheduler(struct task_struct *p, int policy,
4711 struct sched_param *param)
4713 return __sched_setscheduler(p, policy, param, true);
4715 EXPORT_SYMBOL_GPL(sched_setscheduler);
4718 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4719 * @p: the task in question.
4720 * @policy: new policy.
4721 * @param: structure containing the new RT priority.
4723 * Just like sched_setscheduler, only don't bother checking if the
4724 * current context has permission. For example, this is needed in
4725 * stop_machine(): we create temporary high priority worker threads,
4726 * but our caller might not have that capability.
4728 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4729 struct sched_param *param)
4731 return __sched_setscheduler(p, policy, param, false);
4735 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4737 struct sched_param lparam;
4738 struct task_struct *p;
4741 if (!param || pid < 0)
4743 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4748 p = find_process_by_pid(pid);
4750 retval = sched_setscheduler(p, policy, &lparam);
4757 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4758 * @pid: the pid in question.
4759 * @policy: new policy.
4760 * @param: structure containing the new RT priority.
4762 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4763 struct sched_param __user *, param)
4765 /* negative values for policy are not valid */
4769 return do_sched_setscheduler(pid, policy, param);
4773 * sys_sched_setparam - set/change the RT priority of a thread
4774 * @pid: the pid in question.
4775 * @param: structure containing the new RT priority.
4777 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4779 return do_sched_setscheduler(pid, -1, param);
4783 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4784 * @pid: the pid in question.
4786 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4788 struct task_struct *p;
4796 p = find_process_by_pid(pid);
4798 retval = security_task_getscheduler(p);
4801 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4808 * sys_sched_getparam - get the RT priority of a thread
4809 * @pid: the pid in question.
4810 * @param: structure containing the RT priority.
4812 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4814 struct sched_param lp;
4815 struct task_struct *p;
4818 if (!param || pid < 0)
4822 p = find_process_by_pid(pid);
4827 retval = security_task_getscheduler(p);
4831 lp.sched_priority = p->rt_priority;
4835 * This one might sleep, we cannot do it with a spinlock held ...
4837 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4846 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4848 cpumask_var_t cpus_allowed, new_mask;
4849 struct task_struct *p;
4855 p = find_process_by_pid(pid);
4862 /* Prevent p going away */
4866 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4870 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4872 goto out_free_cpus_allowed;
4875 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
4878 retval = security_task_setscheduler(p, 0, NULL);
4882 cpuset_cpus_allowed(p, cpus_allowed);
4883 cpumask_and(new_mask, in_mask, cpus_allowed);
4885 retval = set_cpus_allowed_ptr(p, new_mask);
4888 cpuset_cpus_allowed(p, cpus_allowed);
4889 if (!cpumask_subset(new_mask, cpus_allowed)) {
4891 * We must have raced with a concurrent cpuset
4892 * update. Just reset the cpus_allowed to the
4893 * cpuset's cpus_allowed
4895 cpumask_copy(new_mask, cpus_allowed);
4900 free_cpumask_var(new_mask);
4901 out_free_cpus_allowed:
4902 free_cpumask_var(cpus_allowed);
4909 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4910 struct cpumask *new_mask)
4912 if (len < cpumask_size())
4913 cpumask_clear(new_mask);
4914 else if (len > cpumask_size())
4915 len = cpumask_size();
4917 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4921 * sys_sched_setaffinity - set the cpu affinity of a process
4922 * @pid: pid of the process
4923 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4924 * @user_mask_ptr: user-space pointer to the new cpu mask
4926 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4927 unsigned long __user *, user_mask_ptr)
4929 cpumask_var_t new_mask;
4932 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4935 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4937 retval = sched_setaffinity(pid, new_mask);
4938 free_cpumask_var(new_mask);
4942 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4944 struct task_struct *p;
4945 unsigned long flags;
4953 p = find_process_by_pid(pid);
4957 retval = security_task_getscheduler(p);
4961 rq = task_rq_lock(p, &flags);
4962 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4963 task_rq_unlock(rq, &flags);
4973 * sys_sched_getaffinity - get the cpu affinity of a process
4974 * @pid: pid of the process
4975 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4976 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4978 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4979 unsigned long __user *, user_mask_ptr)
4984 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4986 if (len & (sizeof(unsigned long)-1))
4989 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4992 ret = sched_getaffinity(pid, mask);
4994 size_t retlen = min_t(size_t, len, cpumask_size());
4996 if (copy_to_user(user_mask_ptr, mask, retlen))
5001 free_cpumask_var(mask);
5007 * sys_sched_yield - yield the current processor to other threads.
5009 * This function yields the current CPU to other tasks. If there are no
5010 * other threads running on this CPU then this function will return.
5012 SYSCALL_DEFINE0(sched_yield)
5014 struct rq *rq = this_rq_lock();
5016 schedstat_inc(rq, yld_count);
5017 current->sched_class->yield_task(rq);
5020 * Since we are going to call schedule() anyway, there's
5021 * no need to preempt or enable interrupts:
5023 __release(rq->lock);
5024 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5025 do_raw_spin_unlock(&rq->lock);
5026 preempt_enable_no_resched();
5033 static inline int should_resched(void)
5035 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5038 static void __cond_resched(void)
5040 add_preempt_count(PREEMPT_ACTIVE);
5042 sub_preempt_count(PREEMPT_ACTIVE);
5045 int __sched _cond_resched(void)
5047 if (should_resched()) {
5053 EXPORT_SYMBOL(_cond_resched);
5056 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5057 * call schedule, and on return reacquire the lock.
5059 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5060 * operations here to prevent schedule() from being called twice (once via
5061 * spin_unlock(), once by hand).
5063 int __cond_resched_lock(spinlock_t *lock)
5065 int resched = should_resched();
5068 lockdep_assert_held(lock);
5070 if (spin_needbreak(lock) || resched) {
5081 EXPORT_SYMBOL(__cond_resched_lock);
5083 int __sched __cond_resched_softirq(void)
5085 BUG_ON(!in_softirq());
5087 if (should_resched()) {
5095 EXPORT_SYMBOL(__cond_resched_softirq);
5098 * yield - yield the current processor to other threads.
5100 * This is a shortcut for kernel-space yielding - it marks the
5101 * thread runnable and calls sys_sched_yield().
5103 void __sched yield(void)
5105 set_current_state(TASK_RUNNING);
5108 EXPORT_SYMBOL(yield);
5111 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5112 * that process accounting knows that this is a task in IO wait state.
5114 void __sched io_schedule(void)
5116 struct rq *rq = raw_rq();
5118 delayacct_blkio_start();
5119 atomic_inc(&rq->nr_iowait);
5120 current->in_iowait = 1;
5122 current->in_iowait = 0;
5123 atomic_dec(&rq->nr_iowait);
5124 delayacct_blkio_end();
5126 EXPORT_SYMBOL(io_schedule);
5128 long __sched io_schedule_timeout(long timeout)
5130 struct rq *rq = raw_rq();
5133 delayacct_blkio_start();
5134 atomic_inc(&rq->nr_iowait);
5135 current->in_iowait = 1;
5136 ret = schedule_timeout(timeout);
5137 current->in_iowait = 0;
5138 atomic_dec(&rq->nr_iowait);
5139 delayacct_blkio_end();
5144 * sys_sched_get_priority_max - return maximum RT priority.
5145 * @policy: scheduling class.
5147 * this syscall returns the maximum rt_priority that can be used
5148 * by a given scheduling class.
5150 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5157 ret = MAX_USER_RT_PRIO-1;
5169 * sys_sched_get_priority_min - return minimum RT priority.
5170 * @policy: scheduling class.
5172 * this syscall returns the minimum rt_priority that can be used
5173 * by a given scheduling class.
5175 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5193 * sys_sched_rr_get_interval - return the default timeslice of a process.
5194 * @pid: pid of the process.
5195 * @interval: userspace pointer to the timeslice value.
5197 * this syscall writes the default timeslice value of a given process
5198 * into the user-space timespec buffer. A value of '0' means infinity.
5200 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5201 struct timespec __user *, interval)
5203 struct task_struct *p;
5204 unsigned int time_slice;
5205 unsigned long flags;
5215 p = find_process_by_pid(pid);
5219 retval = security_task_getscheduler(p);
5223 rq = task_rq_lock(p, &flags);
5224 time_slice = p->sched_class->get_rr_interval(rq, p);
5225 task_rq_unlock(rq, &flags);
5228 jiffies_to_timespec(time_slice, &t);
5229 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5237 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5239 void sched_show_task(struct task_struct *p)
5241 unsigned long free = 0;
5244 state = p->state ? __ffs(p->state) + 1 : 0;
5245 printk(KERN_INFO "%-13.13s %c", p->comm,
5246 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5247 #if BITS_PER_LONG == 32
5248 if (state == TASK_RUNNING)
5249 printk(KERN_CONT " running ");
5251 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5253 if (state == TASK_RUNNING)
5254 printk(KERN_CONT " running task ");
5256 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5258 #ifdef CONFIG_DEBUG_STACK_USAGE
5259 free = stack_not_used(p);
5261 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5262 task_pid_nr(p), task_pid_nr(p->real_parent),
5263 (unsigned long)task_thread_info(p)->flags);
5265 show_stack(p, NULL);
5268 void show_state_filter(unsigned long state_filter)
5270 struct task_struct *g, *p;
5272 #if BITS_PER_LONG == 32
5274 " task PC stack pid father\n");
5277 " task PC stack pid father\n");
5279 read_lock(&tasklist_lock);
5280 do_each_thread(g, p) {
5282 * reset the NMI-timeout, listing all files on a slow
5283 * console might take alot of time:
5285 touch_nmi_watchdog();
5286 if (!state_filter || (p->state & state_filter))
5288 } while_each_thread(g, p);
5290 touch_all_softlockup_watchdogs();
5292 #ifdef CONFIG_SCHED_DEBUG
5293 sysrq_sched_debug_show();
5295 read_unlock(&tasklist_lock);
5297 * Only show locks if all tasks are dumped:
5300 debug_show_all_locks();
5303 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5305 idle->sched_class = &idle_sched_class;
5309 * init_idle - set up an idle thread for a given CPU
5310 * @idle: task in question
5311 * @cpu: cpu the idle task belongs to
5313 * NOTE: this function does not set the idle thread's NEED_RESCHED
5314 * flag, to make booting more robust.
5316 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5318 struct rq *rq = cpu_rq(cpu);
5319 unsigned long flags;
5321 raw_spin_lock_irqsave(&rq->lock, flags);
5324 idle->state = TASK_RUNNING;
5325 idle->se.exec_start = sched_clock();
5327 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
5328 __set_task_cpu(idle, cpu);
5330 rq->curr = rq->idle = idle;
5331 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5334 raw_spin_unlock_irqrestore(&rq->lock, flags);
5336 /* Set the preempt count _outside_ the spinlocks! */
5337 #if defined(CONFIG_PREEMPT)
5338 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5340 task_thread_info(idle)->preempt_count = 0;
5343 * The idle tasks have their own, simple scheduling class:
5345 idle->sched_class = &idle_sched_class;
5346 ftrace_graph_init_task(idle);
5350 * In a system that switches off the HZ timer nohz_cpu_mask
5351 * indicates which cpus entered this state. This is used
5352 * in the rcu update to wait only for active cpus. For system
5353 * which do not switch off the HZ timer nohz_cpu_mask should
5354 * always be CPU_BITS_NONE.
5356 cpumask_var_t nohz_cpu_mask;
5359 * Increase the granularity value when there are more CPUs,
5360 * because with more CPUs the 'effective latency' as visible
5361 * to users decreases. But the relationship is not linear,
5362 * so pick a second-best guess by going with the log2 of the
5365 * This idea comes from the SD scheduler of Con Kolivas:
5367 static int get_update_sysctl_factor(void)
5369 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5370 unsigned int factor;
5372 switch (sysctl_sched_tunable_scaling) {
5373 case SCHED_TUNABLESCALING_NONE:
5376 case SCHED_TUNABLESCALING_LINEAR:
5379 case SCHED_TUNABLESCALING_LOG:
5381 factor = 1 + ilog2(cpus);
5388 static void update_sysctl(void)
5390 unsigned int factor = get_update_sysctl_factor();
5392 #define SET_SYSCTL(name) \
5393 (sysctl_##name = (factor) * normalized_sysctl_##name)
5394 SET_SYSCTL(sched_min_granularity);
5395 SET_SYSCTL(sched_latency);
5396 SET_SYSCTL(sched_wakeup_granularity);
5397 SET_SYSCTL(sched_shares_ratelimit);
5401 static inline void sched_init_granularity(void)
5408 * This is how migration works:
5410 * 1) we invoke migration_cpu_stop() on the target CPU using
5412 * 2) stopper starts to run (implicitly forcing the migrated thread
5414 * 3) it checks whether the migrated task is still in the wrong runqueue.
5415 * 4) if it's in the wrong runqueue then the migration thread removes
5416 * it and puts it into the right queue.
5417 * 5) stopper completes and stop_one_cpu() returns and the migration
5422 * Change a given task's CPU affinity. Migrate the thread to a
5423 * proper CPU and schedule it away if the CPU it's executing on
5424 * is removed from the allowed bitmask.
5426 * NOTE: the caller must have a valid reference to the task, the
5427 * task must not exit() & deallocate itself prematurely. The
5428 * call is not atomic; no spinlocks may be held.
5430 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5432 unsigned long flags;
5434 unsigned int dest_cpu;
5438 * Serialize against TASK_WAKING so that ttwu() and wunt() can
5439 * drop the rq->lock and still rely on ->cpus_allowed.
5442 while (task_is_waking(p))
5444 rq = task_rq_lock(p, &flags);
5445 if (task_is_waking(p)) {
5446 task_rq_unlock(rq, &flags);
5450 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5455 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5456 !cpumask_equal(&p->cpus_allowed, new_mask))) {
5461 if (p->sched_class->set_cpus_allowed)
5462 p->sched_class->set_cpus_allowed(p, new_mask);
5464 cpumask_copy(&p->cpus_allowed, new_mask);
5465 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5468 /* Can the task run on the task's current CPU? If so, we're done */
5469 if (cpumask_test_cpu(task_cpu(p), new_mask))
5472 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
5473 if (migrate_task(p, dest_cpu)) {
5474 struct migration_arg arg = { p, dest_cpu };
5475 /* Need help from migration thread: drop lock and wait. */
5476 task_rq_unlock(rq, &flags);
5477 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
5478 tlb_migrate_finish(p->mm);
5482 task_rq_unlock(rq, &flags);
5486 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5489 * Move (not current) task off this cpu, onto dest cpu. We're doing
5490 * this because either it can't run here any more (set_cpus_allowed()
5491 * away from this CPU, or CPU going down), or because we're
5492 * attempting to rebalance this task on exec (sched_exec).
5494 * So we race with normal scheduler movements, but that's OK, as long
5495 * as the task is no longer on this CPU.
5497 * Returns non-zero if task was successfully migrated.
5499 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5501 struct rq *rq_dest, *rq_src;
5504 if (unlikely(!cpu_active(dest_cpu)))
5507 rq_src = cpu_rq(src_cpu);
5508 rq_dest = cpu_rq(dest_cpu);
5510 double_rq_lock(rq_src, rq_dest);
5511 /* Already moved. */
5512 if (task_cpu(p) != src_cpu)
5514 /* Affinity changed (again). */
5515 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
5519 * If we're not on a rq, the next wake-up will ensure we're
5523 deactivate_task(rq_src, p, 0);
5524 set_task_cpu(p, dest_cpu);
5525 activate_task(rq_dest, p, 0);
5526 check_preempt_curr(rq_dest, p, 0);
5531 double_rq_unlock(rq_src, rq_dest);
5536 * migration_cpu_stop - this will be executed by a highprio stopper thread
5537 * and performs thread migration by bumping thread off CPU then
5538 * 'pushing' onto another runqueue.
5540 static int migration_cpu_stop(void *data)
5542 struct migration_arg *arg = data;
5545 * The original target cpu might have gone down and we might
5546 * be on another cpu but it doesn't matter.
5548 local_irq_disable();
5549 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
5554 #ifdef CONFIG_HOTPLUG_CPU
5556 * Figure out where task on dead CPU should go, use force if necessary.
5558 void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5560 struct rq *rq = cpu_rq(dead_cpu);
5561 int needs_cpu, uninitialized_var(dest_cpu);
5562 unsigned long flags;
5564 local_irq_save(flags);
5566 raw_spin_lock(&rq->lock);
5567 needs_cpu = (task_cpu(p) == dead_cpu) && (p->state != TASK_WAKING);
5569 dest_cpu = select_fallback_rq(dead_cpu, p);
5570 raw_spin_unlock(&rq->lock);
5572 * It can only fail if we race with set_cpus_allowed(),
5573 * in the racer should migrate the task anyway.
5576 __migrate_task(p, dead_cpu, dest_cpu);
5577 local_irq_restore(flags);
5581 * While a dead CPU has no uninterruptible tasks queued at this point,
5582 * it might still have a nonzero ->nr_uninterruptible counter, because
5583 * for performance reasons the counter is not stricly tracking tasks to
5584 * their home CPUs. So we just add the counter to another CPU's counter,
5585 * to keep the global sum constant after CPU-down:
5587 static void migrate_nr_uninterruptible(struct rq *rq_src)
5589 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5590 unsigned long flags;
5592 local_irq_save(flags);
5593 double_rq_lock(rq_src, rq_dest);
5594 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5595 rq_src->nr_uninterruptible = 0;
5596 double_rq_unlock(rq_src, rq_dest);
5597 local_irq_restore(flags);
5600 /* Run through task list and migrate tasks from the dead cpu. */
5601 static void migrate_live_tasks(int src_cpu)
5603 struct task_struct *p, *t;
5605 read_lock(&tasklist_lock);
5607 do_each_thread(t, p) {
5611 if (task_cpu(p) == src_cpu)
5612 move_task_off_dead_cpu(src_cpu, p);
5613 } while_each_thread(t, p);
5615 read_unlock(&tasklist_lock);
5619 * Schedules idle task to be the next runnable task on current CPU.
5620 * It does so by boosting its priority to highest possible.
5621 * Used by CPU offline code.
5623 void sched_idle_next(void)
5625 int this_cpu = smp_processor_id();
5626 struct rq *rq = cpu_rq(this_cpu);
5627 struct task_struct *p = rq->idle;
5628 unsigned long flags;
5630 /* cpu has to be offline */
5631 BUG_ON(cpu_online(this_cpu));
5634 * Strictly not necessary since rest of the CPUs are stopped by now
5635 * and interrupts disabled on the current cpu.
5637 raw_spin_lock_irqsave(&rq->lock, flags);
5639 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5641 activate_task(rq, p, 0);
5643 raw_spin_unlock_irqrestore(&rq->lock, flags);
5647 * Ensures that the idle task is using init_mm right before its cpu goes
5650 void idle_task_exit(void)
5652 struct mm_struct *mm = current->active_mm;
5654 BUG_ON(cpu_online(smp_processor_id()));
5657 switch_mm(mm, &init_mm, current);
5661 /* called under rq->lock with disabled interrupts */
5662 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5664 struct rq *rq = cpu_rq(dead_cpu);
5666 /* Must be exiting, otherwise would be on tasklist. */
5667 BUG_ON(!p->exit_state);
5669 /* Cannot have done final schedule yet: would have vanished. */
5670 BUG_ON(p->state == TASK_DEAD);
5675 * Drop lock around migration; if someone else moves it,
5676 * that's OK. No task can be added to this CPU, so iteration is
5679 raw_spin_unlock_irq(&rq->lock);
5680 move_task_off_dead_cpu(dead_cpu, p);
5681 raw_spin_lock_irq(&rq->lock);
5686 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5687 static void migrate_dead_tasks(unsigned int dead_cpu)
5689 struct rq *rq = cpu_rq(dead_cpu);
5690 struct task_struct *next;
5693 if (!rq->nr_running)
5695 next = pick_next_task(rq);
5698 next->sched_class->put_prev_task(rq, next);
5699 migrate_dead(dead_cpu, next);
5705 * remove the tasks which were accounted by rq from calc_load_tasks.
5707 static void calc_global_load_remove(struct rq *rq)
5709 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5710 rq->calc_load_active = 0;
5712 #endif /* CONFIG_HOTPLUG_CPU */
5714 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5716 static struct ctl_table sd_ctl_dir[] = {
5718 .procname = "sched_domain",
5724 static struct ctl_table sd_ctl_root[] = {
5726 .procname = "kernel",
5728 .child = sd_ctl_dir,
5733 static struct ctl_table *sd_alloc_ctl_entry(int n)
5735 struct ctl_table *entry =
5736 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5741 static void sd_free_ctl_entry(struct ctl_table **tablep)
5743 struct ctl_table *entry;
5746 * In the intermediate directories, both the child directory and
5747 * procname are dynamically allocated and could fail but the mode
5748 * will always be set. In the lowest directory the names are
5749 * static strings and all have proc handlers.
5751 for (entry = *tablep; entry->mode; entry++) {
5753 sd_free_ctl_entry(&entry->child);
5754 if (entry->proc_handler == NULL)
5755 kfree(entry->procname);
5763 set_table_entry(struct ctl_table *entry,
5764 const char *procname, void *data, int maxlen,
5765 mode_t mode, proc_handler *proc_handler)
5767 entry->procname = procname;
5769 entry->maxlen = maxlen;
5771 entry->proc_handler = proc_handler;
5774 static struct ctl_table *
5775 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5777 struct ctl_table *table = sd_alloc_ctl_entry(13);
5782 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5783 sizeof(long), 0644, proc_doulongvec_minmax);
5784 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5785 sizeof(long), 0644, proc_doulongvec_minmax);
5786 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5787 sizeof(int), 0644, proc_dointvec_minmax);
5788 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5789 sizeof(int), 0644, proc_dointvec_minmax);
5790 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5791 sizeof(int), 0644, proc_dointvec_minmax);
5792 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5793 sizeof(int), 0644, proc_dointvec_minmax);
5794 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5795 sizeof(int), 0644, proc_dointvec_minmax);
5796 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5797 sizeof(int), 0644, proc_dointvec_minmax);
5798 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5799 sizeof(int), 0644, proc_dointvec_minmax);
5800 set_table_entry(&table[9], "cache_nice_tries",
5801 &sd->cache_nice_tries,
5802 sizeof(int), 0644, proc_dointvec_minmax);
5803 set_table_entry(&table[10], "flags", &sd->flags,
5804 sizeof(int), 0644, proc_dointvec_minmax);
5805 set_table_entry(&table[11], "name", sd->name,
5806 CORENAME_MAX_SIZE, 0444, proc_dostring);
5807 /* &table[12] is terminator */
5812 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5814 struct ctl_table *entry, *table;
5815 struct sched_domain *sd;
5816 int domain_num = 0, i;
5819 for_each_domain(cpu, sd)
5821 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5826 for_each_domain(cpu, sd) {
5827 snprintf(buf, 32, "domain%d", i);
5828 entry->procname = kstrdup(buf, GFP_KERNEL);
5830 entry->child = sd_alloc_ctl_domain_table(sd);
5837 static struct ctl_table_header *sd_sysctl_header;
5838 static void register_sched_domain_sysctl(void)
5840 int i, cpu_num = num_possible_cpus();
5841 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5844 WARN_ON(sd_ctl_dir[0].child);
5845 sd_ctl_dir[0].child = entry;
5850 for_each_possible_cpu(i) {
5851 snprintf(buf, 32, "cpu%d", i);
5852 entry->procname = kstrdup(buf, GFP_KERNEL);
5854 entry->child = sd_alloc_ctl_cpu_table(i);
5858 WARN_ON(sd_sysctl_header);
5859 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5862 /* may be called multiple times per register */
5863 static void unregister_sched_domain_sysctl(void)
5865 if (sd_sysctl_header)
5866 unregister_sysctl_table(sd_sysctl_header);
5867 sd_sysctl_header = NULL;
5868 if (sd_ctl_dir[0].child)
5869 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5872 static void register_sched_domain_sysctl(void)
5875 static void unregister_sched_domain_sysctl(void)
5880 static void set_rq_online(struct rq *rq)
5883 const struct sched_class *class;
5885 cpumask_set_cpu(rq->cpu, rq->rd->online);
5888 for_each_class(class) {
5889 if (class->rq_online)
5890 class->rq_online(rq);
5895 static void set_rq_offline(struct rq *rq)
5898 const struct sched_class *class;
5900 for_each_class(class) {
5901 if (class->rq_offline)
5902 class->rq_offline(rq);
5905 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5911 * migration_call - callback that gets triggered when a CPU is added.
5912 * Here we can start up the necessary migration thread for the new CPU.
5914 static int __cpuinit
5915 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5917 int cpu = (long)hcpu;
5918 unsigned long flags;
5919 struct rq *rq = cpu_rq(cpu);
5923 case CPU_UP_PREPARE:
5924 case CPU_UP_PREPARE_FROZEN:
5925 rq->calc_load_update = calc_load_update;
5929 case CPU_ONLINE_FROZEN:
5930 /* Update our root-domain */
5931 raw_spin_lock_irqsave(&rq->lock, flags);
5933 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5937 raw_spin_unlock_irqrestore(&rq->lock, flags);
5940 #ifdef CONFIG_HOTPLUG_CPU
5942 case CPU_DEAD_FROZEN:
5943 migrate_live_tasks(cpu);
5944 /* Idle task back to normal (off runqueue, low prio) */
5945 raw_spin_lock_irq(&rq->lock);
5946 deactivate_task(rq, rq->idle, 0);
5947 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5948 rq->idle->sched_class = &idle_sched_class;
5949 migrate_dead_tasks(cpu);
5950 raw_spin_unlock_irq(&rq->lock);
5951 migrate_nr_uninterruptible(rq);
5952 BUG_ON(rq->nr_running != 0);
5953 calc_global_load_remove(rq);
5957 case CPU_DYING_FROZEN:
5958 /* Update our root-domain */
5959 raw_spin_lock_irqsave(&rq->lock, flags);
5961 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5964 raw_spin_unlock_irqrestore(&rq->lock, flags);
5972 * Register at high priority so that task migration (migrate_all_tasks)
5973 * happens before everything else. This has to be lower priority than
5974 * the notifier in the perf_event subsystem, though.
5976 static struct notifier_block __cpuinitdata migration_notifier = {
5977 .notifier_call = migration_call,
5978 .priority = CPU_PRI_MIGRATION,
5981 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
5982 unsigned long action, void *hcpu)
5984 switch (action & ~CPU_TASKS_FROZEN) {
5986 case CPU_DOWN_FAILED:
5987 set_cpu_active((long)hcpu, true);
5994 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
5995 unsigned long action, void *hcpu)
5997 switch (action & ~CPU_TASKS_FROZEN) {
5998 case CPU_DOWN_PREPARE:
5999 set_cpu_active((long)hcpu, false);
6006 static int __init migration_init(void)
6008 void *cpu = (void *)(long)smp_processor_id();
6011 /* Initialize migration for the boot CPU */
6012 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6013 BUG_ON(err == NOTIFY_BAD);
6014 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6015 register_cpu_notifier(&migration_notifier);
6017 /* Register cpu active notifiers */
6018 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
6019 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
6023 early_initcall(migration_init);
6028 #ifdef CONFIG_SCHED_DEBUG
6030 static __read_mostly int sched_domain_debug_enabled;
6032 static int __init sched_domain_debug_setup(char *str)
6034 sched_domain_debug_enabled = 1;
6038 early_param("sched_debug", sched_domain_debug_setup);
6040 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6041 struct cpumask *groupmask)
6043 struct sched_group *group = sd->groups;
6046 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6047 cpumask_clear(groupmask);
6049 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6051 if (!(sd->flags & SD_LOAD_BALANCE)) {
6052 printk("does not load-balance\n");
6054 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6059 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6061 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6062 printk(KERN_ERR "ERROR: domain->span does not contain "
6065 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6066 printk(KERN_ERR "ERROR: domain->groups does not contain"
6070 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6074 printk(KERN_ERR "ERROR: group is NULL\n");
6078 if (!group->cpu_power) {
6079 printk(KERN_CONT "\n");
6080 printk(KERN_ERR "ERROR: domain->cpu_power not "
6085 if (!cpumask_weight(sched_group_cpus(group))) {
6086 printk(KERN_CONT "\n");
6087 printk(KERN_ERR "ERROR: empty group\n");
6091 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6092 printk(KERN_CONT "\n");
6093 printk(KERN_ERR "ERROR: repeated CPUs\n");
6097 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6099 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6101 printk(KERN_CONT " %s", str);
6102 if (group->cpu_power != SCHED_LOAD_SCALE) {
6103 printk(KERN_CONT " (cpu_power = %d)",
6107 group = group->next;
6108 } while (group != sd->groups);
6109 printk(KERN_CONT "\n");
6111 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6112 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6115 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6116 printk(KERN_ERR "ERROR: parent span is not a superset "
6117 "of domain->span\n");
6121 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6123 cpumask_var_t groupmask;
6126 if (!sched_domain_debug_enabled)
6130 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6134 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6136 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6137 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6142 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6149 free_cpumask_var(groupmask);
6151 #else /* !CONFIG_SCHED_DEBUG */
6152 # define sched_domain_debug(sd, cpu) do { } while (0)
6153 #endif /* CONFIG_SCHED_DEBUG */
6155 static int sd_degenerate(struct sched_domain *sd)
6157 if (cpumask_weight(sched_domain_span(sd)) == 1)
6160 /* Following flags need at least 2 groups */
6161 if (sd->flags & (SD_LOAD_BALANCE |
6162 SD_BALANCE_NEWIDLE |
6166 SD_SHARE_PKG_RESOURCES)) {
6167 if (sd->groups != sd->groups->next)
6171 /* Following flags don't use groups */
6172 if (sd->flags & (SD_WAKE_AFFINE))
6179 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6181 unsigned long cflags = sd->flags, pflags = parent->flags;
6183 if (sd_degenerate(parent))
6186 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6189 /* Flags needing groups don't count if only 1 group in parent */
6190 if (parent->groups == parent->groups->next) {
6191 pflags &= ~(SD_LOAD_BALANCE |
6192 SD_BALANCE_NEWIDLE |
6196 SD_SHARE_PKG_RESOURCES);
6197 if (nr_node_ids == 1)
6198 pflags &= ~SD_SERIALIZE;
6200 if (~cflags & pflags)
6206 static void free_rootdomain(struct root_domain *rd)
6208 synchronize_sched();
6210 cpupri_cleanup(&rd->cpupri);
6212 free_cpumask_var(rd->rto_mask);
6213 free_cpumask_var(rd->online);
6214 free_cpumask_var(rd->span);
6218 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6220 struct root_domain *old_rd = NULL;
6221 unsigned long flags;
6223 raw_spin_lock_irqsave(&rq->lock, flags);
6228 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6231 cpumask_clear_cpu(rq->cpu, old_rd->span);
6234 * If we dont want to free the old_rt yet then
6235 * set old_rd to NULL to skip the freeing later
6238 if (!atomic_dec_and_test(&old_rd->refcount))
6242 atomic_inc(&rd->refcount);
6245 cpumask_set_cpu(rq->cpu, rd->span);
6246 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6249 raw_spin_unlock_irqrestore(&rq->lock, flags);
6252 free_rootdomain(old_rd);
6255 static int init_rootdomain(struct root_domain *rd, bool bootmem)
6257 gfp_t gfp = GFP_KERNEL;
6259 memset(rd, 0, sizeof(*rd));
6264 if (!alloc_cpumask_var(&rd->span, gfp))
6266 if (!alloc_cpumask_var(&rd->online, gfp))
6268 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
6271 if (cpupri_init(&rd->cpupri, bootmem) != 0)
6276 free_cpumask_var(rd->rto_mask);
6278 free_cpumask_var(rd->online);
6280 free_cpumask_var(rd->span);
6285 static void init_defrootdomain(void)
6287 init_rootdomain(&def_root_domain, true);
6289 atomic_set(&def_root_domain.refcount, 1);
6292 static struct root_domain *alloc_rootdomain(void)
6294 struct root_domain *rd;
6296 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6300 if (init_rootdomain(rd, false) != 0) {
6309 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6310 * hold the hotplug lock.
6313 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6315 struct rq *rq = cpu_rq(cpu);
6316 struct sched_domain *tmp;
6318 for (tmp = sd; tmp; tmp = tmp->parent)
6319 tmp->span_weight = cpumask_weight(sched_domain_span(tmp));
6321 /* Remove the sched domains which do not contribute to scheduling. */
6322 for (tmp = sd; tmp; ) {
6323 struct sched_domain *parent = tmp->parent;
6327 if (sd_parent_degenerate(tmp, parent)) {
6328 tmp->parent = parent->parent;
6330 parent->parent->child = tmp;
6335 if (sd && sd_degenerate(sd)) {
6341 sched_domain_debug(sd, cpu);
6343 rq_attach_root(rq, rd);
6344 rcu_assign_pointer(rq->sd, sd);
6347 /* cpus with isolated domains */
6348 static cpumask_var_t cpu_isolated_map;
6350 /* Setup the mask of cpus configured for isolated domains */
6351 static int __init isolated_cpu_setup(char *str)
6353 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6354 cpulist_parse(str, cpu_isolated_map);
6358 __setup("isolcpus=", isolated_cpu_setup);
6361 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6362 * to a function which identifies what group(along with sched group) a CPU
6363 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6364 * (due to the fact that we keep track of groups covered with a struct cpumask).
6366 * init_sched_build_groups will build a circular linked list of the groups
6367 * covered by the given span, and will set each group's ->cpumask correctly,
6368 * and ->cpu_power to 0.
6371 init_sched_build_groups(const struct cpumask *span,
6372 const struct cpumask *cpu_map,
6373 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
6374 struct sched_group **sg,
6375 struct cpumask *tmpmask),
6376 struct cpumask *covered, struct cpumask *tmpmask)
6378 struct sched_group *first = NULL, *last = NULL;
6381 cpumask_clear(covered);
6383 for_each_cpu(i, span) {
6384 struct sched_group *sg;
6385 int group = group_fn(i, cpu_map, &sg, tmpmask);
6388 if (cpumask_test_cpu(i, covered))
6391 cpumask_clear(sched_group_cpus(sg));
6394 for_each_cpu(j, span) {
6395 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6398 cpumask_set_cpu(j, covered);
6399 cpumask_set_cpu(j, sched_group_cpus(sg));
6410 #define SD_NODES_PER_DOMAIN 16
6415 * find_next_best_node - find the next node to include in a sched_domain
6416 * @node: node whose sched_domain we're building
6417 * @used_nodes: nodes already in the sched_domain
6419 * Find the next node to include in a given scheduling domain. Simply
6420 * finds the closest node not already in the @used_nodes map.
6422 * Should use nodemask_t.
6424 static int find_next_best_node(int node, nodemask_t *used_nodes)
6426 int i, n, val, min_val, best_node = 0;
6430 for (i = 0; i < nr_node_ids; i++) {
6431 /* Start at @node */
6432 n = (node + i) % nr_node_ids;
6434 if (!nr_cpus_node(n))
6437 /* Skip already used nodes */
6438 if (node_isset(n, *used_nodes))
6441 /* Simple min distance search */
6442 val = node_distance(node, n);
6444 if (val < min_val) {
6450 node_set(best_node, *used_nodes);
6455 * sched_domain_node_span - get a cpumask for a node's sched_domain
6456 * @node: node whose cpumask we're constructing
6457 * @span: resulting cpumask
6459 * Given a node, construct a good cpumask for its sched_domain to span. It
6460 * should be one that prevents unnecessary balancing, but also spreads tasks
6463 static void sched_domain_node_span(int node, struct cpumask *span)
6465 nodemask_t used_nodes;
6468 cpumask_clear(span);
6469 nodes_clear(used_nodes);
6471 cpumask_or(span, span, cpumask_of_node(node));
6472 node_set(node, used_nodes);
6474 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6475 int next_node = find_next_best_node(node, &used_nodes);
6477 cpumask_or(span, span, cpumask_of_node(next_node));
6480 #endif /* CONFIG_NUMA */
6482 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6485 * The cpus mask in sched_group and sched_domain hangs off the end.
6487 * ( See the the comments in include/linux/sched.h:struct sched_group
6488 * and struct sched_domain. )
6490 struct static_sched_group {
6491 struct sched_group sg;
6492 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
6495 struct static_sched_domain {
6496 struct sched_domain sd;
6497 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
6503 cpumask_var_t domainspan;
6504 cpumask_var_t covered;
6505 cpumask_var_t notcovered;
6507 cpumask_var_t nodemask;
6508 cpumask_var_t this_sibling_map;
6509 cpumask_var_t this_core_map;
6510 cpumask_var_t send_covered;
6511 cpumask_var_t tmpmask;
6512 struct sched_group **sched_group_nodes;
6513 struct root_domain *rd;
6517 sa_sched_groups = 0,
6522 sa_this_sibling_map,
6524 sa_sched_group_nodes,
6534 * SMT sched-domains:
6536 #ifdef CONFIG_SCHED_SMT
6537 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
6538 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
6541 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
6542 struct sched_group **sg, struct cpumask *unused)
6545 *sg = &per_cpu(sched_groups, cpu).sg;
6548 #endif /* CONFIG_SCHED_SMT */
6551 * multi-core sched-domains:
6553 #ifdef CONFIG_SCHED_MC
6554 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
6555 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
6556 #endif /* CONFIG_SCHED_MC */
6558 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6560 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6561 struct sched_group **sg, struct cpumask *mask)
6565 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6566 group = cpumask_first(mask);
6568 *sg = &per_cpu(sched_group_core, group).sg;
6571 #elif defined(CONFIG_SCHED_MC)
6573 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6574 struct sched_group **sg, struct cpumask *unused)
6577 *sg = &per_cpu(sched_group_core, cpu).sg;
6582 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
6583 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
6586 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
6587 struct sched_group **sg, struct cpumask *mask)
6590 #ifdef CONFIG_SCHED_MC
6591 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6592 group = cpumask_first(mask);
6593 #elif defined(CONFIG_SCHED_SMT)
6594 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6595 group = cpumask_first(mask);
6600 *sg = &per_cpu(sched_group_phys, group).sg;
6606 * The init_sched_build_groups can't handle what we want to do with node
6607 * groups, so roll our own. Now each node has its own list of groups which
6608 * gets dynamically allocated.
6610 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
6611 static struct sched_group ***sched_group_nodes_bycpu;
6613 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
6614 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
6616 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
6617 struct sched_group **sg,
6618 struct cpumask *nodemask)
6622 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
6623 group = cpumask_first(nodemask);
6626 *sg = &per_cpu(sched_group_allnodes, group).sg;
6630 static void init_numa_sched_groups_power(struct sched_group *group_head)
6632 struct sched_group *sg = group_head;
6638 for_each_cpu(j, sched_group_cpus(sg)) {
6639 struct sched_domain *sd;
6641 sd = &per_cpu(phys_domains, j).sd;
6642 if (j != group_first_cpu(sd->groups)) {
6644 * Only add "power" once for each
6650 sg->cpu_power += sd->groups->cpu_power;
6653 } while (sg != group_head);
6656 static int build_numa_sched_groups(struct s_data *d,
6657 const struct cpumask *cpu_map, int num)
6659 struct sched_domain *sd;
6660 struct sched_group *sg, *prev;
6663 cpumask_clear(d->covered);
6664 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
6665 if (cpumask_empty(d->nodemask)) {
6666 d->sched_group_nodes[num] = NULL;
6670 sched_domain_node_span(num, d->domainspan);
6671 cpumask_and(d->domainspan, d->domainspan, cpu_map);
6673 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6676 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
6680 d->sched_group_nodes[num] = sg;
6682 for_each_cpu(j, d->nodemask) {
6683 sd = &per_cpu(node_domains, j).sd;
6688 cpumask_copy(sched_group_cpus(sg), d->nodemask);
6690 cpumask_or(d->covered, d->covered, d->nodemask);
6693 for (j = 0; j < nr_node_ids; j++) {
6694 n = (num + j) % nr_node_ids;
6695 cpumask_complement(d->notcovered, d->covered);
6696 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
6697 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
6698 if (cpumask_empty(d->tmpmask))
6700 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
6701 if (cpumask_empty(d->tmpmask))
6703 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6707 "Can not alloc domain group for node %d\n", j);
6711 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
6712 sg->next = prev->next;
6713 cpumask_or(d->covered, d->covered, d->tmpmask);
6720 #endif /* CONFIG_NUMA */
6723 /* Free memory allocated for various sched_group structures */
6724 static void free_sched_groups(const struct cpumask *cpu_map,
6725 struct cpumask *nodemask)
6729 for_each_cpu(cpu, cpu_map) {
6730 struct sched_group **sched_group_nodes
6731 = sched_group_nodes_bycpu[cpu];
6733 if (!sched_group_nodes)
6736 for (i = 0; i < nr_node_ids; i++) {
6737 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6739 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
6740 if (cpumask_empty(nodemask))
6750 if (oldsg != sched_group_nodes[i])
6753 kfree(sched_group_nodes);
6754 sched_group_nodes_bycpu[cpu] = NULL;
6757 #else /* !CONFIG_NUMA */
6758 static void free_sched_groups(const struct cpumask *cpu_map,
6759 struct cpumask *nodemask)
6762 #endif /* CONFIG_NUMA */
6765 * Initialize sched groups cpu_power.
6767 * cpu_power indicates the capacity of sched group, which is used while
6768 * distributing the load between different sched groups in a sched domain.
6769 * Typically cpu_power for all the groups in a sched domain will be same unless
6770 * there are asymmetries in the topology. If there are asymmetries, group
6771 * having more cpu_power will pickup more load compared to the group having
6774 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6776 struct sched_domain *child;
6777 struct sched_group *group;
6781 WARN_ON(!sd || !sd->groups);
6783 if (cpu != group_first_cpu(sd->groups))
6788 sd->groups->cpu_power = 0;
6791 power = SCHED_LOAD_SCALE;
6792 weight = cpumask_weight(sched_domain_span(sd));
6794 * SMT siblings share the power of a single core.
6795 * Usually multiple threads get a better yield out of
6796 * that one core than a single thread would have,
6797 * reflect that in sd->smt_gain.
6799 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
6800 power *= sd->smt_gain;
6802 power >>= SCHED_LOAD_SHIFT;
6804 sd->groups->cpu_power += power;
6809 * Add cpu_power of each child group to this groups cpu_power.
6811 group = child->groups;
6813 sd->groups->cpu_power += group->cpu_power;
6814 group = group->next;
6815 } while (group != child->groups);
6819 * Initializers for schedule domains
6820 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6823 #ifdef CONFIG_SCHED_DEBUG
6824 # define SD_INIT_NAME(sd, type) sd->name = #type
6826 # define SD_INIT_NAME(sd, type) do { } while (0)
6829 #define SD_INIT(sd, type) sd_init_##type(sd)
6831 #define SD_INIT_FUNC(type) \
6832 static noinline void sd_init_##type(struct sched_domain *sd) \
6834 memset(sd, 0, sizeof(*sd)); \
6835 *sd = SD_##type##_INIT; \
6836 sd->level = SD_LV_##type; \
6837 SD_INIT_NAME(sd, type); \
6842 SD_INIT_FUNC(ALLNODES)
6845 #ifdef CONFIG_SCHED_SMT
6846 SD_INIT_FUNC(SIBLING)
6848 #ifdef CONFIG_SCHED_MC
6852 static int default_relax_domain_level = -1;
6854 static int __init setup_relax_domain_level(char *str)
6858 val = simple_strtoul(str, NULL, 0);
6859 if (val < SD_LV_MAX)
6860 default_relax_domain_level = val;
6864 __setup("relax_domain_level=", setup_relax_domain_level);
6866 static void set_domain_attribute(struct sched_domain *sd,
6867 struct sched_domain_attr *attr)
6871 if (!attr || attr->relax_domain_level < 0) {
6872 if (default_relax_domain_level < 0)
6875 request = default_relax_domain_level;
6877 request = attr->relax_domain_level;
6878 if (request < sd->level) {
6879 /* turn off idle balance on this domain */
6880 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6882 /* turn on idle balance on this domain */
6883 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6887 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6888 const struct cpumask *cpu_map)
6891 case sa_sched_groups:
6892 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
6893 d->sched_group_nodes = NULL;
6895 free_rootdomain(d->rd); /* fall through */
6897 free_cpumask_var(d->tmpmask); /* fall through */
6898 case sa_send_covered:
6899 free_cpumask_var(d->send_covered); /* fall through */
6900 case sa_this_core_map:
6901 free_cpumask_var(d->this_core_map); /* fall through */
6902 case sa_this_sibling_map:
6903 free_cpumask_var(d->this_sibling_map); /* fall through */
6905 free_cpumask_var(d->nodemask); /* fall through */
6906 case sa_sched_group_nodes:
6908 kfree(d->sched_group_nodes); /* fall through */
6910 free_cpumask_var(d->notcovered); /* fall through */
6912 free_cpumask_var(d->covered); /* fall through */
6914 free_cpumask_var(d->domainspan); /* fall through */
6921 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6922 const struct cpumask *cpu_map)
6925 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
6927 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
6928 return sa_domainspan;
6929 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
6931 /* Allocate the per-node list of sched groups */
6932 d->sched_group_nodes = kcalloc(nr_node_ids,
6933 sizeof(struct sched_group *), GFP_KERNEL);
6934 if (!d->sched_group_nodes) {
6935 printk(KERN_WARNING "Can not alloc sched group node list\n");
6936 return sa_notcovered;
6938 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
6940 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
6941 return sa_sched_group_nodes;
6942 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
6944 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
6945 return sa_this_sibling_map;
6946 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
6947 return sa_this_core_map;
6948 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
6949 return sa_send_covered;
6950 d->rd = alloc_rootdomain();
6952 printk(KERN_WARNING "Cannot alloc root domain\n");
6955 return sa_rootdomain;
6958 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
6959 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
6961 struct sched_domain *sd = NULL;
6963 struct sched_domain *parent;
6966 if (cpumask_weight(cpu_map) >
6967 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
6968 sd = &per_cpu(allnodes_domains, i).sd;
6969 SD_INIT(sd, ALLNODES);
6970 set_domain_attribute(sd, attr);
6971 cpumask_copy(sched_domain_span(sd), cpu_map);
6972 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
6977 sd = &per_cpu(node_domains, i).sd;
6979 set_domain_attribute(sd, attr);
6980 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
6981 sd->parent = parent;
6984 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
6989 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
6990 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6991 struct sched_domain *parent, int i)
6993 struct sched_domain *sd;
6994 sd = &per_cpu(phys_domains, i).sd;
6996 set_domain_attribute(sd, attr);
6997 cpumask_copy(sched_domain_span(sd), d->nodemask);
6998 sd->parent = parent;
7001 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
7005 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
7006 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7007 struct sched_domain *parent, int i)
7009 struct sched_domain *sd = parent;
7010 #ifdef CONFIG_SCHED_MC
7011 sd = &per_cpu(core_domains, i).sd;
7013 set_domain_attribute(sd, attr);
7014 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
7015 sd->parent = parent;
7017 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
7022 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
7023 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7024 struct sched_domain *parent, int i)
7026 struct sched_domain *sd = parent;
7027 #ifdef CONFIG_SCHED_SMT
7028 sd = &per_cpu(cpu_domains, i).sd;
7029 SD_INIT(sd, SIBLING);
7030 set_domain_attribute(sd, attr);
7031 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
7032 sd->parent = parent;
7034 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
7039 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
7040 const struct cpumask *cpu_map, int cpu)
7043 #ifdef CONFIG_SCHED_SMT
7044 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
7045 cpumask_and(d->this_sibling_map, cpu_map,
7046 topology_thread_cpumask(cpu));
7047 if (cpu == cpumask_first(d->this_sibling_map))
7048 init_sched_build_groups(d->this_sibling_map, cpu_map,
7050 d->send_covered, d->tmpmask);
7053 #ifdef CONFIG_SCHED_MC
7054 case SD_LV_MC: /* set up multi-core groups */
7055 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
7056 if (cpu == cpumask_first(d->this_core_map))
7057 init_sched_build_groups(d->this_core_map, cpu_map,
7059 d->send_covered, d->tmpmask);
7062 case SD_LV_CPU: /* set up physical groups */
7063 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
7064 if (!cpumask_empty(d->nodemask))
7065 init_sched_build_groups(d->nodemask, cpu_map,
7067 d->send_covered, d->tmpmask);
7070 case SD_LV_ALLNODES:
7071 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
7072 d->send_covered, d->tmpmask);
7081 * Build sched domains for a given set of cpus and attach the sched domains
7082 * to the individual cpus
7084 static int __build_sched_domains(const struct cpumask *cpu_map,
7085 struct sched_domain_attr *attr)
7087 enum s_alloc alloc_state = sa_none;
7089 struct sched_domain *sd;
7095 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7096 if (alloc_state != sa_rootdomain)
7098 alloc_state = sa_sched_groups;
7101 * Set up domains for cpus specified by the cpu_map.
7103 for_each_cpu(i, cpu_map) {
7104 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
7107 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
7108 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
7109 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
7110 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
7113 for_each_cpu(i, cpu_map) {
7114 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
7115 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
7118 /* Set up physical groups */
7119 for (i = 0; i < nr_node_ids; i++)
7120 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
7123 /* Set up node groups */
7125 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
7127 for (i = 0; i < nr_node_ids; i++)
7128 if (build_numa_sched_groups(&d, cpu_map, i))
7132 /* Calculate CPU power for physical packages and nodes */
7133 #ifdef CONFIG_SCHED_SMT
7134 for_each_cpu(i, cpu_map) {
7135 sd = &per_cpu(cpu_domains, i).sd;
7136 init_sched_groups_power(i, sd);
7139 #ifdef CONFIG_SCHED_MC
7140 for_each_cpu(i, cpu_map) {
7141 sd = &per_cpu(core_domains, i).sd;
7142 init_sched_groups_power(i, sd);
7146 for_each_cpu(i, cpu_map) {
7147 sd = &per_cpu(phys_domains, i).sd;
7148 init_sched_groups_power(i, sd);
7152 for (i = 0; i < nr_node_ids; i++)
7153 init_numa_sched_groups_power(d.sched_group_nodes[i]);
7155 if (d.sd_allnodes) {
7156 struct sched_group *sg;
7158 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7160 init_numa_sched_groups_power(sg);
7164 /* Attach the domains */
7165 for_each_cpu(i, cpu_map) {
7166 #ifdef CONFIG_SCHED_SMT
7167 sd = &per_cpu(cpu_domains, i).sd;
7168 #elif defined(CONFIG_SCHED_MC)
7169 sd = &per_cpu(core_domains, i).sd;
7171 sd = &per_cpu(phys_domains, i).sd;
7173 cpu_attach_domain(sd, d.rd, i);
7176 d.sched_group_nodes = NULL; /* don't free this we still need it */
7177 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
7181 __free_domain_allocs(&d, alloc_state, cpu_map);
7185 static int build_sched_domains(const struct cpumask *cpu_map)
7187 return __build_sched_domains(cpu_map, NULL);
7190 static cpumask_var_t *doms_cur; /* current sched domains */
7191 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7192 static struct sched_domain_attr *dattr_cur;
7193 /* attribues of custom domains in 'doms_cur' */
7196 * Special case: If a kmalloc of a doms_cur partition (array of
7197 * cpumask) fails, then fallback to a single sched domain,
7198 * as determined by the single cpumask fallback_doms.
7200 static cpumask_var_t fallback_doms;
7203 * arch_update_cpu_topology lets virtualized architectures update the
7204 * cpu core maps. It is supposed to return 1 if the topology changed
7205 * or 0 if it stayed the same.
7207 int __attribute__((weak)) arch_update_cpu_topology(void)
7212 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7215 cpumask_var_t *doms;
7217 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7220 for (i = 0; i < ndoms; i++) {
7221 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7222 free_sched_domains(doms, i);
7229 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7232 for (i = 0; i < ndoms; i++)
7233 free_cpumask_var(doms[i]);
7238 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7239 * For now this just excludes isolated cpus, but could be used to
7240 * exclude other special cases in the future.
7242 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7246 arch_update_cpu_topology();
7248 doms_cur = alloc_sched_domains(ndoms_cur);
7250 doms_cur = &fallback_doms;
7251 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7253 err = build_sched_domains(doms_cur[0]);
7254 register_sched_domain_sysctl();
7259 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7260 struct cpumask *tmpmask)
7262 free_sched_groups(cpu_map, tmpmask);
7266 * Detach sched domains from a group of cpus specified in cpu_map
7267 * These cpus will now be attached to the NULL domain
7269 static void detach_destroy_domains(const struct cpumask *cpu_map)
7271 /* Save because hotplug lock held. */
7272 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7275 for_each_cpu(i, cpu_map)
7276 cpu_attach_domain(NULL, &def_root_domain, i);
7277 synchronize_sched();
7278 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7281 /* handle null as "default" */
7282 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7283 struct sched_domain_attr *new, int idx_new)
7285 struct sched_domain_attr tmp;
7292 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7293 new ? (new + idx_new) : &tmp,
7294 sizeof(struct sched_domain_attr));
7298 * Partition sched domains as specified by the 'ndoms_new'
7299 * cpumasks in the array doms_new[] of cpumasks. This compares
7300 * doms_new[] to the current sched domain partitioning, doms_cur[].
7301 * It destroys each deleted domain and builds each new domain.
7303 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7304 * The masks don't intersect (don't overlap.) We should setup one
7305 * sched domain for each mask. CPUs not in any of the cpumasks will
7306 * not be load balanced. If the same cpumask appears both in the
7307 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7310 * The passed in 'doms_new' should be allocated using
7311 * alloc_sched_domains. This routine takes ownership of it and will
7312 * free_sched_domains it when done with it. If the caller failed the
7313 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7314 * and partition_sched_domains() will fallback to the single partition
7315 * 'fallback_doms', it also forces the domains to be rebuilt.
7317 * If doms_new == NULL it will be replaced with cpu_online_mask.
7318 * ndoms_new == 0 is a special case for destroying existing domains,
7319 * and it will not create the default domain.
7321 * Call with hotplug lock held
7323 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7324 struct sched_domain_attr *dattr_new)
7329 mutex_lock(&sched_domains_mutex);
7331 /* always unregister in case we don't destroy any domains */
7332 unregister_sched_domain_sysctl();
7334 /* Let architecture update cpu core mappings. */
7335 new_topology = arch_update_cpu_topology();
7337 n = doms_new ? ndoms_new : 0;
7339 /* Destroy deleted domains */
7340 for (i = 0; i < ndoms_cur; i++) {
7341 for (j = 0; j < n && !new_topology; j++) {
7342 if (cpumask_equal(doms_cur[i], doms_new[j])
7343 && dattrs_equal(dattr_cur, i, dattr_new, j))
7346 /* no match - a current sched domain not in new doms_new[] */
7347 detach_destroy_domains(doms_cur[i]);
7352 if (doms_new == NULL) {
7354 doms_new = &fallback_doms;
7355 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7356 WARN_ON_ONCE(dattr_new);
7359 /* Build new domains */
7360 for (i = 0; i < ndoms_new; i++) {
7361 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7362 if (cpumask_equal(doms_new[i], doms_cur[j])
7363 && dattrs_equal(dattr_new, i, dattr_cur, j))
7366 /* no match - add a new doms_new */
7367 __build_sched_domains(doms_new[i],
7368 dattr_new ? dattr_new + i : NULL);
7373 /* Remember the new sched domains */
7374 if (doms_cur != &fallback_doms)
7375 free_sched_domains(doms_cur, ndoms_cur);
7376 kfree(dattr_cur); /* kfree(NULL) is safe */
7377 doms_cur = doms_new;
7378 dattr_cur = dattr_new;
7379 ndoms_cur = ndoms_new;
7381 register_sched_domain_sysctl();
7383 mutex_unlock(&sched_domains_mutex);
7386 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7387 static void arch_reinit_sched_domains(void)
7391 /* Destroy domains first to force the rebuild */
7392 partition_sched_domains(0, NULL, NULL);
7394 rebuild_sched_domains();
7398 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7400 unsigned int level = 0;
7402 if (sscanf(buf, "%u", &level) != 1)
7406 * level is always be positive so don't check for
7407 * level < POWERSAVINGS_BALANCE_NONE which is 0
7408 * What happens on 0 or 1 byte write,
7409 * need to check for count as well?
7412 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7416 sched_smt_power_savings = level;
7418 sched_mc_power_savings = level;
7420 arch_reinit_sched_domains();
7425 #ifdef CONFIG_SCHED_MC
7426 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7427 struct sysdev_class_attribute *attr,
7430 return sprintf(page, "%u\n", sched_mc_power_savings);
7432 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7433 struct sysdev_class_attribute *attr,
7434 const char *buf, size_t count)
7436 return sched_power_savings_store(buf, count, 0);
7438 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7439 sched_mc_power_savings_show,
7440 sched_mc_power_savings_store);
7443 #ifdef CONFIG_SCHED_SMT
7444 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7445 struct sysdev_class_attribute *attr,
7448 return sprintf(page, "%u\n", sched_smt_power_savings);
7450 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7451 struct sysdev_class_attribute *attr,
7452 const char *buf, size_t count)
7454 return sched_power_savings_store(buf, count, 1);
7456 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7457 sched_smt_power_savings_show,
7458 sched_smt_power_savings_store);
7461 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7465 #ifdef CONFIG_SCHED_SMT
7467 err = sysfs_create_file(&cls->kset.kobj,
7468 &attr_sched_smt_power_savings.attr);
7470 #ifdef CONFIG_SCHED_MC
7471 if (!err && mc_capable())
7472 err = sysfs_create_file(&cls->kset.kobj,
7473 &attr_sched_mc_power_savings.attr);
7477 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7480 * Update cpusets according to cpu_active mask. If cpusets are
7481 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7482 * around partition_sched_domains().
7484 static int __cpuexit cpuset_cpu_active(struct notifier_block *nfb,
7485 unsigned long action, void *hcpu)
7487 switch (action & ~CPU_TASKS_FROZEN) {
7489 case CPU_DOWN_FAILED:
7490 cpuset_update_active_cpus();
7497 static int __cpuexit cpuset_cpu_inactive(struct notifier_block *nfb,
7498 unsigned long action, void *hcpu)
7500 switch (action & ~CPU_TASKS_FROZEN) {
7501 case CPU_DOWN_PREPARE:
7502 cpuset_update_active_cpus();
7509 static int update_runtime(struct notifier_block *nfb,
7510 unsigned long action, void *hcpu)
7512 int cpu = (int)(long)hcpu;
7515 case CPU_DOWN_PREPARE:
7516 case CPU_DOWN_PREPARE_FROZEN:
7517 disable_runtime(cpu_rq(cpu));
7520 case CPU_DOWN_FAILED:
7521 case CPU_DOWN_FAILED_FROZEN:
7523 case CPU_ONLINE_FROZEN:
7524 enable_runtime(cpu_rq(cpu));
7532 void __init sched_init_smp(void)
7534 cpumask_var_t non_isolated_cpus;
7536 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7537 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7539 #if defined(CONFIG_NUMA)
7540 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7542 BUG_ON(sched_group_nodes_bycpu == NULL);
7545 mutex_lock(&sched_domains_mutex);
7546 arch_init_sched_domains(cpu_active_mask);
7547 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7548 if (cpumask_empty(non_isolated_cpus))
7549 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7550 mutex_unlock(&sched_domains_mutex);
7553 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7554 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7556 /* RT runtime code needs to handle some hotplug events */
7557 hotcpu_notifier(update_runtime, 0);
7561 /* Move init over to a non-isolated CPU */
7562 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7564 sched_init_granularity();
7565 free_cpumask_var(non_isolated_cpus);
7567 init_sched_rt_class();
7570 void __init sched_init_smp(void)
7572 sched_init_granularity();
7574 #endif /* CONFIG_SMP */
7576 const_debug unsigned int sysctl_timer_migration = 1;
7578 int in_sched_functions(unsigned long addr)
7580 return in_lock_functions(addr) ||
7581 (addr >= (unsigned long)__sched_text_start
7582 && addr < (unsigned long)__sched_text_end);
7585 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7587 cfs_rq->tasks_timeline = RB_ROOT;
7588 INIT_LIST_HEAD(&cfs_rq->tasks);
7589 #ifdef CONFIG_FAIR_GROUP_SCHED
7592 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7595 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7597 struct rt_prio_array *array;
7600 array = &rt_rq->active;
7601 for (i = 0; i < MAX_RT_PRIO; i++) {
7602 INIT_LIST_HEAD(array->queue + i);
7603 __clear_bit(i, array->bitmap);
7605 /* delimiter for bitsearch: */
7606 __set_bit(MAX_RT_PRIO, array->bitmap);
7608 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7609 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7611 rt_rq->highest_prio.next = MAX_RT_PRIO;
7615 rt_rq->rt_nr_migratory = 0;
7616 rt_rq->overloaded = 0;
7617 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
7621 rt_rq->rt_throttled = 0;
7622 rt_rq->rt_runtime = 0;
7623 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7625 #ifdef CONFIG_RT_GROUP_SCHED
7626 rt_rq->rt_nr_boosted = 0;
7631 #ifdef CONFIG_FAIR_GROUP_SCHED
7632 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7633 struct sched_entity *se, int cpu, int add,
7634 struct sched_entity *parent)
7636 struct rq *rq = cpu_rq(cpu);
7637 tg->cfs_rq[cpu] = cfs_rq;
7638 init_cfs_rq(cfs_rq, rq);
7641 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7644 /* se could be NULL for init_task_group */
7649 se->cfs_rq = &rq->cfs;
7651 se->cfs_rq = parent->my_q;
7654 se->load.weight = tg->shares;
7655 se->load.inv_weight = 0;
7656 se->parent = parent;
7660 #ifdef CONFIG_RT_GROUP_SCHED
7661 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7662 struct sched_rt_entity *rt_se, int cpu, int add,
7663 struct sched_rt_entity *parent)
7665 struct rq *rq = cpu_rq(cpu);
7667 tg->rt_rq[cpu] = rt_rq;
7668 init_rt_rq(rt_rq, rq);
7670 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7672 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7674 tg->rt_se[cpu] = rt_se;
7679 rt_se->rt_rq = &rq->rt;
7681 rt_se->rt_rq = parent->my_q;
7683 rt_se->my_q = rt_rq;
7684 rt_se->parent = parent;
7685 INIT_LIST_HEAD(&rt_se->run_list);
7689 void __init sched_init(void)
7692 unsigned long alloc_size = 0, ptr;
7694 #ifdef CONFIG_FAIR_GROUP_SCHED
7695 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7697 #ifdef CONFIG_RT_GROUP_SCHED
7698 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7700 #ifdef CONFIG_CPUMASK_OFFSTACK
7701 alloc_size += num_possible_cpus() * cpumask_size();
7704 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7706 #ifdef CONFIG_FAIR_GROUP_SCHED
7707 init_task_group.se = (struct sched_entity **)ptr;
7708 ptr += nr_cpu_ids * sizeof(void **);
7710 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7711 ptr += nr_cpu_ids * sizeof(void **);
7713 #endif /* CONFIG_FAIR_GROUP_SCHED */
7714 #ifdef CONFIG_RT_GROUP_SCHED
7715 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7716 ptr += nr_cpu_ids * sizeof(void **);
7718 init_task_group.rt_rq = (struct rt_rq **)ptr;
7719 ptr += nr_cpu_ids * sizeof(void **);
7721 #endif /* CONFIG_RT_GROUP_SCHED */
7722 #ifdef CONFIG_CPUMASK_OFFSTACK
7723 for_each_possible_cpu(i) {
7724 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
7725 ptr += cpumask_size();
7727 #endif /* CONFIG_CPUMASK_OFFSTACK */
7731 init_defrootdomain();
7734 init_rt_bandwidth(&def_rt_bandwidth,
7735 global_rt_period(), global_rt_runtime());
7737 #ifdef CONFIG_RT_GROUP_SCHED
7738 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7739 global_rt_period(), global_rt_runtime());
7740 #endif /* CONFIG_RT_GROUP_SCHED */
7742 #ifdef CONFIG_CGROUP_SCHED
7743 list_add(&init_task_group.list, &task_groups);
7744 INIT_LIST_HEAD(&init_task_group.children);
7746 #endif /* CONFIG_CGROUP_SCHED */
7748 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
7749 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
7750 __alignof__(unsigned long));
7752 for_each_possible_cpu(i) {
7756 raw_spin_lock_init(&rq->lock);
7758 rq->calc_load_active = 0;
7759 rq->calc_load_update = jiffies + LOAD_FREQ;
7760 init_cfs_rq(&rq->cfs, rq);
7761 init_rt_rq(&rq->rt, rq);
7762 #ifdef CONFIG_FAIR_GROUP_SCHED
7763 init_task_group.shares = init_task_group_load;
7764 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7765 #ifdef CONFIG_CGROUP_SCHED
7767 * How much cpu bandwidth does init_task_group get?
7769 * In case of task-groups formed thr' the cgroup filesystem, it
7770 * gets 100% of the cpu resources in the system. This overall
7771 * system cpu resource is divided among the tasks of
7772 * init_task_group and its child task-groups in a fair manner,
7773 * based on each entity's (task or task-group's) weight
7774 * (se->load.weight).
7776 * In other words, if init_task_group has 10 tasks of weight
7777 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7778 * then A0's share of the cpu resource is:
7780 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7782 * We achieve this by letting init_task_group's tasks sit
7783 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7785 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
7787 #endif /* CONFIG_FAIR_GROUP_SCHED */
7789 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7790 #ifdef CONFIG_RT_GROUP_SCHED
7791 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7792 #ifdef CONFIG_CGROUP_SCHED
7793 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
7797 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7798 rq->cpu_load[j] = 0;
7800 rq->last_load_update_tick = jiffies;
7805 rq->cpu_power = SCHED_LOAD_SCALE;
7806 rq->post_schedule = 0;
7807 rq->active_balance = 0;
7808 rq->next_balance = jiffies;
7813 rq->avg_idle = 2*sysctl_sched_migration_cost;
7814 rq_attach_root(rq, &def_root_domain);
7816 rq->nohz_balance_kick = 0;
7817 init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb, i));
7821 atomic_set(&rq->nr_iowait, 0);
7824 set_load_weight(&init_task);
7826 #ifdef CONFIG_PREEMPT_NOTIFIERS
7827 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7831 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7834 #ifdef CONFIG_RT_MUTEXES
7835 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
7839 * The boot idle thread does lazy MMU switching as well:
7841 atomic_inc(&init_mm.mm_count);
7842 enter_lazy_tlb(&init_mm, current);
7845 * Make us the idle thread. Technically, schedule() should not be
7846 * called from this thread, however somewhere below it might be,
7847 * but because we are the idle thread, we just pick up running again
7848 * when this runqueue becomes "idle".
7850 init_idle(current, smp_processor_id());
7852 calc_load_update = jiffies + LOAD_FREQ;
7855 * During early bootup we pretend to be a normal task:
7857 current->sched_class = &fair_sched_class;
7859 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
7860 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
7863 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
7864 alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT);
7865 atomic_set(&nohz.load_balancer, nr_cpu_ids);
7866 atomic_set(&nohz.first_pick_cpu, nr_cpu_ids);
7867 atomic_set(&nohz.second_pick_cpu, nr_cpu_ids);
7869 /* May be allocated at isolcpus cmdline parse time */
7870 if (cpu_isolated_map == NULL)
7871 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7876 scheduler_running = 1;
7879 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7880 static inline int preempt_count_equals(int preempt_offset)
7882 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7884 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
7887 void __might_sleep(const char *file, int line, int preempt_offset)
7890 static unsigned long prev_jiffy; /* ratelimiting */
7892 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7893 system_state != SYSTEM_RUNNING || oops_in_progress)
7895 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7897 prev_jiffy = jiffies;
7900 "BUG: sleeping function called from invalid context at %s:%d\n",
7903 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7904 in_atomic(), irqs_disabled(),
7905 current->pid, current->comm);
7907 debug_show_held_locks(current);
7908 if (irqs_disabled())
7909 print_irqtrace_events(current);
7913 EXPORT_SYMBOL(__might_sleep);
7916 #ifdef CONFIG_MAGIC_SYSRQ
7917 static void normalize_task(struct rq *rq, struct task_struct *p)
7921 on_rq = p->se.on_rq;
7923 deactivate_task(rq, p, 0);
7924 __setscheduler(rq, p, SCHED_NORMAL, 0);
7926 activate_task(rq, p, 0);
7927 resched_task(rq->curr);
7931 void normalize_rt_tasks(void)
7933 struct task_struct *g, *p;
7934 unsigned long flags;
7937 read_lock_irqsave(&tasklist_lock, flags);
7938 do_each_thread(g, p) {
7940 * Only normalize user tasks:
7945 p->se.exec_start = 0;
7946 #ifdef CONFIG_SCHEDSTATS
7947 p->se.statistics.wait_start = 0;
7948 p->se.statistics.sleep_start = 0;
7949 p->se.statistics.block_start = 0;
7954 * Renice negative nice level userspace
7957 if (TASK_NICE(p) < 0 && p->mm)
7958 set_user_nice(p, 0);
7962 raw_spin_lock(&p->pi_lock);
7963 rq = __task_rq_lock(p);
7965 normalize_task(rq, p);
7967 __task_rq_unlock(rq);
7968 raw_spin_unlock(&p->pi_lock);
7969 } while_each_thread(g, p);
7971 read_unlock_irqrestore(&tasklist_lock, flags);
7974 #endif /* CONFIG_MAGIC_SYSRQ */
7976 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7978 * These functions are only useful for the IA64 MCA handling, or kdb.
7980 * They can only be called when the whole system has been
7981 * stopped - every CPU needs to be quiescent, and no scheduling
7982 * activity can take place. Using them for anything else would
7983 * be a serious bug, and as a result, they aren't even visible
7984 * under any other configuration.
7988 * curr_task - return the current task for a given cpu.
7989 * @cpu: the processor in question.
7991 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7993 struct task_struct *curr_task(int cpu)
7995 return cpu_curr(cpu);
7998 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8002 * set_curr_task - set the current task for a given cpu.
8003 * @cpu: the processor in question.
8004 * @p: the task pointer to set.
8006 * Description: This function must only be used when non-maskable interrupts
8007 * are serviced on a separate stack. It allows the architecture to switch the
8008 * notion of the current task on a cpu in a non-blocking manner. This function
8009 * must be called with all CPU's synchronized, and interrupts disabled, the
8010 * and caller must save the original value of the current task (see
8011 * curr_task() above) and restore that value before reenabling interrupts and
8012 * re-starting the system.
8014 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8016 void set_curr_task(int cpu, struct task_struct *p)
8023 #ifdef CONFIG_FAIR_GROUP_SCHED
8024 static void free_fair_sched_group(struct task_group *tg)
8028 for_each_possible_cpu(i) {
8030 kfree(tg->cfs_rq[i]);
8040 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8042 struct cfs_rq *cfs_rq;
8043 struct sched_entity *se;
8047 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8050 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8054 tg->shares = NICE_0_LOAD;
8056 for_each_possible_cpu(i) {
8059 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8060 GFP_KERNEL, cpu_to_node(i));
8064 se = kzalloc_node(sizeof(struct sched_entity),
8065 GFP_KERNEL, cpu_to_node(i));
8069 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
8080 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8082 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8083 &cpu_rq(cpu)->leaf_cfs_rq_list);
8086 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8088 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8090 #else /* !CONFG_FAIR_GROUP_SCHED */
8091 static inline void free_fair_sched_group(struct task_group *tg)
8096 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8101 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8105 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8108 #endif /* CONFIG_FAIR_GROUP_SCHED */
8110 #ifdef CONFIG_RT_GROUP_SCHED
8111 static void free_rt_sched_group(struct task_group *tg)
8115 destroy_rt_bandwidth(&tg->rt_bandwidth);
8117 for_each_possible_cpu(i) {
8119 kfree(tg->rt_rq[i]);
8121 kfree(tg->rt_se[i]);
8129 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8131 struct rt_rq *rt_rq;
8132 struct sched_rt_entity *rt_se;
8136 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8139 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8143 init_rt_bandwidth(&tg->rt_bandwidth,
8144 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8146 for_each_possible_cpu(i) {
8149 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8150 GFP_KERNEL, cpu_to_node(i));
8154 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8155 GFP_KERNEL, cpu_to_node(i));
8159 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
8170 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8172 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8173 &cpu_rq(cpu)->leaf_rt_rq_list);
8176 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8178 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8180 #else /* !CONFIG_RT_GROUP_SCHED */
8181 static inline void free_rt_sched_group(struct task_group *tg)
8186 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8191 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8195 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8198 #endif /* CONFIG_RT_GROUP_SCHED */
8200 #ifdef CONFIG_CGROUP_SCHED
8201 static void free_sched_group(struct task_group *tg)
8203 free_fair_sched_group(tg);
8204 free_rt_sched_group(tg);
8208 /* allocate runqueue etc for a new task group */
8209 struct task_group *sched_create_group(struct task_group *parent)
8211 struct task_group *tg;
8212 unsigned long flags;
8215 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8217 return ERR_PTR(-ENOMEM);
8219 if (!alloc_fair_sched_group(tg, parent))
8222 if (!alloc_rt_sched_group(tg, parent))
8225 spin_lock_irqsave(&task_group_lock, flags);
8226 for_each_possible_cpu(i) {
8227 register_fair_sched_group(tg, i);
8228 register_rt_sched_group(tg, i);
8230 list_add_rcu(&tg->list, &task_groups);
8232 WARN_ON(!parent); /* root should already exist */
8234 tg->parent = parent;
8235 INIT_LIST_HEAD(&tg->children);
8236 list_add_rcu(&tg->siblings, &parent->children);
8237 spin_unlock_irqrestore(&task_group_lock, flags);
8242 free_sched_group(tg);
8243 return ERR_PTR(-ENOMEM);
8246 /* rcu callback to free various structures associated with a task group */
8247 static void free_sched_group_rcu(struct rcu_head *rhp)
8249 /* now it should be safe to free those cfs_rqs */
8250 free_sched_group(container_of(rhp, struct task_group, rcu));
8253 /* Destroy runqueue etc associated with a task group */
8254 void sched_destroy_group(struct task_group *tg)
8256 unsigned long flags;
8259 spin_lock_irqsave(&task_group_lock, flags);
8260 for_each_possible_cpu(i) {
8261 unregister_fair_sched_group(tg, i);
8262 unregister_rt_sched_group(tg, i);
8264 list_del_rcu(&tg->list);
8265 list_del_rcu(&tg->siblings);
8266 spin_unlock_irqrestore(&task_group_lock, flags);
8268 /* wait for possible concurrent references to cfs_rqs complete */
8269 call_rcu(&tg->rcu, free_sched_group_rcu);
8272 /* change task's runqueue when it moves between groups.
8273 * The caller of this function should have put the task in its new group
8274 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8275 * reflect its new group.
8277 void sched_move_task(struct task_struct *tsk)
8280 unsigned long flags;
8283 rq = task_rq_lock(tsk, &flags);
8285 running = task_current(rq, tsk);
8286 on_rq = tsk->se.on_rq;
8289 dequeue_task(rq, tsk, 0);
8290 if (unlikely(running))
8291 tsk->sched_class->put_prev_task(rq, tsk);
8293 set_task_rq(tsk, task_cpu(tsk));
8295 #ifdef CONFIG_FAIR_GROUP_SCHED
8296 if (tsk->sched_class->moved_group)
8297 tsk->sched_class->moved_group(tsk, on_rq);
8300 if (unlikely(running))
8301 tsk->sched_class->set_curr_task(rq);
8303 enqueue_task(rq, tsk, 0);
8305 task_rq_unlock(rq, &flags);
8307 #endif /* CONFIG_CGROUP_SCHED */
8309 #ifdef CONFIG_FAIR_GROUP_SCHED
8310 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8312 struct cfs_rq *cfs_rq = se->cfs_rq;
8317 dequeue_entity(cfs_rq, se, 0);
8319 se->load.weight = shares;
8320 se->load.inv_weight = 0;
8323 enqueue_entity(cfs_rq, se, 0);
8326 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8328 struct cfs_rq *cfs_rq = se->cfs_rq;
8329 struct rq *rq = cfs_rq->rq;
8330 unsigned long flags;
8332 raw_spin_lock_irqsave(&rq->lock, flags);
8333 __set_se_shares(se, shares);
8334 raw_spin_unlock_irqrestore(&rq->lock, flags);
8337 static DEFINE_MUTEX(shares_mutex);
8339 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8342 unsigned long flags;
8345 * We can't change the weight of the root cgroup.
8350 if (shares < MIN_SHARES)
8351 shares = MIN_SHARES;
8352 else if (shares > MAX_SHARES)
8353 shares = MAX_SHARES;
8355 mutex_lock(&shares_mutex);
8356 if (tg->shares == shares)
8359 spin_lock_irqsave(&task_group_lock, flags);
8360 for_each_possible_cpu(i)
8361 unregister_fair_sched_group(tg, i);
8362 list_del_rcu(&tg->siblings);
8363 spin_unlock_irqrestore(&task_group_lock, flags);
8365 /* wait for any ongoing reference to this group to finish */
8366 synchronize_sched();
8369 * Now we are free to modify the group's share on each cpu
8370 * w/o tripping rebalance_share or load_balance_fair.
8372 tg->shares = shares;
8373 for_each_possible_cpu(i) {
8377 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8378 set_se_shares(tg->se[i], shares);
8382 * Enable load balance activity on this group, by inserting it back on
8383 * each cpu's rq->leaf_cfs_rq_list.
8385 spin_lock_irqsave(&task_group_lock, flags);
8386 for_each_possible_cpu(i)
8387 register_fair_sched_group(tg, i);
8388 list_add_rcu(&tg->siblings, &tg->parent->children);
8389 spin_unlock_irqrestore(&task_group_lock, flags);
8391 mutex_unlock(&shares_mutex);
8395 unsigned long sched_group_shares(struct task_group *tg)
8401 #ifdef CONFIG_RT_GROUP_SCHED
8403 * Ensure that the real time constraints are schedulable.
8405 static DEFINE_MUTEX(rt_constraints_mutex);
8407 static unsigned long to_ratio(u64 period, u64 runtime)
8409 if (runtime == RUNTIME_INF)
8412 return div64_u64(runtime << 20, period);
8415 /* Must be called with tasklist_lock held */
8416 static inline int tg_has_rt_tasks(struct task_group *tg)
8418 struct task_struct *g, *p;
8420 do_each_thread(g, p) {
8421 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8423 } while_each_thread(g, p);
8428 struct rt_schedulable_data {
8429 struct task_group *tg;
8434 static int tg_schedulable(struct task_group *tg, void *data)
8436 struct rt_schedulable_data *d = data;
8437 struct task_group *child;
8438 unsigned long total, sum = 0;
8439 u64 period, runtime;
8441 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8442 runtime = tg->rt_bandwidth.rt_runtime;
8445 period = d->rt_period;
8446 runtime = d->rt_runtime;
8450 * Cannot have more runtime than the period.
8452 if (runtime > period && runtime != RUNTIME_INF)
8456 * Ensure we don't starve existing RT tasks.
8458 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8461 total = to_ratio(period, runtime);
8464 * Nobody can have more than the global setting allows.
8466 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8470 * The sum of our children's runtime should not exceed our own.
8472 list_for_each_entry_rcu(child, &tg->children, siblings) {
8473 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8474 runtime = child->rt_bandwidth.rt_runtime;
8476 if (child == d->tg) {
8477 period = d->rt_period;
8478 runtime = d->rt_runtime;
8481 sum += to_ratio(period, runtime);
8490 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8492 struct rt_schedulable_data data = {
8494 .rt_period = period,
8495 .rt_runtime = runtime,
8498 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8501 static int tg_set_bandwidth(struct task_group *tg,
8502 u64 rt_period, u64 rt_runtime)
8506 mutex_lock(&rt_constraints_mutex);
8507 read_lock(&tasklist_lock);
8508 err = __rt_schedulable(tg, rt_period, rt_runtime);
8512 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8513 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8514 tg->rt_bandwidth.rt_runtime = rt_runtime;
8516 for_each_possible_cpu(i) {
8517 struct rt_rq *rt_rq = tg->rt_rq[i];
8519 raw_spin_lock(&rt_rq->rt_runtime_lock);
8520 rt_rq->rt_runtime = rt_runtime;
8521 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8523 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8525 read_unlock(&tasklist_lock);
8526 mutex_unlock(&rt_constraints_mutex);
8531 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8533 u64 rt_runtime, rt_period;
8535 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8536 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8537 if (rt_runtime_us < 0)
8538 rt_runtime = RUNTIME_INF;
8540 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8543 long sched_group_rt_runtime(struct task_group *tg)
8547 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8550 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8551 do_div(rt_runtime_us, NSEC_PER_USEC);
8552 return rt_runtime_us;
8555 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8557 u64 rt_runtime, rt_period;
8559 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8560 rt_runtime = tg->rt_bandwidth.rt_runtime;
8565 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8568 long sched_group_rt_period(struct task_group *tg)
8572 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8573 do_div(rt_period_us, NSEC_PER_USEC);
8574 return rt_period_us;
8577 static int sched_rt_global_constraints(void)
8579 u64 runtime, period;
8582 if (sysctl_sched_rt_period <= 0)
8585 runtime = global_rt_runtime();
8586 period = global_rt_period();
8589 * Sanity check on the sysctl variables.
8591 if (runtime > period && runtime != RUNTIME_INF)
8594 mutex_lock(&rt_constraints_mutex);
8595 read_lock(&tasklist_lock);
8596 ret = __rt_schedulable(NULL, 0, 0);
8597 read_unlock(&tasklist_lock);
8598 mutex_unlock(&rt_constraints_mutex);
8603 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8605 /* Don't accept realtime tasks when there is no way for them to run */
8606 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8612 #else /* !CONFIG_RT_GROUP_SCHED */
8613 static int sched_rt_global_constraints(void)
8615 unsigned long flags;
8618 if (sysctl_sched_rt_period <= 0)
8622 * There's always some RT tasks in the root group
8623 * -- migration, kstopmachine etc..
8625 if (sysctl_sched_rt_runtime == 0)
8628 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8629 for_each_possible_cpu(i) {
8630 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8632 raw_spin_lock(&rt_rq->rt_runtime_lock);
8633 rt_rq->rt_runtime = global_rt_runtime();
8634 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8636 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8640 #endif /* CONFIG_RT_GROUP_SCHED */
8642 int sched_rt_handler(struct ctl_table *table, int write,
8643 void __user *buffer, size_t *lenp,
8647 int old_period, old_runtime;
8648 static DEFINE_MUTEX(mutex);
8651 old_period = sysctl_sched_rt_period;
8652 old_runtime = sysctl_sched_rt_runtime;
8654 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8656 if (!ret && write) {
8657 ret = sched_rt_global_constraints();
8659 sysctl_sched_rt_period = old_period;
8660 sysctl_sched_rt_runtime = old_runtime;
8662 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8663 def_rt_bandwidth.rt_period =
8664 ns_to_ktime(global_rt_period());
8667 mutex_unlock(&mutex);
8672 #ifdef CONFIG_CGROUP_SCHED
8674 /* return corresponding task_group object of a cgroup */
8675 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8677 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8678 struct task_group, css);
8681 static struct cgroup_subsys_state *
8682 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8684 struct task_group *tg, *parent;
8686 if (!cgrp->parent) {
8687 /* This is early initialization for the top cgroup */
8688 return &init_task_group.css;
8691 parent = cgroup_tg(cgrp->parent);
8692 tg = sched_create_group(parent);
8694 return ERR_PTR(-ENOMEM);
8700 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8702 struct task_group *tg = cgroup_tg(cgrp);
8704 sched_destroy_group(tg);
8708 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8710 #ifdef CONFIG_RT_GROUP_SCHED
8711 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
8714 /* We don't support RT-tasks being in separate groups */
8715 if (tsk->sched_class != &fair_sched_class)
8722 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8723 struct task_struct *tsk, bool threadgroup)
8725 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
8729 struct task_struct *c;
8731 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8732 retval = cpu_cgroup_can_attach_task(cgrp, c);
8744 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8745 struct cgroup *old_cont, struct task_struct *tsk,
8748 sched_move_task(tsk);
8750 struct task_struct *c;
8752 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8759 #ifdef CONFIG_FAIR_GROUP_SCHED
8760 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8763 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8766 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8768 struct task_group *tg = cgroup_tg(cgrp);
8770 return (u64) tg->shares;
8772 #endif /* CONFIG_FAIR_GROUP_SCHED */
8774 #ifdef CONFIG_RT_GROUP_SCHED
8775 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8778 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8781 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8783 return sched_group_rt_runtime(cgroup_tg(cgrp));
8786 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8789 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8792 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8794 return sched_group_rt_period(cgroup_tg(cgrp));
8796 #endif /* CONFIG_RT_GROUP_SCHED */
8798 static struct cftype cpu_files[] = {
8799 #ifdef CONFIG_FAIR_GROUP_SCHED
8802 .read_u64 = cpu_shares_read_u64,
8803 .write_u64 = cpu_shares_write_u64,
8806 #ifdef CONFIG_RT_GROUP_SCHED
8808 .name = "rt_runtime_us",
8809 .read_s64 = cpu_rt_runtime_read,
8810 .write_s64 = cpu_rt_runtime_write,
8813 .name = "rt_period_us",
8814 .read_u64 = cpu_rt_period_read_uint,
8815 .write_u64 = cpu_rt_period_write_uint,
8820 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8822 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8825 struct cgroup_subsys cpu_cgroup_subsys = {
8827 .create = cpu_cgroup_create,
8828 .destroy = cpu_cgroup_destroy,
8829 .can_attach = cpu_cgroup_can_attach,
8830 .attach = cpu_cgroup_attach,
8831 .populate = cpu_cgroup_populate,
8832 .subsys_id = cpu_cgroup_subsys_id,
8836 #endif /* CONFIG_CGROUP_SCHED */
8838 #ifdef CONFIG_CGROUP_CPUACCT
8841 * CPU accounting code for task groups.
8843 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8844 * (balbir@in.ibm.com).
8847 /* track cpu usage of a group of tasks and its child groups */
8849 struct cgroup_subsys_state css;
8850 /* cpuusage holds pointer to a u64-type object on every cpu */
8851 u64 __percpu *cpuusage;
8852 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
8853 struct cpuacct *parent;
8856 struct cgroup_subsys cpuacct_subsys;
8858 /* return cpu accounting group corresponding to this container */
8859 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8861 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8862 struct cpuacct, css);
8865 /* return cpu accounting group to which this task belongs */
8866 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8868 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8869 struct cpuacct, css);
8872 /* create a new cpu accounting group */
8873 static struct cgroup_subsys_state *cpuacct_create(
8874 struct cgroup_subsys *ss, struct cgroup *cgrp)
8876 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8882 ca->cpuusage = alloc_percpu(u64);
8886 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8887 if (percpu_counter_init(&ca->cpustat[i], 0))
8888 goto out_free_counters;
8891 ca->parent = cgroup_ca(cgrp->parent);
8897 percpu_counter_destroy(&ca->cpustat[i]);
8898 free_percpu(ca->cpuusage);
8902 return ERR_PTR(-ENOMEM);
8905 /* destroy an existing cpu accounting group */
8907 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8909 struct cpuacct *ca = cgroup_ca(cgrp);
8912 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8913 percpu_counter_destroy(&ca->cpustat[i]);
8914 free_percpu(ca->cpuusage);
8918 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
8920 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8923 #ifndef CONFIG_64BIT
8925 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8927 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8929 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8937 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
8939 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8941 #ifndef CONFIG_64BIT
8943 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8945 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8947 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8953 /* return total cpu usage (in nanoseconds) of a group */
8954 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8956 struct cpuacct *ca = cgroup_ca(cgrp);
8957 u64 totalcpuusage = 0;
8960 for_each_present_cpu(i)
8961 totalcpuusage += cpuacct_cpuusage_read(ca, i);
8963 return totalcpuusage;
8966 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8969 struct cpuacct *ca = cgroup_ca(cgrp);
8978 for_each_present_cpu(i)
8979 cpuacct_cpuusage_write(ca, i, 0);
8985 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
8988 struct cpuacct *ca = cgroup_ca(cgroup);
8992 for_each_present_cpu(i) {
8993 percpu = cpuacct_cpuusage_read(ca, i);
8994 seq_printf(m, "%llu ", (unsigned long long) percpu);
8996 seq_printf(m, "\n");
9000 static const char *cpuacct_stat_desc[] = {
9001 [CPUACCT_STAT_USER] = "user",
9002 [CPUACCT_STAT_SYSTEM] = "system",
9005 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
9006 struct cgroup_map_cb *cb)
9008 struct cpuacct *ca = cgroup_ca(cgrp);
9011 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
9012 s64 val = percpu_counter_read(&ca->cpustat[i]);
9013 val = cputime64_to_clock_t(val);
9014 cb->fill(cb, cpuacct_stat_desc[i], val);
9019 static struct cftype files[] = {
9022 .read_u64 = cpuusage_read,
9023 .write_u64 = cpuusage_write,
9026 .name = "usage_percpu",
9027 .read_seq_string = cpuacct_percpu_seq_read,
9031 .read_map = cpuacct_stats_show,
9035 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9037 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9041 * charge this task's execution time to its accounting group.
9043 * called with rq->lock held.
9045 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9050 if (unlikely(!cpuacct_subsys.active))
9053 cpu = task_cpu(tsk);
9059 for (; ca; ca = ca->parent) {
9060 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9061 *cpuusage += cputime;
9068 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9069 * in cputime_t units. As a result, cpuacct_update_stats calls
9070 * percpu_counter_add with values large enough to always overflow the
9071 * per cpu batch limit causing bad SMP scalability.
9073 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9074 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9075 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9078 #define CPUACCT_BATCH \
9079 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9081 #define CPUACCT_BATCH 0
9085 * Charge the system/user time to the task's accounting group.
9087 static void cpuacct_update_stats(struct task_struct *tsk,
9088 enum cpuacct_stat_index idx, cputime_t val)
9091 int batch = CPUACCT_BATCH;
9093 if (unlikely(!cpuacct_subsys.active))
9100 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9106 struct cgroup_subsys cpuacct_subsys = {
9108 .create = cpuacct_create,
9109 .destroy = cpuacct_destroy,
9110 .populate = cpuacct_populate,
9111 .subsys_id = cpuacct_subsys_id,
9113 #endif /* CONFIG_CGROUP_CPUACCT */
9117 void synchronize_sched_expedited(void)
9121 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9123 #else /* #ifndef CONFIG_SMP */
9125 static atomic_t synchronize_sched_expedited_count = ATOMIC_INIT(0);
9127 static int synchronize_sched_expedited_cpu_stop(void *data)
9130 * There must be a full memory barrier on each affected CPU
9131 * between the time that try_stop_cpus() is called and the
9132 * time that it returns.
9134 * In the current initial implementation of cpu_stop, the
9135 * above condition is already met when the control reaches
9136 * this point and the following smp_mb() is not strictly
9137 * necessary. Do smp_mb() anyway for documentation and
9138 * robustness against future implementation changes.
9140 smp_mb(); /* See above comment block. */
9145 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
9146 * approach to force grace period to end quickly. This consumes
9147 * significant time on all CPUs, and is thus not recommended for
9148 * any sort of common-case code.
9150 * Note that it is illegal to call this function while holding any
9151 * lock that is acquired by a CPU-hotplug notifier. Failing to
9152 * observe this restriction will result in deadlock.
9154 void synchronize_sched_expedited(void)
9156 int snap, trycount = 0;
9158 smp_mb(); /* ensure prior mod happens before capturing snap. */
9159 snap = atomic_read(&synchronize_sched_expedited_count) + 1;
9161 while (try_stop_cpus(cpu_online_mask,
9162 synchronize_sched_expedited_cpu_stop,
9165 if (trycount++ < 10)
9166 udelay(trycount * num_online_cpus());
9168 synchronize_sched();
9171 if (atomic_read(&synchronize_sched_expedited_count) - snap > 0) {
9172 smp_mb(); /* ensure test happens before caller kfree */
9177 atomic_inc(&synchronize_sched_expedited_count);
9178 smp_mb__after_atomic_inc(); /* ensure post-GP actions seen after GP. */
9181 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9183 #endif /* #else #ifndef CONFIG_SMP */