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;
429 struct cpupri cpupri;
433 * By default the system creates a single root-domain with all cpus as
434 * members (mimicking the global state we have today).
436 static struct root_domain def_root_domain;
438 #endif /* CONFIG_SMP */
441 * This is the main, per-CPU runqueue data structure.
443 * Locking rule: those places that want to lock multiple runqueues
444 * (such as the load balancing or the thread migration code), lock
445 * acquire operations must be ordered by ascending &runqueue.
452 * nr_running and cpu_load should be in the same cacheline because
453 * remote CPUs use both these fields when doing load calculation.
455 unsigned long nr_running;
456 #define CPU_LOAD_IDX_MAX 5
457 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
458 unsigned long last_load_update_tick;
461 unsigned char nohz_balance_kick;
463 unsigned int skip_clock_update;
465 /* capture load from *all* tasks on this cpu: */
466 struct load_weight load;
467 unsigned long nr_load_updates;
473 #ifdef CONFIG_FAIR_GROUP_SCHED
474 /* list of leaf cfs_rq on this cpu: */
475 struct list_head leaf_cfs_rq_list;
477 #ifdef CONFIG_RT_GROUP_SCHED
478 struct list_head leaf_rt_rq_list;
482 * This is part of a global counter where only the total sum
483 * over all CPUs matters. A task can increase this counter on
484 * one CPU and if it got migrated afterwards it may decrease
485 * it on another CPU. Always updated under the runqueue lock:
487 unsigned long nr_uninterruptible;
489 struct task_struct *curr, *idle, *stop;
490 unsigned long next_balance;
491 struct mm_struct *prev_mm;
498 struct root_domain *rd;
499 struct sched_domain *sd;
501 unsigned long cpu_power;
503 unsigned char idle_at_tick;
504 /* For active balancing */
508 struct cpu_stop_work active_balance_work;
509 /* cpu of this runqueue: */
513 unsigned long avg_load_per_task;
521 /* calc_load related fields */
522 unsigned long calc_load_update;
523 long calc_load_active;
525 #ifdef CONFIG_SCHED_HRTICK
527 int hrtick_csd_pending;
528 struct call_single_data hrtick_csd;
530 struct hrtimer hrtick_timer;
533 #ifdef CONFIG_SCHEDSTATS
535 struct sched_info rq_sched_info;
536 unsigned long long rq_cpu_time;
537 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
539 /* sys_sched_yield() stats */
540 unsigned int yld_count;
542 /* schedule() stats */
543 unsigned int sched_switch;
544 unsigned int sched_count;
545 unsigned int sched_goidle;
547 /* try_to_wake_up() stats */
548 unsigned int ttwu_count;
549 unsigned int ttwu_local;
552 unsigned int bkl_count;
556 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
559 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
561 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
564 * A queue event has occurred, and we're going to schedule. In
565 * this case, we can save a useless back to back clock update.
567 if (test_tsk_need_resched(p))
568 rq->skip_clock_update = 1;
571 static inline int cpu_of(struct rq *rq)
580 #define rcu_dereference_check_sched_domain(p) \
581 rcu_dereference_check((p), \
582 rcu_read_lock_sched_held() || \
583 lockdep_is_held(&sched_domains_mutex))
586 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
587 * See detach_destroy_domains: synchronize_sched for details.
589 * The domain tree of any CPU may only be accessed from within
590 * preempt-disabled sections.
592 #define for_each_domain(cpu, __sd) \
593 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
595 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
596 #define this_rq() (&__get_cpu_var(runqueues))
597 #define task_rq(p) cpu_rq(task_cpu(p))
598 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
599 #define raw_rq() (&__raw_get_cpu_var(runqueues))
601 #ifdef CONFIG_CGROUP_SCHED
604 * Return the group to which this tasks belongs.
606 * We use task_subsys_state_check() and extend the RCU verification
607 * with lockdep_is_held(&task_rq(p)->lock) because cpu_cgroup_attach()
608 * holds that lock for each task it moves into the cgroup. Therefore
609 * by holding that lock, we pin the task to the current cgroup.
611 static inline struct task_group *task_group(struct task_struct *p)
613 struct cgroup_subsys_state *css;
615 css = task_subsys_state_check(p, cpu_cgroup_subsys_id,
616 lockdep_is_held(&task_rq(p)->lock));
617 return container_of(css, struct task_group, css);
620 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
621 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
623 #ifdef CONFIG_FAIR_GROUP_SCHED
624 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
625 p->se.parent = task_group(p)->se[cpu];
628 #ifdef CONFIG_RT_GROUP_SCHED
629 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
630 p->rt.parent = task_group(p)->rt_se[cpu];
634 #else /* CONFIG_CGROUP_SCHED */
636 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
637 static inline struct task_group *task_group(struct task_struct *p)
642 #endif /* CONFIG_CGROUP_SCHED */
644 inline void update_rq_clock(struct rq *rq)
646 if (!rq->skip_clock_update)
647 rq->clock = sched_clock_cpu(cpu_of(rq));
651 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
653 #ifdef CONFIG_SCHED_DEBUG
654 # define const_debug __read_mostly
656 # define const_debug static const
661 * @cpu: the processor in question.
663 * Returns true if the current cpu runqueue is locked.
664 * This interface allows printk to be called with the runqueue lock
665 * held and know whether or not it is OK to wake up the klogd.
667 int runqueue_is_locked(int cpu)
669 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
673 * Debugging: various feature bits
676 #define SCHED_FEAT(name, enabled) \
677 __SCHED_FEAT_##name ,
680 #include "sched_features.h"
685 #define SCHED_FEAT(name, enabled) \
686 (1UL << __SCHED_FEAT_##name) * enabled |
688 const_debug unsigned int sysctl_sched_features =
689 #include "sched_features.h"
694 #ifdef CONFIG_SCHED_DEBUG
695 #define SCHED_FEAT(name, enabled) \
698 static __read_mostly char *sched_feat_names[] = {
699 #include "sched_features.h"
705 static int sched_feat_show(struct seq_file *m, void *v)
709 for (i = 0; sched_feat_names[i]; i++) {
710 if (!(sysctl_sched_features & (1UL << i)))
712 seq_printf(m, "%s ", sched_feat_names[i]);
720 sched_feat_write(struct file *filp, const char __user *ubuf,
721 size_t cnt, loff_t *ppos)
731 if (copy_from_user(&buf, ubuf, cnt))
737 if (strncmp(buf, "NO_", 3) == 0) {
742 for (i = 0; sched_feat_names[i]; i++) {
743 if (strcmp(cmp, sched_feat_names[i]) == 0) {
745 sysctl_sched_features &= ~(1UL << i);
747 sysctl_sched_features |= (1UL << i);
752 if (!sched_feat_names[i])
760 static int sched_feat_open(struct inode *inode, struct file *filp)
762 return single_open(filp, sched_feat_show, NULL);
765 static const struct file_operations sched_feat_fops = {
766 .open = sched_feat_open,
767 .write = sched_feat_write,
770 .release = single_release,
773 static __init int sched_init_debug(void)
775 debugfs_create_file("sched_features", 0644, NULL, NULL,
780 late_initcall(sched_init_debug);
784 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
787 * Number of tasks to iterate in a single balance run.
788 * Limited because this is done with IRQs disabled.
790 const_debug unsigned int sysctl_sched_nr_migrate = 32;
793 * ratelimit for updating the group shares.
796 unsigned int sysctl_sched_shares_ratelimit = 250000;
797 unsigned int normalized_sysctl_sched_shares_ratelimit = 250000;
800 * Inject some fuzzyness into changing the per-cpu group shares
801 * this avoids remote rq-locks at the expense of fairness.
804 unsigned int sysctl_sched_shares_thresh = 4;
807 * period over which we average the RT time consumption, measured
812 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
815 * period over which we measure -rt task cpu usage in us.
818 unsigned int sysctl_sched_rt_period = 1000000;
820 static __read_mostly int scheduler_running;
823 * part of the period that we allow rt tasks to run in us.
826 int sysctl_sched_rt_runtime = 950000;
828 static inline u64 global_rt_period(void)
830 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
833 static inline u64 global_rt_runtime(void)
835 if (sysctl_sched_rt_runtime < 0)
838 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
841 #ifndef prepare_arch_switch
842 # define prepare_arch_switch(next) do { } while (0)
844 #ifndef finish_arch_switch
845 # define finish_arch_switch(prev) do { } while (0)
848 static inline int task_current(struct rq *rq, struct task_struct *p)
850 return rq->curr == p;
853 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
854 static inline int task_running(struct rq *rq, struct task_struct *p)
856 return task_current(rq, p);
859 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
863 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
865 #ifdef CONFIG_DEBUG_SPINLOCK
866 /* this is a valid case when another task releases the spinlock */
867 rq->lock.owner = current;
870 * If we are tracking spinlock dependencies then we have to
871 * fix up the runqueue lock - which gets 'carried over' from
874 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
876 raw_spin_unlock_irq(&rq->lock);
879 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
880 static inline int task_running(struct rq *rq, struct task_struct *p)
885 return task_current(rq, p);
889 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
893 * We can optimise this out completely for !SMP, because the
894 * SMP rebalancing from interrupt is the only thing that cares
899 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
900 raw_spin_unlock_irq(&rq->lock);
902 raw_spin_unlock(&rq->lock);
906 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
910 * After ->oncpu is cleared, the task can be moved to a different CPU.
911 * We must ensure this doesn't happen until the switch is completely
917 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
921 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
924 * Check whether the task is waking, we use this to synchronize ->cpus_allowed
927 static inline int task_is_waking(struct task_struct *p)
929 return unlikely(p->state == TASK_WAKING);
933 * __task_rq_lock - lock the runqueue a given task resides on.
934 * Must be called interrupts disabled.
936 static inline struct rq *__task_rq_lock(struct task_struct *p)
943 raw_spin_lock(&rq->lock);
944 if (likely(rq == task_rq(p)))
946 raw_spin_unlock(&rq->lock);
951 * task_rq_lock - lock the runqueue a given task resides on and disable
952 * interrupts. Note the ordering: we can safely lookup the task_rq without
953 * explicitly disabling preemption.
955 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
961 local_irq_save(*flags);
963 raw_spin_lock(&rq->lock);
964 if (likely(rq == task_rq(p)))
966 raw_spin_unlock_irqrestore(&rq->lock, *flags);
970 static void __task_rq_unlock(struct rq *rq)
973 raw_spin_unlock(&rq->lock);
976 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
979 raw_spin_unlock_irqrestore(&rq->lock, *flags);
983 * this_rq_lock - lock this runqueue and disable interrupts.
985 static struct rq *this_rq_lock(void)
992 raw_spin_lock(&rq->lock);
997 #ifdef CONFIG_SCHED_HRTICK
999 * Use HR-timers to deliver accurate preemption points.
1001 * Its all a bit involved since we cannot program an hrt while holding the
1002 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1005 * When we get rescheduled we reprogram the hrtick_timer outside of the
1011 * - enabled by features
1012 * - hrtimer is actually high res
1014 static inline int hrtick_enabled(struct rq *rq)
1016 if (!sched_feat(HRTICK))
1018 if (!cpu_active(cpu_of(rq)))
1020 return hrtimer_is_hres_active(&rq->hrtick_timer);
1023 static void hrtick_clear(struct rq *rq)
1025 if (hrtimer_active(&rq->hrtick_timer))
1026 hrtimer_cancel(&rq->hrtick_timer);
1030 * High-resolution timer tick.
1031 * Runs from hardirq context with interrupts disabled.
1033 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1035 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1037 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1039 raw_spin_lock(&rq->lock);
1040 update_rq_clock(rq);
1041 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1042 raw_spin_unlock(&rq->lock);
1044 return HRTIMER_NORESTART;
1049 * called from hardirq (IPI) context
1051 static void __hrtick_start(void *arg)
1053 struct rq *rq = arg;
1055 raw_spin_lock(&rq->lock);
1056 hrtimer_restart(&rq->hrtick_timer);
1057 rq->hrtick_csd_pending = 0;
1058 raw_spin_unlock(&rq->lock);
1062 * Called to set the hrtick timer state.
1064 * called with rq->lock held and irqs disabled
1066 static void hrtick_start(struct rq *rq, u64 delay)
1068 struct hrtimer *timer = &rq->hrtick_timer;
1069 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1071 hrtimer_set_expires(timer, time);
1073 if (rq == this_rq()) {
1074 hrtimer_restart(timer);
1075 } else if (!rq->hrtick_csd_pending) {
1076 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1077 rq->hrtick_csd_pending = 1;
1082 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1084 int cpu = (int)(long)hcpu;
1087 case CPU_UP_CANCELED:
1088 case CPU_UP_CANCELED_FROZEN:
1089 case CPU_DOWN_PREPARE:
1090 case CPU_DOWN_PREPARE_FROZEN:
1092 case CPU_DEAD_FROZEN:
1093 hrtick_clear(cpu_rq(cpu));
1100 static __init void init_hrtick(void)
1102 hotcpu_notifier(hotplug_hrtick, 0);
1106 * Called to set the hrtick timer state.
1108 * called with rq->lock held and irqs disabled
1110 static void hrtick_start(struct rq *rq, u64 delay)
1112 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1113 HRTIMER_MODE_REL_PINNED, 0);
1116 static inline void init_hrtick(void)
1119 #endif /* CONFIG_SMP */
1121 static void init_rq_hrtick(struct rq *rq)
1124 rq->hrtick_csd_pending = 0;
1126 rq->hrtick_csd.flags = 0;
1127 rq->hrtick_csd.func = __hrtick_start;
1128 rq->hrtick_csd.info = rq;
1131 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1132 rq->hrtick_timer.function = hrtick;
1134 #else /* CONFIG_SCHED_HRTICK */
1135 static inline void hrtick_clear(struct rq *rq)
1139 static inline void init_rq_hrtick(struct rq *rq)
1143 static inline void init_hrtick(void)
1146 #endif /* CONFIG_SCHED_HRTICK */
1149 * resched_task - mark a task 'to be rescheduled now'.
1151 * On UP this means the setting of the need_resched flag, on SMP it
1152 * might also involve a cross-CPU call to trigger the scheduler on
1157 #ifndef tsk_is_polling
1158 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1161 static void resched_task(struct task_struct *p)
1165 assert_raw_spin_locked(&task_rq(p)->lock);
1167 if (test_tsk_need_resched(p))
1170 set_tsk_need_resched(p);
1173 if (cpu == smp_processor_id())
1176 /* NEED_RESCHED must be visible before we test polling */
1178 if (!tsk_is_polling(p))
1179 smp_send_reschedule(cpu);
1182 static void resched_cpu(int cpu)
1184 struct rq *rq = cpu_rq(cpu);
1185 unsigned long flags;
1187 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1189 resched_task(cpu_curr(cpu));
1190 raw_spin_unlock_irqrestore(&rq->lock, flags);
1195 * In the semi idle case, use the nearest busy cpu for migrating timers
1196 * from an idle cpu. This is good for power-savings.
1198 * We don't do similar optimization for completely idle system, as
1199 * selecting an idle cpu will add more delays to the timers than intended
1200 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1202 int get_nohz_timer_target(void)
1204 int cpu = smp_processor_id();
1206 struct sched_domain *sd;
1208 for_each_domain(cpu, sd) {
1209 for_each_cpu(i, sched_domain_span(sd))
1216 * When add_timer_on() enqueues a timer into the timer wheel of an
1217 * idle CPU then this timer might expire before the next timer event
1218 * which is scheduled to wake up that CPU. In case of a completely
1219 * idle system the next event might even be infinite time into the
1220 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1221 * leaves the inner idle loop so the newly added timer is taken into
1222 * account when the CPU goes back to idle and evaluates the timer
1223 * wheel for the next timer event.
1225 void wake_up_idle_cpu(int cpu)
1227 struct rq *rq = cpu_rq(cpu);
1229 if (cpu == smp_processor_id())
1233 * This is safe, as this function is called with the timer
1234 * wheel base lock of (cpu) held. When the CPU is on the way
1235 * to idle and has not yet set rq->curr to idle then it will
1236 * be serialized on the timer wheel base lock and take the new
1237 * timer into account automatically.
1239 if (rq->curr != rq->idle)
1243 * We can set TIF_RESCHED on the idle task of the other CPU
1244 * lockless. The worst case is that the other CPU runs the
1245 * idle task through an additional NOOP schedule()
1247 set_tsk_need_resched(rq->idle);
1249 /* NEED_RESCHED must be visible before we test polling */
1251 if (!tsk_is_polling(rq->idle))
1252 smp_send_reschedule(cpu);
1255 #endif /* CONFIG_NO_HZ */
1257 static u64 sched_avg_period(void)
1259 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1262 static void sched_avg_update(struct rq *rq)
1264 s64 period = sched_avg_period();
1266 while ((s64)(rq->clock - rq->age_stamp) > period) {
1268 * Inline assembly required to prevent the compiler
1269 * optimising this loop into a divmod call.
1270 * See __iter_div_u64_rem() for another example of this.
1272 asm("" : "+rm" (rq->age_stamp));
1273 rq->age_stamp += period;
1278 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1280 rq->rt_avg += rt_delta;
1281 sched_avg_update(rq);
1284 #else /* !CONFIG_SMP */
1285 static void resched_task(struct task_struct *p)
1287 assert_raw_spin_locked(&task_rq(p)->lock);
1288 set_tsk_need_resched(p);
1291 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1295 static void sched_avg_update(struct rq *rq)
1298 #endif /* CONFIG_SMP */
1300 #if BITS_PER_LONG == 32
1301 # define WMULT_CONST (~0UL)
1303 # define WMULT_CONST (1UL << 32)
1306 #define WMULT_SHIFT 32
1309 * Shift right and round:
1311 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1314 * delta *= weight / lw
1316 static unsigned long
1317 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1318 struct load_weight *lw)
1322 if (!lw->inv_weight) {
1323 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1326 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1330 tmp = (u64)delta_exec * weight;
1332 * Check whether we'd overflow the 64-bit multiplication:
1334 if (unlikely(tmp > WMULT_CONST))
1335 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1338 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1340 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1343 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1349 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1356 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1357 * of tasks with abnormal "nice" values across CPUs the contribution that
1358 * each task makes to its run queue's load is weighted according to its
1359 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1360 * scaled version of the new time slice allocation that they receive on time
1364 #define WEIGHT_IDLEPRIO 3
1365 #define WMULT_IDLEPRIO 1431655765
1368 * Nice levels are multiplicative, with a gentle 10% change for every
1369 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1370 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1371 * that remained on nice 0.
1373 * The "10% effect" is relative and cumulative: from _any_ nice level,
1374 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1375 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1376 * If a task goes up by ~10% and another task goes down by ~10% then
1377 * the relative distance between them is ~25%.)
1379 static const int prio_to_weight[40] = {
1380 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1381 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1382 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1383 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1384 /* 0 */ 1024, 820, 655, 526, 423,
1385 /* 5 */ 335, 272, 215, 172, 137,
1386 /* 10 */ 110, 87, 70, 56, 45,
1387 /* 15 */ 36, 29, 23, 18, 15,
1391 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1393 * In cases where the weight does not change often, we can use the
1394 * precalculated inverse to speed up arithmetics by turning divisions
1395 * into multiplications:
1397 static const u32 prio_to_wmult[40] = {
1398 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1399 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1400 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1401 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1402 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1403 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1404 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1405 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1408 /* Time spent by the tasks of the cpu accounting group executing in ... */
1409 enum cpuacct_stat_index {
1410 CPUACCT_STAT_USER, /* ... user mode */
1411 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1413 CPUACCT_STAT_NSTATS,
1416 #ifdef CONFIG_CGROUP_CPUACCT
1417 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1418 static void cpuacct_update_stats(struct task_struct *tsk,
1419 enum cpuacct_stat_index idx, cputime_t val);
1421 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1422 static inline void cpuacct_update_stats(struct task_struct *tsk,
1423 enum cpuacct_stat_index idx, cputime_t val) {}
1426 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1428 update_load_add(&rq->load, load);
1431 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1433 update_load_sub(&rq->load, load);
1436 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1437 typedef int (*tg_visitor)(struct task_group *, void *);
1440 * Iterate the full tree, calling @down when first entering a node and @up when
1441 * leaving it for the final time.
1443 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1445 struct task_group *parent, *child;
1449 parent = &root_task_group;
1451 ret = (*down)(parent, data);
1454 list_for_each_entry_rcu(child, &parent->children, siblings) {
1461 ret = (*up)(parent, data);
1466 parent = parent->parent;
1475 static int tg_nop(struct task_group *tg, void *data)
1482 /* Used instead of source_load when we know the type == 0 */
1483 static unsigned long weighted_cpuload(const int cpu)
1485 return cpu_rq(cpu)->load.weight;
1489 * Return a low guess at the load of a migration-source cpu weighted
1490 * according to the scheduling class and "nice" value.
1492 * We want to under-estimate the load of migration sources, to
1493 * balance conservatively.
1495 static unsigned long source_load(int cpu, int type)
1497 struct rq *rq = cpu_rq(cpu);
1498 unsigned long total = weighted_cpuload(cpu);
1500 if (type == 0 || !sched_feat(LB_BIAS))
1503 return min(rq->cpu_load[type-1], total);
1507 * Return a high guess at the load of a migration-target cpu weighted
1508 * according to the scheduling class and "nice" value.
1510 static unsigned long target_load(int cpu, int type)
1512 struct rq *rq = cpu_rq(cpu);
1513 unsigned long total = weighted_cpuload(cpu);
1515 if (type == 0 || !sched_feat(LB_BIAS))
1518 return max(rq->cpu_load[type-1], total);
1521 static unsigned long power_of(int cpu)
1523 return cpu_rq(cpu)->cpu_power;
1526 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1528 static unsigned long cpu_avg_load_per_task(int cpu)
1530 struct rq *rq = cpu_rq(cpu);
1531 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1534 rq->avg_load_per_task = rq->load.weight / nr_running;
1536 rq->avg_load_per_task = 0;
1538 return rq->avg_load_per_task;
1541 #ifdef CONFIG_FAIR_GROUP_SCHED
1543 static __read_mostly unsigned long __percpu *update_shares_data;
1545 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1548 * Calculate and set the cpu's group shares.
1550 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1551 unsigned long sd_shares,
1552 unsigned long sd_rq_weight,
1553 unsigned long *usd_rq_weight)
1555 unsigned long shares, rq_weight;
1558 rq_weight = usd_rq_weight[cpu];
1561 rq_weight = NICE_0_LOAD;
1565 * \Sum_j shares_j * rq_weight_i
1566 * shares_i = -----------------------------
1567 * \Sum_j rq_weight_j
1569 shares = (sd_shares * rq_weight) / sd_rq_weight;
1570 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1572 if (abs(shares - tg->se[cpu]->load.weight) >
1573 sysctl_sched_shares_thresh) {
1574 struct rq *rq = cpu_rq(cpu);
1575 unsigned long flags;
1577 raw_spin_lock_irqsave(&rq->lock, flags);
1578 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1579 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1580 __set_se_shares(tg->se[cpu], shares);
1581 raw_spin_unlock_irqrestore(&rq->lock, flags);
1586 * Re-compute the task group their per cpu shares over the given domain.
1587 * This needs to be done in a bottom-up fashion because the rq weight of a
1588 * parent group depends on the shares of its child groups.
1590 static int tg_shares_up(struct task_group *tg, void *data)
1592 unsigned long weight, rq_weight = 0, sum_weight = 0, shares = 0;
1593 unsigned long *usd_rq_weight;
1594 struct sched_domain *sd = data;
1595 unsigned long flags;
1601 local_irq_save(flags);
1602 usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id());
1604 for_each_cpu(i, sched_domain_span(sd)) {
1605 weight = tg->cfs_rq[i]->load.weight;
1606 usd_rq_weight[i] = weight;
1608 rq_weight += weight;
1610 * If there are currently no tasks on the cpu pretend there
1611 * is one of average load so that when a new task gets to
1612 * run here it will not get delayed by group starvation.
1615 weight = NICE_0_LOAD;
1617 sum_weight += weight;
1618 shares += tg->cfs_rq[i]->shares;
1622 rq_weight = sum_weight;
1624 if ((!shares && rq_weight) || shares > tg->shares)
1625 shares = tg->shares;
1627 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1628 shares = tg->shares;
1630 for_each_cpu(i, sched_domain_span(sd))
1631 update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight);
1633 local_irq_restore(flags);
1639 * Compute the cpu's hierarchical load factor for each task group.
1640 * This needs to be done in a top-down fashion because the load of a child
1641 * group is a fraction of its parents load.
1643 static int tg_load_down(struct task_group *tg, void *data)
1646 long cpu = (long)data;
1649 load = cpu_rq(cpu)->load.weight;
1651 load = tg->parent->cfs_rq[cpu]->h_load;
1652 load *= tg->cfs_rq[cpu]->shares;
1653 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1656 tg->cfs_rq[cpu]->h_load = load;
1661 static void update_shares(struct sched_domain *sd)
1666 if (root_task_group_empty())
1669 now = local_clock();
1670 elapsed = now - sd->last_update;
1672 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1673 sd->last_update = now;
1674 walk_tg_tree(tg_nop, tg_shares_up, sd);
1678 static void update_h_load(long cpu)
1680 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1685 static inline void update_shares(struct sched_domain *sd)
1691 #ifdef CONFIG_PREEMPT
1693 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1696 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1697 * way at the expense of forcing extra atomic operations in all
1698 * invocations. This assures that the double_lock is acquired using the
1699 * same underlying policy as the spinlock_t on this architecture, which
1700 * reduces latency compared to the unfair variant below. However, it
1701 * also adds more overhead and therefore may reduce throughput.
1703 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1704 __releases(this_rq->lock)
1705 __acquires(busiest->lock)
1706 __acquires(this_rq->lock)
1708 raw_spin_unlock(&this_rq->lock);
1709 double_rq_lock(this_rq, busiest);
1716 * Unfair double_lock_balance: Optimizes throughput at the expense of
1717 * latency by eliminating extra atomic operations when the locks are
1718 * already in proper order on entry. This favors lower cpu-ids and will
1719 * grant the double lock to lower cpus over higher ids under contention,
1720 * regardless of entry order into the function.
1722 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1723 __releases(this_rq->lock)
1724 __acquires(busiest->lock)
1725 __acquires(this_rq->lock)
1729 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1730 if (busiest < this_rq) {
1731 raw_spin_unlock(&this_rq->lock);
1732 raw_spin_lock(&busiest->lock);
1733 raw_spin_lock_nested(&this_rq->lock,
1734 SINGLE_DEPTH_NESTING);
1737 raw_spin_lock_nested(&busiest->lock,
1738 SINGLE_DEPTH_NESTING);
1743 #endif /* CONFIG_PREEMPT */
1746 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1748 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1750 if (unlikely(!irqs_disabled())) {
1751 /* printk() doesn't work good under rq->lock */
1752 raw_spin_unlock(&this_rq->lock);
1756 return _double_lock_balance(this_rq, busiest);
1759 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1760 __releases(busiest->lock)
1762 raw_spin_unlock(&busiest->lock);
1763 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1767 * double_rq_lock - safely lock two runqueues
1769 * Note this does not disable interrupts like task_rq_lock,
1770 * you need to do so manually before calling.
1772 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1773 __acquires(rq1->lock)
1774 __acquires(rq2->lock)
1776 BUG_ON(!irqs_disabled());
1778 raw_spin_lock(&rq1->lock);
1779 __acquire(rq2->lock); /* Fake it out ;) */
1782 raw_spin_lock(&rq1->lock);
1783 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1785 raw_spin_lock(&rq2->lock);
1786 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1792 * double_rq_unlock - safely unlock two runqueues
1794 * Note this does not restore interrupts like task_rq_unlock,
1795 * you need to do so manually after calling.
1797 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1798 __releases(rq1->lock)
1799 __releases(rq2->lock)
1801 raw_spin_unlock(&rq1->lock);
1803 raw_spin_unlock(&rq2->lock);
1805 __release(rq2->lock);
1810 #ifdef CONFIG_FAIR_GROUP_SCHED
1811 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1814 cfs_rq->shares = shares;
1819 static void calc_load_account_idle(struct rq *this_rq);
1820 static void update_sysctl(void);
1821 static int get_update_sysctl_factor(void);
1822 static void update_cpu_load(struct rq *this_rq);
1824 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1826 set_task_rq(p, cpu);
1829 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1830 * successfuly executed on another CPU. We must ensure that updates of
1831 * per-task data have been completed by this moment.
1834 task_thread_info(p)->cpu = cpu;
1838 static const struct sched_class rt_sched_class;
1840 #define sched_class_highest (&stop_sched_class)
1841 #define for_each_class(class) \
1842 for (class = sched_class_highest; class; class = class->next)
1844 #include "sched_stats.h"
1846 static void inc_nr_running(struct rq *rq)
1851 static void dec_nr_running(struct rq *rq)
1856 static void set_load_weight(struct task_struct *p)
1859 * SCHED_IDLE tasks get minimal weight:
1861 if (p->policy == SCHED_IDLE) {
1862 p->se.load.weight = WEIGHT_IDLEPRIO;
1863 p->se.load.inv_weight = WMULT_IDLEPRIO;
1867 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1868 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1871 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1873 update_rq_clock(rq);
1874 sched_info_queued(p);
1875 p->sched_class->enqueue_task(rq, p, flags);
1879 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1881 update_rq_clock(rq);
1882 sched_info_dequeued(p);
1883 p->sched_class->dequeue_task(rq, p, flags);
1888 * activate_task - move a task to the runqueue.
1890 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1892 if (task_contributes_to_load(p))
1893 rq->nr_uninterruptible--;
1895 enqueue_task(rq, p, flags);
1900 * deactivate_task - remove a task from the runqueue.
1902 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1904 if (task_contributes_to_load(p))
1905 rq->nr_uninterruptible++;
1907 dequeue_task(rq, p, flags);
1911 #include "sched_idletask.c"
1912 #include "sched_fair.c"
1913 #include "sched_rt.c"
1914 #include "sched_stoptask.c"
1915 #ifdef CONFIG_SCHED_DEBUG
1916 # include "sched_debug.c"
1919 void sched_set_stop_task(int cpu, struct task_struct *stop)
1921 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
1922 struct task_struct *old_stop = cpu_rq(cpu)->stop;
1926 * Make it appear like a SCHED_FIFO task, its something
1927 * userspace knows about and won't get confused about.
1929 * Also, it will make PI more or less work without too
1930 * much confusion -- but then, stop work should not
1931 * rely on PI working anyway.
1933 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
1935 stop->sched_class = &stop_sched_class;
1938 cpu_rq(cpu)->stop = stop;
1942 * Reset it back to a normal scheduling class so that
1943 * it can die in pieces.
1945 old_stop->sched_class = &rt_sched_class;
1950 * __normal_prio - return the priority that is based on the static prio
1952 static inline int __normal_prio(struct task_struct *p)
1954 return p->static_prio;
1958 * Calculate the expected normal priority: i.e. priority
1959 * without taking RT-inheritance into account. Might be
1960 * boosted by interactivity modifiers. Changes upon fork,
1961 * setprio syscalls, and whenever the interactivity
1962 * estimator recalculates.
1964 static inline int normal_prio(struct task_struct *p)
1968 if (task_has_rt_policy(p))
1969 prio = MAX_RT_PRIO-1 - p->rt_priority;
1971 prio = __normal_prio(p);
1976 * Calculate the current priority, i.e. the priority
1977 * taken into account by the scheduler. This value might
1978 * be boosted by RT tasks, or might be boosted by
1979 * interactivity modifiers. Will be RT if the task got
1980 * RT-boosted. If not then it returns p->normal_prio.
1982 static int effective_prio(struct task_struct *p)
1984 p->normal_prio = normal_prio(p);
1986 * If we are RT tasks or we were boosted to RT priority,
1987 * keep the priority unchanged. Otherwise, update priority
1988 * to the normal priority:
1990 if (!rt_prio(p->prio))
1991 return p->normal_prio;
1996 * task_curr - is this task currently executing on a CPU?
1997 * @p: the task in question.
1999 inline int task_curr(const struct task_struct *p)
2001 return cpu_curr(task_cpu(p)) == p;
2004 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2005 const struct sched_class *prev_class,
2006 int oldprio, int running)
2008 if (prev_class != p->sched_class) {
2009 if (prev_class->switched_from)
2010 prev_class->switched_from(rq, p, running);
2011 p->sched_class->switched_to(rq, p, running);
2013 p->sched_class->prio_changed(rq, p, oldprio, running);
2018 * Is this task likely cache-hot:
2021 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2025 if (p->sched_class != &fair_sched_class)
2029 * Buddy candidates are cache hot:
2031 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2032 (&p->se == cfs_rq_of(&p->se)->next ||
2033 &p->se == cfs_rq_of(&p->se)->last))
2036 if (sysctl_sched_migration_cost == -1)
2038 if (sysctl_sched_migration_cost == 0)
2041 delta = now - p->se.exec_start;
2043 return delta < (s64)sysctl_sched_migration_cost;
2046 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2048 #ifdef CONFIG_SCHED_DEBUG
2050 * We should never call set_task_cpu() on a blocked task,
2051 * ttwu() will sort out the placement.
2053 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2054 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2057 trace_sched_migrate_task(p, new_cpu);
2059 if (task_cpu(p) != new_cpu) {
2060 p->se.nr_migrations++;
2061 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2064 __set_task_cpu(p, new_cpu);
2067 struct migration_arg {
2068 struct task_struct *task;
2072 static int migration_cpu_stop(void *data);
2075 * The task's runqueue lock must be held.
2076 * Returns true if you have to wait for migration thread.
2078 static bool migrate_task(struct task_struct *p, int dest_cpu)
2080 struct rq *rq = task_rq(p);
2083 * If the task is not on a runqueue (and not running), then
2084 * the next wake-up will properly place the task.
2086 return p->se.on_rq || task_running(rq, p);
2090 * wait_task_inactive - wait for a thread to unschedule.
2092 * If @match_state is nonzero, it's the @p->state value just checked and
2093 * not expected to change. If it changes, i.e. @p might have woken up,
2094 * then return zero. When we succeed in waiting for @p to be off its CPU,
2095 * we return a positive number (its total switch count). If a second call
2096 * a short while later returns the same number, the caller can be sure that
2097 * @p has remained unscheduled the whole time.
2099 * The caller must ensure that the task *will* unschedule sometime soon,
2100 * else this function might spin for a *long* time. This function can't
2101 * be called with interrupts off, or it may introduce deadlock with
2102 * smp_call_function() if an IPI is sent by the same process we are
2103 * waiting to become inactive.
2105 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2107 unsigned long flags;
2114 * We do the initial early heuristics without holding
2115 * any task-queue locks at all. We'll only try to get
2116 * the runqueue lock when things look like they will
2122 * If the task is actively running on another CPU
2123 * still, just relax and busy-wait without holding
2126 * NOTE! Since we don't hold any locks, it's not
2127 * even sure that "rq" stays as the right runqueue!
2128 * But we don't care, since "task_running()" will
2129 * return false if the runqueue has changed and p
2130 * is actually now running somewhere else!
2132 while (task_running(rq, p)) {
2133 if (match_state && unlikely(p->state != match_state))
2139 * Ok, time to look more closely! We need the rq
2140 * lock now, to be *sure*. If we're wrong, we'll
2141 * just go back and repeat.
2143 rq = task_rq_lock(p, &flags);
2144 trace_sched_wait_task(p);
2145 running = task_running(rq, p);
2146 on_rq = p->se.on_rq;
2148 if (!match_state || p->state == match_state)
2149 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2150 task_rq_unlock(rq, &flags);
2153 * If it changed from the expected state, bail out now.
2155 if (unlikely(!ncsw))
2159 * Was it really running after all now that we
2160 * checked with the proper locks actually held?
2162 * Oops. Go back and try again..
2164 if (unlikely(running)) {
2170 * It's not enough that it's not actively running,
2171 * it must be off the runqueue _entirely_, and not
2174 * So if it was still runnable (but just not actively
2175 * running right now), it's preempted, and we should
2176 * yield - it could be a while.
2178 if (unlikely(on_rq)) {
2179 schedule_timeout_uninterruptible(1);
2184 * Ahh, all good. It wasn't running, and it wasn't
2185 * runnable, which means that it will never become
2186 * running in the future either. We're all done!
2195 * kick_process - kick a running thread to enter/exit the kernel
2196 * @p: the to-be-kicked thread
2198 * Cause a process which is running on another CPU to enter
2199 * kernel-mode, without any delay. (to get signals handled.)
2201 * NOTE: this function doesnt have to take the runqueue lock,
2202 * because all it wants to ensure is that the remote task enters
2203 * the kernel. If the IPI races and the task has been migrated
2204 * to another CPU then no harm is done and the purpose has been
2207 void kick_process(struct task_struct *p)
2213 if ((cpu != smp_processor_id()) && task_curr(p))
2214 smp_send_reschedule(cpu);
2217 EXPORT_SYMBOL_GPL(kick_process);
2218 #endif /* CONFIG_SMP */
2221 * task_oncpu_function_call - call a function on the cpu on which a task runs
2222 * @p: the task to evaluate
2223 * @func: the function to be called
2224 * @info: the function call argument
2226 * Calls the function @func when the task is currently running. This might
2227 * be on the current CPU, which just calls the function directly
2229 void task_oncpu_function_call(struct task_struct *p,
2230 void (*func) (void *info), void *info)
2237 smp_call_function_single(cpu, func, info, 1);
2243 * ->cpus_allowed is protected by either TASK_WAKING or rq->lock held.
2245 static int select_fallback_rq(int cpu, struct task_struct *p)
2248 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2250 /* Look for allowed, online CPU in same node. */
2251 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2252 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2255 /* Any allowed, online CPU? */
2256 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2257 if (dest_cpu < nr_cpu_ids)
2260 /* No more Mr. Nice Guy. */
2261 if (unlikely(dest_cpu >= nr_cpu_ids)) {
2262 dest_cpu = cpuset_cpus_allowed_fallback(p);
2264 * Don't tell them about moving exiting tasks or
2265 * kernel threads (both mm NULL), since they never
2268 if (p->mm && printk_ratelimit()) {
2269 printk(KERN_INFO "process %d (%s) no "
2270 "longer affine to cpu%d\n",
2271 task_pid_nr(p), p->comm, cpu);
2279 * The caller (fork, wakeup) owns TASK_WAKING, ->cpus_allowed is stable.
2282 int select_task_rq(struct rq *rq, struct task_struct *p, int sd_flags, int wake_flags)
2284 int cpu = p->sched_class->select_task_rq(rq, p, sd_flags, wake_flags);
2287 * In order not to call set_task_cpu() on a blocking task we need
2288 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2291 * Since this is common to all placement strategies, this lives here.
2293 * [ this allows ->select_task() to simply return task_cpu(p) and
2294 * not worry about this generic constraint ]
2296 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2298 cpu = select_fallback_rq(task_cpu(p), p);
2303 static void update_avg(u64 *avg, u64 sample)
2305 s64 diff = sample - *avg;
2310 static inline void ttwu_activate(struct task_struct *p, struct rq *rq,
2311 bool is_sync, bool is_migrate, bool is_local,
2312 unsigned long en_flags)
2314 schedstat_inc(p, se.statistics.nr_wakeups);
2316 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2318 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2320 schedstat_inc(p, se.statistics.nr_wakeups_local);
2322 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2324 activate_task(rq, p, en_flags);
2327 static inline void ttwu_post_activation(struct task_struct *p, struct rq *rq,
2328 int wake_flags, bool success)
2330 trace_sched_wakeup(p, success);
2331 check_preempt_curr(rq, p, wake_flags);
2333 p->state = TASK_RUNNING;
2335 if (p->sched_class->task_woken)
2336 p->sched_class->task_woken(rq, p);
2338 if (unlikely(rq->idle_stamp)) {
2339 u64 delta = rq->clock - rq->idle_stamp;
2340 u64 max = 2*sysctl_sched_migration_cost;
2345 update_avg(&rq->avg_idle, delta);
2349 /* if a worker is waking up, notify workqueue */
2350 if ((p->flags & PF_WQ_WORKER) && success)
2351 wq_worker_waking_up(p, cpu_of(rq));
2355 * try_to_wake_up - wake up a thread
2356 * @p: the thread to be awakened
2357 * @state: the mask of task states that can be woken
2358 * @wake_flags: wake modifier flags (WF_*)
2360 * Put it on the run-queue if it's not already there. The "current"
2361 * thread is always on the run-queue (except when the actual
2362 * re-schedule is in progress), and as such you're allowed to do
2363 * the simpler "current->state = TASK_RUNNING" to mark yourself
2364 * runnable without the overhead of this.
2366 * Returns %true if @p was woken up, %false if it was already running
2367 * or @state didn't match @p's state.
2369 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2372 int cpu, orig_cpu, this_cpu, success = 0;
2373 unsigned long flags;
2374 unsigned long en_flags = ENQUEUE_WAKEUP;
2377 this_cpu = get_cpu();
2380 rq = task_rq_lock(p, &flags);
2381 if (!(p->state & state))
2391 if (unlikely(task_running(rq, p)))
2395 * In order to handle concurrent wakeups and release the rq->lock
2396 * we put the task in TASK_WAKING state.
2398 * First fix up the nr_uninterruptible count:
2400 if (task_contributes_to_load(p)) {
2401 if (likely(cpu_online(orig_cpu)))
2402 rq->nr_uninterruptible--;
2404 this_rq()->nr_uninterruptible--;
2406 p->state = TASK_WAKING;
2408 if (p->sched_class->task_waking) {
2409 p->sched_class->task_waking(rq, p);
2410 en_flags |= ENQUEUE_WAKING;
2413 cpu = select_task_rq(rq, p, SD_BALANCE_WAKE, wake_flags);
2414 if (cpu != orig_cpu)
2415 set_task_cpu(p, cpu);
2416 __task_rq_unlock(rq);
2419 raw_spin_lock(&rq->lock);
2422 * We migrated the task without holding either rq->lock, however
2423 * since the task is not on the task list itself, nobody else
2424 * will try and migrate the task, hence the rq should match the
2425 * cpu we just moved it to.
2427 WARN_ON(task_cpu(p) != cpu);
2428 WARN_ON(p->state != TASK_WAKING);
2430 #ifdef CONFIG_SCHEDSTATS
2431 schedstat_inc(rq, ttwu_count);
2432 if (cpu == this_cpu)
2433 schedstat_inc(rq, ttwu_local);
2435 struct sched_domain *sd;
2436 for_each_domain(this_cpu, sd) {
2437 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2438 schedstat_inc(sd, ttwu_wake_remote);
2443 #endif /* CONFIG_SCHEDSTATS */
2446 #endif /* CONFIG_SMP */
2447 ttwu_activate(p, rq, wake_flags & WF_SYNC, orig_cpu != cpu,
2448 cpu == this_cpu, en_flags);
2451 ttwu_post_activation(p, rq, wake_flags, success);
2453 task_rq_unlock(rq, &flags);
2460 * try_to_wake_up_local - try to wake up a local task with rq lock held
2461 * @p: the thread to be awakened
2463 * Put @p on the run-queue if it's not alredy there. The caller must
2464 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2465 * the current task. this_rq() stays locked over invocation.
2467 static void try_to_wake_up_local(struct task_struct *p)
2469 struct rq *rq = task_rq(p);
2470 bool success = false;
2472 BUG_ON(rq != this_rq());
2473 BUG_ON(p == current);
2474 lockdep_assert_held(&rq->lock);
2476 if (!(p->state & TASK_NORMAL))
2480 if (likely(!task_running(rq, p))) {
2481 schedstat_inc(rq, ttwu_count);
2482 schedstat_inc(rq, ttwu_local);
2484 ttwu_activate(p, rq, false, false, true, ENQUEUE_WAKEUP);
2487 ttwu_post_activation(p, rq, 0, success);
2491 * wake_up_process - Wake up a specific process
2492 * @p: The process to be woken up.
2494 * Attempt to wake up the nominated process and move it to the set of runnable
2495 * processes. Returns 1 if the process was woken up, 0 if it was already
2498 * It may be assumed that this function implies a write memory barrier before
2499 * changing the task state if and only if any tasks are woken up.
2501 int wake_up_process(struct task_struct *p)
2503 return try_to_wake_up(p, TASK_ALL, 0);
2505 EXPORT_SYMBOL(wake_up_process);
2507 int wake_up_state(struct task_struct *p, unsigned int state)
2509 return try_to_wake_up(p, state, 0);
2513 * Perform scheduler related setup for a newly forked process p.
2514 * p is forked by current.
2516 * __sched_fork() is basic setup used by init_idle() too:
2518 static void __sched_fork(struct task_struct *p)
2520 p->se.exec_start = 0;
2521 p->se.sum_exec_runtime = 0;
2522 p->se.prev_sum_exec_runtime = 0;
2523 p->se.nr_migrations = 0;
2525 #ifdef CONFIG_SCHEDSTATS
2526 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2529 INIT_LIST_HEAD(&p->rt.run_list);
2531 INIT_LIST_HEAD(&p->se.group_node);
2533 #ifdef CONFIG_PREEMPT_NOTIFIERS
2534 INIT_HLIST_HEAD(&p->preempt_notifiers);
2539 * fork()/clone()-time setup:
2541 void sched_fork(struct task_struct *p, int clone_flags)
2543 int cpu = get_cpu();
2547 * We mark the process as running here. This guarantees that
2548 * nobody will actually run it, and a signal or other external
2549 * event cannot wake it up and insert it on the runqueue either.
2551 p->state = TASK_RUNNING;
2554 * Revert to default priority/policy on fork if requested.
2556 if (unlikely(p->sched_reset_on_fork)) {
2557 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2558 p->policy = SCHED_NORMAL;
2559 p->normal_prio = p->static_prio;
2562 if (PRIO_TO_NICE(p->static_prio) < 0) {
2563 p->static_prio = NICE_TO_PRIO(0);
2564 p->normal_prio = p->static_prio;
2569 * We don't need the reset flag anymore after the fork. It has
2570 * fulfilled its duty:
2572 p->sched_reset_on_fork = 0;
2576 * Make sure we do not leak PI boosting priority to the child.
2578 p->prio = current->normal_prio;
2580 if (!rt_prio(p->prio))
2581 p->sched_class = &fair_sched_class;
2583 if (p->sched_class->task_fork)
2584 p->sched_class->task_fork(p);
2587 * The child is not yet in the pid-hash so no cgroup attach races,
2588 * and the cgroup is pinned to this child due to cgroup_fork()
2589 * is ran before sched_fork().
2591 * Silence PROVE_RCU.
2594 set_task_cpu(p, cpu);
2597 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2598 if (likely(sched_info_on()))
2599 memset(&p->sched_info, 0, sizeof(p->sched_info));
2601 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2604 #ifdef CONFIG_PREEMPT
2605 /* Want to start with kernel preemption disabled. */
2606 task_thread_info(p)->preempt_count = 1;
2608 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2614 * wake_up_new_task - wake up a newly created task for the first time.
2616 * This function will do some initial scheduler statistics housekeeping
2617 * that must be done for every newly created context, then puts the task
2618 * on the runqueue and wakes it.
2620 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2622 unsigned long flags;
2624 int cpu __maybe_unused = get_cpu();
2627 rq = task_rq_lock(p, &flags);
2628 p->state = TASK_WAKING;
2631 * Fork balancing, do it here and not earlier because:
2632 * - cpus_allowed can change in the fork path
2633 * - any previously selected cpu might disappear through hotplug
2635 * We set TASK_WAKING so that select_task_rq() can drop rq->lock
2636 * without people poking at ->cpus_allowed.
2638 cpu = select_task_rq(rq, p, SD_BALANCE_FORK, 0);
2639 set_task_cpu(p, cpu);
2641 p->state = TASK_RUNNING;
2642 task_rq_unlock(rq, &flags);
2645 rq = task_rq_lock(p, &flags);
2646 activate_task(rq, p, 0);
2647 trace_sched_wakeup_new(p, 1);
2648 check_preempt_curr(rq, p, WF_FORK);
2650 if (p->sched_class->task_woken)
2651 p->sched_class->task_woken(rq, p);
2653 task_rq_unlock(rq, &flags);
2657 #ifdef CONFIG_PREEMPT_NOTIFIERS
2660 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2661 * @notifier: notifier struct to register
2663 void preempt_notifier_register(struct preempt_notifier *notifier)
2665 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2667 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2670 * preempt_notifier_unregister - no longer interested in preemption notifications
2671 * @notifier: notifier struct to unregister
2673 * This is safe to call from within a preemption notifier.
2675 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2677 hlist_del(¬ifier->link);
2679 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2681 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2683 struct preempt_notifier *notifier;
2684 struct hlist_node *node;
2686 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2687 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2691 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2692 struct task_struct *next)
2694 struct preempt_notifier *notifier;
2695 struct hlist_node *node;
2697 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2698 notifier->ops->sched_out(notifier, next);
2701 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2703 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2708 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2709 struct task_struct *next)
2713 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2716 * prepare_task_switch - prepare to switch tasks
2717 * @rq: the runqueue preparing to switch
2718 * @prev: the current task that is being switched out
2719 * @next: the task we are going to switch to.
2721 * This is called with the rq lock held and interrupts off. It must
2722 * be paired with a subsequent finish_task_switch after the context
2725 * prepare_task_switch sets up locking and calls architecture specific
2729 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2730 struct task_struct *next)
2732 fire_sched_out_preempt_notifiers(prev, next);
2733 prepare_lock_switch(rq, next);
2734 prepare_arch_switch(next);
2738 * finish_task_switch - clean up after a task-switch
2739 * @rq: runqueue associated with task-switch
2740 * @prev: the thread we just switched away from.
2742 * finish_task_switch must be called after the context switch, paired
2743 * with a prepare_task_switch call before the context switch.
2744 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2745 * and do any other architecture-specific cleanup actions.
2747 * Note that we may have delayed dropping an mm in context_switch(). If
2748 * so, we finish that here outside of the runqueue lock. (Doing it
2749 * with the lock held can cause deadlocks; see schedule() for
2752 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2753 __releases(rq->lock)
2755 struct mm_struct *mm = rq->prev_mm;
2761 * A task struct has one reference for the use as "current".
2762 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2763 * schedule one last time. The schedule call will never return, and
2764 * the scheduled task must drop that reference.
2765 * The test for TASK_DEAD must occur while the runqueue locks are
2766 * still held, otherwise prev could be scheduled on another cpu, die
2767 * there before we look at prev->state, and then the reference would
2769 * Manfred Spraul <manfred@colorfullife.com>
2771 prev_state = prev->state;
2772 finish_arch_switch(prev);
2773 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2774 local_irq_disable();
2775 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2776 perf_event_task_sched_in(current);
2777 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2779 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2780 finish_lock_switch(rq, prev);
2782 fire_sched_in_preempt_notifiers(current);
2785 if (unlikely(prev_state == TASK_DEAD)) {
2787 * Remove function-return probe instances associated with this
2788 * task and put them back on the free list.
2790 kprobe_flush_task(prev);
2791 put_task_struct(prev);
2797 /* assumes rq->lock is held */
2798 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2800 if (prev->sched_class->pre_schedule)
2801 prev->sched_class->pre_schedule(rq, prev);
2804 /* rq->lock is NOT held, but preemption is disabled */
2805 static inline void post_schedule(struct rq *rq)
2807 if (rq->post_schedule) {
2808 unsigned long flags;
2810 raw_spin_lock_irqsave(&rq->lock, flags);
2811 if (rq->curr->sched_class->post_schedule)
2812 rq->curr->sched_class->post_schedule(rq);
2813 raw_spin_unlock_irqrestore(&rq->lock, flags);
2815 rq->post_schedule = 0;
2821 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2825 static inline void post_schedule(struct rq *rq)
2832 * schedule_tail - first thing a freshly forked thread must call.
2833 * @prev: the thread we just switched away from.
2835 asmlinkage void schedule_tail(struct task_struct *prev)
2836 __releases(rq->lock)
2838 struct rq *rq = this_rq();
2840 finish_task_switch(rq, prev);
2843 * FIXME: do we need to worry about rq being invalidated by the
2848 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2849 /* In this case, finish_task_switch does not reenable preemption */
2852 if (current->set_child_tid)
2853 put_user(task_pid_vnr(current), current->set_child_tid);
2857 * context_switch - switch to the new MM and the new
2858 * thread's register state.
2861 context_switch(struct rq *rq, struct task_struct *prev,
2862 struct task_struct *next)
2864 struct mm_struct *mm, *oldmm;
2866 prepare_task_switch(rq, prev, next);
2867 trace_sched_switch(prev, next);
2869 oldmm = prev->active_mm;
2871 * For paravirt, this is coupled with an exit in switch_to to
2872 * combine the page table reload and the switch backend into
2875 arch_start_context_switch(prev);
2878 next->active_mm = oldmm;
2879 atomic_inc(&oldmm->mm_count);
2880 enter_lazy_tlb(oldmm, next);
2882 switch_mm(oldmm, mm, next);
2885 prev->active_mm = NULL;
2886 rq->prev_mm = oldmm;
2889 * Since the runqueue lock will be released by the next
2890 * task (which is an invalid locking op but in the case
2891 * of the scheduler it's an obvious special-case), so we
2892 * do an early lockdep release here:
2894 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2895 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2898 /* Here we just switch the register state and the stack. */
2899 switch_to(prev, next, prev);
2903 * this_rq must be evaluated again because prev may have moved
2904 * CPUs since it called schedule(), thus the 'rq' on its stack
2905 * frame will be invalid.
2907 finish_task_switch(this_rq(), prev);
2911 * nr_running, nr_uninterruptible and nr_context_switches:
2913 * externally visible scheduler statistics: current number of runnable
2914 * threads, current number of uninterruptible-sleeping threads, total
2915 * number of context switches performed since bootup.
2917 unsigned long nr_running(void)
2919 unsigned long i, sum = 0;
2921 for_each_online_cpu(i)
2922 sum += cpu_rq(i)->nr_running;
2927 unsigned long nr_uninterruptible(void)
2929 unsigned long i, sum = 0;
2931 for_each_possible_cpu(i)
2932 sum += cpu_rq(i)->nr_uninterruptible;
2935 * Since we read the counters lockless, it might be slightly
2936 * inaccurate. Do not allow it to go below zero though:
2938 if (unlikely((long)sum < 0))
2944 unsigned long long nr_context_switches(void)
2947 unsigned long long sum = 0;
2949 for_each_possible_cpu(i)
2950 sum += cpu_rq(i)->nr_switches;
2955 unsigned long nr_iowait(void)
2957 unsigned long i, sum = 0;
2959 for_each_possible_cpu(i)
2960 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2965 unsigned long nr_iowait_cpu(int cpu)
2967 struct rq *this = cpu_rq(cpu);
2968 return atomic_read(&this->nr_iowait);
2971 unsigned long this_cpu_load(void)
2973 struct rq *this = this_rq();
2974 return this->cpu_load[0];
2978 /* Variables and functions for calc_load */
2979 static atomic_long_t calc_load_tasks;
2980 static unsigned long calc_load_update;
2981 unsigned long avenrun[3];
2982 EXPORT_SYMBOL(avenrun);
2984 static long calc_load_fold_active(struct rq *this_rq)
2986 long nr_active, delta = 0;
2988 nr_active = this_rq->nr_running;
2989 nr_active += (long) this_rq->nr_uninterruptible;
2991 if (nr_active != this_rq->calc_load_active) {
2992 delta = nr_active - this_rq->calc_load_active;
2993 this_rq->calc_load_active = nr_active;
3001 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
3003 * When making the ILB scale, we should try to pull this in as well.
3005 static atomic_long_t calc_load_tasks_idle;
3007 static void calc_load_account_idle(struct rq *this_rq)
3011 delta = calc_load_fold_active(this_rq);
3013 atomic_long_add(delta, &calc_load_tasks_idle);
3016 static long calc_load_fold_idle(void)
3021 * Its got a race, we don't care...
3023 if (atomic_long_read(&calc_load_tasks_idle))
3024 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
3029 static void calc_load_account_idle(struct rq *this_rq)
3033 static inline long calc_load_fold_idle(void)
3040 * get_avenrun - get the load average array
3041 * @loads: pointer to dest load array
3042 * @offset: offset to add
3043 * @shift: shift count to shift the result left
3045 * These values are estimates at best, so no need for locking.
3047 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3049 loads[0] = (avenrun[0] + offset) << shift;
3050 loads[1] = (avenrun[1] + offset) << shift;
3051 loads[2] = (avenrun[2] + offset) << shift;
3054 static unsigned long
3055 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3058 load += active * (FIXED_1 - exp);
3059 return load >> FSHIFT;
3063 * calc_load - update the avenrun load estimates 10 ticks after the
3064 * CPUs have updated calc_load_tasks.
3066 void calc_global_load(void)
3068 unsigned long upd = calc_load_update + 10;
3071 if (time_before(jiffies, upd))
3074 active = atomic_long_read(&calc_load_tasks);
3075 active = active > 0 ? active * FIXED_1 : 0;
3077 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3078 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3079 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3081 calc_load_update += LOAD_FREQ;
3085 * Called from update_cpu_load() to periodically update this CPU's
3088 static void calc_load_account_active(struct rq *this_rq)
3092 if (time_before(jiffies, this_rq->calc_load_update))
3095 delta = calc_load_fold_active(this_rq);
3096 delta += calc_load_fold_idle();
3098 atomic_long_add(delta, &calc_load_tasks);
3100 this_rq->calc_load_update += LOAD_FREQ;
3104 * The exact cpuload at various idx values, calculated at every tick would be
3105 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3107 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3108 * on nth tick when cpu may be busy, then we have:
3109 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3110 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3112 * decay_load_missed() below does efficient calculation of
3113 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3114 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3116 * The calculation is approximated on a 128 point scale.
3117 * degrade_zero_ticks is the number of ticks after which load at any
3118 * particular idx is approximated to be zero.
3119 * degrade_factor is a precomputed table, a row for each load idx.
3120 * Each column corresponds to degradation factor for a power of two ticks,
3121 * based on 128 point scale.
3123 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3124 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3126 * With this power of 2 load factors, we can degrade the load n times
3127 * by looking at 1 bits in n and doing as many mult/shift instead of
3128 * n mult/shifts needed by the exact degradation.
3130 #define DEGRADE_SHIFT 7
3131 static const unsigned char
3132 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3133 static const unsigned char
3134 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3135 {0, 0, 0, 0, 0, 0, 0, 0},
3136 {64, 32, 8, 0, 0, 0, 0, 0},
3137 {96, 72, 40, 12, 1, 0, 0},
3138 {112, 98, 75, 43, 15, 1, 0},
3139 {120, 112, 98, 76, 45, 16, 2} };
3142 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3143 * would be when CPU is idle and so we just decay the old load without
3144 * adding any new load.
3146 static unsigned long
3147 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3151 if (!missed_updates)
3154 if (missed_updates >= degrade_zero_ticks[idx])
3158 return load >> missed_updates;
3160 while (missed_updates) {
3161 if (missed_updates % 2)
3162 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3164 missed_updates >>= 1;
3171 * Update rq->cpu_load[] statistics. This function is usually called every
3172 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3173 * every tick. We fix it up based on jiffies.
3175 static void update_cpu_load(struct rq *this_rq)
3177 unsigned long this_load = this_rq->load.weight;
3178 unsigned long curr_jiffies = jiffies;
3179 unsigned long pending_updates;
3182 this_rq->nr_load_updates++;
3184 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3185 if (curr_jiffies == this_rq->last_load_update_tick)
3188 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3189 this_rq->last_load_update_tick = curr_jiffies;
3191 /* Update our load: */
3192 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3193 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3194 unsigned long old_load, new_load;
3196 /* scale is effectively 1 << i now, and >> i divides by scale */
3198 old_load = this_rq->cpu_load[i];
3199 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3200 new_load = this_load;
3202 * Round up the averaging division if load is increasing. This
3203 * prevents us from getting stuck on 9 if the load is 10, for
3206 if (new_load > old_load)
3207 new_load += scale - 1;
3209 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3212 sched_avg_update(this_rq);
3215 static void update_cpu_load_active(struct rq *this_rq)
3217 update_cpu_load(this_rq);
3219 calc_load_account_active(this_rq);
3225 * sched_exec - execve() is a valuable balancing opportunity, because at
3226 * this point the task has the smallest effective memory and cache footprint.
3228 void sched_exec(void)
3230 struct task_struct *p = current;
3231 unsigned long flags;
3235 rq = task_rq_lock(p, &flags);
3236 dest_cpu = p->sched_class->select_task_rq(rq, p, SD_BALANCE_EXEC, 0);
3237 if (dest_cpu == smp_processor_id())
3241 * select_task_rq() can race against ->cpus_allowed
3243 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed) &&
3244 likely(cpu_active(dest_cpu)) && migrate_task(p, dest_cpu)) {
3245 struct migration_arg arg = { p, dest_cpu };
3247 task_rq_unlock(rq, &flags);
3248 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
3252 task_rq_unlock(rq, &flags);
3257 DEFINE_PER_CPU(struct kernel_stat, kstat);
3259 EXPORT_PER_CPU_SYMBOL(kstat);
3262 * Return any ns on the sched_clock that have not yet been accounted in
3263 * @p in case that task is currently running.
3265 * Called with task_rq_lock() held on @rq.
3267 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3271 if (task_current(rq, p)) {
3272 update_rq_clock(rq);
3273 ns = rq->clock - p->se.exec_start;
3281 unsigned long long task_delta_exec(struct task_struct *p)
3283 unsigned long flags;
3287 rq = task_rq_lock(p, &flags);
3288 ns = do_task_delta_exec(p, rq);
3289 task_rq_unlock(rq, &flags);
3295 * Return accounted runtime for the task.
3296 * In case the task is currently running, return the runtime plus current's
3297 * pending runtime that have not been accounted yet.
3299 unsigned long long task_sched_runtime(struct task_struct *p)
3301 unsigned long flags;
3305 rq = task_rq_lock(p, &flags);
3306 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3307 task_rq_unlock(rq, &flags);
3313 * Return sum_exec_runtime for the thread group.
3314 * In case the task is currently running, return the sum plus current's
3315 * pending runtime that have not been accounted yet.
3317 * Note that the thread group might have other running tasks as well,
3318 * so the return value not includes other pending runtime that other
3319 * running tasks might have.
3321 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3323 struct task_cputime totals;
3324 unsigned long flags;
3328 rq = task_rq_lock(p, &flags);
3329 thread_group_cputime(p, &totals);
3330 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3331 task_rq_unlock(rq, &flags);
3337 * Account user cpu time to a process.
3338 * @p: the process that the cpu time gets accounted to
3339 * @cputime: the cpu time spent in user space since the last update
3340 * @cputime_scaled: cputime scaled by cpu frequency
3342 void account_user_time(struct task_struct *p, cputime_t cputime,
3343 cputime_t cputime_scaled)
3345 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3348 /* Add user time to process. */
3349 p->utime = cputime_add(p->utime, cputime);
3350 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3351 account_group_user_time(p, cputime);
3353 /* Add user time to cpustat. */
3354 tmp = cputime_to_cputime64(cputime);
3355 if (TASK_NICE(p) > 0)
3356 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3358 cpustat->user = cputime64_add(cpustat->user, tmp);
3360 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3361 /* Account for user time used */
3362 acct_update_integrals(p);
3366 * Account guest cpu time to a process.
3367 * @p: the process that the cpu time gets accounted to
3368 * @cputime: the cpu time spent in virtual machine since the last update
3369 * @cputime_scaled: cputime scaled by cpu frequency
3371 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3372 cputime_t cputime_scaled)
3375 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3377 tmp = cputime_to_cputime64(cputime);
3379 /* Add guest time to process. */
3380 p->utime = cputime_add(p->utime, cputime);
3381 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3382 account_group_user_time(p, cputime);
3383 p->gtime = cputime_add(p->gtime, cputime);
3385 /* Add guest time to cpustat. */
3386 if (TASK_NICE(p) > 0) {
3387 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3388 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3390 cpustat->user = cputime64_add(cpustat->user, tmp);
3391 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3396 * Account system cpu time to a process.
3397 * @p: the process that the cpu time gets accounted to
3398 * @hardirq_offset: the offset to subtract from hardirq_count()
3399 * @cputime: the cpu time spent in kernel space since the last update
3400 * @cputime_scaled: cputime scaled by cpu frequency
3402 void account_system_time(struct task_struct *p, int hardirq_offset,
3403 cputime_t cputime, cputime_t cputime_scaled)
3405 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3408 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3409 account_guest_time(p, cputime, cputime_scaled);
3413 /* Add system time to process. */
3414 p->stime = cputime_add(p->stime, cputime);
3415 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3416 account_group_system_time(p, cputime);
3418 /* Add system time to cpustat. */
3419 tmp = cputime_to_cputime64(cputime);
3420 if (hardirq_count() - hardirq_offset)
3421 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3422 else if (softirq_count())
3423 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3425 cpustat->system = cputime64_add(cpustat->system, tmp);
3427 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3429 /* Account for system time used */
3430 acct_update_integrals(p);
3434 * Account for involuntary wait time.
3435 * @steal: the cpu time spent in involuntary wait
3437 void account_steal_time(cputime_t cputime)
3439 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3440 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3442 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3446 * Account for idle time.
3447 * @cputime: the cpu time spent in idle wait
3449 void account_idle_time(cputime_t cputime)
3451 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3452 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3453 struct rq *rq = this_rq();
3455 if (atomic_read(&rq->nr_iowait) > 0)
3456 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3458 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3461 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3464 * Account a single tick of cpu time.
3465 * @p: the process that the cpu time gets accounted to
3466 * @user_tick: indicates if the tick is a user or a system tick
3468 void account_process_tick(struct task_struct *p, int user_tick)
3470 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3471 struct rq *rq = this_rq();
3474 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3475 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3476 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3479 account_idle_time(cputime_one_jiffy);
3483 * Account multiple ticks of steal time.
3484 * @p: the process from which the cpu time has been stolen
3485 * @ticks: number of stolen ticks
3487 void account_steal_ticks(unsigned long ticks)
3489 account_steal_time(jiffies_to_cputime(ticks));
3493 * Account multiple ticks of idle time.
3494 * @ticks: number of stolen ticks
3496 void account_idle_ticks(unsigned long ticks)
3498 account_idle_time(jiffies_to_cputime(ticks));
3504 * Use precise platform statistics if available:
3506 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3507 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3513 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3515 struct task_cputime cputime;
3517 thread_group_cputime(p, &cputime);
3519 *ut = cputime.utime;
3520 *st = cputime.stime;
3524 #ifndef nsecs_to_cputime
3525 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3528 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3530 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
3533 * Use CFS's precise accounting:
3535 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3541 do_div(temp, total);
3542 utime = (cputime_t)temp;
3547 * Compare with previous values, to keep monotonicity:
3549 p->prev_utime = max(p->prev_utime, utime);
3550 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
3552 *ut = p->prev_utime;
3553 *st = p->prev_stime;
3557 * Must be called with siglock held.
3559 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3561 struct signal_struct *sig = p->signal;
3562 struct task_cputime cputime;
3563 cputime_t rtime, utime, total;
3565 thread_group_cputime(p, &cputime);
3567 total = cputime_add(cputime.utime, cputime.stime);
3568 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3573 temp *= cputime.utime;
3574 do_div(temp, total);
3575 utime = (cputime_t)temp;
3579 sig->prev_utime = max(sig->prev_utime, utime);
3580 sig->prev_stime = max(sig->prev_stime,
3581 cputime_sub(rtime, sig->prev_utime));
3583 *ut = sig->prev_utime;
3584 *st = sig->prev_stime;
3589 * This function gets called by the timer code, with HZ frequency.
3590 * We call it with interrupts disabled.
3592 * It also gets called by the fork code, when changing the parent's
3595 void scheduler_tick(void)
3597 int cpu = smp_processor_id();
3598 struct rq *rq = cpu_rq(cpu);
3599 struct task_struct *curr = rq->curr;
3603 raw_spin_lock(&rq->lock);
3604 update_rq_clock(rq);
3605 update_cpu_load_active(rq);
3606 curr->sched_class->task_tick(rq, curr, 0);
3607 raw_spin_unlock(&rq->lock);
3609 perf_event_task_tick(curr);
3612 rq->idle_at_tick = idle_cpu(cpu);
3613 trigger_load_balance(rq, cpu);
3617 notrace unsigned long get_parent_ip(unsigned long addr)
3619 if (in_lock_functions(addr)) {
3620 addr = CALLER_ADDR2;
3621 if (in_lock_functions(addr))
3622 addr = CALLER_ADDR3;
3627 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3628 defined(CONFIG_PREEMPT_TRACER))
3630 void __kprobes add_preempt_count(int val)
3632 #ifdef CONFIG_DEBUG_PREEMPT
3636 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3639 preempt_count() += val;
3640 #ifdef CONFIG_DEBUG_PREEMPT
3642 * Spinlock count overflowing soon?
3644 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3647 if (preempt_count() == val)
3648 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3650 EXPORT_SYMBOL(add_preempt_count);
3652 void __kprobes sub_preempt_count(int val)
3654 #ifdef CONFIG_DEBUG_PREEMPT
3658 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3661 * Is the spinlock portion underflowing?
3663 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3664 !(preempt_count() & PREEMPT_MASK)))
3668 if (preempt_count() == val)
3669 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3670 preempt_count() -= val;
3672 EXPORT_SYMBOL(sub_preempt_count);
3677 * Print scheduling while atomic bug:
3679 static noinline void __schedule_bug(struct task_struct *prev)
3681 struct pt_regs *regs = get_irq_regs();
3683 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3684 prev->comm, prev->pid, preempt_count());
3686 debug_show_held_locks(prev);
3688 if (irqs_disabled())
3689 print_irqtrace_events(prev);
3698 * Various schedule()-time debugging checks and statistics:
3700 static inline void schedule_debug(struct task_struct *prev)
3703 * Test if we are atomic. Since do_exit() needs to call into
3704 * schedule() atomically, we ignore that path for now.
3705 * Otherwise, whine if we are scheduling when we should not be.
3707 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3708 __schedule_bug(prev);
3710 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3712 schedstat_inc(this_rq(), sched_count);
3713 #ifdef CONFIG_SCHEDSTATS
3714 if (unlikely(prev->lock_depth >= 0)) {
3715 schedstat_inc(this_rq(), bkl_count);
3716 schedstat_inc(prev, sched_info.bkl_count);
3721 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3724 update_rq_clock(rq);
3725 rq->skip_clock_update = 0;
3726 prev->sched_class->put_prev_task(rq, prev);
3730 * Pick up the highest-prio task:
3732 static inline struct task_struct *
3733 pick_next_task(struct rq *rq)
3735 const struct sched_class *class;
3736 struct task_struct *p;
3739 * Optimization: we know that if all tasks are in
3740 * the fair class we can call that function directly:
3742 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3743 p = fair_sched_class.pick_next_task(rq);
3748 for_each_class(class) {
3749 p = class->pick_next_task(rq);
3754 BUG(); /* the idle class will always have a runnable task */
3758 * schedule() is the main scheduler function.
3760 asmlinkage void __sched schedule(void)
3762 struct task_struct *prev, *next;
3763 unsigned long *switch_count;
3769 cpu = smp_processor_id();
3771 rcu_note_context_switch(cpu);
3774 release_kernel_lock(prev);
3775 need_resched_nonpreemptible:
3777 schedule_debug(prev);
3779 if (sched_feat(HRTICK))
3782 raw_spin_lock_irq(&rq->lock);
3783 clear_tsk_need_resched(prev);
3785 switch_count = &prev->nivcsw;
3786 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3787 if (unlikely(signal_pending_state(prev->state, prev))) {
3788 prev->state = TASK_RUNNING;
3791 * If a worker is going to sleep, notify and
3792 * ask workqueue whether it wants to wake up a
3793 * task to maintain concurrency. If so, wake
3796 if (prev->flags & PF_WQ_WORKER) {
3797 struct task_struct *to_wakeup;
3799 to_wakeup = wq_worker_sleeping(prev, cpu);
3801 try_to_wake_up_local(to_wakeup);
3803 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3805 switch_count = &prev->nvcsw;
3808 pre_schedule(rq, prev);
3810 if (unlikely(!rq->nr_running))
3811 idle_balance(cpu, rq);
3813 put_prev_task(rq, prev);
3814 next = pick_next_task(rq);
3816 if (likely(prev != next)) {
3817 sched_info_switch(prev, next);
3818 perf_event_task_sched_out(prev, next);
3824 context_switch(rq, prev, next); /* unlocks the rq */
3826 * The context switch have flipped the stack from under us
3827 * and restored the local variables which were saved when
3828 * this task called schedule() in the past. prev == current
3829 * is still correct, but it can be moved to another cpu/rq.
3831 cpu = smp_processor_id();
3834 raw_spin_unlock_irq(&rq->lock);
3838 if (unlikely(reacquire_kernel_lock(prev)))
3839 goto need_resched_nonpreemptible;
3841 preempt_enable_no_resched();
3845 EXPORT_SYMBOL(schedule);
3847 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3849 * Look out! "owner" is an entirely speculative pointer
3850 * access and not reliable.
3852 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
3857 if (!sched_feat(OWNER_SPIN))
3860 #ifdef CONFIG_DEBUG_PAGEALLOC
3862 * Need to access the cpu field knowing that
3863 * DEBUG_PAGEALLOC could have unmapped it if
3864 * the mutex owner just released it and exited.
3866 if (probe_kernel_address(&owner->cpu, cpu))
3873 * Even if the access succeeded (likely case),
3874 * the cpu field may no longer be valid.
3876 if (cpu >= nr_cpumask_bits)
3880 * We need to validate that we can do a
3881 * get_cpu() and that we have the percpu area.
3883 if (!cpu_online(cpu))
3890 * Owner changed, break to re-assess state.
3892 if (lock->owner != owner) {
3894 * If the lock has switched to a different owner,
3895 * we likely have heavy contention. Return 0 to quit
3896 * optimistic spinning and not contend further:
3904 * Is that owner really running on that cpu?
3906 if (task_thread_info(rq->curr) != owner || need_resched())
3916 #ifdef CONFIG_PREEMPT
3918 * this is the entry point to schedule() from in-kernel preemption
3919 * off of preempt_enable. Kernel preemptions off return from interrupt
3920 * occur there and call schedule directly.
3922 asmlinkage void __sched notrace preempt_schedule(void)
3924 struct thread_info *ti = current_thread_info();
3927 * If there is a non-zero preempt_count or interrupts are disabled,
3928 * we do not want to preempt the current task. Just return..
3930 if (likely(ti->preempt_count || irqs_disabled()))
3934 add_preempt_count_notrace(PREEMPT_ACTIVE);
3936 sub_preempt_count_notrace(PREEMPT_ACTIVE);
3939 * Check again in case we missed a preemption opportunity
3940 * between schedule and now.
3943 } while (need_resched());
3945 EXPORT_SYMBOL(preempt_schedule);
3948 * this is the entry point to schedule() from kernel preemption
3949 * off of irq context.
3950 * Note, that this is called and return with irqs disabled. This will
3951 * protect us against recursive calling from irq.
3953 asmlinkage void __sched preempt_schedule_irq(void)
3955 struct thread_info *ti = current_thread_info();
3957 /* Catch callers which need to be fixed */
3958 BUG_ON(ti->preempt_count || !irqs_disabled());
3961 add_preempt_count(PREEMPT_ACTIVE);
3964 local_irq_disable();
3965 sub_preempt_count(PREEMPT_ACTIVE);
3968 * Check again in case we missed a preemption opportunity
3969 * between schedule and now.
3972 } while (need_resched());
3975 #endif /* CONFIG_PREEMPT */
3977 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3980 return try_to_wake_up(curr->private, mode, wake_flags);
3982 EXPORT_SYMBOL(default_wake_function);
3985 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3986 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3987 * number) then we wake all the non-exclusive tasks and one exclusive task.
3989 * There are circumstances in which we can try to wake a task which has already
3990 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3991 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3993 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3994 int nr_exclusive, int wake_flags, void *key)
3996 wait_queue_t *curr, *next;
3998 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3999 unsigned flags = curr->flags;
4001 if (curr->func(curr, mode, wake_flags, key) &&
4002 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4008 * __wake_up - wake up threads blocked on a waitqueue.
4010 * @mode: which threads
4011 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4012 * @key: is directly passed to the wakeup function
4014 * It may be assumed that this function implies a write memory barrier before
4015 * changing the task state if and only if any tasks are woken up.
4017 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4018 int nr_exclusive, void *key)
4020 unsigned long flags;
4022 spin_lock_irqsave(&q->lock, flags);
4023 __wake_up_common(q, mode, nr_exclusive, 0, key);
4024 spin_unlock_irqrestore(&q->lock, flags);
4026 EXPORT_SYMBOL(__wake_up);
4029 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4031 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4033 __wake_up_common(q, mode, 1, 0, NULL);
4035 EXPORT_SYMBOL_GPL(__wake_up_locked);
4037 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
4039 __wake_up_common(q, mode, 1, 0, key);
4043 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4045 * @mode: which threads
4046 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4047 * @key: opaque value to be passed to wakeup targets
4049 * The sync wakeup differs that the waker knows that it will schedule
4050 * away soon, so while the target thread will be woken up, it will not
4051 * be migrated to another CPU - ie. the two threads are 'synchronized'
4052 * with each other. This can prevent needless bouncing between CPUs.
4054 * On UP it can prevent extra preemption.
4056 * It may be assumed that this function implies a write memory barrier before
4057 * changing the task state if and only if any tasks are woken up.
4059 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4060 int nr_exclusive, void *key)
4062 unsigned long flags;
4063 int wake_flags = WF_SYNC;
4068 if (unlikely(!nr_exclusive))
4071 spin_lock_irqsave(&q->lock, flags);
4072 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4073 spin_unlock_irqrestore(&q->lock, flags);
4075 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4078 * __wake_up_sync - see __wake_up_sync_key()
4080 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4082 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4084 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4087 * complete: - signals a single thread waiting on this completion
4088 * @x: holds the state of this particular completion
4090 * This will wake up a single thread waiting on this completion. Threads will be
4091 * awakened in the same order in which they were queued.
4093 * See also complete_all(), wait_for_completion() and related routines.
4095 * It may be assumed that this function implies a write memory barrier before
4096 * changing the task state if and only if any tasks are woken up.
4098 void complete(struct completion *x)
4100 unsigned long flags;
4102 spin_lock_irqsave(&x->wait.lock, flags);
4104 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4105 spin_unlock_irqrestore(&x->wait.lock, flags);
4107 EXPORT_SYMBOL(complete);
4110 * complete_all: - signals all threads waiting on this completion
4111 * @x: holds the state of this particular completion
4113 * This will wake up all threads waiting on this particular completion event.
4115 * It may be assumed that this function implies a write memory barrier before
4116 * changing the task state if and only if any tasks are woken up.
4118 void complete_all(struct completion *x)
4120 unsigned long flags;
4122 spin_lock_irqsave(&x->wait.lock, flags);
4123 x->done += UINT_MAX/2;
4124 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4125 spin_unlock_irqrestore(&x->wait.lock, flags);
4127 EXPORT_SYMBOL(complete_all);
4129 static inline long __sched
4130 do_wait_for_common(struct completion *x, long timeout, int state)
4133 DECLARE_WAITQUEUE(wait, current);
4135 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4137 if (signal_pending_state(state, current)) {
4138 timeout = -ERESTARTSYS;
4141 __set_current_state(state);
4142 spin_unlock_irq(&x->wait.lock);
4143 timeout = schedule_timeout(timeout);
4144 spin_lock_irq(&x->wait.lock);
4145 } while (!x->done && timeout);
4146 __remove_wait_queue(&x->wait, &wait);
4151 return timeout ?: 1;
4155 wait_for_common(struct completion *x, long timeout, int state)
4159 spin_lock_irq(&x->wait.lock);
4160 timeout = do_wait_for_common(x, timeout, state);
4161 spin_unlock_irq(&x->wait.lock);
4166 * wait_for_completion: - waits for completion of a task
4167 * @x: holds the state of this particular completion
4169 * This waits to be signaled for completion of a specific task. It is NOT
4170 * interruptible and there is no timeout.
4172 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4173 * and interrupt capability. Also see complete().
4175 void __sched wait_for_completion(struct completion *x)
4177 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4179 EXPORT_SYMBOL(wait_for_completion);
4182 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4183 * @x: holds the state of this particular completion
4184 * @timeout: timeout value in jiffies
4186 * This waits for either a completion of a specific task to be signaled or for a
4187 * specified timeout to expire. The timeout is in jiffies. It is not
4190 unsigned long __sched
4191 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4193 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4195 EXPORT_SYMBOL(wait_for_completion_timeout);
4198 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4199 * @x: holds the state of this particular completion
4201 * This waits for completion of a specific task to be signaled. It is
4204 int __sched wait_for_completion_interruptible(struct completion *x)
4206 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4207 if (t == -ERESTARTSYS)
4211 EXPORT_SYMBOL(wait_for_completion_interruptible);
4214 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4215 * @x: holds the state of this particular completion
4216 * @timeout: timeout value in jiffies
4218 * This waits for either a completion of a specific task to be signaled or for a
4219 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4221 unsigned long __sched
4222 wait_for_completion_interruptible_timeout(struct completion *x,
4223 unsigned long timeout)
4225 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4227 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4230 * wait_for_completion_killable: - waits for completion of a task (killable)
4231 * @x: holds the state of this particular completion
4233 * This waits to be signaled for completion of a specific task. It can be
4234 * interrupted by a kill signal.
4236 int __sched wait_for_completion_killable(struct completion *x)
4238 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4239 if (t == -ERESTARTSYS)
4243 EXPORT_SYMBOL(wait_for_completion_killable);
4246 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4247 * @x: holds the state of this particular completion
4248 * @timeout: timeout value in jiffies
4250 * This waits for either a completion of a specific task to be
4251 * signaled or for a specified timeout to expire. It can be
4252 * interrupted by a kill signal. The timeout is in jiffies.
4254 unsigned long __sched
4255 wait_for_completion_killable_timeout(struct completion *x,
4256 unsigned long timeout)
4258 return wait_for_common(x, timeout, TASK_KILLABLE);
4260 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
4263 * try_wait_for_completion - try to decrement a completion without blocking
4264 * @x: completion structure
4266 * Returns: 0 if a decrement cannot be done without blocking
4267 * 1 if a decrement succeeded.
4269 * If a completion is being used as a counting completion,
4270 * attempt to decrement the counter without blocking. This
4271 * enables us to avoid waiting if the resource the completion
4272 * is protecting is not available.
4274 bool try_wait_for_completion(struct completion *x)
4276 unsigned long flags;
4279 spin_lock_irqsave(&x->wait.lock, flags);
4284 spin_unlock_irqrestore(&x->wait.lock, flags);
4287 EXPORT_SYMBOL(try_wait_for_completion);
4290 * completion_done - Test to see if a completion has any waiters
4291 * @x: completion structure
4293 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4294 * 1 if there are no waiters.
4297 bool completion_done(struct completion *x)
4299 unsigned long flags;
4302 spin_lock_irqsave(&x->wait.lock, flags);
4305 spin_unlock_irqrestore(&x->wait.lock, flags);
4308 EXPORT_SYMBOL(completion_done);
4311 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4313 unsigned long flags;
4316 init_waitqueue_entry(&wait, current);
4318 __set_current_state(state);
4320 spin_lock_irqsave(&q->lock, flags);
4321 __add_wait_queue(q, &wait);
4322 spin_unlock(&q->lock);
4323 timeout = schedule_timeout(timeout);
4324 spin_lock_irq(&q->lock);
4325 __remove_wait_queue(q, &wait);
4326 spin_unlock_irqrestore(&q->lock, flags);
4331 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4333 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4335 EXPORT_SYMBOL(interruptible_sleep_on);
4338 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4340 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4342 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4344 void __sched sleep_on(wait_queue_head_t *q)
4346 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4348 EXPORT_SYMBOL(sleep_on);
4350 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4352 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4354 EXPORT_SYMBOL(sleep_on_timeout);
4356 #ifdef CONFIG_RT_MUTEXES
4359 * rt_mutex_setprio - set the current priority of a task
4361 * @prio: prio value (kernel-internal form)
4363 * This function changes the 'effective' priority of a task. It does
4364 * not touch ->normal_prio like __setscheduler().
4366 * Used by the rt_mutex code to implement priority inheritance logic.
4368 void rt_mutex_setprio(struct task_struct *p, int prio)
4370 unsigned long flags;
4371 int oldprio, on_rq, running;
4373 const struct sched_class *prev_class;
4375 BUG_ON(prio < 0 || prio > MAX_PRIO);
4377 rq = task_rq_lock(p, &flags);
4379 trace_sched_pi_setprio(p, prio);
4381 prev_class = p->sched_class;
4382 on_rq = p->se.on_rq;
4383 running = task_current(rq, p);
4385 dequeue_task(rq, p, 0);
4387 p->sched_class->put_prev_task(rq, p);
4390 p->sched_class = &rt_sched_class;
4392 p->sched_class = &fair_sched_class;
4397 p->sched_class->set_curr_task(rq);
4399 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4401 check_class_changed(rq, p, prev_class, oldprio, running);
4403 task_rq_unlock(rq, &flags);
4408 void set_user_nice(struct task_struct *p, long nice)
4410 int old_prio, delta, on_rq;
4411 unsigned long flags;
4414 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4417 * We have to be careful, if called from sys_setpriority(),
4418 * the task might be in the middle of scheduling on another CPU.
4420 rq = task_rq_lock(p, &flags);
4422 * The RT priorities are set via sched_setscheduler(), but we still
4423 * allow the 'normal' nice value to be set - but as expected
4424 * it wont have any effect on scheduling until the task is
4425 * SCHED_FIFO/SCHED_RR:
4427 if (task_has_rt_policy(p)) {
4428 p->static_prio = NICE_TO_PRIO(nice);
4431 on_rq = p->se.on_rq;
4433 dequeue_task(rq, p, 0);
4435 p->static_prio = NICE_TO_PRIO(nice);
4438 p->prio = effective_prio(p);
4439 delta = p->prio - old_prio;
4442 enqueue_task(rq, p, 0);
4444 * If the task increased its priority or is running and
4445 * lowered its priority, then reschedule its CPU:
4447 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4448 resched_task(rq->curr);
4451 task_rq_unlock(rq, &flags);
4453 EXPORT_SYMBOL(set_user_nice);
4456 * can_nice - check if a task can reduce its nice value
4460 int can_nice(const struct task_struct *p, const int nice)
4462 /* convert nice value [19,-20] to rlimit style value [1,40] */
4463 int nice_rlim = 20 - nice;
4465 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4466 capable(CAP_SYS_NICE));
4469 #ifdef __ARCH_WANT_SYS_NICE
4472 * sys_nice - change the priority of the current process.
4473 * @increment: priority increment
4475 * sys_setpriority is a more generic, but much slower function that
4476 * does similar things.
4478 SYSCALL_DEFINE1(nice, int, increment)
4483 * Setpriority might change our priority at the same moment.
4484 * We don't have to worry. Conceptually one call occurs first
4485 * and we have a single winner.
4487 if (increment < -40)
4492 nice = TASK_NICE(current) + increment;
4498 if (increment < 0 && !can_nice(current, nice))
4501 retval = security_task_setnice(current, nice);
4505 set_user_nice(current, nice);
4512 * task_prio - return the priority value of a given task.
4513 * @p: the task in question.
4515 * This is the priority value as seen by users in /proc.
4516 * RT tasks are offset by -200. Normal tasks are centered
4517 * around 0, value goes from -16 to +15.
4519 int task_prio(const struct task_struct *p)
4521 return p->prio - MAX_RT_PRIO;
4525 * task_nice - return the nice value of a given task.
4526 * @p: the task in question.
4528 int task_nice(const struct task_struct *p)
4530 return TASK_NICE(p);
4532 EXPORT_SYMBOL(task_nice);
4535 * idle_cpu - is a given cpu idle currently?
4536 * @cpu: the processor in question.
4538 int idle_cpu(int cpu)
4540 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4544 * idle_task - return the idle task for a given cpu.
4545 * @cpu: the processor in question.
4547 struct task_struct *idle_task(int cpu)
4549 return cpu_rq(cpu)->idle;
4553 * find_process_by_pid - find a process with a matching PID value.
4554 * @pid: the pid in question.
4556 static struct task_struct *find_process_by_pid(pid_t pid)
4558 return pid ? find_task_by_vpid(pid) : current;
4561 /* Actually do priority change: must hold rq lock. */
4563 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4565 BUG_ON(p->se.on_rq);
4568 p->rt_priority = prio;
4569 p->normal_prio = normal_prio(p);
4570 /* we are holding p->pi_lock already */
4571 p->prio = rt_mutex_getprio(p);
4572 if (rt_prio(p->prio))
4573 p->sched_class = &rt_sched_class;
4575 p->sched_class = &fair_sched_class;
4580 * check the target process has a UID that matches the current process's
4582 static bool check_same_owner(struct task_struct *p)
4584 const struct cred *cred = current_cred(), *pcred;
4588 pcred = __task_cred(p);
4589 match = (cred->euid == pcred->euid ||
4590 cred->euid == pcred->uid);
4595 static int __sched_setscheduler(struct task_struct *p, int policy,
4596 struct sched_param *param, bool user)
4598 int retval, oldprio, oldpolicy = -1, on_rq, running;
4599 unsigned long flags;
4600 const struct sched_class *prev_class;
4604 /* may grab non-irq protected spin_locks */
4605 BUG_ON(in_interrupt());
4607 /* double check policy once rq lock held */
4609 reset_on_fork = p->sched_reset_on_fork;
4610 policy = oldpolicy = p->policy;
4612 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4613 policy &= ~SCHED_RESET_ON_FORK;
4615 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4616 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4617 policy != SCHED_IDLE)
4622 * Valid priorities for SCHED_FIFO and SCHED_RR are
4623 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4624 * SCHED_BATCH and SCHED_IDLE is 0.
4626 if (param->sched_priority < 0 ||
4627 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4628 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4630 if (rt_policy(policy) != (param->sched_priority != 0))
4634 * Allow unprivileged RT tasks to decrease priority:
4636 if (user && !capable(CAP_SYS_NICE)) {
4637 if (rt_policy(policy)) {
4638 unsigned long rlim_rtprio =
4639 task_rlimit(p, RLIMIT_RTPRIO);
4641 /* can't set/change the rt policy */
4642 if (policy != p->policy && !rlim_rtprio)
4645 /* can't increase priority */
4646 if (param->sched_priority > p->rt_priority &&
4647 param->sched_priority > rlim_rtprio)
4651 * Like positive nice levels, dont allow tasks to
4652 * move out of SCHED_IDLE either:
4654 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4657 /* can't change other user's priorities */
4658 if (!check_same_owner(p))
4661 /* Normal users shall not reset the sched_reset_on_fork flag */
4662 if (p->sched_reset_on_fork && !reset_on_fork)
4667 retval = security_task_setscheduler(p, policy, param);
4673 * make sure no PI-waiters arrive (or leave) while we are
4674 * changing the priority of the task:
4676 raw_spin_lock_irqsave(&p->pi_lock, flags);
4678 * To be able to change p->policy safely, the apropriate
4679 * runqueue lock must be held.
4681 rq = __task_rq_lock(p);
4684 * Changing the policy of the stop threads its a very bad idea
4686 if (p == rq->stop) {
4687 __task_rq_unlock(rq);
4688 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4692 #ifdef CONFIG_RT_GROUP_SCHED
4695 * Do not allow realtime tasks into groups that have no runtime
4698 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4699 task_group(p)->rt_bandwidth.rt_runtime == 0) {
4700 __task_rq_unlock(rq);
4701 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4707 /* recheck policy now with rq lock held */
4708 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4709 policy = oldpolicy = -1;
4710 __task_rq_unlock(rq);
4711 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4714 on_rq = p->se.on_rq;
4715 running = task_current(rq, p);
4717 deactivate_task(rq, p, 0);
4719 p->sched_class->put_prev_task(rq, p);
4721 p->sched_reset_on_fork = reset_on_fork;
4724 prev_class = p->sched_class;
4725 __setscheduler(rq, p, policy, param->sched_priority);
4728 p->sched_class->set_curr_task(rq);
4730 activate_task(rq, p, 0);
4732 check_class_changed(rq, p, prev_class, oldprio, running);
4734 __task_rq_unlock(rq);
4735 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4737 rt_mutex_adjust_pi(p);
4743 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4744 * @p: the task in question.
4745 * @policy: new policy.
4746 * @param: structure containing the new RT priority.
4748 * NOTE that the task may be already dead.
4750 int sched_setscheduler(struct task_struct *p, int policy,
4751 struct sched_param *param)
4753 return __sched_setscheduler(p, policy, param, true);
4755 EXPORT_SYMBOL_GPL(sched_setscheduler);
4758 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4759 * @p: the task in question.
4760 * @policy: new policy.
4761 * @param: structure containing the new RT priority.
4763 * Just like sched_setscheduler, only don't bother checking if the
4764 * current context has permission. For example, this is needed in
4765 * stop_machine(): we create temporary high priority worker threads,
4766 * but our caller might not have that capability.
4768 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4769 struct sched_param *param)
4771 return __sched_setscheduler(p, policy, param, false);
4775 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4777 struct sched_param lparam;
4778 struct task_struct *p;
4781 if (!param || pid < 0)
4783 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4788 p = find_process_by_pid(pid);
4790 retval = sched_setscheduler(p, policy, &lparam);
4797 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4798 * @pid: the pid in question.
4799 * @policy: new policy.
4800 * @param: structure containing the new RT priority.
4802 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4803 struct sched_param __user *, param)
4805 /* negative values for policy are not valid */
4809 return do_sched_setscheduler(pid, policy, param);
4813 * sys_sched_setparam - set/change the RT priority of a thread
4814 * @pid: the pid in question.
4815 * @param: structure containing the new RT priority.
4817 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4819 return do_sched_setscheduler(pid, -1, param);
4823 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4824 * @pid: the pid in question.
4826 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4828 struct task_struct *p;
4836 p = find_process_by_pid(pid);
4838 retval = security_task_getscheduler(p);
4841 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4848 * sys_sched_getparam - get the RT priority of a thread
4849 * @pid: the pid in question.
4850 * @param: structure containing the RT priority.
4852 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4854 struct sched_param lp;
4855 struct task_struct *p;
4858 if (!param || pid < 0)
4862 p = find_process_by_pid(pid);
4867 retval = security_task_getscheduler(p);
4871 lp.sched_priority = p->rt_priority;
4875 * This one might sleep, we cannot do it with a spinlock held ...
4877 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4886 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4888 cpumask_var_t cpus_allowed, new_mask;
4889 struct task_struct *p;
4895 p = find_process_by_pid(pid);
4902 /* Prevent p going away */
4906 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4910 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4912 goto out_free_cpus_allowed;
4915 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
4918 retval = security_task_setscheduler(p, 0, NULL);
4922 cpuset_cpus_allowed(p, cpus_allowed);
4923 cpumask_and(new_mask, in_mask, cpus_allowed);
4925 retval = set_cpus_allowed_ptr(p, new_mask);
4928 cpuset_cpus_allowed(p, cpus_allowed);
4929 if (!cpumask_subset(new_mask, cpus_allowed)) {
4931 * We must have raced with a concurrent cpuset
4932 * update. Just reset the cpus_allowed to the
4933 * cpuset's cpus_allowed
4935 cpumask_copy(new_mask, cpus_allowed);
4940 free_cpumask_var(new_mask);
4941 out_free_cpus_allowed:
4942 free_cpumask_var(cpus_allowed);
4949 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4950 struct cpumask *new_mask)
4952 if (len < cpumask_size())
4953 cpumask_clear(new_mask);
4954 else if (len > cpumask_size())
4955 len = cpumask_size();
4957 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4961 * sys_sched_setaffinity - set the cpu affinity of a process
4962 * @pid: pid of the process
4963 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4964 * @user_mask_ptr: user-space pointer to the new cpu mask
4966 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4967 unsigned long __user *, user_mask_ptr)
4969 cpumask_var_t new_mask;
4972 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4975 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4977 retval = sched_setaffinity(pid, new_mask);
4978 free_cpumask_var(new_mask);
4982 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4984 struct task_struct *p;
4985 unsigned long flags;
4993 p = find_process_by_pid(pid);
4997 retval = security_task_getscheduler(p);
5001 rq = task_rq_lock(p, &flags);
5002 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5003 task_rq_unlock(rq, &flags);
5013 * sys_sched_getaffinity - get the cpu affinity of a process
5014 * @pid: pid of the process
5015 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5016 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5018 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5019 unsigned long __user *, user_mask_ptr)
5024 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5026 if (len & (sizeof(unsigned long)-1))
5029 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5032 ret = sched_getaffinity(pid, mask);
5034 size_t retlen = min_t(size_t, len, cpumask_size());
5036 if (copy_to_user(user_mask_ptr, mask, retlen))
5041 free_cpumask_var(mask);
5047 * sys_sched_yield - yield the current processor to other threads.
5049 * This function yields the current CPU to other tasks. If there are no
5050 * other threads running on this CPU then this function will return.
5052 SYSCALL_DEFINE0(sched_yield)
5054 struct rq *rq = this_rq_lock();
5056 schedstat_inc(rq, yld_count);
5057 current->sched_class->yield_task(rq);
5060 * Since we are going to call schedule() anyway, there's
5061 * no need to preempt or enable interrupts:
5063 __release(rq->lock);
5064 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5065 do_raw_spin_unlock(&rq->lock);
5066 preempt_enable_no_resched();
5073 static inline int should_resched(void)
5075 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5078 static void __cond_resched(void)
5080 add_preempt_count(PREEMPT_ACTIVE);
5082 sub_preempt_count(PREEMPT_ACTIVE);
5085 int __sched _cond_resched(void)
5087 if (should_resched()) {
5093 EXPORT_SYMBOL(_cond_resched);
5096 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5097 * call schedule, and on return reacquire the lock.
5099 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5100 * operations here to prevent schedule() from being called twice (once via
5101 * spin_unlock(), once by hand).
5103 int __cond_resched_lock(spinlock_t *lock)
5105 int resched = should_resched();
5108 lockdep_assert_held(lock);
5110 if (spin_needbreak(lock) || resched) {
5121 EXPORT_SYMBOL(__cond_resched_lock);
5123 int __sched __cond_resched_softirq(void)
5125 BUG_ON(!in_softirq());
5127 if (should_resched()) {
5135 EXPORT_SYMBOL(__cond_resched_softirq);
5138 * yield - yield the current processor to other threads.
5140 * This is a shortcut for kernel-space yielding - it marks the
5141 * thread runnable and calls sys_sched_yield().
5143 void __sched yield(void)
5145 set_current_state(TASK_RUNNING);
5148 EXPORT_SYMBOL(yield);
5151 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5152 * that process accounting knows that this is a task in IO wait state.
5154 void __sched io_schedule(void)
5156 struct rq *rq = raw_rq();
5158 delayacct_blkio_start();
5159 atomic_inc(&rq->nr_iowait);
5160 current->in_iowait = 1;
5162 current->in_iowait = 0;
5163 atomic_dec(&rq->nr_iowait);
5164 delayacct_blkio_end();
5166 EXPORT_SYMBOL(io_schedule);
5168 long __sched io_schedule_timeout(long timeout)
5170 struct rq *rq = raw_rq();
5173 delayacct_blkio_start();
5174 atomic_inc(&rq->nr_iowait);
5175 current->in_iowait = 1;
5176 ret = schedule_timeout(timeout);
5177 current->in_iowait = 0;
5178 atomic_dec(&rq->nr_iowait);
5179 delayacct_blkio_end();
5184 * sys_sched_get_priority_max - return maximum RT priority.
5185 * @policy: scheduling class.
5187 * this syscall returns the maximum rt_priority that can be used
5188 * by a given scheduling class.
5190 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5197 ret = MAX_USER_RT_PRIO-1;
5209 * sys_sched_get_priority_min - return minimum RT priority.
5210 * @policy: scheduling class.
5212 * this syscall returns the minimum rt_priority that can be used
5213 * by a given scheduling class.
5215 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5233 * sys_sched_rr_get_interval - return the default timeslice of a process.
5234 * @pid: pid of the process.
5235 * @interval: userspace pointer to the timeslice value.
5237 * this syscall writes the default timeslice value of a given process
5238 * into the user-space timespec buffer. A value of '0' means infinity.
5240 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5241 struct timespec __user *, interval)
5243 struct task_struct *p;
5244 unsigned int time_slice;
5245 unsigned long flags;
5255 p = find_process_by_pid(pid);
5259 retval = security_task_getscheduler(p);
5263 rq = task_rq_lock(p, &flags);
5264 time_slice = p->sched_class->get_rr_interval(rq, p);
5265 task_rq_unlock(rq, &flags);
5268 jiffies_to_timespec(time_slice, &t);
5269 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5277 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5279 void sched_show_task(struct task_struct *p)
5281 unsigned long free = 0;
5284 state = p->state ? __ffs(p->state) + 1 : 0;
5285 printk(KERN_INFO "%-13.13s %c", p->comm,
5286 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5287 #if BITS_PER_LONG == 32
5288 if (state == TASK_RUNNING)
5289 printk(KERN_CONT " running ");
5291 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5293 if (state == TASK_RUNNING)
5294 printk(KERN_CONT " running task ");
5296 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5298 #ifdef CONFIG_DEBUG_STACK_USAGE
5299 free = stack_not_used(p);
5301 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5302 task_pid_nr(p), task_pid_nr(p->real_parent),
5303 (unsigned long)task_thread_info(p)->flags);
5305 show_stack(p, NULL);
5308 void show_state_filter(unsigned long state_filter)
5310 struct task_struct *g, *p;
5312 #if BITS_PER_LONG == 32
5314 " task PC stack pid father\n");
5317 " task PC stack pid father\n");
5319 read_lock(&tasklist_lock);
5320 do_each_thread(g, p) {
5322 * reset the NMI-timeout, listing all files on a slow
5323 * console might take alot of time:
5325 touch_nmi_watchdog();
5326 if (!state_filter || (p->state & state_filter))
5328 } while_each_thread(g, p);
5330 touch_all_softlockup_watchdogs();
5332 #ifdef CONFIG_SCHED_DEBUG
5333 sysrq_sched_debug_show();
5335 read_unlock(&tasklist_lock);
5337 * Only show locks if all tasks are dumped:
5340 debug_show_all_locks();
5343 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5345 idle->sched_class = &idle_sched_class;
5349 * init_idle - set up an idle thread for a given CPU
5350 * @idle: task in question
5351 * @cpu: cpu the idle task belongs to
5353 * NOTE: this function does not set the idle thread's NEED_RESCHED
5354 * flag, to make booting more robust.
5356 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5358 struct rq *rq = cpu_rq(cpu);
5359 unsigned long flags;
5361 raw_spin_lock_irqsave(&rq->lock, flags);
5364 idle->state = TASK_RUNNING;
5365 idle->se.exec_start = sched_clock();
5367 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
5368 __set_task_cpu(idle, cpu);
5370 rq->curr = rq->idle = idle;
5371 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5374 raw_spin_unlock_irqrestore(&rq->lock, flags);
5376 /* Set the preempt count _outside_ the spinlocks! */
5377 #if defined(CONFIG_PREEMPT)
5378 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5380 task_thread_info(idle)->preempt_count = 0;
5383 * The idle tasks have their own, simple scheduling class:
5385 idle->sched_class = &idle_sched_class;
5386 ftrace_graph_init_task(idle);
5390 * In a system that switches off the HZ timer nohz_cpu_mask
5391 * indicates which cpus entered this state. This is used
5392 * in the rcu update to wait only for active cpus. For system
5393 * which do not switch off the HZ timer nohz_cpu_mask should
5394 * always be CPU_BITS_NONE.
5396 cpumask_var_t nohz_cpu_mask;
5399 * Increase the granularity value when there are more CPUs,
5400 * because with more CPUs the 'effective latency' as visible
5401 * to users decreases. But the relationship is not linear,
5402 * so pick a second-best guess by going with the log2 of the
5405 * This idea comes from the SD scheduler of Con Kolivas:
5407 static int get_update_sysctl_factor(void)
5409 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5410 unsigned int factor;
5412 switch (sysctl_sched_tunable_scaling) {
5413 case SCHED_TUNABLESCALING_NONE:
5416 case SCHED_TUNABLESCALING_LINEAR:
5419 case SCHED_TUNABLESCALING_LOG:
5421 factor = 1 + ilog2(cpus);
5428 static void update_sysctl(void)
5430 unsigned int factor = get_update_sysctl_factor();
5432 #define SET_SYSCTL(name) \
5433 (sysctl_##name = (factor) * normalized_sysctl_##name)
5434 SET_SYSCTL(sched_min_granularity);
5435 SET_SYSCTL(sched_latency);
5436 SET_SYSCTL(sched_wakeup_granularity);
5437 SET_SYSCTL(sched_shares_ratelimit);
5441 static inline void sched_init_granularity(void)
5448 * This is how migration works:
5450 * 1) we invoke migration_cpu_stop() on the target CPU using
5452 * 2) stopper starts to run (implicitly forcing the migrated thread
5454 * 3) it checks whether the migrated task is still in the wrong runqueue.
5455 * 4) if it's in the wrong runqueue then the migration thread removes
5456 * it and puts it into the right queue.
5457 * 5) stopper completes and stop_one_cpu() returns and the migration
5462 * Change a given task's CPU affinity. Migrate the thread to a
5463 * proper CPU and schedule it away if the CPU it's executing on
5464 * is removed from the allowed bitmask.
5466 * NOTE: the caller must have a valid reference to the task, the
5467 * task must not exit() & deallocate itself prematurely. The
5468 * call is not atomic; no spinlocks may be held.
5470 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5472 unsigned long flags;
5474 unsigned int dest_cpu;
5478 * Serialize against TASK_WAKING so that ttwu() and wunt() can
5479 * drop the rq->lock and still rely on ->cpus_allowed.
5482 while (task_is_waking(p))
5484 rq = task_rq_lock(p, &flags);
5485 if (task_is_waking(p)) {
5486 task_rq_unlock(rq, &flags);
5490 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5495 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5496 !cpumask_equal(&p->cpus_allowed, new_mask))) {
5501 if (p->sched_class->set_cpus_allowed)
5502 p->sched_class->set_cpus_allowed(p, new_mask);
5504 cpumask_copy(&p->cpus_allowed, new_mask);
5505 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5508 /* Can the task run on the task's current CPU? If so, we're done */
5509 if (cpumask_test_cpu(task_cpu(p), new_mask))
5512 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
5513 if (migrate_task(p, dest_cpu)) {
5514 struct migration_arg arg = { p, dest_cpu };
5515 /* Need help from migration thread: drop lock and wait. */
5516 task_rq_unlock(rq, &flags);
5517 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
5518 tlb_migrate_finish(p->mm);
5522 task_rq_unlock(rq, &flags);
5526 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5529 * Move (not current) task off this cpu, onto dest cpu. We're doing
5530 * this because either it can't run here any more (set_cpus_allowed()
5531 * away from this CPU, or CPU going down), or because we're
5532 * attempting to rebalance this task on exec (sched_exec).
5534 * So we race with normal scheduler movements, but that's OK, as long
5535 * as the task is no longer on this CPU.
5537 * Returns non-zero if task was successfully migrated.
5539 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5541 struct rq *rq_dest, *rq_src;
5544 if (unlikely(!cpu_active(dest_cpu)))
5547 rq_src = cpu_rq(src_cpu);
5548 rq_dest = cpu_rq(dest_cpu);
5550 double_rq_lock(rq_src, rq_dest);
5551 /* Already moved. */
5552 if (task_cpu(p) != src_cpu)
5554 /* Affinity changed (again). */
5555 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
5559 * If we're not on a rq, the next wake-up will ensure we're
5563 deactivate_task(rq_src, p, 0);
5564 set_task_cpu(p, dest_cpu);
5565 activate_task(rq_dest, p, 0);
5566 check_preempt_curr(rq_dest, p, 0);
5571 double_rq_unlock(rq_src, rq_dest);
5576 * migration_cpu_stop - this will be executed by a highprio stopper thread
5577 * and performs thread migration by bumping thread off CPU then
5578 * 'pushing' onto another runqueue.
5580 static int migration_cpu_stop(void *data)
5582 struct migration_arg *arg = data;
5585 * The original target cpu might have gone down and we might
5586 * be on another cpu but it doesn't matter.
5588 local_irq_disable();
5589 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
5594 #ifdef CONFIG_HOTPLUG_CPU
5596 * Figure out where task on dead CPU should go, use force if necessary.
5598 void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5600 struct rq *rq = cpu_rq(dead_cpu);
5601 int needs_cpu, uninitialized_var(dest_cpu);
5602 unsigned long flags;
5604 local_irq_save(flags);
5606 raw_spin_lock(&rq->lock);
5607 needs_cpu = (task_cpu(p) == dead_cpu) && (p->state != TASK_WAKING);
5609 dest_cpu = select_fallback_rq(dead_cpu, p);
5610 raw_spin_unlock(&rq->lock);
5612 * It can only fail if we race with set_cpus_allowed(),
5613 * in the racer should migrate the task anyway.
5616 __migrate_task(p, dead_cpu, dest_cpu);
5617 local_irq_restore(flags);
5621 * While a dead CPU has no uninterruptible tasks queued at this point,
5622 * it might still have a nonzero ->nr_uninterruptible counter, because
5623 * for performance reasons the counter is not stricly tracking tasks to
5624 * their home CPUs. So we just add the counter to another CPU's counter,
5625 * to keep the global sum constant after CPU-down:
5627 static void migrate_nr_uninterruptible(struct rq *rq_src)
5629 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5630 unsigned long flags;
5632 local_irq_save(flags);
5633 double_rq_lock(rq_src, rq_dest);
5634 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5635 rq_src->nr_uninterruptible = 0;
5636 double_rq_unlock(rq_src, rq_dest);
5637 local_irq_restore(flags);
5640 /* Run through task list and migrate tasks from the dead cpu. */
5641 static void migrate_live_tasks(int src_cpu)
5643 struct task_struct *p, *t;
5645 read_lock(&tasklist_lock);
5647 do_each_thread(t, p) {
5651 if (task_cpu(p) == src_cpu)
5652 move_task_off_dead_cpu(src_cpu, p);
5653 } while_each_thread(t, p);
5655 read_unlock(&tasklist_lock);
5659 * Schedules idle task to be the next runnable task on current CPU.
5660 * It does so by boosting its priority to highest possible.
5661 * Used by CPU offline code.
5663 void sched_idle_next(void)
5665 int this_cpu = smp_processor_id();
5666 struct rq *rq = cpu_rq(this_cpu);
5667 struct task_struct *p = rq->idle;
5668 unsigned long flags;
5670 /* cpu has to be offline */
5671 BUG_ON(cpu_online(this_cpu));
5674 * Strictly not necessary since rest of the CPUs are stopped by now
5675 * and interrupts disabled on the current cpu.
5677 raw_spin_lock_irqsave(&rq->lock, flags);
5679 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5681 activate_task(rq, p, 0);
5683 raw_spin_unlock_irqrestore(&rq->lock, flags);
5687 * Ensures that the idle task is using init_mm right before its cpu goes
5690 void idle_task_exit(void)
5692 struct mm_struct *mm = current->active_mm;
5694 BUG_ON(cpu_online(smp_processor_id()));
5697 switch_mm(mm, &init_mm, current);
5701 /* called under rq->lock with disabled interrupts */
5702 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5704 struct rq *rq = cpu_rq(dead_cpu);
5706 /* Must be exiting, otherwise would be on tasklist. */
5707 BUG_ON(!p->exit_state);
5709 /* Cannot have done final schedule yet: would have vanished. */
5710 BUG_ON(p->state == TASK_DEAD);
5715 * Drop lock around migration; if someone else moves it,
5716 * that's OK. No task can be added to this CPU, so iteration is
5719 raw_spin_unlock_irq(&rq->lock);
5720 move_task_off_dead_cpu(dead_cpu, p);
5721 raw_spin_lock_irq(&rq->lock);
5726 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5727 static void migrate_dead_tasks(unsigned int dead_cpu)
5729 struct rq *rq = cpu_rq(dead_cpu);
5730 struct task_struct *next;
5733 if (!rq->nr_running)
5735 next = pick_next_task(rq);
5738 next->sched_class->put_prev_task(rq, next);
5739 migrate_dead(dead_cpu, next);
5745 * remove the tasks which were accounted by rq from calc_load_tasks.
5747 static void calc_global_load_remove(struct rq *rq)
5749 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5750 rq->calc_load_active = 0;
5752 #endif /* CONFIG_HOTPLUG_CPU */
5754 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5756 static struct ctl_table sd_ctl_dir[] = {
5758 .procname = "sched_domain",
5764 static struct ctl_table sd_ctl_root[] = {
5766 .procname = "kernel",
5768 .child = sd_ctl_dir,
5773 static struct ctl_table *sd_alloc_ctl_entry(int n)
5775 struct ctl_table *entry =
5776 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5781 static void sd_free_ctl_entry(struct ctl_table **tablep)
5783 struct ctl_table *entry;
5786 * In the intermediate directories, both the child directory and
5787 * procname are dynamically allocated and could fail but the mode
5788 * will always be set. In the lowest directory the names are
5789 * static strings and all have proc handlers.
5791 for (entry = *tablep; entry->mode; entry++) {
5793 sd_free_ctl_entry(&entry->child);
5794 if (entry->proc_handler == NULL)
5795 kfree(entry->procname);
5803 set_table_entry(struct ctl_table *entry,
5804 const char *procname, void *data, int maxlen,
5805 mode_t mode, proc_handler *proc_handler)
5807 entry->procname = procname;
5809 entry->maxlen = maxlen;
5811 entry->proc_handler = proc_handler;
5814 static struct ctl_table *
5815 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5817 struct ctl_table *table = sd_alloc_ctl_entry(13);
5822 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5823 sizeof(long), 0644, proc_doulongvec_minmax);
5824 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5825 sizeof(long), 0644, proc_doulongvec_minmax);
5826 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5827 sizeof(int), 0644, proc_dointvec_minmax);
5828 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5829 sizeof(int), 0644, proc_dointvec_minmax);
5830 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5831 sizeof(int), 0644, proc_dointvec_minmax);
5832 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5833 sizeof(int), 0644, proc_dointvec_minmax);
5834 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5835 sizeof(int), 0644, proc_dointvec_minmax);
5836 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5837 sizeof(int), 0644, proc_dointvec_minmax);
5838 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5839 sizeof(int), 0644, proc_dointvec_minmax);
5840 set_table_entry(&table[9], "cache_nice_tries",
5841 &sd->cache_nice_tries,
5842 sizeof(int), 0644, proc_dointvec_minmax);
5843 set_table_entry(&table[10], "flags", &sd->flags,
5844 sizeof(int), 0644, proc_dointvec_minmax);
5845 set_table_entry(&table[11], "name", sd->name,
5846 CORENAME_MAX_SIZE, 0444, proc_dostring);
5847 /* &table[12] is terminator */
5852 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5854 struct ctl_table *entry, *table;
5855 struct sched_domain *sd;
5856 int domain_num = 0, i;
5859 for_each_domain(cpu, sd)
5861 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5866 for_each_domain(cpu, sd) {
5867 snprintf(buf, 32, "domain%d", i);
5868 entry->procname = kstrdup(buf, GFP_KERNEL);
5870 entry->child = sd_alloc_ctl_domain_table(sd);
5877 static struct ctl_table_header *sd_sysctl_header;
5878 static void register_sched_domain_sysctl(void)
5880 int i, cpu_num = num_possible_cpus();
5881 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5884 WARN_ON(sd_ctl_dir[0].child);
5885 sd_ctl_dir[0].child = entry;
5890 for_each_possible_cpu(i) {
5891 snprintf(buf, 32, "cpu%d", i);
5892 entry->procname = kstrdup(buf, GFP_KERNEL);
5894 entry->child = sd_alloc_ctl_cpu_table(i);
5898 WARN_ON(sd_sysctl_header);
5899 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5902 /* may be called multiple times per register */
5903 static void unregister_sched_domain_sysctl(void)
5905 if (sd_sysctl_header)
5906 unregister_sysctl_table(sd_sysctl_header);
5907 sd_sysctl_header = NULL;
5908 if (sd_ctl_dir[0].child)
5909 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5912 static void register_sched_domain_sysctl(void)
5915 static void unregister_sched_domain_sysctl(void)
5920 static void set_rq_online(struct rq *rq)
5923 const struct sched_class *class;
5925 cpumask_set_cpu(rq->cpu, rq->rd->online);
5928 for_each_class(class) {
5929 if (class->rq_online)
5930 class->rq_online(rq);
5935 static void set_rq_offline(struct rq *rq)
5938 const struct sched_class *class;
5940 for_each_class(class) {
5941 if (class->rq_offline)
5942 class->rq_offline(rq);
5945 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5951 * migration_call - callback that gets triggered when a CPU is added.
5952 * Here we can start up the necessary migration thread for the new CPU.
5954 static int __cpuinit
5955 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5957 int cpu = (long)hcpu;
5958 unsigned long flags;
5959 struct rq *rq = cpu_rq(cpu);
5963 case CPU_UP_PREPARE:
5964 case CPU_UP_PREPARE_FROZEN:
5965 rq->calc_load_update = calc_load_update;
5969 case CPU_ONLINE_FROZEN:
5970 /* Update our root-domain */
5971 raw_spin_lock_irqsave(&rq->lock, flags);
5973 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5977 raw_spin_unlock_irqrestore(&rq->lock, flags);
5980 #ifdef CONFIG_HOTPLUG_CPU
5982 case CPU_DEAD_FROZEN:
5983 migrate_live_tasks(cpu);
5984 /* Idle task back to normal (off runqueue, low prio) */
5985 raw_spin_lock_irq(&rq->lock);
5986 deactivate_task(rq, rq->idle, 0);
5987 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5988 rq->idle->sched_class = &idle_sched_class;
5989 migrate_dead_tasks(cpu);
5990 raw_spin_unlock_irq(&rq->lock);
5991 migrate_nr_uninterruptible(rq);
5992 BUG_ON(rq->nr_running != 0);
5993 calc_global_load_remove(rq);
5997 case CPU_DYING_FROZEN:
5998 /* Update our root-domain */
5999 raw_spin_lock_irqsave(&rq->lock, flags);
6001 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6004 raw_spin_unlock_irqrestore(&rq->lock, flags);
6012 * Register at high priority so that task migration (migrate_all_tasks)
6013 * happens before everything else. This has to be lower priority than
6014 * the notifier in the perf_event subsystem, though.
6016 static struct notifier_block __cpuinitdata migration_notifier = {
6017 .notifier_call = migration_call,
6018 .priority = CPU_PRI_MIGRATION,
6021 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
6022 unsigned long action, void *hcpu)
6024 switch (action & ~CPU_TASKS_FROZEN) {
6026 case CPU_DOWN_FAILED:
6027 set_cpu_active((long)hcpu, true);
6034 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
6035 unsigned long action, void *hcpu)
6037 switch (action & ~CPU_TASKS_FROZEN) {
6038 case CPU_DOWN_PREPARE:
6039 set_cpu_active((long)hcpu, false);
6046 static int __init migration_init(void)
6048 void *cpu = (void *)(long)smp_processor_id();
6051 /* Initialize migration for the boot CPU */
6052 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6053 BUG_ON(err == NOTIFY_BAD);
6054 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6055 register_cpu_notifier(&migration_notifier);
6057 /* Register cpu active notifiers */
6058 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
6059 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
6063 early_initcall(migration_init);
6068 #ifdef CONFIG_SCHED_DEBUG
6070 static __read_mostly int sched_domain_debug_enabled;
6072 static int __init sched_domain_debug_setup(char *str)
6074 sched_domain_debug_enabled = 1;
6078 early_param("sched_debug", sched_domain_debug_setup);
6080 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6081 struct cpumask *groupmask)
6083 struct sched_group *group = sd->groups;
6086 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6087 cpumask_clear(groupmask);
6089 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6091 if (!(sd->flags & SD_LOAD_BALANCE)) {
6092 printk("does not load-balance\n");
6094 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6099 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6101 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6102 printk(KERN_ERR "ERROR: domain->span does not contain "
6105 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6106 printk(KERN_ERR "ERROR: domain->groups does not contain"
6110 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6114 printk(KERN_ERR "ERROR: group is NULL\n");
6118 if (!group->cpu_power) {
6119 printk(KERN_CONT "\n");
6120 printk(KERN_ERR "ERROR: domain->cpu_power not "
6125 if (!cpumask_weight(sched_group_cpus(group))) {
6126 printk(KERN_CONT "\n");
6127 printk(KERN_ERR "ERROR: empty group\n");
6131 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6132 printk(KERN_CONT "\n");
6133 printk(KERN_ERR "ERROR: repeated CPUs\n");
6137 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6139 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6141 printk(KERN_CONT " %s", str);
6142 if (group->cpu_power != SCHED_LOAD_SCALE) {
6143 printk(KERN_CONT " (cpu_power = %d)",
6147 group = group->next;
6148 } while (group != sd->groups);
6149 printk(KERN_CONT "\n");
6151 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6152 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6155 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6156 printk(KERN_ERR "ERROR: parent span is not a superset "
6157 "of domain->span\n");
6161 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6163 cpumask_var_t groupmask;
6166 if (!sched_domain_debug_enabled)
6170 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6174 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6176 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6177 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6182 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6189 free_cpumask_var(groupmask);
6191 #else /* !CONFIG_SCHED_DEBUG */
6192 # define sched_domain_debug(sd, cpu) do { } while (0)
6193 #endif /* CONFIG_SCHED_DEBUG */
6195 static int sd_degenerate(struct sched_domain *sd)
6197 if (cpumask_weight(sched_domain_span(sd)) == 1)
6200 /* Following flags need at least 2 groups */
6201 if (sd->flags & (SD_LOAD_BALANCE |
6202 SD_BALANCE_NEWIDLE |
6206 SD_SHARE_PKG_RESOURCES)) {
6207 if (sd->groups != sd->groups->next)
6211 /* Following flags don't use groups */
6212 if (sd->flags & (SD_WAKE_AFFINE))
6219 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6221 unsigned long cflags = sd->flags, pflags = parent->flags;
6223 if (sd_degenerate(parent))
6226 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6229 /* Flags needing groups don't count if only 1 group in parent */
6230 if (parent->groups == parent->groups->next) {
6231 pflags &= ~(SD_LOAD_BALANCE |
6232 SD_BALANCE_NEWIDLE |
6236 SD_SHARE_PKG_RESOURCES);
6237 if (nr_node_ids == 1)
6238 pflags &= ~SD_SERIALIZE;
6240 if (~cflags & pflags)
6246 static void free_rootdomain(struct root_domain *rd)
6248 synchronize_sched();
6250 cpupri_cleanup(&rd->cpupri);
6252 free_cpumask_var(rd->rto_mask);
6253 free_cpumask_var(rd->online);
6254 free_cpumask_var(rd->span);
6258 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6260 struct root_domain *old_rd = NULL;
6261 unsigned long flags;
6263 raw_spin_lock_irqsave(&rq->lock, flags);
6268 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6271 cpumask_clear_cpu(rq->cpu, old_rd->span);
6274 * If we dont want to free the old_rt yet then
6275 * set old_rd to NULL to skip the freeing later
6278 if (!atomic_dec_and_test(&old_rd->refcount))
6282 atomic_inc(&rd->refcount);
6285 cpumask_set_cpu(rq->cpu, rd->span);
6286 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6289 raw_spin_unlock_irqrestore(&rq->lock, flags);
6292 free_rootdomain(old_rd);
6295 static int init_rootdomain(struct root_domain *rd)
6297 memset(rd, 0, sizeof(*rd));
6299 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
6301 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
6303 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6306 if (cpupri_init(&rd->cpupri) != 0)
6311 free_cpumask_var(rd->rto_mask);
6313 free_cpumask_var(rd->online);
6315 free_cpumask_var(rd->span);
6320 static void init_defrootdomain(void)
6322 init_rootdomain(&def_root_domain);
6324 atomic_set(&def_root_domain.refcount, 1);
6327 static struct root_domain *alloc_rootdomain(void)
6329 struct root_domain *rd;
6331 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6335 if (init_rootdomain(rd) != 0) {
6344 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6345 * hold the hotplug lock.
6348 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6350 struct rq *rq = cpu_rq(cpu);
6351 struct sched_domain *tmp;
6353 for (tmp = sd; tmp; tmp = tmp->parent)
6354 tmp->span_weight = cpumask_weight(sched_domain_span(tmp));
6356 /* Remove the sched domains which do not contribute to scheduling. */
6357 for (tmp = sd; tmp; ) {
6358 struct sched_domain *parent = tmp->parent;
6362 if (sd_parent_degenerate(tmp, parent)) {
6363 tmp->parent = parent->parent;
6365 parent->parent->child = tmp;
6370 if (sd && sd_degenerate(sd)) {
6376 sched_domain_debug(sd, cpu);
6378 rq_attach_root(rq, rd);
6379 rcu_assign_pointer(rq->sd, sd);
6382 /* cpus with isolated domains */
6383 static cpumask_var_t cpu_isolated_map;
6385 /* Setup the mask of cpus configured for isolated domains */
6386 static int __init isolated_cpu_setup(char *str)
6388 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6389 cpulist_parse(str, cpu_isolated_map);
6393 __setup("isolcpus=", isolated_cpu_setup);
6396 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6397 * to a function which identifies what group(along with sched group) a CPU
6398 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6399 * (due to the fact that we keep track of groups covered with a struct cpumask).
6401 * init_sched_build_groups will build a circular linked list of the groups
6402 * covered by the given span, and will set each group's ->cpumask correctly,
6403 * and ->cpu_power to 0.
6406 init_sched_build_groups(const struct cpumask *span,
6407 const struct cpumask *cpu_map,
6408 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
6409 struct sched_group **sg,
6410 struct cpumask *tmpmask),
6411 struct cpumask *covered, struct cpumask *tmpmask)
6413 struct sched_group *first = NULL, *last = NULL;
6416 cpumask_clear(covered);
6418 for_each_cpu(i, span) {
6419 struct sched_group *sg;
6420 int group = group_fn(i, cpu_map, &sg, tmpmask);
6423 if (cpumask_test_cpu(i, covered))
6426 cpumask_clear(sched_group_cpus(sg));
6429 for_each_cpu(j, span) {
6430 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6433 cpumask_set_cpu(j, covered);
6434 cpumask_set_cpu(j, sched_group_cpus(sg));
6445 #define SD_NODES_PER_DOMAIN 16
6450 * find_next_best_node - find the next node to include in a sched_domain
6451 * @node: node whose sched_domain we're building
6452 * @used_nodes: nodes already in the sched_domain
6454 * Find the next node to include in a given scheduling domain. Simply
6455 * finds the closest node not already in the @used_nodes map.
6457 * Should use nodemask_t.
6459 static int find_next_best_node(int node, nodemask_t *used_nodes)
6461 int i, n, val, min_val, best_node = 0;
6465 for (i = 0; i < nr_node_ids; i++) {
6466 /* Start at @node */
6467 n = (node + i) % nr_node_ids;
6469 if (!nr_cpus_node(n))
6472 /* Skip already used nodes */
6473 if (node_isset(n, *used_nodes))
6476 /* Simple min distance search */
6477 val = node_distance(node, n);
6479 if (val < min_val) {
6485 node_set(best_node, *used_nodes);
6490 * sched_domain_node_span - get a cpumask for a node's sched_domain
6491 * @node: node whose cpumask we're constructing
6492 * @span: resulting cpumask
6494 * Given a node, construct a good cpumask for its sched_domain to span. It
6495 * should be one that prevents unnecessary balancing, but also spreads tasks
6498 static void sched_domain_node_span(int node, struct cpumask *span)
6500 nodemask_t used_nodes;
6503 cpumask_clear(span);
6504 nodes_clear(used_nodes);
6506 cpumask_or(span, span, cpumask_of_node(node));
6507 node_set(node, used_nodes);
6509 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6510 int next_node = find_next_best_node(node, &used_nodes);
6512 cpumask_or(span, span, cpumask_of_node(next_node));
6515 #endif /* CONFIG_NUMA */
6517 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6520 * The cpus mask in sched_group and sched_domain hangs off the end.
6522 * ( See the the comments in include/linux/sched.h:struct sched_group
6523 * and struct sched_domain. )
6525 struct static_sched_group {
6526 struct sched_group sg;
6527 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
6530 struct static_sched_domain {
6531 struct sched_domain sd;
6532 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
6538 cpumask_var_t domainspan;
6539 cpumask_var_t covered;
6540 cpumask_var_t notcovered;
6542 cpumask_var_t nodemask;
6543 cpumask_var_t this_sibling_map;
6544 cpumask_var_t this_core_map;
6545 cpumask_var_t this_book_map;
6546 cpumask_var_t send_covered;
6547 cpumask_var_t tmpmask;
6548 struct sched_group **sched_group_nodes;
6549 struct root_domain *rd;
6553 sa_sched_groups = 0,
6559 sa_this_sibling_map,
6561 sa_sched_group_nodes,
6571 * SMT sched-domains:
6573 #ifdef CONFIG_SCHED_SMT
6574 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
6575 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
6578 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
6579 struct sched_group **sg, struct cpumask *unused)
6582 *sg = &per_cpu(sched_groups, cpu).sg;
6585 #endif /* CONFIG_SCHED_SMT */
6588 * multi-core sched-domains:
6590 #ifdef CONFIG_SCHED_MC
6591 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
6592 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
6595 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6596 struct sched_group **sg, struct cpumask *mask)
6599 #ifdef CONFIG_SCHED_SMT
6600 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6601 group = cpumask_first(mask);
6606 *sg = &per_cpu(sched_group_core, group).sg;
6609 #endif /* CONFIG_SCHED_MC */
6612 * book sched-domains:
6614 #ifdef CONFIG_SCHED_BOOK
6615 static DEFINE_PER_CPU(struct static_sched_domain, book_domains);
6616 static DEFINE_PER_CPU(struct static_sched_group, sched_group_book);
6619 cpu_to_book_group(int cpu, const struct cpumask *cpu_map,
6620 struct sched_group **sg, struct cpumask *mask)
6623 #ifdef CONFIG_SCHED_MC
6624 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6625 group = cpumask_first(mask);
6626 #elif defined(CONFIG_SCHED_SMT)
6627 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6628 group = cpumask_first(mask);
6631 *sg = &per_cpu(sched_group_book, group).sg;
6634 #endif /* CONFIG_SCHED_BOOK */
6636 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
6637 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
6640 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
6641 struct sched_group **sg, struct cpumask *mask)
6644 #ifdef CONFIG_SCHED_BOOK
6645 cpumask_and(mask, cpu_book_mask(cpu), cpu_map);
6646 group = cpumask_first(mask);
6647 #elif defined(CONFIG_SCHED_MC)
6648 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6649 group = cpumask_first(mask);
6650 #elif defined(CONFIG_SCHED_SMT)
6651 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6652 group = cpumask_first(mask);
6657 *sg = &per_cpu(sched_group_phys, group).sg;
6663 * The init_sched_build_groups can't handle what we want to do with node
6664 * groups, so roll our own. Now each node has its own list of groups which
6665 * gets dynamically allocated.
6667 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
6668 static struct sched_group ***sched_group_nodes_bycpu;
6670 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
6671 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
6673 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
6674 struct sched_group **sg,
6675 struct cpumask *nodemask)
6679 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
6680 group = cpumask_first(nodemask);
6683 *sg = &per_cpu(sched_group_allnodes, group).sg;
6687 static void init_numa_sched_groups_power(struct sched_group *group_head)
6689 struct sched_group *sg = group_head;
6695 for_each_cpu(j, sched_group_cpus(sg)) {
6696 struct sched_domain *sd;
6698 sd = &per_cpu(phys_domains, j).sd;
6699 if (j != group_first_cpu(sd->groups)) {
6701 * Only add "power" once for each
6707 sg->cpu_power += sd->groups->cpu_power;
6710 } while (sg != group_head);
6713 static int build_numa_sched_groups(struct s_data *d,
6714 const struct cpumask *cpu_map, int num)
6716 struct sched_domain *sd;
6717 struct sched_group *sg, *prev;
6720 cpumask_clear(d->covered);
6721 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
6722 if (cpumask_empty(d->nodemask)) {
6723 d->sched_group_nodes[num] = NULL;
6727 sched_domain_node_span(num, d->domainspan);
6728 cpumask_and(d->domainspan, d->domainspan, cpu_map);
6730 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6733 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
6737 d->sched_group_nodes[num] = sg;
6739 for_each_cpu(j, d->nodemask) {
6740 sd = &per_cpu(node_domains, j).sd;
6745 cpumask_copy(sched_group_cpus(sg), d->nodemask);
6747 cpumask_or(d->covered, d->covered, d->nodemask);
6750 for (j = 0; j < nr_node_ids; j++) {
6751 n = (num + j) % nr_node_ids;
6752 cpumask_complement(d->notcovered, d->covered);
6753 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
6754 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
6755 if (cpumask_empty(d->tmpmask))
6757 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
6758 if (cpumask_empty(d->tmpmask))
6760 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6764 "Can not alloc domain group for node %d\n", j);
6768 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
6769 sg->next = prev->next;
6770 cpumask_or(d->covered, d->covered, d->tmpmask);
6777 #endif /* CONFIG_NUMA */
6780 /* Free memory allocated for various sched_group structures */
6781 static void free_sched_groups(const struct cpumask *cpu_map,
6782 struct cpumask *nodemask)
6786 for_each_cpu(cpu, cpu_map) {
6787 struct sched_group **sched_group_nodes
6788 = sched_group_nodes_bycpu[cpu];
6790 if (!sched_group_nodes)
6793 for (i = 0; i < nr_node_ids; i++) {
6794 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6796 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
6797 if (cpumask_empty(nodemask))
6807 if (oldsg != sched_group_nodes[i])
6810 kfree(sched_group_nodes);
6811 sched_group_nodes_bycpu[cpu] = NULL;
6814 #else /* !CONFIG_NUMA */
6815 static void free_sched_groups(const struct cpumask *cpu_map,
6816 struct cpumask *nodemask)
6819 #endif /* CONFIG_NUMA */
6822 * Initialize sched groups cpu_power.
6824 * cpu_power indicates the capacity of sched group, which is used while
6825 * distributing the load between different sched groups in a sched domain.
6826 * Typically cpu_power for all the groups in a sched domain will be same unless
6827 * there are asymmetries in the topology. If there are asymmetries, group
6828 * having more cpu_power will pickup more load compared to the group having
6831 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6833 struct sched_domain *child;
6834 struct sched_group *group;
6838 WARN_ON(!sd || !sd->groups);
6840 if (cpu != group_first_cpu(sd->groups))
6845 sd->groups->cpu_power = 0;
6848 power = SCHED_LOAD_SCALE;
6849 weight = cpumask_weight(sched_domain_span(sd));
6851 * SMT siblings share the power of a single core.
6852 * Usually multiple threads get a better yield out of
6853 * that one core than a single thread would have,
6854 * reflect that in sd->smt_gain.
6856 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
6857 power *= sd->smt_gain;
6859 power >>= SCHED_LOAD_SHIFT;
6861 sd->groups->cpu_power += power;
6866 * Add cpu_power of each child group to this groups cpu_power.
6868 group = child->groups;
6870 sd->groups->cpu_power += group->cpu_power;
6871 group = group->next;
6872 } while (group != child->groups);
6876 * Initializers for schedule domains
6877 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6880 #ifdef CONFIG_SCHED_DEBUG
6881 # define SD_INIT_NAME(sd, type) sd->name = #type
6883 # define SD_INIT_NAME(sd, type) do { } while (0)
6886 #define SD_INIT(sd, type) sd_init_##type(sd)
6888 #define SD_INIT_FUNC(type) \
6889 static noinline void sd_init_##type(struct sched_domain *sd) \
6891 memset(sd, 0, sizeof(*sd)); \
6892 *sd = SD_##type##_INIT; \
6893 sd->level = SD_LV_##type; \
6894 SD_INIT_NAME(sd, type); \
6899 SD_INIT_FUNC(ALLNODES)
6902 #ifdef CONFIG_SCHED_SMT
6903 SD_INIT_FUNC(SIBLING)
6905 #ifdef CONFIG_SCHED_MC
6908 #ifdef CONFIG_SCHED_BOOK
6912 static int default_relax_domain_level = -1;
6914 static int __init setup_relax_domain_level(char *str)
6918 val = simple_strtoul(str, NULL, 0);
6919 if (val < SD_LV_MAX)
6920 default_relax_domain_level = val;
6924 __setup("relax_domain_level=", setup_relax_domain_level);
6926 static void set_domain_attribute(struct sched_domain *sd,
6927 struct sched_domain_attr *attr)
6931 if (!attr || attr->relax_domain_level < 0) {
6932 if (default_relax_domain_level < 0)
6935 request = default_relax_domain_level;
6937 request = attr->relax_domain_level;
6938 if (request < sd->level) {
6939 /* turn off idle balance on this domain */
6940 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6942 /* turn on idle balance on this domain */
6943 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6947 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6948 const struct cpumask *cpu_map)
6951 case sa_sched_groups:
6952 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
6953 d->sched_group_nodes = NULL;
6955 free_rootdomain(d->rd); /* fall through */
6957 free_cpumask_var(d->tmpmask); /* fall through */
6958 case sa_send_covered:
6959 free_cpumask_var(d->send_covered); /* fall through */
6960 case sa_this_book_map:
6961 free_cpumask_var(d->this_book_map); /* fall through */
6962 case sa_this_core_map:
6963 free_cpumask_var(d->this_core_map); /* fall through */
6964 case sa_this_sibling_map:
6965 free_cpumask_var(d->this_sibling_map); /* fall through */
6967 free_cpumask_var(d->nodemask); /* fall through */
6968 case sa_sched_group_nodes:
6970 kfree(d->sched_group_nodes); /* fall through */
6972 free_cpumask_var(d->notcovered); /* fall through */
6974 free_cpumask_var(d->covered); /* fall through */
6976 free_cpumask_var(d->domainspan); /* fall through */
6983 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6984 const struct cpumask *cpu_map)
6987 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
6989 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
6990 return sa_domainspan;
6991 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
6993 /* Allocate the per-node list of sched groups */
6994 d->sched_group_nodes = kcalloc(nr_node_ids,
6995 sizeof(struct sched_group *), GFP_KERNEL);
6996 if (!d->sched_group_nodes) {
6997 printk(KERN_WARNING "Can not alloc sched group node list\n");
6998 return sa_notcovered;
7000 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
7002 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
7003 return sa_sched_group_nodes;
7004 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
7006 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
7007 return sa_this_sibling_map;
7008 if (!alloc_cpumask_var(&d->this_book_map, GFP_KERNEL))
7009 return sa_this_core_map;
7010 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
7011 return sa_this_book_map;
7012 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
7013 return sa_send_covered;
7014 d->rd = alloc_rootdomain();
7016 printk(KERN_WARNING "Cannot alloc root domain\n");
7019 return sa_rootdomain;
7022 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
7023 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
7025 struct sched_domain *sd = NULL;
7027 struct sched_domain *parent;
7030 if (cpumask_weight(cpu_map) >
7031 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
7032 sd = &per_cpu(allnodes_domains, i).sd;
7033 SD_INIT(sd, ALLNODES);
7034 set_domain_attribute(sd, attr);
7035 cpumask_copy(sched_domain_span(sd), cpu_map);
7036 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
7041 sd = &per_cpu(node_domains, i).sd;
7043 set_domain_attribute(sd, attr);
7044 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
7045 sd->parent = parent;
7048 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
7053 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
7054 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7055 struct sched_domain *parent, int i)
7057 struct sched_domain *sd;
7058 sd = &per_cpu(phys_domains, i).sd;
7060 set_domain_attribute(sd, attr);
7061 cpumask_copy(sched_domain_span(sd), d->nodemask);
7062 sd->parent = parent;
7065 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
7069 static struct sched_domain *__build_book_sched_domain(struct s_data *d,
7070 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7071 struct sched_domain *parent, int i)
7073 struct sched_domain *sd = parent;
7074 #ifdef CONFIG_SCHED_BOOK
7075 sd = &per_cpu(book_domains, i).sd;
7077 set_domain_attribute(sd, attr);
7078 cpumask_and(sched_domain_span(sd), cpu_map, cpu_book_mask(i));
7079 sd->parent = parent;
7081 cpu_to_book_group(i, cpu_map, &sd->groups, d->tmpmask);
7086 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
7087 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7088 struct sched_domain *parent, int i)
7090 struct sched_domain *sd = parent;
7091 #ifdef CONFIG_SCHED_MC
7092 sd = &per_cpu(core_domains, i).sd;
7094 set_domain_attribute(sd, attr);
7095 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
7096 sd->parent = parent;
7098 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
7103 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
7104 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7105 struct sched_domain *parent, int i)
7107 struct sched_domain *sd = parent;
7108 #ifdef CONFIG_SCHED_SMT
7109 sd = &per_cpu(cpu_domains, i).sd;
7110 SD_INIT(sd, SIBLING);
7111 set_domain_attribute(sd, attr);
7112 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
7113 sd->parent = parent;
7115 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
7120 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
7121 const struct cpumask *cpu_map, int cpu)
7124 #ifdef CONFIG_SCHED_SMT
7125 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
7126 cpumask_and(d->this_sibling_map, cpu_map,
7127 topology_thread_cpumask(cpu));
7128 if (cpu == cpumask_first(d->this_sibling_map))
7129 init_sched_build_groups(d->this_sibling_map, cpu_map,
7131 d->send_covered, d->tmpmask);
7134 #ifdef CONFIG_SCHED_MC
7135 case SD_LV_MC: /* set up multi-core groups */
7136 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
7137 if (cpu == cpumask_first(d->this_core_map))
7138 init_sched_build_groups(d->this_core_map, cpu_map,
7140 d->send_covered, d->tmpmask);
7143 #ifdef CONFIG_SCHED_BOOK
7144 case SD_LV_BOOK: /* set up book groups */
7145 cpumask_and(d->this_book_map, cpu_map, cpu_book_mask(cpu));
7146 if (cpu == cpumask_first(d->this_book_map))
7147 init_sched_build_groups(d->this_book_map, cpu_map,
7149 d->send_covered, d->tmpmask);
7152 case SD_LV_CPU: /* set up physical groups */
7153 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
7154 if (!cpumask_empty(d->nodemask))
7155 init_sched_build_groups(d->nodemask, cpu_map,
7157 d->send_covered, d->tmpmask);
7160 case SD_LV_ALLNODES:
7161 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
7162 d->send_covered, d->tmpmask);
7171 * Build sched domains for a given set of cpus and attach the sched domains
7172 * to the individual cpus
7174 static int __build_sched_domains(const struct cpumask *cpu_map,
7175 struct sched_domain_attr *attr)
7177 enum s_alloc alloc_state = sa_none;
7179 struct sched_domain *sd;
7185 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7186 if (alloc_state != sa_rootdomain)
7188 alloc_state = sa_sched_groups;
7191 * Set up domains for cpus specified by the cpu_map.
7193 for_each_cpu(i, cpu_map) {
7194 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
7197 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
7198 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
7199 sd = __build_book_sched_domain(&d, cpu_map, attr, sd, i);
7200 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
7201 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
7204 for_each_cpu(i, cpu_map) {
7205 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
7206 build_sched_groups(&d, SD_LV_BOOK, cpu_map, i);
7207 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
7210 /* Set up physical groups */
7211 for (i = 0; i < nr_node_ids; i++)
7212 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
7215 /* Set up node groups */
7217 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
7219 for (i = 0; i < nr_node_ids; i++)
7220 if (build_numa_sched_groups(&d, cpu_map, i))
7224 /* Calculate CPU power for physical packages and nodes */
7225 #ifdef CONFIG_SCHED_SMT
7226 for_each_cpu(i, cpu_map) {
7227 sd = &per_cpu(cpu_domains, i).sd;
7228 init_sched_groups_power(i, sd);
7231 #ifdef CONFIG_SCHED_MC
7232 for_each_cpu(i, cpu_map) {
7233 sd = &per_cpu(core_domains, i).sd;
7234 init_sched_groups_power(i, sd);
7237 #ifdef CONFIG_SCHED_BOOK
7238 for_each_cpu(i, cpu_map) {
7239 sd = &per_cpu(book_domains, i).sd;
7240 init_sched_groups_power(i, sd);
7244 for_each_cpu(i, cpu_map) {
7245 sd = &per_cpu(phys_domains, i).sd;
7246 init_sched_groups_power(i, sd);
7250 for (i = 0; i < nr_node_ids; i++)
7251 init_numa_sched_groups_power(d.sched_group_nodes[i]);
7253 if (d.sd_allnodes) {
7254 struct sched_group *sg;
7256 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7258 init_numa_sched_groups_power(sg);
7262 /* Attach the domains */
7263 for_each_cpu(i, cpu_map) {
7264 #ifdef CONFIG_SCHED_SMT
7265 sd = &per_cpu(cpu_domains, i).sd;
7266 #elif defined(CONFIG_SCHED_MC)
7267 sd = &per_cpu(core_domains, i).sd;
7268 #elif defined(CONFIG_SCHED_BOOK)
7269 sd = &per_cpu(book_domains, i).sd;
7271 sd = &per_cpu(phys_domains, i).sd;
7273 cpu_attach_domain(sd, d.rd, i);
7276 d.sched_group_nodes = NULL; /* don't free this we still need it */
7277 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
7281 __free_domain_allocs(&d, alloc_state, cpu_map);
7285 static int build_sched_domains(const struct cpumask *cpu_map)
7287 return __build_sched_domains(cpu_map, NULL);
7290 static cpumask_var_t *doms_cur; /* current sched domains */
7291 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7292 static struct sched_domain_attr *dattr_cur;
7293 /* attribues of custom domains in 'doms_cur' */
7296 * Special case: If a kmalloc of a doms_cur partition (array of
7297 * cpumask) fails, then fallback to a single sched domain,
7298 * as determined by the single cpumask fallback_doms.
7300 static cpumask_var_t fallback_doms;
7303 * arch_update_cpu_topology lets virtualized architectures update the
7304 * cpu core maps. It is supposed to return 1 if the topology changed
7305 * or 0 if it stayed the same.
7307 int __attribute__((weak)) arch_update_cpu_topology(void)
7312 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7315 cpumask_var_t *doms;
7317 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7320 for (i = 0; i < ndoms; i++) {
7321 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7322 free_sched_domains(doms, i);
7329 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7332 for (i = 0; i < ndoms; i++)
7333 free_cpumask_var(doms[i]);
7338 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7339 * For now this just excludes isolated cpus, but could be used to
7340 * exclude other special cases in the future.
7342 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7346 arch_update_cpu_topology();
7348 doms_cur = alloc_sched_domains(ndoms_cur);
7350 doms_cur = &fallback_doms;
7351 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7353 err = build_sched_domains(doms_cur[0]);
7354 register_sched_domain_sysctl();
7359 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7360 struct cpumask *tmpmask)
7362 free_sched_groups(cpu_map, tmpmask);
7366 * Detach sched domains from a group of cpus specified in cpu_map
7367 * These cpus will now be attached to the NULL domain
7369 static void detach_destroy_domains(const struct cpumask *cpu_map)
7371 /* Save because hotplug lock held. */
7372 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7375 for_each_cpu(i, cpu_map)
7376 cpu_attach_domain(NULL, &def_root_domain, i);
7377 synchronize_sched();
7378 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7381 /* handle null as "default" */
7382 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7383 struct sched_domain_attr *new, int idx_new)
7385 struct sched_domain_attr tmp;
7392 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7393 new ? (new + idx_new) : &tmp,
7394 sizeof(struct sched_domain_attr));
7398 * Partition sched domains as specified by the 'ndoms_new'
7399 * cpumasks in the array doms_new[] of cpumasks. This compares
7400 * doms_new[] to the current sched domain partitioning, doms_cur[].
7401 * It destroys each deleted domain and builds each new domain.
7403 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7404 * The masks don't intersect (don't overlap.) We should setup one
7405 * sched domain for each mask. CPUs not in any of the cpumasks will
7406 * not be load balanced. If the same cpumask appears both in the
7407 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7410 * The passed in 'doms_new' should be allocated using
7411 * alloc_sched_domains. This routine takes ownership of it and will
7412 * free_sched_domains it when done with it. If the caller failed the
7413 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7414 * and partition_sched_domains() will fallback to the single partition
7415 * 'fallback_doms', it also forces the domains to be rebuilt.
7417 * If doms_new == NULL it will be replaced with cpu_online_mask.
7418 * ndoms_new == 0 is a special case for destroying existing domains,
7419 * and it will not create the default domain.
7421 * Call with hotplug lock held
7423 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7424 struct sched_domain_attr *dattr_new)
7429 mutex_lock(&sched_domains_mutex);
7431 /* always unregister in case we don't destroy any domains */
7432 unregister_sched_domain_sysctl();
7434 /* Let architecture update cpu core mappings. */
7435 new_topology = arch_update_cpu_topology();
7437 n = doms_new ? ndoms_new : 0;
7439 /* Destroy deleted domains */
7440 for (i = 0; i < ndoms_cur; i++) {
7441 for (j = 0; j < n && !new_topology; j++) {
7442 if (cpumask_equal(doms_cur[i], doms_new[j])
7443 && dattrs_equal(dattr_cur, i, dattr_new, j))
7446 /* no match - a current sched domain not in new doms_new[] */
7447 detach_destroy_domains(doms_cur[i]);
7452 if (doms_new == NULL) {
7454 doms_new = &fallback_doms;
7455 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7456 WARN_ON_ONCE(dattr_new);
7459 /* Build new domains */
7460 for (i = 0; i < ndoms_new; i++) {
7461 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7462 if (cpumask_equal(doms_new[i], doms_cur[j])
7463 && dattrs_equal(dattr_new, i, dattr_cur, j))
7466 /* no match - add a new doms_new */
7467 __build_sched_domains(doms_new[i],
7468 dattr_new ? dattr_new + i : NULL);
7473 /* Remember the new sched domains */
7474 if (doms_cur != &fallback_doms)
7475 free_sched_domains(doms_cur, ndoms_cur);
7476 kfree(dattr_cur); /* kfree(NULL) is safe */
7477 doms_cur = doms_new;
7478 dattr_cur = dattr_new;
7479 ndoms_cur = ndoms_new;
7481 register_sched_domain_sysctl();
7483 mutex_unlock(&sched_domains_mutex);
7486 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7487 static void arch_reinit_sched_domains(void)
7491 /* Destroy domains first to force the rebuild */
7492 partition_sched_domains(0, NULL, NULL);
7494 rebuild_sched_domains();
7498 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7500 unsigned int level = 0;
7502 if (sscanf(buf, "%u", &level) != 1)
7506 * level is always be positive so don't check for
7507 * level < POWERSAVINGS_BALANCE_NONE which is 0
7508 * What happens on 0 or 1 byte write,
7509 * need to check for count as well?
7512 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7516 sched_smt_power_savings = level;
7518 sched_mc_power_savings = level;
7520 arch_reinit_sched_domains();
7525 #ifdef CONFIG_SCHED_MC
7526 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7527 struct sysdev_class_attribute *attr,
7530 return sprintf(page, "%u\n", sched_mc_power_savings);
7532 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7533 struct sysdev_class_attribute *attr,
7534 const char *buf, size_t count)
7536 return sched_power_savings_store(buf, count, 0);
7538 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7539 sched_mc_power_savings_show,
7540 sched_mc_power_savings_store);
7543 #ifdef CONFIG_SCHED_SMT
7544 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7545 struct sysdev_class_attribute *attr,
7548 return sprintf(page, "%u\n", sched_smt_power_savings);
7550 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7551 struct sysdev_class_attribute *attr,
7552 const char *buf, size_t count)
7554 return sched_power_savings_store(buf, count, 1);
7556 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7557 sched_smt_power_savings_show,
7558 sched_smt_power_savings_store);
7561 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7565 #ifdef CONFIG_SCHED_SMT
7567 err = sysfs_create_file(&cls->kset.kobj,
7568 &attr_sched_smt_power_savings.attr);
7570 #ifdef CONFIG_SCHED_MC
7571 if (!err && mc_capable())
7572 err = sysfs_create_file(&cls->kset.kobj,
7573 &attr_sched_mc_power_savings.attr);
7577 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7580 * Update cpusets according to cpu_active mask. If cpusets are
7581 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7582 * around partition_sched_domains().
7584 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7587 switch (action & ~CPU_TASKS_FROZEN) {
7589 case CPU_DOWN_FAILED:
7590 cpuset_update_active_cpus();
7597 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7600 switch (action & ~CPU_TASKS_FROZEN) {
7601 case CPU_DOWN_PREPARE:
7602 cpuset_update_active_cpus();
7609 static int update_runtime(struct notifier_block *nfb,
7610 unsigned long action, void *hcpu)
7612 int cpu = (int)(long)hcpu;
7615 case CPU_DOWN_PREPARE:
7616 case CPU_DOWN_PREPARE_FROZEN:
7617 disable_runtime(cpu_rq(cpu));
7620 case CPU_DOWN_FAILED:
7621 case CPU_DOWN_FAILED_FROZEN:
7623 case CPU_ONLINE_FROZEN:
7624 enable_runtime(cpu_rq(cpu));
7632 void __init sched_init_smp(void)
7634 cpumask_var_t non_isolated_cpus;
7636 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7637 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7639 #if defined(CONFIG_NUMA)
7640 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7642 BUG_ON(sched_group_nodes_bycpu == NULL);
7645 mutex_lock(&sched_domains_mutex);
7646 arch_init_sched_domains(cpu_active_mask);
7647 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7648 if (cpumask_empty(non_isolated_cpus))
7649 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7650 mutex_unlock(&sched_domains_mutex);
7653 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7654 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7656 /* RT runtime code needs to handle some hotplug events */
7657 hotcpu_notifier(update_runtime, 0);
7661 /* Move init over to a non-isolated CPU */
7662 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7664 sched_init_granularity();
7665 free_cpumask_var(non_isolated_cpus);
7667 init_sched_rt_class();
7670 void __init sched_init_smp(void)
7672 sched_init_granularity();
7674 #endif /* CONFIG_SMP */
7676 const_debug unsigned int sysctl_timer_migration = 1;
7678 int in_sched_functions(unsigned long addr)
7680 return in_lock_functions(addr) ||
7681 (addr >= (unsigned long)__sched_text_start
7682 && addr < (unsigned long)__sched_text_end);
7685 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7687 cfs_rq->tasks_timeline = RB_ROOT;
7688 INIT_LIST_HEAD(&cfs_rq->tasks);
7689 #ifdef CONFIG_FAIR_GROUP_SCHED
7692 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7695 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7697 struct rt_prio_array *array;
7700 array = &rt_rq->active;
7701 for (i = 0; i < MAX_RT_PRIO; i++) {
7702 INIT_LIST_HEAD(array->queue + i);
7703 __clear_bit(i, array->bitmap);
7705 /* delimiter for bitsearch: */
7706 __set_bit(MAX_RT_PRIO, array->bitmap);
7708 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7709 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7711 rt_rq->highest_prio.next = MAX_RT_PRIO;
7715 rt_rq->rt_nr_migratory = 0;
7716 rt_rq->overloaded = 0;
7717 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
7721 rt_rq->rt_throttled = 0;
7722 rt_rq->rt_runtime = 0;
7723 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7725 #ifdef CONFIG_RT_GROUP_SCHED
7726 rt_rq->rt_nr_boosted = 0;
7731 #ifdef CONFIG_FAIR_GROUP_SCHED
7732 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7733 struct sched_entity *se, int cpu, int add,
7734 struct sched_entity *parent)
7736 struct rq *rq = cpu_rq(cpu);
7737 tg->cfs_rq[cpu] = cfs_rq;
7738 init_cfs_rq(cfs_rq, rq);
7741 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7744 /* se could be NULL for init_task_group */
7749 se->cfs_rq = &rq->cfs;
7751 se->cfs_rq = parent->my_q;
7754 se->load.weight = tg->shares;
7755 se->load.inv_weight = 0;
7756 se->parent = parent;
7760 #ifdef CONFIG_RT_GROUP_SCHED
7761 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7762 struct sched_rt_entity *rt_se, int cpu, int add,
7763 struct sched_rt_entity *parent)
7765 struct rq *rq = cpu_rq(cpu);
7767 tg->rt_rq[cpu] = rt_rq;
7768 init_rt_rq(rt_rq, rq);
7770 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7772 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7774 tg->rt_se[cpu] = rt_se;
7779 rt_se->rt_rq = &rq->rt;
7781 rt_se->rt_rq = parent->my_q;
7783 rt_se->my_q = rt_rq;
7784 rt_se->parent = parent;
7785 INIT_LIST_HEAD(&rt_se->run_list);
7789 void __init sched_init(void)
7792 unsigned long alloc_size = 0, ptr;
7794 #ifdef CONFIG_FAIR_GROUP_SCHED
7795 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7797 #ifdef CONFIG_RT_GROUP_SCHED
7798 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7800 #ifdef CONFIG_CPUMASK_OFFSTACK
7801 alloc_size += num_possible_cpus() * cpumask_size();
7804 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7806 #ifdef CONFIG_FAIR_GROUP_SCHED
7807 init_task_group.se = (struct sched_entity **)ptr;
7808 ptr += nr_cpu_ids * sizeof(void **);
7810 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7811 ptr += nr_cpu_ids * sizeof(void **);
7813 #endif /* CONFIG_FAIR_GROUP_SCHED */
7814 #ifdef CONFIG_RT_GROUP_SCHED
7815 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7816 ptr += nr_cpu_ids * sizeof(void **);
7818 init_task_group.rt_rq = (struct rt_rq **)ptr;
7819 ptr += nr_cpu_ids * sizeof(void **);
7821 #endif /* CONFIG_RT_GROUP_SCHED */
7822 #ifdef CONFIG_CPUMASK_OFFSTACK
7823 for_each_possible_cpu(i) {
7824 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
7825 ptr += cpumask_size();
7827 #endif /* CONFIG_CPUMASK_OFFSTACK */
7831 init_defrootdomain();
7834 init_rt_bandwidth(&def_rt_bandwidth,
7835 global_rt_period(), global_rt_runtime());
7837 #ifdef CONFIG_RT_GROUP_SCHED
7838 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7839 global_rt_period(), global_rt_runtime());
7840 #endif /* CONFIG_RT_GROUP_SCHED */
7842 #ifdef CONFIG_CGROUP_SCHED
7843 list_add(&init_task_group.list, &task_groups);
7844 INIT_LIST_HEAD(&init_task_group.children);
7846 #endif /* CONFIG_CGROUP_SCHED */
7848 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
7849 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
7850 __alignof__(unsigned long));
7852 for_each_possible_cpu(i) {
7856 raw_spin_lock_init(&rq->lock);
7858 rq->calc_load_active = 0;
7859 rq->calc_load_update = jiffies + LOAD_FREQ;
7860 init_cfs_rq(&rq->cfs, rq);
7861 init_rt_rq(&rq->rt, rq);
7862 #ifdef CONFIG_FAIR_GROUP_SCHED
7863 init_task_group.shares = init_task_group_load;
7864 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7865 #ifdef CONFIG_CGROUP_SCHED
7867 * How much cpu bandwidth does init_task_group get?
7869 * In case of task-groups formed thr' the cgroup filesystem, it
7870 * gets 100% of the cpu resources in the system. This overall
7871 * system cpu resource is divided among the tasks of
7872 * init_task_group and its child task-groups in a fair manner,
7873 * based on each entity's (task or task-group's) weight
7874 * (se->load.weight).
7876 * In other words, if init_task_group has 10 tasks of weight
7877 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7878 * then A0's share of the cpu resource is:
7880 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7882 * We achieve this by letting init_task_group's tasks sit
7883 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7885 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
7887 #endif /* CONFIG_FAIR_GROUP_SCHED */
7889 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7890 #ifdef CONFIG_RT_GROUP_SCHED
7891 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7892 #ifdef CONFIG_CGROUP_SCHED
7893 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
7897 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7898 rq->cpu_load[j] = 0;
7900 rq->last_load_update_tick = jiffies;
7905 rq->cpu_power = SCHED_LOAD_SCALE;
7906 rq->post_schedule = 0;
7907 rq->active_balance = 0;
7908 rq->next_balance = jiffies;
7913 rq->avg_idle = 2*sysctl_sched_migration_cost;
7914 rq_attach_root(rq, &def_root_domain);
7916 rq->nohz_balance_kick = 0;
7917 init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb, i));
7921 atomic_set(&rq->nr_iowait, 0);
7924 set_load_weight(&init_task);
7926 #ifdef CONFIG_PREEMPT_NOTIFIERS
7927 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7931 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7934 #ifdef CONFIG_RT_MUTEXES
7935 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
7939 * The boot idle thread does lazy MMU switching as well:
7941 atomic_inc(&init_mm.mm_count);
7942 enter_lazy_tlb(&init_mm, current);
7945 * Make us the idle thread. Technically, schedule() should not be
7946 * called from this thread, however somewhere below it might be,
7947 * but because we are the idle thread, we just pick up running again
7948 * when this runqueue becomes "idle".
7950 init_idle(current, smp_processor_id());
7952 calc_load_update = jiffies + LOAD_FREQ;
7955 * During early bootup we pretend to be a normal task:
7957 current->sched_class = &fair_sched_class;
7959 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
7960 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
7963 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
7964 alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT);
7965 atomic_set(&nohz.load_balancer, nr_cpu_ids);
7966 atomic_set(&nohz.first_pick_cpu, nr_cpu_ids);
7967 atomic_set(&nohz.second_pick_cpu, nr_cpu_ids);
7969 /* May be allocated at isolcpus cmdline parse time */
7970 if (cpu_isolated_map == NULL)
7971 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7976 scheduler_running = 1;
7979 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7980 static inline int preempt_count_equals(int preempt_offset)
7982 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7984 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
7987 void __might_sleep(const char *file, int line, int preempt_offset)
7990 static unsigned long prev_jiffy; /* ratelimiting */
7992 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7993 system_state != SYSTEM_RUNNING || oops_in_progress)
7995 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7997 prev_jiffy = jiffies;
8000 "BUG: sleeping function called from invalid context at %s:%d\n",
8003 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8004 in_atomic(), irqs_disabled(),
8005 current->pid, current->comm);
8007 debug_show_held_locks(current);
8008 if (irqs_disabled())
8009 print_irqtrace_events(current);
8013 EXPORT_SYMBOL(__might_sleep);
8016 #ifdef CONFIG_MAGIC_SYSRQ
8017 static void normalize_task(struct rq *rq, struct task_struct *p)
8021 on_rq = p->se.on_rq;
8023 deactivate_task(rq, p, 0);
8024 __setscheduler(rq, p, SCHED_NORMAL, 0);
8026 activate_task(rq, p, 0);
8027 resched_task(rq->curr);
8031 void normalize_rt_tasks(void)
8033 struct task_struct *g, *p;
8034 unsigned long flags;
8037 read_lock_irqsave(&tasklist_lock, flags);
8038 do_each_thread(g, p) {
8040 * Only normalize user tasks:
8045 p->se.exec_start = 0;
8046 #ifdef CONFIG_SCHEDSTATS
8047 p->se.statistics.wait_start = 0;
8048 p->se.statistics.sleep_start = 0;
8049 p->se.statistics.block_start = 0;
8054 * Renice negative nice level userspace
8057 if (TASK_NICE(p) < 0 && p->mm)
8058 set_user_nice(p, 0);
8062 raw_spin_lock(&p->pi_lock);
8063 rq = __task_rq_lock(p);
8065 normalize_task(rq, p);
8067 __task_rq_unlock(rq);
8068 raw_spin_unlock(&p->pi_lock);
8069 } while_each_thread(g, p);
8071 read_unlock_irqrestore(&tasklist_lock, flags);
8074 #endif /* CONFIG_MAGIC_SYSRQ */
8076 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8078 * These functions are only useful for the IA64 MCA handling, or kdb.
8080 * They can only be called when the whole system has been
8081 * stopped - every CPU needs to be quiescent, and no scheduling
8082 * activity can take place. Using them for anything else would
8083 * be a serious bug, and as a result, they aren't even visible
8084 * under any other configuration.
8088 * curr_task - return the current task for a given cpu.
8089 * @cpu: the processor in question.
8091 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8093 struct task_struct *curr_task(int cpu)
8095 return cpu_curr(cpu);
8098 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8102 * set_curr_task - set the current task for a given cpu.
8103 * @cpu: the processor in question.
8104 * @p: the task pointer to set.
8106 * Description: This function must only be used when non-maskable interrupts
8107 * are serviced on a separate stack. It allows the architecture to switch the
8108 * notion of the current task on a cpu in a non-blocking manner. This function
8109 * must be called with all CPU's synchronized, and interrupts disabled, the
8110 * and caller must save the original value of the current task (see
8111 * curr_task() above) and restore that value before reenabling interrupts and
8112 * re-starting the system.
8114 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8116 void set_curr_task(int cpu, struct task_struct *p)
8123 #ifdef CONFIG_FAIR_GROUP_SCHED
8124 static void free_fair_sched_group(struct task_group *tg)
8128 for_each_possible_cpu(i) {
8130 kfree(tg->cfs_rq[i]);
8140 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8142 struct cfs_rq *cfs_rq;
8143 struct sched_entity *se;
8147 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8150 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8154 tg->shares = NICE_0_LOAD;
8156 for_each_possible_cpu(i) {
8159 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8160 GFP_KERNEL, cpu_to_node(i));
8164 se = kzalloc_node(sizeof(struct sched_entity),
8165 GFP_KERNEL, cpu_to_node(i));
8169 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
8180 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8182 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8183 &cpu_rq(cpu)->leaf_cfs_rq_list);
8186 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8188 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8190 #else /* !CONFG_FAIR_GROUP_SCHED */
8191 static inline void free_fair_sched_group(struct task_group *tg)
8196 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8201 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8205 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8208 #endif /* CONFIG_FAIR_GROUP_SCHED */
8210 #ifdef CONFIG_RT_GROUP_SCHED
8211 static void free_rt_sched_group(struct task_group *tg)
8215 destroy_rt_bandwidth(&tg->rt_bandwidth);
8217 for_each_possible_cpu(i) {
8219 kfree(tg->rt_rq[i]);
8221 kfree(tg->rt_se[i]);
8229 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8231 struct rt_rq *rt_rq;
8232 struct sched_rt_entity *rt_se;
8236 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8239 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8243 init_rt_bandwidth(&tg->rt_bandwidth,
8244 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8246 for_each_possible_cpu(i) {
8249 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8250 GFP_KERNEL, cpu_to_node(i));
8254 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8255 GFP_KERNEL, cpu_to_node(i));
8259 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
8270 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8272 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8273 &cpu_rq(cpu)->leaf_rt_rq_list);
8276 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8278 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8280 #else /* !CONFIG_RT_GROUP_SCHED */
8281 static inline void free_rt_sched_group(struct task_group *tg)
8286 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8291 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8295 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8298 #endif /* CONFIG_RT_GROUP_SCHED */
8300 #ifdef CONFIG_CGROUP_SCHED
8301 static void free_sched_group(struct task_group *tg)
8303 free_fair_sched_group(tg);
8304 free_rt_sched_group(tg);
8308 /* allocate runqueue etc for a new task group */
8309 struct task_group *sched_create_group(struct task_group *parent)
8311 struct task_group *tg;
8312 unsigned long flags;
8315 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8317 return ERR_PTR(-ENOMEM);
8319 if (!alloc_fair_sched_group(tg, parent))
8322 if (!alloc_rt_sched_group(tg, parent))
8325 spin_lock_irqsave(&task_group_lock, flags);
8326 for_each_possible_cpu(i) {
8327 register_fair_sched_group(tg, i);
8328 register_rt_sched_group(tg, i);
8330 list_add_rcu(&tg->list, &task_groups);
8332 WARN_ON(!parent); /* root should already exist */
8334 tg->parent = parent;
8335 INIT_LIST_HEAD(&tg->children);
8336 list_add_rcu(&tg->siblings, &parent->children);
8337 spin_unlock_irqrestore(&task_group_lock, flags);
8342 free_sched_group(tg);
8343 return ERR_PTR(-ENOMEM);
8346 /* rcu callback to free various structures associated with a task group */
8347 static void free_sched_group_rcu(struct rcu_head *rhp)
8349 /* now it should be safe to free those cfs_rqs */
8350 free_sched_group(container_of(rhp, struct task_group, rcu));
8353 /* Destroy runqueue etc associated with a task group */
8354 void sched_destroy_group(struct task_group *tg)
8356 unsigned long flags;
8359 spin_lock_irqsave(&task_group_lock, flags);
8360 for_each_possible_cpu(i) {
8361 unregister_fair_sched_group(tg, i);
8362 unregister_rt_sched_group(tg, i);
8364 list_del_rcu(&tg->list);
8365 list_del_rcu(&tg->siblings);
8366 spin_unlock_irqrestore(&task_group_lock, flags);
8368 /* wait for possible concurrent references to cfs_rqs complete */
8369 call_rcu(&tg->rcu, free_sched_group_rcu);
8372 /* change task's runqueue when it moves between groups.
8373 * The caller of this function should have put the task in its new group
8374 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8375 * reflect its new group.
8377 void sched_move_task(struct task_struct *tsk)
8380 unsigned long flags;
8383 rq = task_rq_lock(tsk, &flags);
8385 running = task_current(rq, tsk);
8386 on_rq = tsk->se.on_rq;
8389 dequeue_task(rq, tsk, 0);
8390 if (unlikely(running))
8391 tsk->sched_class->put_prev_task(rq, tsk);
8393 set_task_rq(tsk, task_cpu(tsk));
8395 #ifdef CONFIG_FAIR_GROUP_SCHED
8396 if (tsk->sched_class->moved_group)
8397 tsk->sched_class->moved_group(tsk, on_rq);
8400 if (unlikely(running))
8401 tsk->sched_class->set_curr_task(rq);
8403 enqueue_task(rq, tsk, 0);
8405 task_rq_unlock(rq, &flags);
8407 #endif /* CONFIG_CGROUP_SCHED */
8409 #ifdef CONFIG_FAIR_GROUP_SCHED
8410 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8412 struct cfs_rq *cfs_rq = se->cfs_rq;
8417 dequeue_entity(cfs_rq, se, 0);
8419 se->load.weight = shares;
8420 se->load.inv_weight = 0;
8423 enqueue_entity(cfs_rq, se, 0);
8426 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8428 struct cfs_rq *cfs_rq = se->cfs_rq;
8429 struct rq *rq = cfs_rq->rq;
8430 unsigned long flags;
8432 raw_spin_lock_irqsave(&rq->lock, flags);
8433 __set_se_shares(se, shares);
8434 raw_spin_unlock_irqrestore(&rq->lock, flags);
8437 static DEFINE_MUTEX(shares_mutex);
8439 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8442 unsigned long flags;
8445 * We can't change the weight of the root cgroup.
8450 if (shares < MIN_SHARES)
8451 shares = MIN_SHARES;
8452 else if (shares > MAX_SHARES)
8453 shares = MAX_SHARES;
8455 mutex_lock(&shares_mutex);
8456 if (tg->shares == shares)
8459 spin_lock_irqsave(&task_group_lock, flags);
8460 for_each_possible_cpu(i)
8461 unregister_fair_sched_group(tg, i);
8462 list_del_rcu(&tg->siblings);
8463 spin_unlock_irqrestore(&task_group_lock, flags);
8465 /* wait for any ongoing reference to this group to finish */
8466 synchronize_sched();
8469 * Now we are free to modify the group's share on each cpu
8470 * w/o tripping rebalance_share or load_balance_fair.
8472 tg->shares = shares;
8473 for_each_possible_cpu(i) {
8477 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8478 set_se_shares(tg->se[i], shares);
8482 * Enable load balance activity on this group, by inserting it back on
8483 * each cpu's rq->leaf_cfs_rq_list.
8485 spin_lock_irqsave(&task_group_lock, flags);
8486 for_each_possible_cpu(i)
8487 register_fair_sched_group(tg, i);
8488 list_add_rcu(&tg->siblings, &tg->parent->children);
8489 spin_unlock_irqrestore(&task_group_lock, flags);
8491 mutex_unlock(&shares_mutex);
8495 unsigned long sched_group_shares(struct task_group *tg)
8501 #ifdef CONFIG_RT_GROUP_SCHED
8503 * Ensure that the real time constraints are schedulable.
8505 static DEFINE_MUTEX(rt_constraints_mutex);
8507 static unsigned long to_ratio(u64 period, u64 runtime)
8509 if (runtime == RUNTIME_INF)
8512 return div64_u64(runtime << 20, period);
8515 /* Must be called with tasklist_lock held */
8516 static inline int tg_has_rt_tasks(struct task_group *tg)
8518 struct task_struct *g, *p;
8520 do_each_thread(g, p) {
8521 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8523 } while_each_thread(g, p);
8528 struct rt_schedulable_data {
8529 struct task_group *tg;
8534 static int tg_schedulable(struct task_group *tg, void *data)
8536 struct rt_schedulable_data *d = data;
8537 struct task_group *child;
8538 unsigned long total, sum = 0;
8539 u64 period, runtime;
8541 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8542 runtime = tg->rt_bandwidth.rt_runtime;
8545 period = d->rt_period;
8546 runtime = d->rt_runtime;
8550 * Cannot have more runtime than the period.
8552 if (runtime > period && runtime != RUNTIME_INF)
8556 * Ensure we don't starve existing RT tasks.
8558 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8561 total = to_ratio(period, runtime);
8564 * Nobody can have more than the global setting allows.
8566 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8570 * The sum of our children's runtime should not exceed our own.
8572 list_for_each_entry_rcu(child, &tg->children, siblings) {
8573 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8574 runtime = child->rt_bandwidth.rt_runtime;
8576 if (child == d->tg) {
8577 period = d->rt_period;
8578 runtime = d->rt_runtime;
8581 sum += to_ratio(period, runtime);
8590 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8592 struct rt_schedulable_data data = {
8594 .rt_period = period,
8595 .rt_runtime = runtime,
8598 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8601 static int tg_set_bandwidth(struct task_group *tg,
8602 u64 rt_period, u64 rt_runtime)
8606 mutex_lock(&rt_constraints_mutex);
8607 read_lock(&tasklist_lock);
8608 err = __rt_schedulable(tg, rt_period, rt_runtime);
8612 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8613 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8614 tg->rt_bandwidth.rt_runtime = rt_runtime;
8616 for_each_possible_cpu(i) {
8617 struct rt_rq *rt_rq = tg->rt_rq[i];
8619 raw_spin_lock(&rt_rq->rt_runtime_lock);
8620 rt_rq->rt_runtime = rt_runtime;
8621 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8623 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8625 read_unlock(&tasklist_lock);
8626 mutex_unlock(&rt_constraints_mutex);
8631 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8633 u64 rt_runtime, rt_period;
8635 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8636 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8637 if (rt_runtime_us < 0)
8638 rt_runtime = RUNTIME_INF;
8640 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8643 long sched_group_rt_runtime(struct task_group *tg)
8647 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8650 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8651 do_div(rt_runtime_us, NSEC_PER_USEC);
8652 return rt_runtime_us;
8655 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8657 u64 rt_runtime, rt_period;
8659 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8660 rt_runtime = tg->rt_bandwidth.rt_runtime;
8665 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8668 long sched_group_rt_period(struct task_group *tg)
8672 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8673 do_div(rt_period_us, NSEC_PER_USEC);
8674 return rt_period_us;
8677 static int sched_rt_global_constraints(void)
8679 u64 runtime, period;
8682 if (sysctl_sched_rt_period <= 0)
8685 runtime = global_rt_runtime();
8686 period = global_rt_period();
8689 * Sanity check on the sysctl variables.
8691 if (runtime > period && runtime != RUNTIME_INF)
8694 mutex_lock(&rt_constraints_mutex);
8695 read_lock(&tasklist_lock);
8696 ret = __rt_schedulable(NULL, 0, 0);
8697 read_unlock(&tasklist_lock);
8698 mutex_unlock(&rt_constraints_mutex);
8703 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8705 /* Don't accept realtime tasks when there is no way for them to run */
8706 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8712 #else /* !CONFIG_RT_GROUP_SCHED */
8713 static int sched_rt_global_constraints(void)
8715 unsigned long flags;
8718 if (sysctl_sched_rt_period <= 0)
8722 * There's always some RT tasks in the root group
8723 * -- migration, kstopmachine etc..
8725 if (sysctl_sched_rt_runtime == 0)
8728 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8729 for_each_possible_cpu(i) {
8730 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8732 raw_spin_lock(&rt_rq->rt_runtime_lock);
8733 rt_rq->rt_runtime = global_rt_runtime();
8734 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8736 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8740 #endif /* CONFIG_RT_GROUP_SCHED */
8742 int sched_rt_handler(struct ctl_table *table, int write,
8743 void __user *buffer, size_t *lenp,
8747 int old_period, old_runtime;
8748 static DEFINE_MUTEX(mutex);
8751 old_period = sysctl_sched_rt_period;
8752 old_runtime = sysctl_sched_rt_runtime;
8754 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8756 if (!ret && write) {
8757 ret = sched_rt_global_constraints();
8759 sysctl_sched_rt_period = old_period;
8760 sysctl_sched_rt_runtime = old_runtime;
8762 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8763 def_rt_bandwidth.rt_period =
8764 ns_to_ktime(global_rt_period());
8767 mutex_unlock(&mutex);
8772 #ifdef CONFIG_CGROUP_SCHED
8774 /* return corresponding task_group object of a cgroup */
8775 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8777 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8778 struct task_group, css);
8781 static struct cgroup_subsys_state *
8782 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8784 struct task_group *tg, *parent;
8786 if (!cgrp->parent) {
8787 /* This is early initialization for the top cgroup */
8788 return &init_task_group.css;
8791 parent = cgroup_tg(cgrp->parent);
8792 tg = sched_create_group(parent);
8794 return ERR_PTR(-ENOMEM);
8800 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8802 struct task_group *tg = cgroup_tg(cgrp);
8804 sched_destroy_group(tg);
8808 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8810 #ifdef CONFIG_RT_GROUP_SCHED
8811 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
8814 /* We don't support RT-tasks being in separate groups */
8815 if (tsk->sched_class != &fair_sched_class)
8822 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8823 struct task_struct *tsk, bool threadgroup)
8825 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
8829 struct task_struct *c;
8831 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8832 retval = cpu_cgroup_can_attach_task(cgrp, c);
8844 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8845 struct cgroup *old_cont, struct task_struct *tsk,
8848 sched_move_task(tsk);
8850 struct task_struct *c;
8852 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8859 #ifdef CONFIG_FAIR_GROUP_SCHED
8860 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8863 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8866 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8868 struct task_group *tg = cgroup_tg(cgrp);
8870 return (u64) tg->shares;
8872 #endif /* CONFIG_FAIR_GROUP_SCHED */
8874 #ifdef CONFIG_RT_GROUP_SCHED
8875 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8878 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8881 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8883 return sched_group_rt_runtime(cgroup_tg(cgrp));
8886 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8889 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8892 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8894 return sched_group_rt_period(cgroup_tg(cgrp));
8896 #endif /* CONFIG_RT_GROUP_SCHED */
8898 static struct cftype cpu_files[] = {
8899 #ifdef CONFIG_FAIR_GROUP_SCHED
8902 .read_u64 = cpu_shares_read_u64,
8903 .write_u64 = cpu_shares_write_u64,
8906 #ifdef CONFIG_RT_GROUP_SCHED
8908 .name = "rt_runtime_us",
8909 .read_s64 = cpu_rt_runtime_read,
8910 .write_s64 = cpu_rt_runtime_write,
8913 .name = "rt_period_us",
8914 .read_u64 = cpu_rt_period_read_uint,
8915 .write_u64 = cpu_rt_period_write_uint,
8920 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8922 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8925 struct cgroup_subsys cpu_cgroup_subsys = {
8927 .create = cpu_cgroup_create,
8928 .destroy = cpu_cgroup_destroy,
8929 .can_attach = cpu_cgroup_can_attach,
8930 .attach = cpu_cgroup_attach,
8931 .populate = cpu_cgroup_populate,
8932 .subsys_id = cpu_cgroup_subsys_id,
8936 #endif /* CONFIG_CGROUP_SCHED */
8938 #ifdef CONFIG_CGROUP_CPUACCT
8941 * CPU accounting code for task groups.
8943 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8944 * (balbir@in.ibm.com).
8947 /* track cpu usage of a group of tasks and its child groups */
8949 struct cgroup_subsys_state css;
8950 /* cpuusage holds pointer to a u64-type object on every cpu */
8951 u64 __percpu *cpuusage;
8952 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
8953 struct cpuacct *parent;
8956 struct cgroup_subsys cpuacct_subsys;
8958 /* return cpu accounting group corresponding to this container */
8959 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8961 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8962 struct cpuacct, css);
8965 /* return cpu accounting group to which this task belongs */
8966 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8968 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8969 struct cpuacct, css);
8972 /* create a new cpu accounting group */
8973 static struct cgroup_subsys_state *cpuacct_create(
8974 struct cgroup_subsys *ss, struct cgroup *cgrp)
8976 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8982 ca->cpuusage = alloc_percpu(u64);
8986 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8987 if (percpu_counter_init(&ca->cpustat[i], 0))
8988 goto out_free_counters;
8991 ca->parent = cgroup_ca(cgrp->parent);
8997 percpu_counter_destroy(&ca->cpustat[i]);
8998 free_percpu(ca->cpuusage);
9002 return ERR_PTR(-ENOMEM);
9005 /* destroy an existing cpu accounting group */
9007 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9009 struct cpuacct *ca = cgroup_ca(cgrp);
9012 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9013 percpu_counter_destroy(&ca->cpustat[i]);
9014 free_percpu(ca->cpuusage);
9018 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
9020 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9023 #ifndef CONFIG_64BIT
9025 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9027 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9029 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9037 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9039 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9041 #ifndef CONFIG_64BIT
9043 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9045 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9047 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9053 /* return total cpu usage (in nanoseconds) of a group */
9054 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9056 struct cpuacct *ca = cgroup_ca(cgrp);
9057 u64 totalcpuusage = 0;
9060 for_each_present_cpu(i)
9061 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9063 return totalcpuusage;
9066 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9069 struct cpuacct *ca = cgroup_ca(cgrp);
9078 for_each_present_cpu(i)
9079 cpuacct_cpuusage_write(ca, i, 0);
9085 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9088 struct cpuacct *ca = cgroup_ca(cgroup);
9092 for_each_present_cpu(i) {
9093 percpu = cpuacct_cpuusage_read(ca, i);
9094 seq_printf(m, "%llu ", (unsigned long long) percpu);
9096 seq_printf(m, "\n");
9100 static const char *cpuacct_stat_desc[] = {
9101 [CPUACCT_STAT_USER] = "user",
9102 [CPUACCT_STAT_SYSTEM] = "system",
9105 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
9106 struct cgroup_map_cb *cb)
9108 struct cpuacct *ca = cgroup_ca(cgrp);
9111 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
9112 s64 val = percpu_counter_read(&ca->cpustat[i]);
9113 val = cputime64_to_clock_t(val);
9114 cb->fill(cb, cpuacct_stat_desc[i], val);
9119 static struct cftype files[] = {
9122 .read_u64 = cpuusage_read,
9123 .write_u64 = cpuusage_write,
9126 .name = "usage_percpu",
9127 .read_seq_string = cpuacct_percpu_seq_read,
9131 .read_map = cpuacct_stats_show,
9135 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9137 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9141 * charge this task's execution time to its accounting group.
9143 * called with rq->lock held.
9145 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9150 if (unlikely(!cpuacct_subsys.active))
9153 cpu = task_cpu(tsk);
9159 for (; ca; ca = ca->parent) {
9160 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9161 *cpuusage += cputime;
9168 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9169 * in cputime_t units. As a result, cpuacct_update_stats calls
9170 * percpu_counter_add with values large enough to always overflow the
9171 * per cpu batch limit causing bad SMP scalability.
9173 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9174 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9175 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9178 #define CPUACCT_BATCH \
9179 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9181 #define CPUACCT_BATCH 0
9185 * Charge the system/user time to the task's accounting group.
9187 static void cpuacct_update_stats(struct task_struct *tsk,
9188 enum cpuacct_stat_index idx, cputime_t val)
9191 int batch = CPUACCT_BATCH;
9193 if (unlikely(!cpuacct_subsys.active))
9200 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9206 struct cgroup_subsys cpuacct_subsys = {
9208 .create = cpuacct_create,
9209 .destroy = cpuacct_destroy,
9210 .populate = cpuacct_populate,
9211 .subsys_id = cpuacct_subsys_id,
9213 #endif /* CONFIG_CGROUP_CPUACCT */
9217 void synchronize_sched_expedited(void)
9221 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9223 #else /* #ifndef CONFIG_SMP */
9225 static atomic_t synchronize_sched_expedited_count = ATOMIC_INIT(0);
9227 static int synchronize_sched_expedited_cpu_stop(void *data)
9230 * There must be a full memory barrier on each affected CPU
9231 * between the time that try_stop_cpus() is called and the
9232 * time that it returns.
9234 * In the current initial implementation of cpu_stop, the
9235 * above condition is already met when the control reaches
9236 * this point and the following smp_mb() is not strictly
9237 * necessary. Do smp_mb() anyway for documentation and
9238 * robustness against future implementation changes.
9240 smp_mb(); /* See above comment block. */
9245 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
9246 * approach to force grace period to end quickly. This consumes
9247 * significant time on all CPUs, and is thus not recommended for
9248 * any sort of common-case code.
9250 * Note that it is illegal to call this function while holding any
9251 * lock that is acquired by a CPU-hotplug notifier. Failing to
9252 * observe this restriction will result in deadlock.
9254 void synchronize_sched_expedited(void)
9256 int snap, trycount = 0;
9258 smp_mb(); /* ensure prior mod happens before capturing snap. */
9259 snap = atomic_read(&synchronize_sched_expedited_count) + 1;
9261 while (try_stop_cpus(cpu_online_mask,
9262 synchronize_sched_expedited_cpu_stop,
9265 if (trycount++ < 10)
9266 udelay(trycount * num_online_cpus());
9268 synchronize_sched();
9271 if (atomic_read(&synchronize_sched_expedited_count) - snap > 0) {
9272 smp_mb(); /* ensure test happens before caller kfree */
9277 atomic_inc(&synchronize_sched_expedited_count);
9278 smp_mb__after_atomic_inc(); /* ensure post-GP actions seen after GP. */
9281 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9283 #endif /* #else #ifndef CONFIG_SMP */