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/kthread.h>
59 #include <linux/proc_fs.h>
60 #include <linux/seq_file.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>
76 #include <asm/irq_regs.h>
78 #include "sched_cpupri.h"
80 #define CREATE_TRACE_POINTS
81 #include <trace/events/sched.h>
84 * Convert user-nice values [ -20 ... 0 ... 19 ]
85 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
88 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
89 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
90 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
93 * 'User priority' is the nice value converted to something we
94 * can work with better when scaling various scheduler parameters,
95 * it's a [ 0 ... 39 ] range.
97 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
98 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
99 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
102 * Helpers for converting nanosecond timing to jiffy resolution
104 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
106 #define NICE_0_LOAD SCHED_LOAD_SCALE
107 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
110 * These are the 'tuning knobs' of the scheduler:
112 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
113 * Timeslices get refilled after they expire.
115 #define DEF_TIMESLICE (100 * HZ / 1000)
118 * single value that denotes runtime == period, ie unlimited time.
120 #define RUNTIME_INF ((u64)~0ULL)
122 static inline int rt_policy(int policy)
124 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
129 static inline int task_has_rt_policy(struct task_struct *p)
131 return rt_policy(p->policy);
135 * This is the priority-queue data structure of the RT scheduling class:
137 struct rt_prio_array {
138 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
139 struct list_head queue[MAX_RT_PRIO];
142 struct rt_bandwidth {
143 /* nests inside the rq lock: */
144 spinlock_t rt_runtime_lock;
147 struct hrtimer rt_period_timer;
150 static struct rt_bandwidth def_rt_bandwidth;
152 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
154 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
156 struct rt_bandwidth *rt_b =
157 container_of(timer, struct rt_bandwidth, rt_period_timer);
163 now = hrtimer_cb_get_time(timer);
164 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
169 idle = do_sched_rt_period_timer(rt_b, overrun);
172 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
176 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
178 rt_b->rt_period = ns_to_ktime(period);
179 rt_b->rt_runtime = runtime;
181 spin_lock_init(&rt_b->rt_runtime_lock);
183 hrtimer_init(&rt_b->rt_period_timer,
184 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
185 rt_b->rt_period_timer.function = sched_rt_period_timer;
188 static inline int rt_bandwidth_enabled(void)
190 return sysctl_sched_rt_runtime >= 0;
193 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
197 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
200 if (hrtimer_active(&rt_b->rt_period_timer))
203 spin_lock(&rt_b->rt_runtime_lock);
208 if (hrtimer_active(&rt_b->rt_period_timer))
211 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
212 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
214 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
215 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
216 delta = ktime_to_ns(ktime_sub(hard, soft));
217 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
218 HRTIMER_MODE_ABS_PINNED, 0);
220 spin_unlock(&rt_b->rt_runtime_lock);
223 #ifdef CONFIG_RT_GROUP_SCHED
224 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
226 hrtimer_cancel(&rt_b->rt_period_timer);
231 * sched_domains_mutex serializes calls to arch_init_sched_domains,
232 * detach_destroy_domains and partition_sched_domains.
234 static DEFINE_MUTEX(sched_domains_mutex);
236 #ifdef CONFIG_GROUP_SCHED
238 #include <linux/cgroup.h>
242 static LIST_HEAD(task_groups);
244 /* task group related information */
246 #ifdef CONFIG_CGROUP_SCHED
247 struct cgroup_subsys_state css;
250 #ifdef CONFIG_USER_SCHED
254 #ifdef CONFIG_FAIR_GROUP_SCHED
255 /* schedulable entities of this group on each cpu */
256 struct sched_entity **se;
257 /* runqueue "owned" by this group on each cpu */
258 struct cfs_rq **cfs_rq;
259 unsigned long shares;
262 #ifdef CONFIG_RT_GROUP_SCHED
263 struct sched_rt_entity **rt_se;
264 struct rt_rq **rt_rq;
266 struct rt_bandwidth rt_bandwidth;
270 struct list_head list;
272 struct task_group *parent;
273 struct list_head siblings;
274 struct list_head children;
277 #ifdef CONFIG_USER_SCHED
279 /* Helper function to pass uid information to create_sched_user() */
280 void set_tg_uid(struct user_struct *user)
282 user->tg->uid = user->uid;
287 * Every UID task group (including init_task_group aka UID-0) will
288 * be a child to this group.
290 struct task_group root_task_group;
292 #ifdef CONFIG_FAIR_GROUP_SCHED
293 /* Default task group's sched entity on each cpu */
294 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
295 /* Default task group's cfs_rq on each cpu */
296 static DEFINE_PER_CPU_SHARED_ALIGNED(struct cfs_rq, init_tg_cfs_rq);
297 #endif /* CONFIG_FAIR_GROUP_SCHED */
299 #ifdef CONFIG_RT_GROUP_SCHED
300 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
301 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rt_rq, init_rt_rq);
302 #endif /* CONFIG_RT_GROUP_SCHED */
303 #else /* !CONFIG_USER_SCHED */
304 #define root_task_group init_task_group
305 #endif /* CONFIG_USER_SCHED */
307 /* task_group_lock serializes add/remove of task groups and also changes to
308 * a task group's cpu shares.
310 static DEFINE_SPINLOCK(task_group_lock);
312 #ifdef CONFIG_FAIR_GROUP_SCHED
315 static int root_task_group_empty(void)
317 return list_empty(&root_task_group.children);
321 #ifdef CONFIG_USER_SCHED
322 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
323 #else /* !CONFIG_USER_SCHED */
324 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
325 #endif /* CONFIG_USER_SCHED */
328 * A weight of 0 or 1 can cause arithmetics problems.
329 * A weight of a cfs_rq is the sum of weights of which entities
330 * are queued on this cfs_rq, so a weight of a entity should not be
331 * too large, so as the shares value of a task group.
332 * (The default weight is 1024 - so there's no practical
333 * limitation from this.)
336 #define MAX_SHARES (1UL << 18)
338 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
341 /* Default task group.
342 * Every task in system belong to this group at bootup.
344 struct task_group init_task_group;
346 /* return group to which a task belongs */
347 static inline struct task_group *task_group(struct task_struct *p)
349 struct task_group *tg;
351 #ifdef CONFIG_USER_SCHED
353 tg = __task_cred(p)->user->tg;
355 #elif defined(CONFIG_CGROUP_SCHED)
356 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
357 struct task_group, css);
359 tg = &init_task_group;
364 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
365 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
367 #ifdef CONFIG_FAIR_GROUP_SCHED
368 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
369 p->se.parent = task_group(p)->se[cpu];
372 #ifdef CONFIG_RT_GROUP_SCHED
373 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
374 p->rt.parent = task_group(p)->rt_se[cpu];
380 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
381 static inline struct task_group *task_group(struct task_struct *p)
386 #endif /* CONFIG_GROUP_SCHED */
388 /* CFS-related fields in a runqueue */
390 struct load_weight load;
391 unsigned long nr_running;
396 struct rb_root tasks_timeline;
397 struct rb_node *rb_leftmost;
399 struct list_head tasks;
400 struct list_head *balance_iterator;
403 * 'curr' points to currently running entity on this cfs_rq.
404 * It is set to NULL otherwise (i.e when none are currently running).
406 struct sched_entity *curr, *next, *last;
408 unsigned int nr_spread_over;
410 #ifdef CONFIG_FAIR_GROUP_SCHED
411 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
414 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
415 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
416 * (like users, containers etc.)
418 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
419 * list is used during load balance.
421 struct list_head leaf_cfs_rq_list;
422 struct task_group *tg; /* group that "owns" this runqueue */
426 * the part of load.weight contributed by tasks
428 unsigned long task_weight;
431 * h_load = weight * f(tg)
433 * Where f(tg) is the recursive weight fraction assigned to
436 unsigned long h_load;
439 * this cpu's part of tg->shares
441 unsigned long shares;
444 * load.weight at the time we set shares
446 unsigned long rq_weight;
451 /* Real-Time classes' related field in a runqueue: */
453 struct rt_prio_array active;
454 unsigned long rt_nr_running;
455 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
457 int curr; /* highest queued rt task prio */
459 int next; /* next highest */
464 unsigned long rt_nr_migratory;
465 unsigned long rt_nr_total;
467 struct plist_head pushable_tasks;
472 /* Nests inside the rq lock: */
473 spinlock_t rt_runtime_lock;
475 #ifdef CONFIG_RT_GROUP_SCHED
476 unsigned long rt_nr_boosted;
479 struct list_head leaf_rt_rq_list;
480 struct task_group *tg;
481 struct sched_rt_entity *rt_se;
488 * We add the notion of a root-domain which will be used to define per-domain
489 * variables. Each exclusive cpuset essentially defines an island domain by
490 * fully partitioning the member cpus from any other cpuset. Whenever a new
491 * exclusive cpuset is created, we also create and attach a new root-domain
498 cpumask_var_t online;
501 * The "RT overload" flag: it gets set if a CPU has more than
502 * one runnable RT task.
504 cpumask_var_t rto_mask;
507 struct cpupri cpupri;
512 * By default the system creates a single root-domain with all cpus as
513 * members (mimicking the global state we have today).
515 static struct root_domain def_root_domain;
520 * This is the main, per-CPU runqueue data structure.
522 * Locking rule: those places that want to lock multiple runqueues
523 * (such as the load balancing or the thread migration code), lock
524 * acquire operations must be ordered by ascending &runqueue.
531 * nr_running and cpu_load should be in the same cacheline because
532 * remote CPUs use both these fields when doing load calculation.
534 unsigned long nr_running;
535 #define CPU_LOAD_IDX_MAX 5
536 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
538 unsigned long last_tick_seen;
539 unsigned char in_nohz_recently;
541 /* capture load from *all* tasks on this cpu: */
542 struct load_weight load;
543 unsigned long nr_load_updates;
545 u64 nr_migrations_in;
550 #ifdef CONFIG_FAIR_GROUP_SCHED
551 /* list of leaf cfs_rq on this cpu: */
552 struct list_head leaf_cfs_rq_list;
554 #ifdef CONFIG_RT_GROUP_SCHED
555 struct list_head leaf_rt_rq_list;
559 * This is part of a global counter where only the total sum
560 * over all CPUs matters. A task can increase this counter on
561 * one CPU and if it got migrated afterwards it may decrease
562 * it on another CPU. Always updated under the runqueue lock:
564 unsigned long nr_uninterruptible;
566 struct task_struct *curr, *idle;
567 unsigned long next_balance;
568 struct mm_struct *prev_mm;
575 struct root_domain *rd;
576 struct sched_domain *sd;
578 unsigned char idle_at_tick;
579 /* For active balancing */
583 /* cpu of this runqueue: */
587 unsigned long avg_load_per_task;
589 struct task_struct *migration_thread;
590 struct list_head migration_queue;
596 /* calc_load related fields */
597 unsigned long calc_load_update;
598 long calc_load_active;
600 #ifdef CONFIG_SCHED_HRTICK
602 int hrtick_csd_pending;
603 struct call_single_data hrtick_csd;
605 struct hrtimer hrtick_timer;
608 #ifdef CONFIG_SCHEDSTATS
610 struct sched_info rq_sched_info;
611 unsigned long long rq_cpu_time;
612 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
614 /* sys_sched_yield() stats */
615 unsigned int yld_count;
617 /* schedule() stats */
618 unsigned int sched_switch;
619 unsigned int sched_count;
620 unsigned int sched_goidle;
622 /* try_to_wake_up() stats */
623 unsigned int ttwu_count;
624 unsigned int ttwu_local;
627 unsigned int bkl_count;
631 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
634 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
636 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
639 static inline int cpu_of(struct rq *rq)
649 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
650 * See detach_destroy_domains: synchronize_sched for details.
652 * The domain tree of any CPU may only be accessed from within
653 * preempt-disabled sections.
655 #define for_each_domain(cpu, __sd) \
656 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
658 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
659 #define this_rq() (&__get_cpu_var(runqueues))
660 #define task_rq(p) cpu_rq(task_cpu(p))
661 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
662 #define raw_rq() (&__raw_get_cpu_var(runqueues))
664 inline void update_rq_clock(struct rq *rq)
666 rq->clock = sched_clock_cpu(cpu_of(rq));
670 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
672 #ifdef CONFIG_SCHED_DEBUG
673 # define const_debug __read_mostly
675 # define const_debug static const
680 * @cpu: the processor in question.
682 * Returns true if the current cpu runqueue is locked.
683 * This interface allows printk to be called with the runqueue lock
684 * held and know whether or not it is OK to wake up the klogd.
686 int runqueue_is_locked(int cpu)
688 return spin_is_locked(&cpu_rq(cpu)->lock);
692 * Debugging: various feature bits
695 #define SCHED_FEAT(name, enabled) \
696 __SCHED_FEAT_##name ,
699 #include "sched_features.h"
704 #define SCHED_FEAT(name, enabled) \
705 (1UL << __SCHED_FEAT_##name) * enabled |
707 const_debug unsigned int sysctl_sched_features =
708 #include "sched_features.h"
713 #ifdef CONFIG_SCHED_DEBUG
714 #define SCHED_FEAT(name, enabled) \
717 static __read_mostly char *sched_feat_names[] = {
718 #include "sched_features.h"
724 static int sched_feat_show(struct seq_file *m, void *v)
728 for (i = 0; sched_feat_names[i]; i++) {
729 if (!(sysctl_sched_features & (1UL << i)))
731 seq_printf(m, "%s ", sched_feat_names[i]);
739 sched_feat_write(struct file *filp, const char __user *ubuf,
740 size_t cnt, loff_t *ppos)
750 if (copy_from_user(&buf, ubuf, cnt))
755 if (strncmp(buf, "NO_", 3) == 0) {
760 for (i = 0; sched_feat_names[i]; i++) {
761 int len = strlen(sched_feat_names[i]);
763 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
765 sysctl_sched_features &= ~(1UL << i);
767 sysctl_sched_features |= (1UL << i);
772 if (!sched_feat_names[i])
780 static int sched_feat_open(struct inode *inode, struct file *filp)
782 return single_open(filp, sched_feat_show, NULL);
785 static const struct file_operations sched_feat_fops = {
786 .open = sched_feat_open,
787 .write = sched_feat_write,
790 .release = single_release,
793 static __init int sched_init_debug(void)
795 debugfs_create_file("sched_features", 0644, NULL, NULL,
800 late_initcall(sched_init_debug);
804 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
807 * Number of tasks to iterate in a single balance run.
808 * Limited because this is done with IRQs disabled.
810 const_debug unsigned int sysctl_sched_nr_migrate = 32;
813 * ratelimit for updating the group shares.
816 unsigned int sysctl_sched_shares_ratelimit = 250000;
819 * Inject some fuzzyness into changing the per-cpu group shares
820 * this avoids remote rq-locks at the expense of fairness.
823 unsigned int sysctl_sched_shares_thresh = 4;
826 * period over which we average the RT time consumption, measured
831 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
834 * period over which we measure -rt task cpu usage in us.
837 unsigned int sysctl_sched_rt_period = 1000000;
839 static __read_mostly int scheduler_running;
842 * part of the period that we allow rt tasks to run in us.
845 int sysctl_sched_rt_runtime = 950000;
847 static inline u64 global_rt_period(void)
849 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
852 static inline u64 global_rt_runtime(void)
854 if (sysctl_sched_rt_runtime < 0)
857 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
860 #ifndef prepare_arch_switch
861 # define prepare_arch_switch(next) do { } while (0)
863 #ifndef finish_arch_switch
864 # define finish_arch_switch(prev) do { } while (0)
867 static inline int task_current(struct rq *rq, struct task_struct *p)
869 return rq->curr == p;
872 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
873 static inline int task_running(struct rq *rq, struct task_struct *p)
875 return task_current(rq, p);
878 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
882 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
884 #ifdef CONFIG_DEBUG_SPINLOCK
885 /* this is a valid case when another task releases the spinlock */
886 rq->lock.owner = current;
889 * If we are tracking spinlock dependencies then we have to
890 * fix up the runqueue lock - which gets 'carried over' from
893 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
895 spin_unlock_irq(&rq->lock);
898 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
899 static inline int task_running(struct rq *rq, struct task_struct *p)
904 return task_current(rq, p);
908 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
912 * We can optimise this out completely for !SMP, because the
913 * SMP rebalancing from interrupt is the only thing that cares
918 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
919 spin_unlock_irq(&rq->lock);
921 spin_unlock(&rq->lock);
925 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
929 * After ->oncpu is cleared, the task can be moved to a different CPU.
930 * We must ensure this doesn't happen until the switch is completely
936 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
940 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
943 * __task_rq_lock - lock the runqueue a given task resides on.
944 * Must be called interrupts disabled.
946 static inline struct rq *__task_rq_lock(struct task_struct *p)
950 struct rq *rq = task_rq(p);
951 spin_lock(&rq->lock);
952 if (likely(rq == task_rq(p)))
954 spin_unlock(&rq->lock);
959 * task_rq_lock - lock the runqueue a given task resides on and disable
960 * interrupts. Note the ordering: we can safely lookup the task_rq without
961 * explicitly disabling preemption.
963 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
969 local_irq_save(*flags);
971 spin_lock(&rq->lock);
972 if (likely(rq == task_rq(p)))
974 spin_unlock_irqrestore(&rq->lock, *flags);
978 void task_rq_unlock_wait(struct task_struct *p)
980 struct rq *rq = task_rq(p);
982 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
983 spin_unlock_wait(&rq->lock);
986 static void __task_rq_unlock(struct rq *rq)
989 spin_unlock(&rq->lock);
992 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
995 spin_unlock_irqrestore(&rq->lock, *flags);
999 * this_rq_lock - lock this runqueue and disable interrupts.
1001 static struct rq *this_rq_lock(void)
1002 __acquires(rq->lock)
1006 local_irq_disable();
1008 spin_lock(&rq->lock);
1013 #ifdef CONFIG_SCHED_HRTICK
1015 * Use HR-timers to deliver accurate preemption points.
1017 * Its all a bit involved since we cannot program an hrt while holding the
1018 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1021 * When we get rescheduled we reprogram the hrtick_timer outside of the
1027 * - enabled by features
1028 * - hrtimer is actually high res
1030 static inline int hrtick_enabled(struct rq *rq)
1032 if (!sched_feat(HRTICK))
1034 if (!cpu_active(cpu_of(rq)))
1036 return hrtimer_is_hres_active(&rq->hrtick_timer);
1039 static void hrtick_clear(struct rq *rq)
1041 if (hrtimer_active(&rq->hrtick_timer))
1042 hrtimer_cancel(&rq->hrtick_timer);
1046 * High-resolution timer tick.
1047 * Runs from hardirq context with interrupts disabled.
1049 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1051 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1053 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1055 spin_lock(&rq->lock);
1056 update_rq_clock(rq);
1057 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1058 spin_unlock(&rq->lock);
1060 return HRTIMER_NORESTART;
1065 * called from hardirq (IPI) context
1067 static void __hrtick_start(void *arg)
1069 struct rq *rq = arg;
1071 spin_lock(&rq->lock);
1072 hrtimer_restart(&rq->hrtick_timer);
1073 rq->hrtick_csd_pending = 0;
1074 spin_unlock(&rq->lock);
1078 * Called to set the hrtick timer state.
1080 * called with rq->lock held and irqs disabled
1082 static void hrtick_start(struct rq *rq, u64 delay)
1084 struct hrtimer *timer = &rq->hrtick_timer;
1085 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1087 hrtimer_set_expires(timer, time);
1089 if (rq == this_rq()) {
1090 hrtimer_restart(timer);
1091 } else if (!rq->hrtick_csd_pending) {
1092 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1093 rq->hrtick_csd_pending = 1;
1098 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1100 int cpu = (int)(long)hcpu;
1103 case CPU_UP_CANCELED:
1104 case CPU_UP_CANCELED_FROZEN:
1105 case CPU_DOWN_PREPARE:
1106 case CPU_DOWN_PREPARE_FROZEN:
1108 case CPU_DEAD_FROZEN:
1109 hrtick_clear(cpu_rq(cpu));
1116 static __init void init_hrtick(void)
1118 hotcpu_notifier(hotplug_hrtick, 0);
1122 * Called to set the hrtick timer state.
1124 * called with rq->lock held and irqs disabled
1126 static void hrtick_start(struct rq *rq, u64 delay)
1128 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1129 HRTIMER_MODE_REL_PINNED, 0);
1132 static inline void init_hrtick(void)
1135 #endif /* CONFIG_SMP */
1137 static void init_rq_hrtick(struct rq *rq)
1140 rq->hrtick_csd_pending = 0;
1142 rq->hrtick_csd.flags = 0;
1143 rq->hrtick_csd.func = __hrtick_start;
1144 rq->hrtick_csd.info = rq;
1147 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1148 rq->hrtick_timer.function = hrtick;
1150 #else /* CONFIG_SCHED_HRTICK */
1151 static inline void hrtick_clear(struct rq *rq)
1155 static inline void init_rq_hrtick(struct rq *rq)
1159 static inline void init_hrtick(void)
1162 #endif /* CONFIG_SCHED_HRTICK */
1165 * resched_task - mark a task 'to be rescheduled now'.
1167 * On UP this means the setting of the need_resched flag, on SMP it
1168 * might also involve a cross-CPU call to trigger the scheduler on
1173 #ifndef tsk_is_polling
1174 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1177 static void resched_task(struct task_struct *p)
1181 assert_spin_locked(&task_rq(p)->lock);
1183 if (test_tsk_need_resched(p))
1186 set_tsk_need_resched(p);
1189 if (cpu == smp_processor_id())
1192 /* NEED_RESCHED must be visible before we test polling */
1194 if (!tsk_is_polling(p))
1195 smp_send_reschedule(cpu);
1198 static void resched_cpu(int cpu)
1200 struct rq *rq = cpu_rq(cpu);
1201 unsigned long flags;
1203 if (!spin_trylock_irqsave(&rq->lock, flags))
1205 resched_task(cpu_curr(cpu));
1206 spin_unlock_irqrestore(&rq->lock, flags);
1211 * When add_timer_on() enqueues a timer into the timer wheel of an
1212 * idle CPU then this timer might expire before the next timer event
1213 * which is scheduled to wake up that CPU. In case of a completely
1214 * idle system the next event might even be infinite time into the
1215 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1216 * leaves the inner idle loop so the newly added timer is taken into
1217 * account when the CPU goes back to idle and evaluates the timer
1218 * wheel for the next timer event.
1220 void wake_up_idle_cpu(int cpu)
1222 struct rq *rq = cpu_rq(cpu);
1224 if (cpu == smp_processor_id())
1228 * This is safe, as this function is called with the timer
1229 * wheel base lock of (cpu) held. When the CPU is on the way
1230 * to idle and has not yet set rq->curr to idle then it will
1231 * be serialized on the timer wheel base lock and take the new
1232 * timer into account automatically.
1234 if (rq->curr != rq->idle)
1238 * We can set TIF_RESCHED on the idle task of the other CPU
1239 * lockless. The worst case is that the other CPU runs the
1240 * idle task through an additional NOOP schedule()
1242 set_tsk_need_resched(rq->idle);
1244 /* NEED_RESCHED must be visible before we test polling */
1246 if (!tsk_is_polling(rq->idle))
1247 smp_send_reschedule(cpu);
1249 #endif /* CONFIG_NO_HZ */
1251 static u64 sched_avg_period(void)
1253 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1256 static void sched_avg_update(struct rq *rq)
1258 s64 period = sched_avg_period();
1260 while ((s64)(rq->clock - rq->age_stamp) > period) {
1261 rq->age_stamp += period;
1266 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1268 rq->rt_avg += rt_delta;
1269 sched_avg_update(rq);
1272 #else /* !CONFIG_SMP */
1273 static void resched_task(struct task_struct *p)
1275 assert_spin_locked(&task_rq(p)->lock);
1276 set_tsk_need_resched(p);
1279 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1282 #endif /* CONFIG_SMP */
1284 #if BITS_PER_LONG == 32
1285 # define WMULT_CONST (~0UL)
1287 # define WMULT_CONST (1UL << 32)
1290 #define WMULT_SHIFT 32
1293 * Shift right and round:
1295 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1298 * delta *= weight / lw
1300 static unsigned long
1301 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1302 struct load_weight *lw)
1306 if (!lw->inv_weight) {
1307 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1310 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1314 tmp = (u64)delta_exec * weight;
1316 * Check whether we'd overflow the 64-bit multiplication:
1318 if (unlikely(tmp > WMULT_CONST))
1319 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1322 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1324 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1327 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1333 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1340 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1341 * of tasks with abnormal "nice" values across CPUs the contribution that
1342 * each task makes to its run queue's load is weighted according to its
1343 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1344 * scaled version of the new time slice allocation that they receive on time
1348 #define WEIGHT_IDLEPRIO 3
1349 #define WMULT_IDLEPRIO 1431655765
1352 * Nice levels are multiplicative, with a gentle 10% change for every
1353 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1354 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1355 * that remained on nice 0.
1357 * The "10% effect" is relative and cumulative: from _any_ nice level,
1358 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1359 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1360 * If a task goes up by ~10% and another task goes down by ~10% then
1361 * the relative distance between them is ~25%.)
1363 static const int prio_to_weight[40] = {
1364 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1365 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1366 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1367 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1368 /* 0 */ 1024, 820, 655, 526, 423,
1369 /* 5 */ 335, 272, 215, 172, 137,
1370 /* 10 */ 110, 87, 70, 56, 45,
1371 /* 15 */ 36, 29, 23, 18, 15,
1375 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1377 * In cases where the weight does not change often, we can use the
1378 * precalculated inverse to speed up arithmetics by turning divisions
1379 * into multiplications:
1381 static const u32 prio_to_wmult[40] = {
1382 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1383 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1384 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1385 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1386 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1387 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1388 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1389 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1392 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1395 * runqueue iterator, to support SMP load-balancing between different
1396 * scheduling classes, without having to expose their internal data
1397 * structures to the load-balancing proper:
1399 struct rq_iterator {
1401 struct task_struct *(*start)(void *);
1402 struct task_struct *(*next)(void *);
1406 static unsigned long
1407 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1408 unsigned long max_load_move, struct sched_domain *sd,
1409 enum cpu_idle_type idle, int *all_pinned,
1410 int *this_best_prio, struct rq_iterator *iterator);
1413 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1414 struct sched_domain *sd, enum cpu_idle_type idle,
1415 struct rq_iterator *iterator);
1418 /* Time spent by the tasks of the cpu accounting group executing in ... */
1419 enum cpuacct_stat_index {
1420 CPUACCT_STAT_USER, /* ... user mode */
1421 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1423 CPUACCT_STAT_NSTATS,
1426 #ifdef CONFIG_CGROUP_CPUACCT
1427 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1428 static void cpuacct_update_stats(struct task_struct *tsk,
1429 enum cpuacct_stat_index idx, cputime_t val);
1431 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1432 static inline void cpuacct_update_stats(struct task_struct *tsk,
1433 enum cpuacct_stat_index idx, cputime_t val) {}
1436 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1438 update_load_add(&rq->load, load);
1441 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1443 update_load_sub(&rq->load, load);
1446 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1447 typedef int (*tg_visitor)(struct task_group *, void *);
1450 * Iterate the full tree, calling @down when first entering a node and @up when
1451 * leaving it for the final time.
1453 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1455 struct task_group *parent, *child;
1459 parent = &root_task_group;
1461 ret = (*down)(parent, data);
1464 list_for_each_entry_rcu(child, &parent->children, siblings) {
1471 ret = (*up)(parent, data);
1476 parent = parent->parent;
1485 static int tg_nop(struct task_group *tg, void *data)
1492 /* Used instead of source_load when we know the type == 0 */
1493 static unsigned long weighted_cpuload(const int cpu)
1495 return cpu_rq(cpu)->load.weight;
1499 * Return a low guess at the load of a migration-source cpu weighted
1500 * according to the scheduling class and "nice" value.
1502 * We want to under-estimate the load of migration sources, to
1503 * balance conservatively.
1505 static unsigned long source_load(int cpu, int type)
1507 struct rq *rq = cpu_rq(cpu);
1508 unsigned long total = weighted_cpuload(cpu);
1510 if (type == 0 || !sched_feat(LB_BIAS))
1513 return min(rq->cpu_load[type-1], total);
1517 * Return a high guess at the load of a migration-target cpu weighted
1518 * according to the scheduling class and "nice" value.
1520 static unsigned long target_load(int cpu, int type)
1522 struct rq *rq = cpu_rq(cpu);
1523 unsigned long total = weighted_cpuload(cpu);
1525 if (type == 0 || !sched_feat(LB_BIAS))
1528 return max(rq->cpu_load[type-1], total);
1531 static struct sched_group *group_of(int cpu)
1533 struct sched_domain *sd = rcu_dereference(cpu_rq(cpu)->sd);
1541 static unsigned long power_of(int cpu)
1543 struct sched_group *group = group_of(cpu);
1546 return SCHED_LOAD_SCALE;
1548 return group->cpu_power;
1551 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1553 static unsigned long cpu_avg_load_per_task(int cpu)
1555 struct rq *rq = cpu_rq(cpu);
1556 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1559 rq->avg_load_per_task = rq->load.weight / nr_running;
1561 rq->avg_load_per_task = 0;
1563 return rq->avg_load_per_task;
1566 #ifdef CONFIG_FAIR_GROUP_SCHED
1568 static __read_mostly unsigned long *update_shares_data;
1570 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1573 * Calculate and set the cpu's group shares.
1575 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1576 unsigned long sd_shares,
1577 unsigned long sd_rq_weight,
1578 unsigned long *usd_rq_weight)
1580 unsigned long shares, rq_weight;
1583 rq_weight = usd_rq_weight[cpu];
1586 rq_weight = NICE_0_LOAD;
1590 * \Sum_j shares_j * rq_weight_i
1591 * shares_i = -----------------------------
1592 * \Sum_j rq_weight_j
1594 shares = (sd_shares * rq_weight) / sd_rq_weight;
1595 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1597 if (abs(shares - tg->se[cpu]->load.weight) >
1598 sysctl_sched_shares_thresh) {
1599 struct rq *rq = cpu_rq(cpu);
1600 unsigned long flags;
1602 spin_lock_irqsave(&rq->lock, flags);
1603 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1604 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1605 __set_se_shares(tg->se[cpu], shares);
1606 spin_unlock_irqrestore(&rq->lock, flags);
1611 * Re-compute the task group their per cpu shares over the given domain.
1612 * This needs to be done in a bottom-up fashion because the rq weight of a
1613 * parent group depends on the shares of its child groups.
1615 static int tg_shares_up(struct task_group *tg, void *data)
1617 unsigned long weight, rq_weight = 0, shares = 0;
1618 unsigned long *usd_rq_weight;
1619 struct sched_domain *sd = data;
1620 unsigned long flags;
1626 local_irq_save(flags);
1627 usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id());
1629 for_each_cpu(i, sched_domain_span(sd)) {
1630 weight = tg->cfs_rq[i]->load.weight;
1631 usd_rq_weight[i] = weight;
1634 * If there are currently no tasks on the cpu pretend there
1635 * is one of average load so that when a new task gets to
1636 * run here it will not get delayed by group starvation.
1639 weight = NICE_0_LOAD;
1641 rq_weight += weight;
1642 shares += tg->cfs_rq[i]->shares;
1645 if ((!shares && rq_weight) || shares > tg->shares)
1646 shares = tg->shares;
1648 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1649 shares = tg->shares;
1651 for_each_cpu(i, sched_domain_span(sd))
1652 update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight);
1654 local_irq_restore(flags);
1660 * Compute the cpu's hierarchical load factor for each task group.
1661 * This needs to be done in a top-down fashion because the load of a child
1662 * group is a fraction of its parents load.
1664 static int tg_load_down(struct task_group *tg, void *data)
1667 long cpu = (long)data;
1670 load = cpu_rq(cpu)->load.weight;
1672 load = tg->parent->cfs_rq[cpu]->h_load;
1673 load *= tg->cfs_rq[cpu]->shares;
1674 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1677 tg->cfs_rq[cpu]->h_load = load;
1682 static void update_shares(struct sched_domain *sd)
1687 if (root_task_group_empty())
1690 now = cpu_clock(raw_smp_processor_id());
1691 elapsed = now - sd->last_update;
1693 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1694 sd->last_update = now;
1695 walk_tg_tree(tg_nop, tg_shares_up, sd);
1699 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1701 if (root_task_group_empty())
1704 spin_unlock(&rq->lock);
1706 spin_lock(&rq->lock);
1709 static void update_h_load(long cpu)
1711 if (root_task_group_empty())
1714 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1719 static inline void update_shares(struct sched_domain *sd)
1723 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1729 #ifdef CONFIG_PREEMPT
1731 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1734 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1735 * way at the expense of forcing extra atomic operations in all
1736 * invocations. This assures that the double_lock is acquired using the
1737 * same underlying policy as the spinlock_t on this architecture, which
1738 * reduces latency compared to the unfair variant below. However, it
1739 * also adds more overhead and therefore may reduce throughput.
1741 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1742 __releases(this_rq->lock)
1743 __acquires(busiest->lock)
1744 __acquires(this_rq->lock)
1746 spin_unlock(&this_rq->lock);
1747 double_rq_lock(this_rq, busiest);
1754 * Unfair double_lock_balance: Optimizes throughput at the expense of
1755 * latency by eliminating extra atomic operations when the locks are
1756 * already in proper order on entry. This favors lower cpu-ids and will
1757 * grant the double lock to lower cpus over higher ids under contention,
1758 * regardless of entry order into the function.
1760 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1761 __releases(this_rq->lock)
1762 __acquires(busiest->lock)
1763 __acquires(this_rq->lock)
1767 if (unlikely(!spin_trylock(&busiest->lock))) {
1768 if (busiest < this_rq) {
1769 spin_unlock(&this_rq->lock);
1770 spin_lock(&busiest->lock);
1771 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
1774 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
1779 #endif /* CONFIG_PREEMPT */
1782 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1784 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1786 if (unlikely(!irqs_disabled())) {
1787 /* printk() doesn't work good under rq->lock */
1788 spin_unlock(&this_rq->lock);
1792 return _double_lock_balance(this_rq, busiest);
1795 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1796 __releases(busiest->lock)
1798 spin_unlock(&busiest->lock);
1799 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1803 #ifdef CONFIG_FAIR_GROUP_SCHED
1804 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1807 cfs_rq->shares = shares;
1812 static void calc_load_account_active(struct rq *this_rq);
1814 #include "sched_stats.h"
1815 #include "sched_idletask.c"
1816 #include "sched_fair.c"
1817 #include "sched_rt.c"
1818 #ifdef CONFIG_SCHED_DEBUG
1819 # include "sched_debug.c"
1822 #define sched_class_highest (&rt_sched_class)
1823 #define for_each_class(class) \
1824 for (class = sched_class_highest; class; class = class->next)
1826 static void inc_nr_running(struct rq *rq)
1831 static void dec_nr_running(struct rq *rq)
1836 static void set_load_weight(struct task_struct *p)
1838 if (task_has_rt_policy(p)) {
1839 p->se.load.weight = prio_to_weight[0] * 2;
1840 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1845 * SCHED_IDLE tasks get minimal weight:
1847 if (p->policy == SCHED_IDLE) {
1848 p->se.load.weight = WEIGHT_IDLEPRIO;
1849 p->se.load.inv_weight = WMULT_IDLEPRIO;
1853 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1854 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1857 static void update_avg(u64 *avg, u64 sample)
1859 s64 diff = sample - *avg;
1863 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1866 p->se.start_runtime = p->se.sum_exec_runtime;
1868 sched_info_queued(p);
1869 p->sched_class->enqueue_task(rq, p, wakeup);
1873 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1876 if (p->se.last_wakeup) {
1877 update_avg(&p->se.avg_overlap,
1878 p->se.sum_exec_runtime - p->se.last_wakeup);
1879 p->se.last_wakeup = 0;
1881 update_avg(&p->se.avg_wakeup,
1882 sysctl_sched_wakeup_granularity);
1886 sched_info_dequeued(p);
1887 p->sched_class->dequeue_task(rq, p, sleep);
1892 * __normal_prio - return the priority that is based on the static prio
1894 static inline int __normal_prio(struct task_struct *p)
1896 return p->static_prio;
1900 * Calculate the expected normal priority: i.e. priority
1901 * without taking RT-inheritance into account. Might be
1902 * boosted by interactivity modifiers. Changes upon fork,
1903 * setprio syscalls, and whenever the interactivity
1904 * estimator recalculates.
1906 static inline int normal_prio(struct task_struct *p)
1910 if (task_has_rt_policy(p))
1911 prio = MAX_RT_PRIO-1 - p->rt_priority;
1913 prio = __normal_prio(p);
1918 * Calculate the current priority, i.e. the priority
1919 * taken into account by the scheduler. This value might
1920 * be boosted by RT tasks, or might be boosted by
1921 * interactivity modifiers. Will be RT if the task got
1922 * RT-boosted. If not then it returns p->normal_prio.
1924 static int effective_prio(struct task_struct *p)
1926 p->normal_prio = normal_prio(p);
1928 * If we are RT tasks or we were boosted to RT priority,
1929 * keep the priority unchanged. Otherwise, update priority
1930 * to the normal priority:
1932 if (!rt_prio(p->prio))
1933 return p->normal_prio;
1938 * activate_task - move a task to the runqueue.
1940 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1942 if (task_contributes_to_load(p))
1943 rq->nr_uninterruptible--;
1945 enqueue_task(rq, p, wakeup);
1950 * deactivate_task - remove a task from the runqueue.
1952 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1954 if (task_contributes_to_load(p))
1955 rq->nr_uninterruptible++;
1957 dequeue_task(rq, p, sleep);
1962 * task_curr - is this task currently executing on a CPU?
1963 * @p: the task in question.
1965 inline int task_curr(const struct task_struct *p)
1967 return cpu_curr(task_cpu(p)) == p;
1970 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1972 set_task_rq(p, cpu);
1975 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1976 * successfuly executed on another CPU. We must ensure that updates of
1977 * per-task data have been completed by this moment.
1980 task_thread_info(p)->cpu = cpu;
1984 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1985 const struct sched_class *prev_class,
1986 int oldprio, int running)
1988 if (prev_class != p->sched_class) {
1989 if (prev_class->switched_from)
1990 prev_class->switched_from(rq, p, running);
1991 p->sched_class->switched_to(rq, p, running);
1993 p->sched_class->prio_changed(rq, p, oldprio, running);
1997 * kthread_bind - bind a just-created kthread to a cpu.
1998 * @p: thread created by kthread_create().
1999 * @cpu: cpu (might not be online, must be possible) for @k to run on.
2001 * Description: This function is equivalent to set_cpus_allowed(),
2002 * except that @cpu doesn't need to be online, and the thread must be
2003 * stopped (i.e., just returned from kthread_create()).
2005 * Function lives here instead of kthread.c because it messes with
2006 * scheduler internals which require locking.
2008 void kthread_bind(struct task_struct *p, unsigned int cpu)
2010 struct rq *rq = cpu_rq(cpu);
2011 unsigned long flags;
2013 /* Must have done schedule() in kthread() before we set_task_cpu */
2014 if (!wait_task_inactive(p, TASK_UNINTERRUPTIBLE)) {
2019 spin_lock_irqsave(&rq->lock, flags);
2020 set_task_cpu(p, cpu);
2021 p->cpus_allowed = cpumask_of_cpu(cpu);
2022 p->rt.nr_cpus_allowed = 1;
2023 p->flags |= PF_THREAD_BOUND;
2024 spin_unlock_irqrestore(&rq->lock, flags);
2026 EXPORT_SYMBOL(kthread_bind);
2030 * Is this task likely cache-hot:
2033 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2038 * Buddy candidates are cache hot:
2040 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2041 (&p->se == cfs_rq_of(&p->se)->next ||
2042 &p->se == cfs_rq_of(&p->se)->last))
2045 if (p->sched_class != &fair_sched_class)
2048 if (sysctl_sched_migration_cost == -1)
2050 if (sysctl_sched_migration_cost == 0)
2053 delta = now - p->se.exec_start;
2055 return delta < (s64)sysctl_sched_migration_cost;
2059 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2061 int old_cpu = task_cpu(p);
2062 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
2063 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
2064 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
2067 clock_offset = old_rq->clock - new_rq->clock;
2069 trace_sched_migrate_task(p, new_cpu);
2071 #ifdef CONFIG_SCHEDSTATS
2072 if (p->se.wait_start)
2073 p->se.wait_start -= clock_offset;
2074 if (p->se.sleep_start)
2075 p->se.sleep_start -= clock_offset;
2076 if (p->se.block_start)
2077 p->se.block_start -= clock_offset;
2079 if (old_cpu != new_cpu) {
2080 p->se.nr_migrations++;
2081 new_rq->nr_migrations_in++;
2082 #ifdef CONFIG_SCHEDSTATS
2083 if (task_hot(p, old_rq->clock, NULL))
2084 schedstat_inc(p, se.nr_forced2_migrations);
2086 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS,
2089 p->se.vruntime -= old_cfsrq->min_vruntime -
2090 new_cfsrq->min_vruntime;
2092 __set_task_cpu(p, new_cpu);
2095 struct migration_req {
2096 struct list_head list;
2098 struct task_struct *task;
2101 struct completion done;
2105 * The task's runqueue lock must be held.
2106 * Returns true if you have to wait for migration thread.
2109 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2111 struct rq *rq = task_rq(p);
2114 * If the task is not on a runqueue (and not running), then
2115 * it is sufficient to simply update the task's cpu field.
2117 if (!p->se.on_rq && !task_running(rq, p)) {
2118 set_task_cpu(p, dest_cpu);
2122 init_completion(&req->done);
2124 req->dest_cpu = dest_cpu;
2125 list_add(&req->list, &rq->migration_queue);
2131 * wait_task_context_switch - wait for a thread to complete at least one
2134 * @p must not be current.
2136 void wait_task_context_switch(struct task_struct *p)
2138 unsigned long nvcsw, nivcsw, flags;
2146 * The runqueue is assigned before the actual context
2147 * switch. We need to take the runqueue lock.
2149 * We could check initially without the lock but it is
2150 * very likely that we need to take the lock in every
2153 rq = task_rq_lock(p, &flags);
2154 running = task_running(rq, p);
2155 task_rq_unlock(rq, &flags);
2157 if (likely(!running))
2160 * The switch count is incremented before the actual
2161 * context switch. We thus wait for two switches to be
2162 * sure at least one completed.
2164 if ((p->nvcsw - nvcsw) > 1)
2166 if ((p->nivcsw - nivcsw) > 1)
2174 * wait_task_inactive - wait for a thread to unschedule.
2176 * If @match_state is nonzero, it's the @p->state value just checked and
2177 * not expected to change. If it changes, i.e. @p might have woken up,
2178 * then return zero. When we succeed in waiting for @p to be off its CPU,
2179 * we return a positive number (its total switch count). If a second call
2180 * a short while later returns the same number, the caller can be sure that
2181 * @p has remained unscheduled the whole time.
2183 * The caller must ensure that the task *will* unschedule sometime soon,
2184 * else this function might spin for a *long* time. This function can't
2185 * be called with interrupts off, or it may introduce deadlock with
2186 * smp_call_function() if an IPI is sent by the same process we are
2187 * waiting to become inactive.
2189 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2191 unsigned long flags;
2198 * We do the initial early heuristics without holding
2199 * any task-queue locks at all. We'll only try to get
2200 * the runqueue lock when things look like they will
2206 * If the task is actively running on another CPU
2207 * still, just relax and busy-wait without holding
2210 * NOTE! Since we don't hold any locks, it's not
2211 * even sure that "rq" stays as the right runqueue!
2212 * But we don't care, since "task_running()" will
2213 * return false if the runqueue has changed and p
2214 * is actually now running somewhere else!
2216 while (task_running(rq, p)) {
2217 if (match_state && unlikely(p->state != match_state))
2223 * Ok, time to look more closely! We need the rq
2224 * lock now, to be *sure*. If we're wrong, we'll
2225 * just go back and repeat.
2227 rq = task_rq_lock(p, &flags);
2228 trace_sched_wait_task(rq, p);
2229 running = task_running(rq, p);
2230 on_rq = p->se.on_rq;
2232 if (!match_state || p->state == match_state)
2233 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2234 task_rq_unlock(rq, &flags);
2237 * If it changed from the expected state, bail out now.
2239 if (unlikely(!ncsw))
2243 * Was it really running after all now that we
2244 * checked with the proper locks actually held?
2246 * Oops. Go back and try again..
2248 if (unlikely(running)) {
2254 * It's not enough that it's not actively running,
2255 * it must be off the runqueue _entirely_, and not
2258 * So if it was still runnable (but just not actively
2259 * running right now), it's preempted, and we should
2260 * yield - it could be a while.
2262 if (unlikely(on_rq)) {
2263 schedule_timeout_uninterruptible(1);
2268 * Ahh, all good. It wasn't running, and it wasn't
2269 * runnable, which means that it will never become
2270 * running in the future either. We're all done!
2279 * kick_process - kick a running thread to enter/exit the kernel
2280 * @p: the to-be-kicked thread
2282 * Cause a process which is running on another CPU to enter
2283 * kernel-mode, without any delay. (to get signals handled.)
2285 * NOTE: this function doesnt have to take the runqueue lock,
2286 * because all it wants to ensure is that the remote task enters
2287 * the kernel. If the IPI races and the task has been migrated
2288 * to another CPU then no harm is done and the purpose has been
2291 void kick_process(struct task_struct *p)
2297 if ((cpu != smp_processor_id()) && task_curr(p))
2298 smp_send_reschedule(cpu);
2301 EXPORT_SYMBOL_GPL(kick_process);
2302 #endif /* CONFIG_SMP */
2305 * task_oncpu_function_call - call a function on the cpu on which a task runs
2306 * @p: the task to evaluate
2307 * @func: the function to be called
2308 * @info: the function call argument
2310 * Calls the function @func when the task is currently running. This might
2311 * be on the current CPU, which just calls the function directly
2313 void task_oncpu_function_call(struct task_struct *p,
2314 void (*func) (void *info), void *info)
2321 smp_call_function_single(cpu, func, info, 1);
2326 * try_to_wake_up - wake up a thread
2327 * @p: the to-be-woken-up thread
2328 * @state: the mask of task states that can be woken
2329 * @sync: do a synchronous wakeup?
2331 * Put it on the run-queue if it's not already there. The "current"
2332 * thread is always on the run-queue (except when the actual
2333 * re-schedule is in progress), and as such you're allowed to do
2334 * the simpler "current->state = TASK_RUNNING" to mark yourself
2335 * runnable without the overhead of this.
2337 * returns failure only if the task is already active.
2339 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2342 int cpu, orig_cpu, this_cpu, success = 0;
2343 unsigned long flags;
2344 struct rq *rq, *orig_rq;
2346 if (!sched_feat(SYNC_WAKEUPS))
2347 wake_flags &= ~WF_SYNC;
2349 this_cpu = get_cpu();
2352 rq = orig_rq = task_rq_lock(p, &flags);
2353 update_rq_clock(rq);
2354 if (!(p->state & state))
2364 if (unlikely(task_running(rq, p)))
2368 * In order to handle concurrent wakeups and release the rq->lock
2369 * we put the task in TASK_WAKING state.
2371 * First fix up the nr_uninterruptible count:
2373 if (task_contributes_to_load(p))
2374 rq->nr_uninterruptible--;
2375 p->state = TASK_WAKING;
2376 task_rq_unlock(rq, &flags);
2378 cpu = p->sched_class->select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2379 if (cpu != orig_cpu)
2380 set_task_cpu(p, cpu);
2382 rq = task_rq_lock(p, &flags);
2385 update_rq_clock(rq);
2387 WARN_ON(p->state != TASK_WAKING);
2390 #ifdef CONFIG_SCHEDSTATS
2391 schedstat_inc(rq, ttwu_count);
2392 if (cpu == this_cpu)
2393 schedstat_inc(rq, ttwu_local);
2395 struct sched_domain *sd;
2396 for_each_domain(this_cpu, sd) {
2397 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2398 schedstat_inc(sd, ttwu_wake_remote);
2403 #endif /* CONFIG_SCHEDSTATS */
2406 #endif /* CONFIG_SMP */
2407 schedstat_inc(p, se.nr_wakeups);
2408 if (wake_flags & WF_SYNC)
2409 schedstat_inc(p, se.nr_wakeups_sync);
2410 if (orig_cpu != cpu)
2411 schedstat_inc(p, se.nr_wakeups_migrate);
2412 if (cpu == this_cpu)
2413 schedstat_inc(p, se.nr_wakeups_local);
2415 schedstat_inc(p, se.nr_wakeups_remote);
2416 activate_task(rq, p, 1);
2420 * Only attribute actual wakeups done by this task.
2422 if (!in_interrupt()) {
2423 struct sched_entity *se = ¤t->se;
2424 u64 sample = se->sum_exec_runtime;
2426 if (se->last_wakeup)
2427 sample -= se->last_wakeup;
2429 sample -= se->start_runtime;
2430 update_avg(&se->avg_wakeup, sample);
2432 se->last_wakeup = se->sum_exec_runtime;
2436 trace_sched_wakeup(rq, p, success);
2437 check_preempt_curr(rq, p, wake_flags);
2439 p->state = TASK_RUNNING;
2441 if (p->sched_class->task_wake_up)
2442 p->sched_class->task_wake_up(rq, p);
2445 task_rq_unlock(rq, &flags);
2452 * wake_up_process - Wake up a specific process
2453 * @p: The process to be woken up.
2455 * Attempt to wake up the nominated process and move it to the set of runnable
2456 * processes. Returns 1 if the process was woken up, 0 if it was already
2459 * It may be assumed that this function implies a write memory barrier before
2460 * changing the task state if and only if any tasks are woken up.
2462 int wake_up_process(struct task_struct *p)
2464 return try_to_wake_up(p, TASK_ALL, 0);
2466 EXPORT_SYMBOL(wake_up_process);
2468 int wake_up_state(struct task_struct *p, unsigned int state)
2470 return try_to_wake_up(p, state, 0);
2474 * Perform scheduler related setup for a newly forked process p.
2475 * p is forked by current.
2477 * __sched_fork() is basic setup used by init_idle() too:
2479 static void __sched_fork(struct task_struct *p)
2481 p->se.exec_start = 0;
2482 p->se.sum_exec_runtime = 0;
2483 p->se.prev_sum_exec_runtime = 0;
2484 p->se.nr_migrations = 0;
2485 p->se.last_wakeup = 0;
2486 p->se.avg_overlap = 0;
2487 p->se.start_runtime = 0;
2488 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2489 p->se.avg_running = 0;
2491 #ifdef CONFIG_SCHEDSTATS
2492 p->se.wait_start = 0;
2494 p->se.wait_count = 0;
2497 p->se.sleep_start = 0;
2498 p->se.sleep_max = 0;
2499 p->se.sum_sleep_runtime = 0;
2501 p->se.block_start = 0;
2502 p->se.block_max = 0;
2504 p->se.slice_max = 0;
2506 p->se.nr_migrations_cold = 0;
2507 p->se.nr_failed_migrations_affine = 0;
2508 p->se.nr_failed_migrations_running = 0;
2509 p->se.nr_failed_migrations_hot = 0;
2510 p->se.nr_forced_migrations = 0;
2511 p->se.nr_forced2_migrations = 0;
2513 p->se.nr_wakeups = 0;
2514 p->se.nr_wakeups_sync = 0;
2515 p->se.nr_wakeups_migrate = 0;
2516 p->se.nr_wakeups_local = 0;
2517 p->se.nr_wakeups_remote = 0;
2518 p->se.nr_wakeups_affine = 0;
2519 p->se.nr_wakeups_affine_attempts = 0;
2520 p->se.nr_wakeups_passive = 0;
2521 p->se.nr_wakeups_idle = 0;
2525 INIT_LIST_HEAD(&p->rt.run_list);
2527 INIT_LIST_HEAD(&p->se.group_node);
2529 #ifdef CONFIG_PREEMPT_NOTIFIERS
2530 INIT_HLIST_HEAD(&p->preempt_notifiers);
2534 * We mark the process as running here, but have not actually
2535 * inserted it onto the runqueue yet. This guarantees that
2536 * nobody will actually run it, and a signal or other external
2537 * event cannot wake it up and insert it on the runqueue either.
2539 p->state = TASK_RUNNING;
2543 * fork()/clone()-time setup:
2545 void sched_fork(struct task_struct *p, int clone_flags)
2547 int cpu = get_cpu();
2552 * Revert to default priority/policy on fork if requested.
2554 if (unlikely(p->sched_reset_on_fork)) {
2555 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2556 p->policy = SCHED_NORMAL;
2557 p->normal_prio = p->static_prio;
2560 if (PRIO_TO_NICE(p->static_prio) < 0) {
2561 p->static_prio = NICE_TO_PRIO(0);
2562 p->normal_prio = p->static_prio;
2567 * We don't need the reset flag anymore after the fork. It has
2568 * fulfilled its duty:
2570 p->sched_reset_on_fork = 0;
2574 * Make sure we do not leak PI boosting priority to the child.
2576 p->prio = current->normal_prio;
2578 if (!rt_prio(p->prio))
2579 p->sched_class = &fair_sched_class;
2582 cpu = p->sched_class->select_task_rq(p, SD_BALANCE_FORK, 0);
2584 set_task_cpu(p, cpu);
2586 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2587 if (likely(sched_info_on()))
2588 memset(&p->sched_info, 0, sizeof(p->sched_info));
2590 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2593 #ifdef CONFIG_PREEMPT
2594 /* Want to start with kernel preemption disabled. */
2595 task_thread_info(p)->preempt_count = 1;
2597 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2603 * wake_up_new_task - wake up a newly created task for the first time.
2605 * This function will do some initial scheduler statistics housekeeping
2606 * that must be done for every newly created context, then puts the task
2607 * on the runqueue and wakes it.
2609 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2611 unsigned long flags;
2614 rq = task_rq_lock(p, &flags);
2615 BUG_ON(p->state != TASK_RUNNING);
2616 update_rq_clock(rq);
2618 if (!p->sched_class->task_new || !current->se.on_rq) {
2619 activate_task(rq, p, 0);
2622 * Let the scheduling class do new task startup
2623 * management (if any):
2625 p->sched_class->task_new(rq, p);
2628 trace_sched_wakeup_new(rq, p, 1);
2629 check_preempt_curr(rq, p, WF_FORK);
2631 if (p->sched_class->task_wake_up)
2632 p->sched_class->task_wake_up(rq, p);
2634 task_rq_unlock(rq, &flags);
2637 #ifdef CONFIG_PREEMPT_NOTIFIERS
2640 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2641 * @notifier: notifier struct to register
2643 void preempt_notifier_register(struct preempt_notifier *notifier)
2645 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2647 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2650 * preempt_notifier_unregister - no longer interested in preemption notifications
2651 * @notifier: notifier struct to unregister
2653 * This is safe to call from within a preemption notifier.
2655 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2657 hlist_del(¬ifier->link);
2659 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2661 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2663 struct preempt_notifier *notifier;
2664 struct hlist_node *node;
2666 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2667 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2671 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2672 struct task_struct *next)
2674 struct preempt_notifier *notifier;
2675 struct hlist_node *node;
2677 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2678 notifier->ops->sched_out(notifier, next);
2681 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2683 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2688 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2689 struct task_struct *next)
2693 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2696 * prepare_task_switch - prepare to switch tasks
2697 * @rq: the runqueue preparing to switch
2698 * @prev: the current task that is being switched out
2699 * @next: the task we are going to switch to.
2701 * This is called with the rq lock held and interrupts off. It must
2702 * be paired with a subsequent finish_task_switch after the context
2705 * prepare_task_switch sets up locking and calls architecture specific
2709 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2710 struct task_struct *next)
2712 fire_sched_out_preempt_notifiers(prev, next);
2713 prepare_lock_switch(rq, next);
2714 prepare_arch_switch(next);
2718 * finish_task_switch - clean up after a task-switch
2719 * @rq: runqueue associated with task-switch
2720 * @prev: the thread we just switched away from.
2722 * finish_task_switch must be called after the context switch, paired
2723 * with a prepare_task_switch call before the context switch.
2724 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2725 * and do any other architecture-specific cleanup actions.
2727 * Note that we may have delayed dropping an mm in context_switch(). If
2728 * so, we finish that here outside of the runqueue lock. (Doing it
2729 * with the lock held can cause deadlocks; see schedule() for
2732 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2733 __releases(rq->lock)
2735 struct mm_struct *mm = rq->prev_mm;
2741 * A task struct has one reference for the use as "current".
2742 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2743 * schedule one last time. The schedule call will never return, and
2744 * the scheduled task must drop that reference.
2745 * The test for TASK_DEAD must occur while the runqueue locks are
2746 * still held, otherwise prev could be scheduled on another cpu, die
2747 * there before we look at prev->state, and then the reference would
2749 * Manfred Spraul <manfred@colorfullife.com>
2751 prev_state = prev->state;
2752 finish_arch_switch(prev);
2753 perf_event_task_sched_in(current, cpu_of(rq));
2754 finish_lock_switch(rq, prev);
2756 fire_sched_in_preempt_notifiers(current);
2759 if (unlikely(prev_state == TASK_DEAD)) {
2761 * Remove function-return probe instances associated with this
2762 * task and put them back on the free list.
2764 kprobe_flush_task(prev);
2765 put_task_struct(prev);
2771 /* assumes rq->lock is held */
2772 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2774 if (prev->sched_class->pre_schedule)
2775 prev->sched_class->pre_schedule(rq, prev);
2778 /* rq->lock is NOT held, but preemption is disabled */
2779 static inline void post_schedule(struct rq *rq)
2781 if (rq->post_schedule) {
2782 unsigned long flags;
2784 spin_lock_irqsave(&rq->lock, flags);
2785 if (rq->curr->sched_class->post_schedule)
2786 rq->curr->sched_class->post_schedule(rq);
2787 spin_unlock_irqrestore(&rq->lock, flags);
2789 rq->post_schedule = 0;
2795 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2799 static inline void post_schedule(struct rq *rq)
2806 * schedule_tail - first thing a freshly forked thread must call.
2807 * @prev: the thread we just switched away from.
2809 asmlinkage void schedule_tail(struct task_struct *prev)
2810 __releases(rq->lock)
2812 struct rq *rq = this_rq();
2814 finish_task_switch(rq, prev);
2817 * FIXME: do we need to worry about rq being invalidated by the
2822 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2823 /* In this case, finish_task_switch does not reenable preemption */
2826 if (current->set_child_tid)
2827 put_user(task_pid_vnr(current), current->set_child_tid);
2831 * context_switch - switch to the new MM and the new
2832 * thread's register state.
2835 context_switch(struct rq *rq, struct task_struct *prev,
2836 struct task_struct *next)
2838 struct mm_struct *mm, *oldmm;
2840 prepare_task_switch(rq, prev, next);
2841 trace_sched_switch(rq, prev, next);
2843 oldmm = prev->active_mm;
2845 * For paravirt, this is coupled with an exit in switch_to to
2846 * combine the page table reload and the switch backend into
2849 arch_start_context_switch(prev);
2851 if (unlikely(!mm)) {
2852 next->active_mm = oldmm;
2853 atomic_inc(&oldmm->mm_count);
2854 enter_lazy_tlb(oldmm, next);
2856 switch_mm(oldmm, mm, next);
2858 if (unlikely(!prev->mm)) {
2859 prev->active_mm = NULL;
2860 rq->prev_mm = oldmm;
2863 * Since the runqueue lock will be released by the next
2864 * task (which is an invalid locking op but in the case
2865 * of the scheduler it's an obvious special-case), so we
2866 * do an early lockdep release here:
2868 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2869 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2872 /* Here we just switch the register state and the stack. */
2873 switch_to(prev, next, prev);
2877 * this_rq must be evaluated again because prev may have moved
2878 * CPUs since it called schedule(), thus the 'rq' on its stack
2879 * frame will be invalid.
2881 finish_task_switch(this_rq(), prev);
2885 * nr_running, nr_uninterruptible and nr_context_switches:
2887 * externally visible scheduler statistics: current number of runnable
2888 * threads, current number of uninterruptible-sleeping threads, total
2889 * number of context switches performed since bootup.
2891 unsigned long nr_running(void)
2893 unsigned long i, sum = 0;
2895 for_each_online_cpu(i)
2896 sum += cpu_rq(i)->nr_running;
2901 unsigned long nr_uninterruptible(void)
2903 unsigned long i, sum = 0;
2905 for_each_possible_cpu(i)
2906 sum += cpu_rq(i)->nr_uninterruptible;
2909 * Since we read the counters lockless, it might be slightly
2910 * inaccurate. Do not allow it to go below zero though:
2912 if (unlikely((long)sum < 0))
2918 unsigned long long nr_context_switches(void)
2921 unsigned long long sum = 0;
2923 for_each_possible_cpu(i)
2924 sum += cpu_rq(i)->nr_switches;
2929 unsigned long nr_iowait(void)
2931 unsigned long i, sum = 0;
2933 for_each_possible_cpu(i)
2934 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2939 unsigned long nr_iowait_cpu(void)
2941 struct rq *this = this_rq();
2942 return atomic_read(&this->nr_iowait);
2945 unsigned long this_cpu_load(void)
2947 struct rq *this = this_rq();
2948 return this->cpu_load[0];
2952 /* Variables and functions for calc_load */
2953 static atomic_long_t calc_load_tasks;
2954 static unsigned long calc_load_update;
2955 unsigned long avenrun[3];
2956 EXPORT_SYMBOL(avenrun);
2959 * get_avenrun - get the load average array
2960 * @loads: pointer to dest load array
2961 * @offset: offset to add
2962 * @shift: shift count to shift the result left
2964 * These values are estimates at best, so no need for locking.
2966 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2968 loads[0] = (avenrun[0] + offset) << shift;
2969 loads[1] = (avenrun[1] + offset) << shift;
2970 loads[2] = (avenrun[2] + offset) << shift;
2973 static unsigned long
2974 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2977 load += active * (FIXED_1 - exp);
2978 return load >> FSHIFT;
2982 * calc_load - update the avenrun load estimates 10 ticks after the
2983 * CPUs have updated calc_load_tasks.
2985 void calc_global_load(void)
2987 unsigned long upd = calc_load_update + 10;
2990 if (time_before(jiffies, upd))
2993 active = atomic_long_read(&calc_load_tasks);
2994 active = active > 0 ? active * FIXED_1 : 0;
2996 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2997 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2998 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3000 calc_load_update += LOAD_FREQ;
3004 * Either called from update_cpu_load() or from a cpu going idle
3006 static void calc_load_account_active(struct rq *this_rq)
3008 long nr_active, delta;
3010 nr_active = this_rq->nr_running;
3011 nr_active += (long) this_rq->nr_uninterruptible;
3013 if (nr_active != this_rq->calc_load_active) {
3014 delta = nr_active - this_rq->calc_load_active;
3015 this_rq->calc_load_active = nr_active;
3016 atomic_long_add(delta, &calc_load_tasks);
3021 * Externally visible per-cpu scheduler statistics:
3022 * cpu_nr_migrations(cpu) - number of migrations into that cpu
3024 u64 cpu_nr_migrations(int cpu)
3026 return cpu_rq(cpu)->nr_migrations_in;
3030 * Update rq->cpu_load[] statistics. This function is usually called every
3031 * scheduler tick (TICK_NSEC).
3033 static void update_cpu_load(struct rq *this_rq)
3035 unsigned long this_load = this_rq->load.weight;
3038 this_rq->nr_load_updates++;
3040 /* Update our load: */
3041 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3042 unsigned long old_load, new_load;
3044 /* scale is effectively 1 << i now, and >> i divides by scale */
3046 old_load = this_rq->cpu_load[i];
3047 new_load = this_load;
3049 * Round up the averaging division if load is increasing. This
3050 * prevents us from getting stuck on 9 if the load is 10, for
3053 if (new_load > old_load)
3054 new_load += scale-1;
3055 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3058 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3059 this_rq->calc_load_update += LOAD_FREQ;
3060 calc_load_account_active(this_rq);
3067 * double_rq_lock - safely lock two runqueues
3069 * Note this does not disable interrupts like task_rq_lock,
3070 * you need to do so manually before calling.
3072 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
3073 __acquires(rq1->lock)
3074 __acquires(rq2->lock)
3076 BUG_ON(!irqs_disabled());
3078 spin_lock(&rq1->lock);
3079 __acquire(rq2->lock); /* Fake it out ;) */
3082 spin_lock(&rq1->lock);
3083 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
3085 spin_lock(&rq2->lock);
3086 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
3089 update_rq_clock(rq1);
3090 update_rq_clock(rq2);
3094 * double_rq_unlock - safely unlock two runqueues
3096 * Note this does not restore interrupts like task_rq_unlock,
3097 * you need to do so manually after calling.
3099 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
3100 __releases(rq1->lock)
3101 __releases(rq2->lock)
3103 spin_unlock(&rq1->lock);
3105 spin_unlock(&rq2->lock);
3107 __release(rq2->lock);
3111 * If dest_cpu is allowed for this process, migrate the task to it.
3112 * This is accomplished by forcing the cpu_allowed mask to only
3113 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3114 * the cpu_allowed mask is restored.
3116 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
3118 struct migration_req req;
3119 unsigned long flags;
3122 rq = task_rq_lock(p, &flags);
3123 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
3124 || unlikely(!cpu_active(dest_cpu)))
3127 /* force the process onto the specified CPU */
3128 if (migrate_task(p, dest_cpu, &req)) {
3129 /* Need to wait for migration thread (might exit: take ref). */
3130 struct task_struct *mt = rq->migration_thread;
3132 get_task_struct(mt);
3133 task_rq_unlock(rq, &flags);
3134 wake_up_process(mt);
3135 put_task_struct(mt);
3136 wait_for_completion(&req.done);
3141 task_rq_unlock(rq, &flags);
3145 * sched_exec - execve() is a valuable balancing opportunity, because at
3146 * this point the task has the smallest effective memory and cache footprint.
3148 void sched_exec(void)
3150 int new_cpu, this_cpu = get_cpu();
3151 new_cpu = current->sched_class->select_task_rq(current, SD_BALANCE_EXEC, 0);
3153 if (new_cpu != this_cpu)
3154 sched_migrate_task(current, new_cpu);
3158 * pull_task - move a task from a remote runqueue to the local runqueue.
3159 * Both runqueues must be locked.
3161 static void pull_task(struct rq *src_rq, struct task_struct *p,
3162 struct rq *this_rq, int this_cpu)
3164 deactivate_task(src_rq, p, 0);
3165 set_task_cpu(p, this_cpu);
3166 activate_task(this_rq, p, 0);
3168 * Note that idle threads have a prio of MAX_PRIO, for this test
3169 * to be always true for them.
3171 check_preempt_curr(this_rq, p, 0);
3175 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3178 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3179 struct sched_domain *sd, enum cpu_idle_type idle,
3182 int tsk_cache_hot = 0;
3184 * We do not migrate tasks that are:
3185 * 1) running (obviously), or
3186 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3187 * 3) are cache-hot on their current CPU.
3189 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
3190 schedstat_inc(p, se.nr_failed_migrations_affine);
3195 if (task_running(rq, p)) {
3196 schedstat_inc(p, se.nr_failed_migrations_running);
3201 * Aggressive migration if:
3202 * 1) task is cache cold, or
3203 * 2) too many balance attempts have failed.
3206 tsk_cache_hot = task_hot(p, rq->clock, sd);
3207 if (!tsk_cache_hot ||
3208 sd->nr_balance_failed > sd->cache_nice_tries) {
3209 #ifdef CONFIG_SCHEDSTATS
3210 if (tsk_cache_hot) {
3211 schedstat_inc(sd, lb_hot_gained[idle]);
3212 schedstat_inc(p, se.nr_forced_migrations);
3218 if (tsk_cache_hot) {
3219 schedstat_inc(p, se.nr_failed_migrations_hot);
3225 static unsigned long
3226 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3227 unsigned long max_load_move, struct sched_domain *sd,
3228 enum cpu_idle_type idle, int *all_pinned,
3229 int *this_best_prio, struct rq_iterator *iterator)
3231 int loops = 0, pulled = 0, pinned = 0;
3232 struct task_struct *p;
3233 long rem_load_move = max_load_move;
3235 if (max_load_move == 0)
3241 * Start the load-balancing iterator:
3243 p = iterator->start(iterator->arg);
3245 if (!p || loops++ > sysctl_sched_nr_migrate)
3248 if ((p->se.load.weight >> 1) > rem_load_move ||
3249 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3250 p = iterator->next(iterator->arg);
3254 pull_task(busiest, p, this_rq, this_cpu);
3256 rem_load_move -= p->se.load.weight;
3258 #ifdef CONFIG_PREEMPT
3260 * NEWIDLE balancing is a source of latency, so preemptible kernels
3261 * will stop after the first task is pulled to minimize the critical
3264 if (idle == CPU_NEWLY_IDLE)
3269 * We only want to steal up to the prescribed amount of weighted load.
3271 if (rem_load_move > 0) {
3272 if (p->prio < *this_best_prio)
3273 *this_best_prio = p->prio;
3274 p = iterator->next(iterator->arg);
3279 * Right now, this is one of only two places pull_task() is called,
3280 * so we can safely collect pull_task() stats here rather than
3281 * inside pull_task().
3283 schedstat_add(sd, lb_gained[idle], pulled);
3286 *all_pinned = pinned;
3288 return max_load_move - rem_load_move;
3292 * move_tasks tries to move up to max_load_move weighted load from busiest to
3293 * this_rq, as part of a balancing operation within domain "sd".
3294 * Returns 1 if successful and 0 otherwise.
3296 * Called with both runqueues locked.
3298 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3299 unsigned long max_load_move,
3300 struct sched_domain *sd, enum cpu_idle_type idle,
3303 const struct sched_class *class = sched_class_highest;
3304 unsigned long total_load_moved = 0;
3305 int this_best_prio = this_rq->curr->prio;
3309 class->load_balance(this_rq, this_cpu, busiest,
3310 max_load_move - total_load_moved,
3311 sd, idle, all_pinned, &this_best_prio);
3312 class = class->next;
3314 #ifdef CONFIG_PREEMPT
3316 * NEWIDLE balancing is a source of latency, so preemptible
3317 * kernels will stop after the first task is pulled to minimize
3318 * the critical section.
3320 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3323 } while (class && max_load_move > total_load_moved);
3325 return total_load_moved > 0;
3329 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3330 struct sched_domain *sd, enum cpu_idle_type idle,
3331 struct rq_iterator *iterator)
3333 struct task_struct *p = iterator->start(iterator->arg);
3337 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3338 pull_task(busiest, p, this_rq, this_cpu);
3340 * Right now, this is only the second place pull_task()
3341 * is called, so we can safely collect pull_task()
3342 * stats here rather than inside pull_task().
3344 schedstat_inc(sd, lb_gained[idle]);
3348 p = iterator->next(iterator->arg);
3355 * move_one_task tries to move exactly one task from busiest to this_rq, as
3356 * part of active balancing operations within "domain".
3357 * Returns 1 if successful and 0 otherwise.
3359 * Called with both runqueues locked.
3361 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3362 struct sched_domain *sd, enum cpu_idle_type idle)
3364 const struct sched_class *class;
3366 for_each_class(class) {
3367 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3373 /********** Helpers for find_busiest_group ************************/
3375 * sd_lb_stats - Structure to store the statistics of a sched_domain
3376 * during load balancing.
3378 struct sd_lb_stats {
3379 struct sched_group *busiest; /* Busiest group in this sd */
3380 struct sched_group *this; /* Local group in this sd */
3381 unsigned long total_load; /* Total load of all groups in sd */
3382 unsigned long total_pwr; /* Total power of all groups in sd */
3383 unsigned long avg_load; /* Average load across all groups in sd */
3385 /** Statistics of this group */
3386 unsigned long this_load;
3387 unsigned long this_load_per_task;
3388 unsigned long this_nr_running;
3390 /* Statistics of the busiest group */
3391 unsigned long max_load;
3392 unsigned long busiest_load_per_task;
3393 unsigned long busiest_nr_running;
3395 int group_imb; /* Is there imbalance in this sd */
3396 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3397 int power_savings_balance; /* Is powersave balance needed for this sd */
3398 struct sched_group *group_min; /* Least loaded group in sd */
3399 struct sched_group *group_leader; /* Group which relieves group_min */
3400 unsigned long min_load_per_task; /* load_per_task in group_min */
3401 unsigned long leader_nr_running; /* Nr running of group_leader */
3402 unsigned long min_nr_running; /* Nr running of group_min */
3407 * sg_lb_stats - stats of a sched_group required for load_balancing
3409 struct sg_lb_stats {
3410 unsigned long avg_load; /*Avg load across the CPUs of the group */
3411 unsigned long group_load; /* Total load over the CPUs of the group */
3412 unsigned long sum_nr_running; /* Nr tasks running in the group */
3413 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3414 unsigned long group_capacity;
3415 int group_imb; /* Is there an imbalance in the group ? */
3419 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3420 * @group: The group whose first cpu is to be returned.
3422 static inline unsigned int group_first_cpu(struct sched_group *group)
3424 return cpumask_first(sched_group_cpus(group));
3428 * get_sd_load_idx - Obtain the load index for a given sched domain.
3429 * @sd: The sched_domain whose load_idx is to be obtained.
3430 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3432 static inline int get_sd_load_idx(struct sched_domain *sd,
3433 enum cpu_idle_type idle)
3439 load_idx = sd->busy_idx;
3442 case CPU_NEWLY_IDLE:
3443 load_idx = sd->newidle_idx;
3446 load_idx = sd->idle_idx;
3454 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3456 * init_sd_power_savings_stats - Initialize power savings statistics for
3457 * the given sched_domain, during load balancing.
3459 * @sd: Sched domain whose power-savings statistics are to be initialized.
3460 * @sds: Variable containing the statistics for sd.
3461 * @idle: Idle status of the CPU at which we're performing load-balancing.
3463 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3464 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3467 * Busy processors will not participate in power savings
3470 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3471 sds->power_savings_balance = 0;
3473 sds->power_savings_balance = 1;
3474 sds->min_nr_running = ULONG_MAX;
3475 sds->leader_nr_running = 0;
3480 * update_sd_power_savings_stats - Update the power saving stats for a
3481 * sched_domain while performing load balancing.
3483 * @group: sched_group belonging to the sched_domain under consideration.
3484 * @sds: Variable containing the statistics of the sched_domain
3485 * @local_group: Does group contain the CPU for which we're performing
3487 * @sgs: Variable containing the statistics of the group.
3489 static inline void update_sd_power_savings_stats(struct sched_group *group,
3490 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3493 if (!sds->power_savings_balance)
3497 * If the local group is idle or completely loaded
3498 * no need to do power savings balance at this domain
3500 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3501 !sds->this_nr_running))
3502 sds->power_savings_balance = 0;
3505 * If a group is already running at full capacity or idle,
3506 * don't include that group in power savings calculations
3508 if (!sds->power_savings_balance ||
3509 sgs->sum_nr_running >= sgs->group_capacity ||
3510 !sgs->sum_nr_running)
3514 * Calculate the group which has the least non-idle load.
3515 * This is the group from where we need to pick up the load
3518 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3519 (sgs->sum_nr_running == sds->min_nr_running &&
3520 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3521 sds->group_min = group;
3522 sds->min_nr_running = sgs->sum_nr_running;
3523 sds->min_load_per_task = sgs->sum_weighted_load /
3524 sgs->sum_nr_running;
3528 * Calculate the group which is almost near its
3529 * capacity but still has some space to pick up some load
3530 * from other group and save more power
3532 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
3535 if (sgs->sum_nr_running > sds->leader_nr_running ||
3536 (sgs->sum_nr_running == sds->leader_nr_running &&
3537 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3538 sds->group_leader = group;
3539 sds->leader_nr_running = sgs->sum_nr_running;
3544 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3545 * @sds: Variable containing the statistics of the sched_domain
3546 * under consideration.
3547 * @this_cpu: Cpu at which we're currently performing load-balancing.
3548 * @imbalance: Variable to store the imbalance.
3551 * Check if we have potential to perform some power-savings balance.
3552 * If yes, set the busiest group to be the least loaded group in the
3553 * sched_domain, so that it's CPUs can be put to idle.
3555 * Returns 1 if there is potential to perform power-savings balance.
3558 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3559 int this_cpu, unsigned long *imbalance)
3561 if (!sds->power_savings_balance)
3564 if (sds->this != sds->group_leader ||
3565 sds->group_leader == sds->group_min)
3568 *imbalance = sds->min_load_per_task;
3569 sds->busiest = sds->group_min;
3574 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3575 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3576 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3581 static inline void update_sd_power_savings_stats(struct sched_group *group,
3582 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3587 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3588 int this_cpu, unsigned long *imbalance)
3592 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3595 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
3597 return SCHED_LOAD_SCALE;
3600 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
3602 return default_scale_freq_power(sd, cpu);
3605 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
3607 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3608 unsigned long smt_gain = sd->smt_gain;
3615 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3617 return default_scale_smt_power(sd, cpu);
3620 unsigned long scale_rt_power(int cpu)
3622 struct rq *rq = cpu_rq(cpu);
3623 u64 total, available;
3625 sched_avg_update(rq);
3627 total = sched_avg_period() + (rq->clock - rq->age_stamp);
3628 available = total - rq->rt_avg;
3630 if (unlikely((s64)total < SCHED_LOAD_SCALE))
3631 total = SCHED_LOAD_SCALE;
3633 total >>= SCHED_LOAD_SHIFT;
3635 return div_u64(available, total);
3638 static void update_cpu_power(struct sched_domain *sd, int cpu)
3640 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3641 unsigned long power = SCHED_LOAD_SCALE;
3642 struct sched_group *sdg = sd->groups;
3644 if (sched_feat(ARCH_POWER))
3645 power *= arch_scale_freq_power(sd, cpu);
3647 power *= default_scale_freq_power(sd, cpu);
3649 power >>= SCHED_LOAD_SHIFT;
3651 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
3652 if (sched_feat(ARCH_POWER))
3653 power *= arch_scale_smt_power(sd, cpu);
3655 power *= default_scale_smt_power(sd, cpu);
3657 power >>= SCHED_LOAD_SHIFT;
3660 power *= scale_rt_power(cpu);
3661 power >>= SCHED_LOAD_SHIFT;
3666 sdg->cpu_power = power;
3669 static void update_group_power(struct sched_domain *sd, int cpu)
3671 struct sched_domain *child = sd->child;
3672 struct sched_group *group, *sdg = sd->groups;
3673 unsigned long power;
3676 update_cpu_power(sd, cpu);
3682 group = child->groups;
3684 power += group->cpu_power;
3685 group = group->next;
3686 } while (group != child->groups);
3688 sdg->cpu_power = power;
3692 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3693 * @sd: The sched_domain whose statistics are to be updated.
3694 * @group: sched_group whose statistics are to be updated.
3695 * @this_cpu: Cpu for which load balance is currently performed.
3696 * @idle: Idle status of this_cpu
3697 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3698 * @sd_idle: Idle status of the sched_domain containing group.
3699 * @local_group: Does group contain this_cpu.
3700 * @cpus: Set of cpus considered for load balancing.
3701 * @balance: Should we balance.
3702 * @sgs: variable to hold the statistics for this group.
3704 static inline void update_sg_lb_stats(struct sched_domain *sd,
3705 struct sched_group *group, int this_cpu,
3706 enum cpu_idle_type idle, int load_idx, int *sd_idle,
3707 int local_group, const struct cpumask *cpus,
3708 int *balance, struct sg_lb_stats *sgs)
3710 unsigned long load, max_cpu_load, min_cpu_load;
3712 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3713 unsigned long sum_avg_load_per_task;
3714 unsigned long avg_load_per_task;
3717 balance_cpu = group_first_cpu(group);
3718 if (balance_cpu == this_cpu)
3719 update_group_power(sd, this_cpu);
3722 /* Tally up the load of all CPUs in the group */
3723 sum_avg_load_per_task = avg_load_per_task = 0;
3725 min_cpu_load = ~0UL;
3727 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3728 struct rq *rq = cpu_rq(i);
3730 if (*sd_idle && rq->nr_running)
3733 /* Bias balancing toward cpus of our domain */
3735 if (idle_cpu(i) && !first_idle_cpu) {
3740 load = target_load(i, load_idx);
3742 load = source_load(i, load_idx);
3743 if (load > max_cpu_load)
3744 max_cpu_load = load;
3745 if (min_cpu_load > load)
3746 min_cpu_load = load;
3749 sgs->group_load += load;
3750 sgs->sum_nr_running += rq->nr_running;
3751 sgs->sum_weighted_load += weighted_cpuload(i);
3753 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3757 * First idle cpu or the first cpu(busiest) in this sched group
3758 * is eligible for doing load balancing at this and above
3759 * domains. In the newly idle case, we will allow all the cpu's
3760 * to do the newly idle load balance.
3762 if (idle != CPU_NEWLY_IDLE && local_group &&
3763 balance_cpu != this_cpu && balance) {
3768 /* Adjust by relative CPU power of the group */
3769 sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power;
3773 * Consider the group unbalanced when the imbalance is larger
3774 * than the average weight of two tasks.
3776 * APZ: with cgroup the avg task weight can vary wildly and
3777 * might not be a suitable number - should we keep a
3778 * normalized nr_running number somewhere that negates
3781 avg_load_per_task = (sum_avg_load_per_task * SCHED_LOAD_SCALE) /
3784 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3787 sgs->group_capacity =
3788 DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
3792 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3793 * @sd: sched_domain whose statistics are to be updated.
3794 * @this_cpu: Cpu for which load balance is currently performed.
3795 * @idle: Idle status of this_cpu
3796 * @sd_idle: Idle status of the sched_domain containing group.
3797 * @cpus: Set of cpus considered for load balancing.
3798 * @balance: Should we balance.
3799 * @sds: variable to hold the statistics for this sched_domain.
3801 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3802 enum cpu_idle_type idle, int *sd_idle,
3803 const struct cpumask *cpus, int *balance,
3804 struct sd_lb_stats *sds)
3806 struct sched_domain *child = sd->child;
3807 struct sched_group *group = sd->groups;
3808 struct sg_lb_stats sgs;
3809 int load_idx, prefer_sibling = 0;
3811 if (child && child->flags & SD_PREFER_SIBLING)
3814 init_sd_power_savings_stats(sd, sds, idle);
3815 load_idx = get_sd_load_idx(sd, idle);
3820 local_group = cpumask_test_cpu(this_cpu,
3821 sched_group_cpus(group));
3822 memset(&sgs, 0, sizeof(sgs));
3823 update_sg_lb_stats(sd, group, this_cpu, idle, load_idx, sd_idle,
3824 local_group, cpus, balance, &sgs);
3826 if (local_group && balance && !(*balance))
3829 sds->total_load += sgs.group_load;
3830 sds->total_pwr += group->cpu_power;
3833 * In case the child domain prefers tasks go to siblings
3834 * first, lower the group capacity to one so that we'll try
3835 * and move all the excess tasks away.
3838 sgs.group_capacity = min(sgs.group_capacity, 1UL);
3841 sds->this_load = sgs.avg_load;
3843 sds->this_nr_running = sgs.sum_nr_running;
3844 sds->this_load_per_task = sgs.sum_weighted_load;
3845 } else if (sgs.avg_load > sds->max_load &&
3846 (sgs.sum_nr_running > sgs.group_capacity ||
3848 sds->max_load = sgs.avg_load;
3849 sds->busiest = group;
3850 sds->busiest_nr_running = sgs.sum_nr_running;
3851 sds->busiest_load_per_task = sgs.sum_weighted_load;
3852 sds->group_imb = sgs.group_imb;
3855 update_sd_power_savings_stats(group, sds, local_group, &sgs);
3856 group = group->next;
3857 } while (group != sd->groups);
3861 * fix_small_imbalance - Calculate the minor imbalance that exists
3862 * amongst the groups of a sched_domain, during
3864 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3865 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3866 * @imbalance: Variable to store the imbalance.
3868 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3869 int this_cpu, unsigned long *imbalance)
3871 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3872 unsigned int imbn = 2;
3874 if (sds->this_nr_running) {
3875 sds->this_load_per_task /= sds->this_nr_running;
3876 if (sds->busiest_load_per_task >
3877 sds->this_load_per_task)
3880 sds->this_load_per_task =
3881 cpu_avg_load_per_task(this_cpu);
3883 if (sds->max_load - sds->this_load + sds->busiest_load_per_task >=
3884 sds->busiest_load_per_task * imbn) {
3885 *imbalance = sds->busiest_load_per_task;
3890 * OK, we don't have enough imbalance to justify moving tasks,
3891 * however we may be able to increase total CPU power used by
3895 pwr_now += sds->busiest->cpu_power *
3896 min(sds->busiest_load_per_task, sds->max_load);
3897 pwr_now += sds->this->cpu_power *
3898 min(sds->this_load_per_task, sds->this_load);
3899 pwr_now /= SCHED_LOAD_SCALE;
3901 /* Amount of load we'd subtract */
3902 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3903 sds->busiest->cpu_power;
3904 if (sds->max_load > tmp)
3905 pwr_move += sds->busiest->cpu_power *
3906 min(sds->busiest_load_per_task, sds->max_load - tmp);
3908 /* Amount of load we'd add */
3909 if (sds->max_load * sds->busiest->cpu_power <
3910 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
3911 tmp = (sds->max_load * sds->busiest->cpu_power) /
3912 sds->this->cpu_power;
3914 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3915 sds->this->cpu_power;
3916 pwr_move += sds->this->cpu_power *
3917 min(sds->this_load_per_task, sds->this_load + tmp);
3918 pwr_move /= SCHED_LOAD_SCALE;
3920 /* Move if we gain throughput */
3921 if (pwr_move > pwr_now)
3922 *imbalance = sds->busiest_load_per_task;
3926 * calculate_imbalance - Calculate the amount of imbalance present within the
3927 * groups of a given sched_domain during load balance.
3928 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3929 * @this_cpu: Cpu for which currently load balance is being performed.
3930 * @imbalance: The variable to store the imbalance.
3932 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3933 unsigned long *imbalance)
3935 unsigned long max_pull;
3937 * In the presence of smp nice balancing, certain scenarios can have
3938 * max load less than avg load(as we skip the groups at or below
3939 * its cpu_power, while calculating max_load..)
3941 if (sds->max_load < sds->avg_load) {
3943 return fix_small_imbalance(sds, this_cpu, imbalance);
3946 /* Don't want to pull so many tasks that a group would go idle */
3947 max_pull = min(sds->max_load - sds->avg_load,
3948 sds->max_load - sds->busiest_load_per_task);
3950 /* How much load to actually move to equalise the imbalance */
3951 *imbalance = min(max_pull * sds->busiest->cpu_power,
3952 (sds->avg_load - sds->this_load) * sds->this->cpu_power)
3956 * if *imbalance is less than the average load per runnable task
3957 * there is no gaurantee that any tasks will be moved so we'll have
3958 * a think about bumping its value to force at least one task to be
3961 if (*imbalance < sds->busiest_load_per_task)
3962 return fix_small_imbalance(sds, this_cpu, imbalance);
3965 /******* find_busiest_group() helpers end here *********************/
3968 * find_busiest_group - Returns the busiest group within the sched_domain
3969 * if there is an imbalance. If there isn't an imbalance, and
3970 * the user has opted for power-savings, it returns a group whose
3971 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3972 * such a group exists.
3974 * Also calculates the amount of weighted load which should be moved
3975 * to restore balance.
3977 * @sd: The sched_domain whose busiest group is to be returned.
3978 * @this_cpu: The cpu for which load balancing is currently being performed.
3979 * @imbalance: Variable which stores amount of weighted load which should
3980 * be moved to restore balance/put a group to idle.
3981 * @idle: The idle status of this_cpu.
3982 * @sd_idle: The idleness of sd
3983 * @cpus: The set of CPUs under consideration for load-balancing.
3984 * @balance: Pointer to a variable indicating if this_cpu
3985 * is the appropriate cpu to perform load balancing at this_level.
3987 * Returns: - the busiest group if imbalance exists.
3988 * - If no imbalance and user has opted for power-savings balance,
3989 * return the least loaded group whose CPUs can be
3990 * put to idle by rebalancing its tasks onto our group.
3992 static struct sched_group *
3993 find_busiest_group(struct sched_domain *sd, int this_cpu,
3994 unsigned long *imbalance, enum cpu_idle_type idle,
3995 int *sd_idle, const struct cpumask *cpus, int *balance)
3997 struct sd_lb_stats sds;
3999 memset(&sds, 0, sizeof(sds));
4002 * Compute the various statistics relavent for load balancing at
4005 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
4008 /* Cases where imbalance does not exist from POV of this_cpu */
4009 /* 1) this_cpu is not the appropriate cpu to perform load balancing
4011 * 2) There is no busy sibling group to pull from.
4012 * 3) This group is the busiest group.
4013 * 4) This group is more busy than the avg busieness at this
4015 * 5) The imbalance is within the specified limit.
4016 * 6) Any rebalance would lead to ping-pong
4018 if (balance && !(*balance))
4021 if (!sds.busiest || sds.busiest_nr_running == 0)
4024 if (sds.this_load >= sds.max_load)
4027 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
4029 if (sds.this_load >= sds.avg_load)
4032 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
4035 sds.busiest_load_per_task /= sds.busiest_nr_running;
4037 sds.busiest_load_per_task =
4038 min(sds.busiest_load_per_task, sds.avg_load);
4041 * We're trying to get all the cpus to the average_load, so we don't
4042 * want to push ourselves above the average load, nor do we wish to
4043 * reduce the max loaded cpu below the average load, as either of these
4044 * actions would just result in more rebalancing later, and ping-pong
4045 * tasks around. Thus we look for the minimum possible imbalance.
4046 * Negative imbalances (*we* are more loaded than anyone else) will
4047 * be counted as no imbalance for these purposes -- we can't fix that
4048 * by pulling tasks to us. Be careful of negative numbers as they'll
4049 * appear as very large values with unsigned longs.
4051 if (sds.max_load <= sds.busiest_load_per_task)
4054 /* Looks like there is an imbalance. Compute it */
4055 calculate_imbalance(&sds, this_cpu, imbalance);
4060 * There is no obvious imbalance. But check if we can do some balancing
4063 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
4071 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4074 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
4075 unsigned long imbalance, const struct cpumask *cpus)
4077 struct rq *busiest = NULL, *rq;
4078 unsigned long max_load = 0;
4081 for_each_cpu(i, sched_group_cpus(group)) {
4082 unsigned long power = power_of(i);
4083 unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
4086 if (!cpumask_test_cpu(i, cpus))
4090 wl = weighted_cpuload(i) * SCHED_LOAD_SCALE;
4093 if (capacity && rq->nr_running == 1 && wl > imbalance)
4096 if (wl > max_load) {
4106 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4107 * so long as it is large enough.
4109 #define MAX_PINNED_INTERVAL 512
4111 /* Working cpumask for load_balance and load_balance_newidle. */
4112 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4115 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4116 * tasks if there is an imbalance.
4118 static int load_balance(int this_cpu, struct rq *this_rq,
4119 struct sched_domain *sd, enum cpu_idle_type idle,
4122 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
4123 struct sched_group *group;
4124 unsigned long imbalance;
4126 unsigned long flags;
4127 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4129 cpumask_setall(cpus);
4132 * When power savings policy is enabled for the parent domain, idle
4133 * sibling can pick up load irrespective of busy siblings. In this case,
4134 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4135 * portraying it as CPU_NOT_IDLE.
4137 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
4138 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4141 schedstat_inc(sd, lb_count[idle]);
4145 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
4152 schedstat_inc(sd, lb_nobusyg[idle]);
4156 busiest = find_busiest_queue(group, idle, imbalance, cpus);
4158 schedstat_inc(sd, lb_nobusyq[idle]);
4162 BUG_ON(busiest == this_rq);
4164 schedstat_add(sd, lb_imbalance[idle], imbalance);
4167 if (busiest->nr_running > 1) {
4169 * Attempt to move tasks. If find_busiest_group has found
4170 * an imbalance but busiest->nr_running <= 1, the group is
4171 * still unbalanced. ld_moved simply stays zero, so it is
4172 * correctly treated as an imbalance.
4174 local_irq_save(flags);
4175 double_rq_lock(this_rq, busiest);
4176 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4177 imbalance, sd, idle, &all_pinned);
4178 double_rq_unlock(this_rq, busiest);
4179 local_irq_restore(flags);
4182 * some other cpu did the load balance for us.
4184 if (ld_moved && this_cpu != smp_processor_id())
4185 resched_cpu(this_cpu);
4187 /* All tasks on this runqueue were pinned by CPU affinity */
4188 if (unlikely(all_pinned)) {
4189 cpumask_clear_cpu(cpu_of(busiest), cpus);
4190 if (!cpumask_empty(cpus))
4197 schedstat_inc(sd, lb_failed[idle]);
4198 sd->nr_balance_failed++;
4200 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
4202 spin_lock_irqsave(&busiest->lock, flags);
4204 /* don't kick the migration_thread, if the curr
4205 * task on busiest cpu can't be moved to this_cpu
4207 if (!cpumask_test_cpu(this_cpu,
4208 &busiest->curr->cpus_allowed)) {
4209 spin_unlock_irqrestore(&busiest->lock, flags);
4211 goto out_one_pinned;
4214 if (!busiest->active_balance) {
4215 busiest->active_balance = 1;
4216 busiest->push_cpu = this_cpu;
4219 spin_unlock_irqrestore(&busiest->lock, flags);
4221 wake_up_process(busiest->migration_thread);
4224 * We've kicked active balancing, reset the failure
4227 sd->nr_balance_failed = sd->cache_nice_tries+1;
4230 sd->nr_balance_failed = 0;
4232 if (likely(!active_balance)) {
4233 /* We were unbalanced, so reset the balancing interval */
4234 sd->balance_interval = sd->min_interval;
4237 * If we've begun active balancing, start to back off. This
4238 * case may not be covered by the all_pinned logic if there
4239 * is only 1 task on the busy runqueue (because we don't call
4242 if (sd->balance_interval < sd->max_interval)
4243 sd->balance_interval *= 2;
4246 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4247 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4253 schedstat_inc(sd, lb_balanced[idle]);
4255 sd->nr_balance_failed = 0;
4258 /* tune up the balancing interval */
4259 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4260 (sd->balance_interval < sd->max_interval))
4261 sd->balance_interval *= 2;
4263 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4264 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4275 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4276 * tasks if there is an imbalance.
4278 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4279 * this_rq is locked.
4282 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
4284 struct sched_group *group;
4285 struct rq *busiest = NULL;
4286 unsigned long imbalance;
4290 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4292 cpumask_setall(cpus);
4295 * When power savings policy is enabled for the parent domain, idle
4296 * sibling can pick up load irrespective of busy siblings. In this case,
4297 * let the state of idle sibling percolate up as IDLE, instead of
4298 * portraying it as CPU_NOT_IDLE.
4300 if (sd->flags & SD_SHARE_CPUPOWER &&
4301 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4304 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
4306 update_shares_locked(this_rq, sd);
4307 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
4308 &sd_idle, cpus, NULL);
4310 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
4314 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
4316 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
4320 BUG_ON(busiest == this_rq);
4322 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
4325 if (busiest->nr_running > 1) {
4326 /* Attempt to move tasks */
4327 double_lock_balance(this_rq, busiest);
4328 /* this_rq->clock is already updated */
4329 update_rq_clock(busiest);
4330 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4331 imbalance, sd, CPU_NEWLY_IDLE,
4333 double_unlock_balance(this_rq, busiest);
4335 if (unlikely(all_pinned)) {
4336 cpumask_clear_cpu(cpu_of(busiest), cpus);
4337 if (!cpumask_empty(cpus))
4343 int active_balance = 0;
4345 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
4346 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4347 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4350 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4353 if (sd->nr_balance_failed++ < 2)
4357 * The only task running in a non-idle cpu can be moved to this
4358 * cpu in an attempt to completely freeup the other CPU
4359 * package. The same method used to move task in load_balance()
4360 * have been extended for load_balance_newidle() to speedup
4361 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4363 * The package power saving logic comes from
4364 * find_busiest_group(). If there are no imbalance, then
4365 * f_b_g() will return NULL. However when sched_mc={1,2} then
4366 * f_b_g() will select a group from which a running task may be
4367 * pulled to this cpu in order to make the other package idle.
4368 * If there is no opportunity to make a package idle and if
4369 * there are no imbalance, then f_b_g() will return NULL and no
4370 * action will be taken in load_balance_newidle().
4372 * Under normal task pull operation due to imbalance, there
4373 * will be more than one task in the source run queue and
4374 * move_tasks() will succeed. ld_moved will be true and this
4375 * active balance code will not be triggered.
4378 /* Lock busiest in correct order while this_rq is held */
4379 double_lock_balance(this_rq, busiest);
4382 * don't kick the migration_thread, if the curr
4383 * task on busiest cpu can't be moved to this_cpu
4385 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
4386 double_unlock_balance(this_rq, busiest);
4391 if (!busiest->active_balance) {
4392 busiest->active_balance = 1;
4393 busiest->push_cpu = this_cpu;
4397 double_unlock_balance(this_rq, busiest);
4399 * Should not call ttwu while holding a rq->lock
4401 spin_unlock(&this_rq->lock);
4403 wake_up_process(busiest->migration_thread);
4404 spin_lock(&this_rq->lock);
4407 sd->nr_balance_failed = 0;
4409 update_shares_locked(this_rq, sd);
4413 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
4414 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4415 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4417 sd->nr_balance_failed = 0;
4423 * idle_balance is called by schedule() if this_cpu is about to become
4424 * idle. Attempts to pull tasks from other CPUs.
4426 static void idle_balance(int this_cpu, struct rq *this_rq)
4428 struct sched_domain *sd;
4429 int pulled_task = 0;
4430 unsigned long next_balance = jiffies + HZ;
4432 for_each_domain(this_cpu, sd) {
4433 unsigned long interval;
4435 if (!(sd->flags & SD_LOAD_BALANCE))
4438 if (sd->flags & SD_BALANCE_NEWIDLE)
4439 /* If we've pulled tasks over stop searching: */
4440 pulled_task = load_balance_newidle(this_cpu, this_rq,
4443 interval = msecs_to_jiffies(sd->balance_interval);
4444 if (time_after(next_balance, sd->last_balance + interval))
4445 next_balance = sd->last_balance + interval;
4449 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4451 * We are going idle. next_balance may be set based on
4452 * a busy processor. So reset next_balance.
4454 this_rq->next_balance = next_balance;
4459 * active_load_balance is run by migration threads. It pushes running tasks
4460 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4461 * running on each physical CPU where possible, and avoids physical /
4462 * logical imbalances.
4464 * Called with busiest_rq locked.
4466 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
4468 int target_cpu = busiest_rq->push_cpu;
4469 struct sched_domain *sd;
4470 struct rq *target_rq;
4472 /* Is there any task to move? */
4473 if (busiest_rq->nr_running <= 1)
4476 target_rq = cpu_rq(target_cpu);
4479 * This condition is "impossible", if it occurs
4480 * we need to fix it. Originally reported by
4481 * Bjorn Helgaas on a 128-cpu setup.
4483 BUG_ON(busiest_rq == target_rq);
4485 /* move a task from busiest_rq to target_rq */
4486 double_lock_balance(busiest_rq, target_rq);
4487 update_rq_clock(busiest_rq);
4488 update_rq_clock(target_rq);
4490 /* Search for an sd spanning us and the target CPU. */
4491 for_each_domain(target_cpu, sd) {
4492 if ((sd->flags & SD_LOAD_BALANCE) &&
4493 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4498 schedstat_inc(sd, alb_count);
4500 if (move_one_task(target_rq, target_cpu, busiest_rq,
4502 schedstat_inc(sd, alb_pushed);
4504 schedstat_inc(sd, alb_failed);
4506 double_unlock_balance(busiest_rq, target_rq);
4511 atomic_t load_balancer;
4512 cpumask_var_t cpu_mask;
4513 cpumask_var_t ilb_grp_nohz_mask;
4514 } nohz ____cacheline_aligned = {
4515 .load_balancer = ATOMIC_INIT(-1),
4518 int get_nohz_load_balancer(void)
4520 return atomic_read(&nohz.load_balancer);
4523 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4525 * lowest_flag_domain - Return lowest sched_domain containing flag.
4526 * @cpu: The cpu whose lowest level of sched domain is to
4528 * @flag: The flag to check for the lowest sched_domain
4529 * for the given cpu.
4531 * Returns the lowest sched_domain of a cpu which contains the given flag.
4533 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4535 struct sched_domain *sd;
4537 for_each_domain(cpu, sd)
4538 if (sd && (sd->flags & flag))
4545 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4546 * @cpu: The cpu whose domains we're iterating over.
4547 * @sd: variable holding the value of the power_savings_sd
4549 * @flag: The flag to filter the sched_domains to be iterated.
4551 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4552 * set, starting from the lowest sched_domain to the highest.
4554 #define for_each_flag_domain(cpu, sd, flag) \
4555 for (sd = lowest_flag_domain(cpu, flag); \
4556 (sd && (sd->flags & flag)); sd = sd->parent)
4559 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4560 * @ilb_group: group to be checked for semi-idleness
4562 * Returns: 1 if the group is semi-idle. 0 otherwise.
4564 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4565 * and atleast one non-idle CPU. This helper function checks if the given
4566 * sched_group is semi-idle or not.
4568 static inline int is_semi_idle_group(struct sched_group *ilb_group)
4570 cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask,
4571 sched_group_cpus(ilb_group));
4574 * A sched_group is semi-idle when it has atleast one busy cpu
4575 * and atleast one idle cpu.
4577 if (cpumask_empty(nohz.ilb_grp_nohz_mask))
4580 if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group)))
4586 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4587 * @cpu: The cpu which is nominating a new idle_load_balancer.
4589 * Returns: Returns the id of the idle load balancer if it exists,
4590 * Else, returns >= nr_cpu_ids.
4592 * This algorithm picks the idle load balancer such that it belongs to a
4593 * semi-idle powersavings sched_domain. The idea is to try and avoid
4594 * completely idle packages/cores just for the purpose of idle load balancing
4595 * when there are other idle cpu's which are better suited for that job.
4597 static int find_new_ilb(int cpu)
4599 struct sched_domain *sd;
4600 struct sched_group *ilb_group;
4603 * Have idle load balancer selection from semi-idle packages only
4604 * when power-aware load balancing is enabled
4606 if (!(sched_smt_power_savings || sched_mc_power_savings))
4610 * Optimize for the case when we have no idle CPUs or only one
4611 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4613 if (cpumask_weight(nohz.cpu_mask) < 2)
4616 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4617 ilb_group = sd->groups;
4620 if (is_semi_idle_group(ilb_group))
4621 return cpumask_first(nohz.ilb_grp_nohz_mask);
4623 ilb_group = ilb_group->next;
4625 } while (ilb_group != sd->groups);
4629 return cpumask_first(nohz.cpu_mask);
4631 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4632 static inline int find_new_ilb(int call_cpu)
4634 return cpumask_first(nohz.cpu_mask);
4639 * This routine will try to nominate the ilb (idle load balancing)
4640 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4641 * load balancing on behalf of all those cpus. If all the cpus in the system
4642 * go into this tickless mode, then there will be no ilb owner (as there is
4643 * no need for one) and all the cpus will sleep till the next wakeup event
4646 * For the ilb owner, tick is not stopped. And this tick will be used
4647 * for idle load balancing. ilb owner will still be part of
4650 * While stopping the tick, this cpu will become the ilb owner if there
4651 * is no other owner. And will be the owner till that cpu becomes busy
4652 * or if all cpus in the system stop their ticks at which point
4653 * there is no need for ilb owner.
4655 * When the ilb owner becomes busy, it nominates another owner, during the
4656 * next busy scheduler_tick()
4658 int select_nohz_load_balancer(int stop_tick)
4660 int cpu = smp_processor_id();
4663 cpu_rq(cpu)->in_nohz_recently = 1;
4665 if (!cpu_active(cpu)) {
4666 if (atomic_read(&nohz.load_balancer) != cpu)
4670 * If we are going offline and still the leader,
4673 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4679 cpumask_set_cpu(cpu, nohz.cpu_mask);
4681 /* time for ilb owner also to sleep */
4682 if (cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4683 if (atomic_read(&nohz.load_balancer) == cpu)
4684 atomic_set(&nohz.load_balancer, -1);
4688 if (atomic_read(&nohz.load_balancer) == -1) {
4689 /* make me the ilb owner */
4690 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4692 } else if (atomic_read(&nohz.load_balancer) == cpu) {
4695 if (!(sched_smt_power_savings ||
4696 sched_mc_power_savings))
4699 * Check to see if there is a more power-efficient
4702 new_ilb = find_new_ilb(cpu);
4703 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4704 atomic_set(&nohz.load_balancer, -1);
4705 resched_cpu(new_ilb);
4711 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4714 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4716 if (atomic_read(&nohz.load_balancer) == cpu)
4717 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4724 static DEFINE_SPINLOCK(balancing);
4727 * It checks each scheduling domain to see if it is due to be balanced,
4728 * and initiates a balancing operation if so.
4730 * Balancing parameters are set up in arch_init_sched_domains.
4732 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4735 struct rq *rq = cpu_rq(cpu);
4736 unsigned long interval;
4737 struct sched_domain *sd;
4738 /* Earliest time when we have to do rebalance again */
4739 unsigned long next_balance = jiffies + 60*HZ;
4740 int update_next_balance = 0;
4743 for_each_domain(cpu, sd) {
4744 if (!(sd->flags & SD_LOAD_BALANCE))
4747 interval = sd->balance_interval;
4748 if (idle != CPU_IDLE)
4749 interval *= sd->busy_factor;
4751 /* scale ms to jiffies */
4752 interval = msecs_to_jiffies(interval);
4753 if (unlikely(!interval))
4755 if (interval > HZ*NR_CPUS/10)
4756 interval = HZ*NR_CPUS/10;
4758 need_serialize = sd->flags & SD_SERIALIZE;
4760 if (need_serialize) {
4761 if (!spin_trylock(&balancing))
4765 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4766 if (load_balance(cpu, rq, sd, idle, &balance)) {
4768 * We've pulled tasks over so either we're no
4769 * longer idle, or one of our SMT siblings is
4772 idle = CPU_NOT_IDLE;
4774 sd->last_balance = jiffies;
4777 spin_unlock(&balancing);
4779 if (time_after(next_balance, sd->last_balance + interval)) {
4780 next_balance = sd->last_balance + interval;
4781 update_next_balance = 1;
4785 * Stop the load balance at this level. There is another
4786 * CPU in our sched group which is doing load balancing more
4794 * next_balance will be updated only when there is a need.
4795 * When the cpu is attached to null domain for ex, it will not be
4798 if (likely(update_next_balance))
4799 rq->next_balance = next_balance;
4803 * run_rebalance_domains is triggered when needed from the scheduler tick.
4804 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4805 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4807 static void run_rebalance_domains(struct softirq_action *h)
4809 int this_cpu = smp_processor_id();
4810 struct rq *this_rq = cpu_rq(this_cpu);
4811 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4812 CPU_IDLE : CPU_NOT_IDLE;
4814 rebalance_domains(this_cpu, idle);
4818 * If this cpu is the owner for idle load balancing, then do the
4819 * balancing on behalf of the other idle cpus whose ticks are
4822 if (this_rq->idle_at_tick &&
4823 atomic_read(&nohz.load_balancer) == this_cpu) {
4827 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4828 if (balance_cpu == this_cpu)
4832 * If this cpu gets work to do, stop the load balancing
4833 * work being done for other cpus. Next load
4834 * balancing owner will pick it up.
4839 rebalance_domains(balance_cpu, CPU_IDLE);
4841 rq = cpu_rq(balance_cpu);
4842 if (time_after(this_rq->next_balance, rq->next_balance))
4843 this_rq->next_balance = rq->next_balance;
4849 static inline int on_null_domain(int cpu)
4851 return !rcu_dereference(cpu_rq(cpu)->sd);
4855 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4857 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4858 * idle load balancing owner or decide to stop the periodic load balancing,
4859 * if the whole system is idle.
4861 static inline void trigger_load_balance(struct rq *rq, int cpu)
4865 * If we were in the nohz mode recently and busy at the current
4866 * scheduler tick, then check if we need to nominate new idle
4869 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4870 rq->in_nohz_recently = 0;
4872 if (atomic_read(&nohz.load_balancer) == cpu) {
4873 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4874 atomic_set(&nohz.load_balancer, -1);
4877 if (atomic_read(&nohz.load_balancer) == -1) {
4878 int ilb = find_new_ilb(cpu);
4880 if (ilb < nr_cpu_ids)
4886 * If this cpu is idle and doing idle load balancing for all the
4887 * cpus with ticks stopped, is it time for that to stop?
4889 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4890 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4896 * If this cpu is idle and the idle load balancing is done by
4897 * someone else, then no need raise the SCHED_SOFTIRQ
4899 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4900 cpumask_test_cpu(cpu, nohz.cpu_mask))
4903 /* Don't need to rebalance while attached to NULL domain */
4904 if (time_after_eq(jiffies, rq->next_balance) &&
4905 likely(!on_null_domain(cpu)))
4906 raise_softirq(SCHED_SOFTIRQ);
4909 #else /* CONFIG_SMP */
4912 * on UP we do not need to balance between CPUs:
4914 static inline void idle_balance(int cpu, struct rq *rq)
4920 DEFINE_PER_CPU(struct kernel_stat, kstat);
4922 EXPORT_PER_CPU_SYMBOL(kstat);
4925 * Return any ns on the sched_clock that have not yet been accounted in
4926 * @p in case that task is currently running.
4928 * Called with task_rq_lock() held on @rq.
4930 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
4934 if (task_current(rq, p)) {
4935 update_rq_clock(rq);
4936 ns = rq->clock - p->se.exec_start;
4944 unsigned long long task_delta_exec(struct task_struct *p)
4946 unsigned long flags;
4950 rq = task_rq_lock(p, &flags);
4951 ns = do_task_delta_exec(p, rq);
4952 task_rq_unlock(rq, &flags);
4958 * Return accounted runtime for the task.
4959 * In case the task is currently running, return the runtime plus current's
4960 * pending runtime that have not been accounted yet.
4962 unsigned long long task_sched_runtime(struct task_struct *p)
4964 unsigned long flags;
4968 rq = task_rq_lock(p, &flags);
4969 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
4970 task_rq_unlock(rq, &flags);
4976 * Return sum_exec_runtime for the thread group.
4977 * In case the task is currently running, return the sum plus current's
4978 * pending runtime that have not been accounted yet.
4980 * Note that the thread group might have other running tasks as well,
4981 * so the return value not includes other pending runtime that other
4982 * running tasks might have.
4984 unsigned long long thread_group_sched_runtime(struct task_struct *p)
4986 struct task_cputime totals;
4987 unsigned long flags;
4991 rq = task_rq_lock(p, &flags);
4992 thread_group_cputime(p, &totals);
4993 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
4994 task_rq_unlock(rq, &flags);
5000 * Account user cpu time to a process.
5001 * @p: the process that the cpu time gets accounted to
5002 * @cputime: the cpu time spent in user space since the last update
5003 * @cputime_scaled: cputime scaled by cpu frequency
5005 void account_user_time(struct task_struct *p, cputime_t cputime,
5006 cputime_t cputime_scaled)
5008 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5011 /* Add user time to process. */
5012 p->utime = cputime_add(p->utime, cputime);
5013 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5014 account_group_user_time(p, cputime);
5016 /* Add user time to cpustat. */
5017 tmp = cputime_to_cputime64(cputime);
5018 if (TASK_NICE(p) > 0)
5019 cpustat->nice = cputime64_add(cpustat->nice, tmp);
5021 cpustat->user = cputime64_add(cpustat->user, tmp);
5023 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
5024 /* Account for user time used */
5025 acct_update_integrals(p);
5029 * Account guest cpu time to a process.
5030 * @p: the process that the cpu time gets accounted to
5031 * @cputime: the cpu time spent in virtual machine since the last update
5032 * @cputime_scaled: cputime scaled by cpu frequency
5034 static void account_guest_time(struct task_struct *p, cputime_t cputime,
5035 cputime_t cputime_scaled)
5038 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5040 tmp = cputime_to_cputime64(cputime);
5042 /* Add guest time to process. */
5043 p->utime = cputime_add(p->utime, cputime);
5044 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5045 account_group_user_time(p, cputime);
5046 p->gtime = cputime_add(p->gtime, cputime);
5048 /* Add guest time to cpustat. */
5049 cpustat->user = cputime64_add(cpustat->user, tmp);
5050 cpustat->guest = cputime64_add(cpustat->guest, tmp);
5054 * Account system cpu time to a process.
5055 * @p: the process that the cpu time gets accounted to
5056 * @hardirq_offset: the offset to subtract from hardirq_count()
5057 * @cputime: the cpu time spent in kernel space since the last update
5058 * @cputime_scaled: cputime scaled by cpu frequency
5060 void account_system_time(struct task_struct *p, int hardirq_offset,
5061 cputime_t cputime, cputime_t cputime_scaled)
5063 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5066 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
5067 account_guest_time(p, cputime, cputime_scaled);
5071 /* Add system time to process. */
5072 p->stime = cputime_add(p->stime, cputime);
5073 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
5074 account_group_system_time(p, cputime);
5076 /* Add system time to cpustat. */
5077 tmp = cputime_to_cputime64(cputime);
5078 if (hardirq_count() - hardirq_offset)
5079 cpustat->irq = cputime64_add(cpustat->irq, tmp);
5080 else if (softirq_count())
5081 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
5083 cpustat->system = cputime64_add(cpustat->system, tmp);
5085 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
5087 /* Account for system time used */
5088 acct_update_integrals(p);
5092 * Account for involuntary wait time.
5093 * @steal: the cpu time spent in involuntary wait
5095 void account_steal_time(cputime_t cputime)
5097 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5098 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5100 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
5104 * Account for idle time.
5105 * @cputime: the cpu time spent in idle wait
5107 void account_idle_time(cputime_t cputime)
5109 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5110 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5111 struct rq *rq = this_rq();
5113 if (atomic_read(&rq->nr_iowait) > 0)
5114 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
5116 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
5119 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
5122 * Account a single tick of cpu time.
5123 * @p: the process that the cpu time gets accounted to
5124 * @user_tick: indicates if the tick is a user or a system tick
5126 void account_process_tick(struct task_struct *p, int user_tick)
5128 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
5129 struct rq *rq = this_rq();
5132 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
5133 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
5134 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
5137 account_idle_time(cputime_one_jiffy);
5141 * Account multiple ticks of steal time.
5142 * @p: the process from which the cpu time has been stolen
5143 * @ticks: number of stolen ticks
5145 void account_steal_ticks(unsigned long ticks)
5147 account_steal_time(jiffies_to_cputime(ticks));
5151 * Account multiple ticks of idle time.
5152 * @ticks: number of stolen ticks
5154 void account_idle_ticks(unsigned long ticks)
5156 account_idle_time(jiffies_to_cputime(ticks));
5162 * Use precise platform statistics if available:
5164 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
5165 cputime_t task_utime(struct task_struct *p)
5170 cputime_t task_stime(struct task_struct *p)
5175 cputime_t task_utime(struct task_struct *p)
5177 clock_t utime = cputime_to_clock_t(p->utime),
5178 total = utime + cputime_to_clock_t(p->stime);
5182 * Use CFS's precise accounting:
5184 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
5188 do_div(temp, total);
5190 utime = (clock_t)temp;
5192 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
5193 return p->prev_utime;
5196 cputime_t task_stime(struct task_struct *p)
5201 * Use CFS's precise accounting. (we subtract utime from
5202 * the total, to make sure the total observed by userspace
5203 * grows monotonically - apps rely on that):
5205 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
5206 cputime_to_clock_t(task_utime(p));
5209 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
5211 return p->prev_stime;
5215 inline cputime_t task_gtime(struct task_struct *p)
5221 * This function gets called by the timer code, with HZ frequency.
5222 * We call it with interrupts disabled.
5224 * It also gets called by the fork code, when changing the parent's
5227 void scheduler_tick(void)
5229 int cpu = smp_processor_id();
5230 struct rq *rq = cpu_rq(cpu);
5231 struct task_struct *curr = rq->curr;
5235 spin_lock(&rq->lock);
5236 update_rq_clock(rq);
5237 update_cpu_load(rq);
5238 curr->sched_class->task_tick(rq, curr, 0);
5239 spin_unlock(&rq->lock);
5241 perf_event_task_tick(curr, cpu);
5244 rq->idle_at_tick = idle_cpu(cpu);
5245 trigger_load_balance(rq, cpu);
5249 notrace unsigned long get_parent_ip(unsigned long addr)
5251 if (in_lock_functions(addr)) {
5252 addr = CALLER_ADDR2;
5253 if (in_lock_functions(addr))
5254 addr = CALLER_ADDR3;
5259 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5260 defined(CONFIG_PREEMPT_TRACER))
5262 void __kprobes add_preempt_count(int val)
5264 #ifdef CONFIG_DEBUG_PREEMPT
5268 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5271 preempt_count() += val;
5272 #ifdef CONFIG_DEBUG_PREEMPT
5274 * Spinlock count overflowing soon?
5276 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5279 if (preempt_count() == val)
5280 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5282 EXPORT_SYMBOL(add_preempt_count);
5284 void __kprobes sub_preempt_count(int val)
5286 #ifdef CONFIG_DEBUG_PREEMPT
5290 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5293 * Is the spinlock portion underflowing?
5295 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5296 !(preempt_count() & PREEMPT_MASK)))
5300 if (preempt_count() == val)
5301 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5302 preempt_count() -= val;
5304 EXPORT_SYMBOL(sub_preempt_count);
5309 * Print scheduling while atomic bug:
5311 static noinline void __schedule_bug(struct task_struct *prev)
5313 struct pt_regs *regs = get_irq_regs();
5315 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5316 prev->comm, prev->pid, preempt_count());
5318 debug_show_held_locks(prev);
5320 if (irqs_disabled())
5321 print_irqtrace_events(prev);
5330 * Various schedule()-time debugging checks and statistics:
5332 static inline void schedule_debug(struct task_struct *prev)
5335 * Test if we are atomic. Since do_exit() needs to call into
5336 * schedule() atomically, we ignore that path for now.
5337 * Otherwise, whine if we are scheduling when we should not be.
5339 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
5340 __schedule_bug(prev);
5342 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5344 schedstat_inc(this_rq(), sched_count);
5345 #ifdef CONFIG_SCHEDSTATS
5346 if (unlikely(prev->lock_depth >= 0)) {
5347 schedstat_inc(this_rq(), bkl_count);
5348 schedstat_inc(prev, sched_info.bkl_count);
5353 static void put_prev_task(struct rq *rq, struct task_struct *p)
5355 u64 runtime = p->se.sum_exec_runtime - p->se.prev_sum_exec_runtime;
5357 update_avg(&p->se.avg_running, runtime);
5359 if (p->state == TASK_RUNNING) {
5361 * In order to avoid avg_overlap growing stale when we are
5362 * indeed overlapping and hence not getting put to sleep, grow
5363 * the avg_overlap on preemption.
5365 * We use the average preemption runtime because that
5366 * correlates to the amount of cache footprint a task can
5369 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
5370 update_avg(&p->se.avg_overlap, runtime);
5372 update_avg(&p->se.avg_running, 0);
5374 p->sched_class->put_prev_task(rq, p);
5378 * Pick up the highest-prio task:
5380 static inline struct task_struct *
5381 pick_next_task(struct rq *rq)
5383 const struct sched_class *class;
5384 struct task_struct *p;
5387 * Optimization: we know that if all tasks are in
5388 * the fair class we can call that function directly:
5390 if (likely(rq->nr_running == rq->cfs.nr_running)) {
5391 p = fair_sched_class.pick_next_task(rq);
5396 class = sched_class_highest;
5398 p = class->pick_next_task(rq);
5402 * Will never be NULL as the idle class always
5403 * returns a non-NULL p:
5405 class = class->next;
5410 * schedule() is the main scheduler function.
5412 asmlinkage void __sched schedule(void)
5414 struct task_struct *prev, *next;
5415 unsigned long *switch_count;
5421 cpu = smp_processor_id();
5425 switch_count = &prev->nivcsw;
5427 release_kernel_lock(prev);
5428 need_resched_nonpreemptible:
5430 schedule_debug(prev);
5432 if (sched_feat(HRTICK))
5435 spin_lock_irq(&rq->lock);
5436 update_rq_clock(rq);
5437 clear_tsk_need_resched(prev);
5439 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
5440 if (unlikely(signal_pending_state(prev->state, prev)))
5441 prev->state = TASK_RUNNING;
5443 deactivate_task(rq, prev, 1);
5444 switch_count = &prev->nvcsw;
5447 pre_schedule(rq, prev);
5449 if (unlikely(!rq->nr_running))
5450 idle_balance(cpu, rq);
5452 put_prev_task(rq, prev);
5453 next = pick_next_task(rq);
5455 if (likely(prev != next)) {
5456 sched_info_switch(prev, next);
5457 perf_event_task_sched_out(prev, next, cpu);
5463 context_switch(rq, prev, next); /* unlocks the rq */
5465 * the context switch might have flipped the stack from under
5466 * us, hence refresh the local variables.
5468 cpu = smp_processor_id();
5471 spin_unlock_irq(&rq->lock);
5475 if (unlikely(reacquire_kernel_lock(current) < 0))
5476 goto need_resched_nonpreemptible;
5478 preempt_enable_no_resched();
5482 EXPORT_SYMBOL(schedule);
5486 * Look out! "owner" is an entirely speculative pointer
5487 * access and not reliable.
5489 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5494 if (!sched_feat(OWNER_SPIN))
5497 #ifdef CONFIG_DEBUG_PAGEALLOC
5499 * Need to access the cpu field knowing that
5500 * DEBUG_PAGEALLOC could have unmapped it if
5501 * the mutex owner just released it and exited.
5503 if (probe_kernel_address(&owner->cpu, cpu))
5510 * Even if the access succeeded (likely case),
5511 * the cpu field may no longer be valid.
5513 if (cpu >= nr_cpumask_bits)
5517 * We need to validate that we can do a
5518 * get_cpu() and that we have the percpu area.
5520 if (!cpu_online(cpu))
5527 * Owner changed, break to re-assess state.
5529 if (lock->owner != owner)
5533 * Is that owner really running on that cpu?
5535 if (task_thread_info(rq->curr) != owner || need_resched())
5545 #ifdef CONFIG_PREEMPT
5547 * this is the entry point to schedule() from in-kernel preemption
5548 * off of preempt_enable. Kernel preemptions off return from interrupt
5549 * occur there and call schedule directly.
5551 asmlinkage void __sched preempt_schedule(void)
5553 struct thread_info *ti = current_thread_info();
5556 * If there is a non-zero preempt_count or interrupts are disabled,
5557 * we do not want to preempt the current task. Just return..
5559 if (likely(ti->preempt_count || irqs_disabled()))
5563 add_preempt_count(PREEMPT_ACTIVE);
5565 sub_preempt_count(PREEMPT_ACTIVE);
5568 * Check again in case we missed a preemption opportunity
5569 * between schedule and now.
5572 } while (need_resched());
5574 EXPORT_SYMBOL(preempt_schedule);
5577 * this is the entry point to schedule() from kernel preemption
5578 * off of irq context.
5579 * Note, that this is called and return with irqs disabled. This will
5580 * protect us against recursive calling from irq.
5582 asmlinkage void __sched preempt_schedule_irq(void)
5584 struct thread_info *ti = current_thread_info();
5586 /* Catch callers which need to be fixed */
5587 BUG_ON(ti->preempt_count || !irqs_disabled());
5590 add_preempt_count(PREEMPT_ACTIVE);
5593 local_irq_disable();
5594 sub_preempt_count(PREEMPT_ACTIVE);
5597 * Check again in case we missed a preemption opportunity
5598 * between schedule and now.
5601 } while (need_resched());
5604 #endif /* CONFIG_PREEMPT */
5606 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
5609 return try_to_wake_up(curr->private, mode, wake_flags);
5611 EXPORT_SYMBOL(default_wake_function);
5614 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5615 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5616 * number) then we wake all the non-exclusive tasks and one exclusive task.
5618 * There are circumstances in which we can try to wake a task which has already
5619 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5620 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5622 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
5623 int nr_exclusive, int wake_flags, void *key)
5625 wait_queue_t *curr, *next;
5627 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
5628 unsigned flags = curr->flags;
5630 if (curr->func(curr, mode, wake_flags, key) &&
5631 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
5637 * __wake_up - wake up threads blocked on a waitqueue.
5639 * @mode: which threads
5640 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5641 * @key: is directly passed to the wakeup function
5643 * It may be assumed that this function implies a write memory barrier before
5644 * changing the task state if and only if any tasks are woken up.
5646 void __wake_up(wait_queue_head_t *q, unsigned int mode,
5647 int nr_exclusive, void *key)
5649 unsigned long flags;
5651 spin_lock_irqsave(&q->lock, flags);
5652 __wake_up_common(q, mode, nr_exclusive, 0, key);
5653 spin_unlock_irqrestore(&q->lock, flags);
5655 EXPORT_SYMBOL(__wake_up);
5658 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5660 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
5662 __wake_up_common(q, mode, 1, 0, NULL);
5665 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
5667 __wake_up_common(q, mode, 1, 0, key);
5671 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5673 * @mode: which threads
5674 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5675 * @key: opaque value to be passed to wakeup targets
5677 * The sync wakeup differs that the waker knows that it will schedule
5678 * away soon, so while the target thread will be woken up, it will not
5679 * be migrated to another CPU - ie. the two threads are 'synchronized'
5680 * with each other. This can prevent needless bouncing between CPUs.
5682 * On UP it can prevent extra preemption.
5684 * It may be assumed that this function implies a write memory barrier before
5685 * changing the task state if and only if any tasks are woken up.
5687 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
5688 int nr_exclusive, void *key)
5690 unsigned long flags;
5691 int wake_flags = WF_SYNC;
5696 if (unlikely(!nr_exclusive))
5699 spin_lock_irqsave(&q->lock, flags);
5700 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
5701 spin_unlock_irqrestore(&q->lock, flags);
5703 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
5706 * __wake_up_sync - see __wake_up_sync_key()
5708 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
5710 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
5712 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
5715 * complete: - signals a single thread waiting on this completion
5716 * @x: holds the state of this particular completion
5718 * This will wake up a single thread waiting on this completion. Threads will be
5719 * awakened in the same order in which they were queued.
5721 * See also complete_all(), wait_for_completion() and related routines.
5723 * It may be assumed that this function implies a write memory barrier before
5724 * changing the task state if and only if any tasks are woken up.
5726 void complete(struct completion *x)
5728 unsigned long flags;
5730 spin_lock_irqsave(&x->wait.lock, flags);
5732 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
5733 spin_unlock_irqrestore(&x->wait.lock, flags);
5735 EXPORT_SYMBOL(complete);
5738 * complete_all: - signals all threads waiting on this completion
5739 * @x: holds the state of this particular completion
5741 * This will wake up all threads waiting on this particular completion event.
5743 * It may be assumed that this function implies a write memory barrier before
5744 * changing the task state if and only if any tasks are woken up.
5746 void complete_all(struct completion *x)
5748 unsigned long flags;
5750 spin_lock_irqsave(&x->wait.lock, flags);
5751 x->done += UINT_MAX/2;
5752 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
5753 spin_unlock_irqrestore(&x->wait.lock, flags);
5755 EXPORT_SYMBOL(complete_all);
5757 static inline long __sched
5758 do_wait_for_common(struct completion *x, long timeout, int state)
5761 DECLARE_WAITQUEUE(wait, current);
5763 wait.flags |= WQ_FLAG_EXCLUSIVE;
5764 __add_wait_queue_tail(&x->wait, &wait);
5766 if (signal_pending_state(state, current)) {
5767 timeout = -ERESTARTSYS;
5770 __set_current_state(state);
5771 spin_unlock_irq(&x->wait.lock);
5772 timeout = schedule_timeout(timeout);
5773 spin_lock_irq(&x->wait.lock);
5774 } while (!x->done && timeout);
5775 __remove_wait_queue(&x->wait, &wait);
5780 return timeout ?: 1;
5784 wait_for_common(struct completion *x, long timeout, int state)
5788 spin_lock_irq(&x->wait.lock);
5789 timeout = do_wait_for_common(x, timeout, state);
5790 spin_unlock_irq(&x->wait.lock);
5795 * wait_for_completion: - waits for completion of a task
5796 * @x: holds the state of this particular completion
5798 * This waits to be signaled for completion of a specific task. It is NOT
5799 * interruptible and there is no timeout.
5801 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5802 * and interrupt capability. Also see complete().
5804 void __sched wait_for_completion(struct completion *x)
5806 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
5808 EXPORT_SYMBOL(wait_for_completion);
5811 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5812 * @x: holds the state of this particular completion
5813 * @timeout: timeout value in jiffies
5815 * This waits for either a completion of a specific task to be signaled or for a
5816 * specified timeout to expire. The timeout is in jiffies. It is not
5819 unsigned long __sched
5820 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5822 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5824 EXPORT_SYMBOL(wait_for_completion_timeout);
5827 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5828 * @x: holds the state of this particular completion
5830 * This waits for completion of a specific task to be signaled. It is
5833 int __sched wait_for_completion_interruptible(struct completion *x)
5835 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5836 if (t == -ERESTARTSYS)
5840 EXPORT_SYMBOL(wait_for_completion_interruptible);
5843 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5844 * @x: holds the state of this particular completion
5845 * @timeout: timeout value in jiffies
5847 * This waits for either a completion of a specific task to be signaled or for a
5848 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5850 unsigned long __sched
5851 wait_for_completion_interruptible_timeout(struct completion *x,
5852 unsigned long timeout)
5854 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
5856 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5859 * wait_for_completion_killable: - waits for completion of a task (killable)
5860 * @x: holds the state of this particular completion
5862 * This waits to be signaled for completion of a specific task. It can be
5863 * interrupted by a kill signal.
5865 int __sched wait_for_completion_killable(struct completion *x)
5867 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5868 if (t == -ERESTARTSYS)
5872 EXPORT_SYMBOL(wait_for_completion_killable);
5875 * try_wait_for_completion - try to decrement a completion without blocking
5876 * @x: completion structure
5878 * Returns: 0 if a decrement cannot be done without blocking
5879 * 1 if a decrement succeeded.
5881 * If a completion is being used as a counting completion,
5882 * attempt to decrement the counter without blocking. This
5883 * enables us to avoid waiting if the resource the completion
5884 * is protecting is not available.
5886 bool try_wait_for_completion(struct completion *x)
5890 spin_lock_irq(&x->wait.lock);
5895 spin_unlock_irq(&x->wait.lock);
5898 EXPORT_SYMBOL(try_wait_for_completion);
5901 * completion_done - Test to see if a completion has any waiters
5902 * @x: completion structure
5904 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5905 * 1 if there are no waiters.
5908 bool completion_done(struct completion *x)
5912 spin_lock_irq(&x->wait.lock);
5915 spin_unlock_irq(&x->wait.lock);
5918 EXPORT_SYMBOL(completion_done);
5921 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
5923 unsigned long flags;
5926 init_waitqueue_entry(&wait, current);
5928 __set_current_state(state);
5930 spin_lock_irqsave(&q->lock, flags);
5931 __add_wait_queue(q, &wait);
5932 spin_unlock(&q->lock);
5933 timeout = schedule_timeout(timeout);
5934 spin_lock_irq(&q->lock);
5935 __remove_wait_queue(q, &wait);
5936 spin_unlock_irqrestore(&q->lock, flags);
5941 void __sched interruptible_sleep_on(wait_queue_head_t *q)
5943 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5945 EXPORT_SYMBOL(interruptible_sleep_on);
5948 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
5950 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
5952 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
5954 void __sched sleep_on(wait_queue_head_t *q)
5956 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5958 EXPORT_SYMBOL(sleep_on);
5960 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
5962 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
5964 EXPORT_SYMBOL(sleep_on_timeout);
5966 #ifdef CONFIG_RT_MUTEXES
5969 * rt_mutex_setprio - set the current priority of a task
5971 * @prio: prio value (kernel-internal form)
5973 * This function changes the 'effective' priority of a task. It does
5974 * not touch ->normal_prio like __setscheduler().
5976 * Used by the rt_mutex code to implement priority inheritance logic.
5978 void rt_mutex_setprio(struct task_struct *p, int prio)
5980 unsigned long flags;
5981 int oldprio, on_rq, running;
5983 const struct sched_class *prev_class = p->sched_class;
5985 BUG_ON(prio < 0 || prio > MAX_PRIO);
5987 rq = task_rq_lock(p, &flags);
5988 update_rq_clock(rq);
5991 on_rq = p->se.on_rq;
5992 running = task_current(rq, p);
5994 dequeue_task(rq, p, 0);
5996 p->sched_class->put_prev_task(rq, p);
5999 p->sched_class = &rt_sched_class;
6001 p->sched_class = &fair_sched_class;
6006 p->sched_class->set_curr_task(rq);
6008 enqueue_task(rq, p, 0);
6010 check_class_changed(rq, p, prev_class, oldprio, running);
6012 task_rq_unlock(rq, &flags);
6017 void set_user_nice(struct task_struct *p, long nice)
6019 int old_prio, delta, on_rq;
6020 unsigned long flags;
6023 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
6026 * We have to be careful, if called from sys_setpriority(),
6027 * the task might be in the middle of scheduling on another CPU.
6029 rq = task_rq_lock(p, &flags);
6030 update_rq_clock(rq);
6032 * The RT priorities are set via sched_setscheduler(), but we still
6033 * allow the 'normal' nice value to be set - but as expected
6034 * it wont have any effect on scheduling until the task is
6035 * SCHED_FIFO/SCHED_RR:
6037 if (task_has_rt_policy(p)) {
6038 p->static_prio = NICE_TO_PRIO(nice);
6041 on_rq = p->se.on_rq;
6043 dequeue_task(rq, p, 0);
6045 p->static_prio = NICE_TO_PRIO(nice);
6048 p->prio = effective_prio(p);
6049 delta = p->prio - old_prio;
6052 enqueue_task(rq, p, 0);
6054 * If the task increased its priority or is running and
6055 * lowered its priority, then reschedule its CPU:
6057 if (delta < 0 || (delta > 0 && task_running(rq, p)))
6058 resched_task(rq->curr);
6061 task_rq_unlock(rq, &flags);
6063 EXPORT_SYMBOL(set_user_nice);
6066 * can_nice - check if a task can reduce its nice value
6070 int can_nice(const struct task_struct *p, const int nice)
6072 /* convert nice value [19,-20] to rlimit style value [1,40] */
6073 int nice_rlim = 20 - nice;
6075 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
6076 capable(CAP_SYS_NICE));
6079 #ifdef __ARCH_WANT_SYS_NICE
6082 * sys_nice - change the priority of the current process.
6083 * @increment: priority increment
6085 * sys_setpriority is a more generic, but much slower function that
6086 * does similar things.
6088 SYSCALL_DEFINE1(nice, int, increment)
6093 * Setpriority might change our priority at the same moment.
6094 * We don't have to worry. Conceptually one call occurs first
6095 * and we have a single winner.
6097 if (increment < -40)
6102 nice = TASK_NICE(current) + increment;
6108 if (increment < 0 && !can_nice(current, nice))
6111 retval = security_task_setnice(current, nice);
6115 set_user_nice(current, nice);
6122 * task_prio - return the priority value of a given task.
6123 * @p: the task in question.
6125 * This is the priority value as seen by users in /proc.
6126 * RT tasks are offset by -200. Normal tasks are centered
6127 * around 0, value goes from -16 to +15.
6129 int task_prio(const struct task_struct *p)
6131 return p->prio - MAX_RT_PRIO;
6135 * task_nice - return the nice value of a given task.
6136 * @p: the task in question.
6138 int task_nice(const struct task_struct *p)
6140 return TASK_NICE(p);
6142 EXPORT_SYMBOL(task_nice);
6145 * idle_cpu - is a given cpu idle currently?
6146 * @cpu: the processor in question.
6148 int idle_cpu(int cpu)
6150 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
6154 * idle_task - return the idle task for a given cpu.
6155 * @cpu: the processor in question.
6157 struct task_struct *idle_task(int cpu)
6159 return cpu_rq(cpu)->idle;
6163 * find_process_by_pid - find a process with a matching PID value.
6164 * @pid: the pid in question.
6166 static struct task_struct *find_process_by_pid(pid_t pid)
6168 return pid ? find_task_by_vpid(pid) : current;
6171 /* Actually do priority change: must hold rq lock. */
6173 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
6175 BUG_ON(p->se.on_rq);
6178 switch (p->policy) {
6182 p->sched_class = &fair_sched_class;
6186 p->sched_class = &rt_sched_class;
6190 p->rt_priority = prio;
6191 p->normal_prio = normal_prio(p);
6192 /* we are holding p->pi_lock already */
6193 p->prio = rt_mutex_getprio(p);
6198 * check the target process has a UID that matches the current process's
6200 static bool check_same_owner(struct task_struct *p)
6202 const struct cred *cred = current_cred(), *pcred;
6206 pcred = __task_cred(p);
6207 match = (cred->euid == pcred->euid ||
6208 cred->euid == pcred->uid);
6213 static int __sched_setscheduler(struct task_struct *p, int policy,
6214 struct sched_param *param, bool user)
6216 int retval, oldprio, oldpolicy = -1, on_rq, running;
6217 unsigned long flags;
6218 const struct sched_class *prev_class = p->sched_class;
6222 /* may grab non-irq protected spin_locks */
6223 BUG_ON(in_interrupt());
6225 /* double check policy once rq lock held */
6227 reset_on_fork = p->sched_reset_on_fork;
6228 policy = oldpolicy = p->policy;
6230 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
6231 policy &= ~SCHED_RESET_ON_FORK;
6233 if (policy != SCHED_FIFO && policy != SCHED_RR &&
6234 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
6235 policy != SCHED_IDLE)
6240 * Valid priorities for SCHED_FIFO and SCHED_RR are
6241 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6242 * SCHED_BATCH and SCHED_IDLE is 0.
6244 if (param->sched_priority < 0 ||
6245 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
6246 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
6248 if (rt_policy(policy) != (param->sched_priority != 0))
6252 * Allow unprivileged RT tasks to decrease priority:
6254 if (user && !capable(CAP_SYS_NICE)) {
6255 if (rt_policy(policy)) {
6256 unsigned long rlim_rtprio;
6258 if (!lock_task_sighand(p, &flags))
6260 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
6261 unlock_task_sighand(p, &flags);
6263 /* can't set/change the rt policy */
6264 if (policy != p->policy && !rlim_rtprio)
6267 /* can't increase priority */
6268 if (param->sched_priority > p->rt_priority &&
6269 param->sched_priority > rlim_rtprio)
6273 * Like positive nice levels, dont allow tasks to
6274 * move out of SCHED_IDLE either:
6276 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
6279 /* can't change other user's priorities */
6280 if (!check_same_owner(p))
6283 /* Normal users shall not reset the sched_reset_on_fork flag */
6284 if (p->sched_reset_on_fork && !reset_on_fork)
6289 #ifdef CONFIG_RT_GROUP_SCHED
6291 * Do not allow realtime tasks into groups that have no runtime
6294 if (rt_bandwidth_enabled() && rt_policy(policy) &&
6295 task_group(p)->rt_bandwidth.rt_runtime == 0)
6299 retval = security_task_setscheduler(p, policy, param);
6305 * make sure no PI-waiters arrive (or leave) while we are
6306 * changing the priority of the task:
6308 spin_lock_irqsave(&p->pi_lock, flags);
6310 * To be able to change p->policy safely, the apropriate
6311 * runqueue lock must be held.
6313 rq = __task_rq_lock(p);
6314 /* recheck policy now with rq lock held */
6315 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
6316 policy = oldpolicy = -1;
6317 __task_rq_unlock(rq);
6318 spin_unlock_irqrestore(&p->pi_lock, flags);
6321 update_rq_clock(rq);
6322 on_rq = p->se.on_rq;
6323 running = task_current(rq, p);
6325 deactivate_task(rq, p, 0);
6327 p->sched_class->put_prev_task(rq, p);
6329 p->sched_reset_on_fork = reset_on_fork;
6332 __setscheduler(rq, p, policy, param->sched_priority);
6335 p->sched_class->set_curr_task(rq);
6337 activate_task(rq, p, 0);
6339 check_class_changed(rq, p, prev_class, oldprio, running);
6341 __task_rq_unlock(rq);
6342 spin_unlock_irqrestore(&p->pi_lock, flags);
6344 rt_mutex_adjust_pi(p);
6350 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6351 * @p: the task in question.
6352 * @policy: new policy.
6353 * @param: structure containing the new RT priority.
6355 * NOTE that the task may be already dead.
6357 int sched_setscheduler(struct task_struct *p, int policy,
6358 struct sched_param *param)
6360 return __sched_setscheduler(p, policy, param, true);
6362 EXPORT_SYMBOL_GPL(sched_setscheduler);
6365 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6366 * @p: the task in question.
6367 * @policy: new policy.
6368 * @param: structure containing the new RT priority.
6370 * Just like sched_setscheduler, only don't bother checking if the
6371 * current context has permission. For example, this is needed in
6372 * stop_machine(): we create temporary high priority worker threads,
6373 * but our caller might not have that capability.
6375 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
6376 struct sched_param *param)
6378 return __sched_setscheduler(p, policy, param, false);
6382 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
6384 struct sched_param lparam;
6385 struct task_struct *p;
6388 if (!param || pid < 0)
6390 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
6395 p = find_process_by_pid(pid);
6397 retval = sched_setscheduler(p, policy, &lparam);
6404 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6405 * @pid: the pid in question.
6406 * @policy: new policy.
6407 * @param: structure containing the new RT priority.
6409 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
6410 struct sched_param __user *, param)
6412 /* negative values for policy are not valid */
6416 return do_sched_setscheduler(pid, policy, param);
6420 * sys_sched_setparam - set/change the RT priority of a thread
6421 * @pid: the pid in question.
6422 * @param: structure containing the new RT priority.
6424 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
6426 return do_sched_setscheduler(pid, -1, param);
6430 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6431 * @pid: the pid in question.
6433 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
6435 struct task_struct *p;
6442 read_lock(&tasklist_lock);
6443 p = find_process_by_pid(pid);
6445 retval = security_task_getscheduler(p);
6448 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
6450 read_unlock(&tasklist_lock);
6455 * sys_sched_getparam - get the RT priority of a thread
6456 * @pid: the pid in question.
6457 * @param: structure containing the RT priority.
6459 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
6461 struct sched_param lp;
6462 struct task_struct *p;
6465 if (!param || pid < 0)
6468 read_lock(&tasklist_lock);
6469 p = find_process_by_pid(pid);
6474 retval = security_task_getscheduler(p);
6478 lp.sched_priority = p->rt_priority;
6479 read_unlock(&tasklist_lock);
6482 * This one might sleep, we cannot do it with a spinlock held ...
6484 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
6489 read_unlock(&tasklist_lock);
6493 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
6495 cpumask_var_t cpus_allowed, new_mask;
6496 struct task_struct *p;
6500 read_lock(&tasklist_lock);
6502 p = find_process_by_pid(pid);
6504 read_unlock(&tasklist_lock);
6510 * It is not safe to call set_cpus_allowed with the
6511 * tasklist_lock held. We will bump the task_struct's
6512 * usage count and then drop tasklist_lock.
6515 read_unlock(&tasklist_lock);
6517 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6521 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
6523 goto out_free_cpus_allowed;
6526 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
6529 retval = security_task_setscheduler(p, 0, NULL);
6533 cpuset_cpus_allowed(p, cpus_allowed);
6534 cpumask_and(new_mask, in_mask, cpus_allowed);
6536 retval = set_cpus_allowed_ptr(p, new_mask);
6539 cpuset_cpus_allowed(p, cpus_allowed);
6540 if (!cpumask_subset(new_mask, cpus_allowed)) {
6542 * We must have raced with a concurrent cpuset
6543 * update. Just reset the cpus_allowed to the
6544 * cpuset's cpus_allowed
6546 cpumask_copy(new_mask, cpus_allowed);
6551 free_cpumask_var(new_mask);
6552 out_free_cpus_allowed:
6553 free_cpumask_var(cpus_allowed);
6560 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
6561 struct cpumask *new_mask)
6563 if (len < cpumask_size())
6564 cpumask_clear(new_mask);
6565 else if (len > cpumask_size())
6566 len = cpumask_size();
6568 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6572 * sys_sched_setaffinity - set the cpu affinity of a process
6573 * @pid: pid of the process
6574 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6575 * @user_mask_ptr: user-space pointer to the new cpu mask
6577 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6578 unsigned long __user *, user_mask_ptr)
6580 cpumask_var_t new_mask;
6583 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6586 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6588 retval = sched_setaffinity(pid, new_mask);
6589 free_cpumask_var(new_mask);
6593 long sched_getaffinity(pid_t pid, struct cpumask *mask)
6595 struct task_struct *p;
6599 read_lock(&tasklist_lock);
6602 p = find_process_by_pid(pid);
6606 retval = security_task_getscheduler(p);
6610 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
6613 read_unlock(&tasklist_lock);
6620 * sys_sched_getaffinity - get the cpu affinity of a process
6621 * @pid: pid of the process
6622 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6623 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6625 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6626 unsigned long __user *, user_mask_ptr)
6631 if (len < cpumask_size())
6634 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6637 ret = sched_getaffinity(pid, mask);
6639 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
6642 ret = cpumask_size();
6644 free_cpumask_var(mask);
6650 * sys_sched_yield - yield the current processor to other threads.
6652 * This function yields the current CPU to other tasks. If there are no
6653 * other threads running on this CPU then this function will return.
6655 SYSCALL_DEFINE0(sched_yield)
6657 struct rq *rq = this_rq_lock();
6659 schedstat_inc(rq, yld_count);
6660 current->sched_class->yield_task(rq);
6663 * Since we are going to call schedule() anyway, there's
6664 * no need to preempt or enable interrupts:
6666 __release(rq->lock);
6667 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
6668 _raw_spin_unlock(&rq->lock);
6669 preempt_enable_no_resched();
6676 static inline int should_resched(void)
6678 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
6681 static void __cond_resched(void)
6683 add_preempt_count(PREEMPT_ACTIVE);
6685 sub_preempt_count(PREEMPT_ACTIVE);
6688 int __sched _cond_resched(void)
6690 if (should_resched()) {
6696 EXPORT_SYMBOL(_cond_resched);
6699 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
6700 * call schedule, and on return reacquire the lock.
6702 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6703 * operations here to prevent schedule() from being called twice (once via
6704 * spin_unlock(), once by hand).
6706 int __cond_resched_lock(spinlock_t *lock)
6708 int resched = should_resched();
6711 lockdep_assert_held(lock);
6713 if (spin_needbreak(lock) || resched) {
6724 EXPORT_SYMBOL(__cond_resched_lock);
6726 int __sched __cond_resched_softirq(void)
6728 BUG_ON(!in_softirq());
6730 if (should_resched()) {
6738 EXPORT_SYMBOL(__cond_resched_softirq);
6741 * yield - yield the current processor to other threads.
6743 * This is a shortcut for kernel-space yielding - it marks the
6744 * thread runnable and calls sys_sched_yield().
6746 void __sched yield(void)
6748 set_current_state(TASK_RUNNING);
6751 EXPORT_SYMBOL(yield);
6754 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6755 * that process accounting knows that this is a task in IO wait state.
6757 void __sched io_schedule(void)
6759 struct rq *rq = raw_rq();
6761 delayacct_blkio_start();
6762 atomic_inc(&rq->nr_iowait);
6763 current->in_iowait = 1;
6765 current->in_iowait = 0;
6766 atomic_dec(&rq->nr_iowait);
6767 delayacct_blkio_end();
6769 EXPORT_SYMBOL(io_schedule);
6771 long __sched io_schedule_timeout(long timeout)
6773 struct rq *rq = raw_rq();
6776 delayacct_blkio_start();
6777 atomic_inc(&rq->nr_iowait);
6778 current->in_iowait = 1;
6779 ret = schedule_timeout(timeout);
6780 current->in_iowait = 0;
6781 atomic_dec(&rq->nr_iowait);
6782 delayacct_blkio_end();
6787 * sys_sched_get_priority_max - return maximum RT priority.
6788 * @policy: scheduling class.
6790 * this syscall returns the maximum rt_priority that can be used
6791 * by a given scheduling class.
6793 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6800 ret = MAX_USER_RT_PRIO-1;
6812 * sys_sched_get_priority_min - return minimum RT priority.
6813 * @policy: scheduling class.
6815 * this syscall returns the minimum rt_priority that can be used
6816 * by a given scheduling class.
6818 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6836 * sys_sched_rr_get_interval - return the default timeslice of a process.
6837 * @pid: pid of the process.
6838 * @interval: userspace pointer to the timeslice value.
6840 * this syscall writes the default timeslice value of a given process
6841 * into the user-space timespec buffer. A value of '0' means infinity.
6843 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6844 struct timespec __user *, interval)
6846 struct task_struct *p;
6847 unsigned int time_slice;
6855 read_lock(&tasklist_lock);
6856 p = find_process_by_pid(pid);
6860 retval = security_task_getscheduler(p);
6864 time_slice = p->sched_class->get_rr_interval(p);
6866 read_unlock(&tasklist_lock);
6867 jiffies_to_timespec(time_slice, &t);
6868 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
6872 read_unlock(&tasklist_lock);
6876 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
6878 void sched_show_task(struct task_struct *p)
6880 unsigned long free = 0;
6883 state = p->state ? __ffs(p->state) + 1 : 0;
6884 printk(KERN_INFO "%-13.13s %c", p->comm,
6885 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
6886 #if BITS_PER_LONG == 32
6887 if (state == TASK_RUNNING)
6888 printk(KERN_CONT " running ");
6890 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
6892 if (state == TASK_RUNNING)
6893 printk(KERN_CONT " running task ");
6895 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
6897 #ifdef CONFIG_DEBUG_STACK_USAGE
6898 free = stack_not_used(p);
6900 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
6901 task_pid_nr(p), task_pid_nr(p->real_parent),
6902 (unsigned long)task_thread_info(p)->flags);
6904 show_stack(p, NULL);
6907 void show_state_filter(unsigned long state_filter)
6909 struct task_struct *g, *p;
6911 #if BITS_PER_LONG == 32
6913 " task PC stack pid father\n");
6916 " task PC stack pid father\n");
6918 read_lock(&tasklist_lock);
6919 do_each_thread(g, p) {
6921 * reset the NMI-timeout, listing all files on a slow
6922 * console might take alot of time:
6924 touch_nmi_watchdog();
6925 if (!state_filter || (p->state & state_filter))
6927 } while_each_thread(g, p);
6929 touch_all_softlockup_watchdogs();
6931 #ifdef CONFIG_SCHED_DEBUG
6932 sysrq_sched_debug_show();
6934 read_unlock(&tasklist_lock);
6936 * Only show locks if all tasks are dumped:
6938 if (state_filter == -1)
6939 debug_show_all_locks();
6942 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6944 idle->sched_class = &idle_sched_class;
6948 * init_idle - set up an idle thread for a given CPU
6949 * @idle: task in question
6950 * @cpu: cpu the idle task belongs to
6952 * NOTE: this function does not set the idle thread's NEED_RESCHED
6953 * flag, to make booting more robust.
6955 void __cpuinit init_idle(struct task_struct *idle, int cpu)
6957 struct rq *rq = cpu_rq(cpu);
6958 unsigned long flags;
6960 spin_lock_irqsave(&rq->lock, flags);
6963 idle->se.exec_start = sched_clock();
6965 idle->prio = idle->normal_prio = MAX_PRIO;
6966 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
6967 __set_task_cpu(idle, cpu);
6969 rq->curr = rq->idle = idle;
6970 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6973 spin_unlock_irqrestore(&rq->lock, flags);
6975 /* Set the preempt count _outside_ the spinlocks! */
6976 #if defined(CONFIG_PREEMPT)
6977 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
6979 task_thread_info(idle)->preempt_count = 0;
6982 * The idle tasks have their own, simple scheduling class:
6984 idle->sched_class = &idle_sched_class;
6985 ftrace_graph_init_task(idle);
6989 * In a system that switches off the HZ timer nohz_cpu_mask
6990 * indicates which cpus entered this state. This is used
6991 * in the rcu update to wait only for active cpus. For system
6992 * which do not switch off the HZ timer nohz_cpu_mask should
6993 * always be CPU_BITS_NONE.
6995 cpumask_var_t nohz_cpu_mask;
6998 * Increase the granularity value when there are more CPUs,
6999 * because with more CPUs the 'effective latency' as visible
7000 * to users decreases. But the relationship is not linear,
7001 * so pick a second-best guess by going with the log2 of the
7004 * This idea comes from the SD scheduler of Con Kolivas:
7006 static inline void sched_init_granularity(void)
7008 unsigned int factor = 1 + ilog2(num_online_cpus());
7009 const unsigned long limit = 200000000;
7011 sysctl_sched_min_granularity *= factor;
7012 if (sysctl_sched_min_granularity > limit)
7013 sysctl_sched_min_granularity = limit;
7015 sysctl_sched_latency *= factor;
7016 if (sysctl_sched_latency > limit)
7017 sysctl_sched_latency = limit;
7019 sysctl_sched_wakeup_granularity *= factor;
7021 sysctl_sched_shares_ratelimit *= factor;
7026 * This is how migration works:
7028 * 1) we queue a struct migration_req structure in the source CPU's
7029 * runqueue and wake up that CPU's migration thread.
7030 * 2) we down() the locked semaphore => thread blocks.
7031 * 3) migration thread wakes up (implicitly it forces the migrated
7032 * thread off the CPU)
7033 * 4) it gets the migration request and checks whether the migrated
7034 * task is still in the wrong runqueue.
7035 * 5) if it's in the wrong runqueue then the migration thread removes
7036 * it and puts it into the right queue.
7037 * 6) migration thread up()s the semaphore.
7038 * 7) we wake up and the migration is done.
7042 * Change a given task's CPU affinity. Migrate the thread to a
7043 * proper CPU and schedule it away if the CPU it's executing on
7044 * is removed from the allowed bitmask.
7046 * NOTE: the caller must have a valid reference to the task, the
7047 * task must not exit() & deallocate itself prematurely. The
7048 * call is not atomic; no spinlocks may be held.
7050 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
7052 struct migration_req req;
7053 unsigned long flags;
7057 rq = task_rq_lock(p, &flags);
7058 if (!cpumask_intersects(new_mask, cpu_online_mask)) {
7063 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
7064 !cpumask_equal(&p->cpus_allowed, new_mask))) {
7069 if (p->sched_class->set_cpus_allowed)
7070 p->sched_class->set_cpus_allowed(p, new_mask);
7072 cpumask_copy(&p->cpus_allowed, new_mask);
7073 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
7076 /* Can the task run on the task's current CPU? If so, we're done */
7077 if (cpumask_test_cpu(task_cpu(p), new_mask))
7080 if (migrate_task(p, cpumask_any_and(cpu_online_mask, new_mask), &req)) {
7081 /* Need help from migration thread: drop lock and wait. */
7082 struct task_struct *mt = rq->migration_thread;
7084 get_task_struct(mt);
7085 task_rq_unlock(rq, &flags);
7086 wake_up_process(rq->migration_thread);
7087 put_task_struct(mt);
7088 wait_for_completion(&req.done);
7089 tlb_migrate_finish(p->mm);
7093 task_rq_unlock(rq, &flags);
7097 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
7100 * Move (not current) task off this cpu, onto dest cpu. We're doing
7101 * this because either it can't run here any more (set_cpus_allowed()
7102 * away from this CPU, or CPU going down), or because we're
7103 * attempting to rebalance this task on exec (sched_exec).
7105 * So we race with normal scheduler movements, but that's OK, as long
7106 * as the task is no longer on this CPU.
7108 * Returns non-zero if task was successfully migrated.
7110 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
7112 struct rq *rq_dest, *rq_src;
7115 if (unlikely(!cpu_active(dest_cpu)))
7118 rq_src = cpu_rq(src_cpu);
7119 rq_dest = cpu_rq(dest_cpu);
7121 double_rq_lock(rq_src, rq_dest);
7122 /* Already moved. */
7123 if (task_cpu(p) != src_cpu)
7125 /* Affinity changed (again). */
7126 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7129 on_rq = p->se.on_rq;
7131 deactivate_task(rq_src, p, 0);
7133 set_task_cpu(p, dest_cpu);
7135 activate_task(rq_dest, p, 0);
7136 check_preempt_curr(rq_dest, p, 0);
7141 double_rq_unlock(rq_src, rq_dest);
7145 #define RCU_MIGRATION_IDLE 0
7146 #define RCU_MIGRATION_NEED_QS 1
7147 #define RCU_MIGRATION_GOT_QS 2
7148 #define RCU_MIGRATION_MUST_SYNC 3
7151 * migration_thread - this is a highprio system thread that performs
7152 * thread migration by bumping thread off CPU then 'pushing' onto
7155 static int migration_thread(void *data)
7158 int cpu = (long)data;
7162 BUG_ON(rq->migration_thread != current);
7164 set_current_state(TASK_INTERRUPTIBLE);
7165 while (!kthread_should_stop()) {
7166 struct migration_req *req;
7167 struct list_head *head;
7169 spin_lock_irq(&rq->lock);
7171 if (cpu_is_offline(cpu)) {
7172 spin_unlock_irq(&rq->lock);
7176 if (rq->active_balance) {
7177 active_load_balance(rq, cpu);
7178 rq->active_balance = 0;
7181 head = &rq->migration_queue;
7183 if (list_empty(head)) {
7184 spin_unlock_irq(&rq->lock);
7186 set_current_state(TASK_INTERRUPTIBLE);
7189 req = list_entry(head->next, struct migration_req, list);
7190 list_del_init(head->next);
7192 if (req->task != NULL) {
7193 spin_unlock(&rq->lock);
7194 __migrate_task(req->task, cpu, req->dest_cpu);
7195 } else if (likely(cpu == (badcpu = smp_processor_id()))) {
7196 req->dest_cpu = RCU_MIGRATION_GOT_QS;
7197 spin_unlock(&rq->lock);
7199 req->dest_cpu = RCU_MIGRATION_MUST_SYNC;
7200 spin_unlock(&rq->lock);
7201 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu, cpu);
7205 complete(&req->done);
7207 __set_current_state(TASK_RUNNING);
7212 #ifdef CONFIG_HOTPLUG_CPU
7214 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
7218 local_irq_disable();
7219 ret = __migrate_task(p, src_cpu, dest_cpu);
7225 * Figure out where task on dead CPU should go, use force if necessary.
7227 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
7230 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(dead_cpu));
7233 /* Look for allowed, online CPU in same node. */
7234 for_each_cpu_and(dest_cpu, nodemask, cpu_online_mask)
7235 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7238 /* Any allowed, online CPU? */
7239 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_online_mask);
7240 if (dest_cpu < nr_cpu_ids)
7243 /* No more Mr. Nice Guy. */
7244 if (dest_cpu >= nr_cpu_ids) {
7245 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
7246 dest_cpu = cpumask_any_and(cpu_online_mask, &p->cpus_allowed);
7249 * Don't tell them about moving exiting tasks or
7250 * kernel threads (both mm NULL), since they never
7253 if (p->mm && printk_ratelimit()) {
7254 printk(KERN_INFO "process %d (%s) no "
7255 "longer affine to cpu%d\n",
7256 task_pid_nr(p), p->comm, dead_cpu);
7261 /* It can have affinity changed while we were choosing. */
7262 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
7267 * While a dead CPU has no uninterruptible tasks queued at this point,
7268 * it might still have a nonzero ->nr_uninterruptible counter, because
7269 * for performance reasons the counter is not stricly tracking tasks to
7270 * their home CPUs. So we just add the counter to another CPU's counter,
7271 * to keep the global sum constant after CPU-down:
7273 static void migrate_nr_uninterruptible(struct rq *rq_src)
7275 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_online_mask));
7276 unsigned long flags;
7278 local_irq_save(flags);
7279 double_rq_lock(rq_src, rq_dest);
7280 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
7281 rq_src->nr_uninterruptible = 0;
7282 double_rq_unlock(rq_src, rq_dest);
7283 local_irq_restore(flags);
7286 /* Run through task list and migrate tasks from the dead cpu. */
7287 static void migrate_live_tasks(int src_cpu)
7289 struct task_struct *p, *t;
7291 read_lock(&tasklist_lock);
7293 do_each_thread(t, p) {
7297 if (task_cpu(p) == src_cpu)
7298 move_task_off_dead_cpu(src_cpu, p);
7299 } while_each_thread(t, p);
7301 read_unlock(&tasklist_lock);
7305 * Schedules idle task to be the next runnable task on current CPU.
7306 * It does so by boosting its priority to highest possible.
7307 * Used by CPU offline code.
7309 void sched_idle_next(void)
7311 int this_cpu = smp_processor_id();
7312 struct rq *rq = cpu_rq(this_cpu);
7313 struct task_struct *p = rq->idle;
7314 unsigned long flags;
7316 /* cpu has to be offline */
7317 BUG_ON(cpu_online(this_cpu));
7320 * Strictly not necessary since rest of the CPUs are stopped by now
7321 * and interrupts disabled on the current cpu.
7323 spin_lock_irqsave(&rq->lock, flags);
7325 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7327 update_rq_clock(rq);
7328 activate_task(rq, p, 0);
7330 spin_unlock_irqrestore(&rq->lock, flags);
7334 * Ensures that the idle task is using init_mm right before its cpu goes
7337 void idle_task_exit(void)
7339 struct mm_struct *mm = current->active_mm;
7341 BUG_ON(cpu_online(smp_processor_id()));
7344 switch_mm(mm, &init_mm, current);
7348 /* called under rq->lock with disabled interrupts */
7349 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
7351 struct rq *rq = cpu_rq(dead_cpu);
7353 /* Must be exiting, otherwise would be on tasklist. */
7354 BUG_ON(!p->exit_state);
7356 /* Cannot have done final schedule yet: would have vanished. */
7357 BUG_ON(p->state == TASK_DEAD);
7362 * Drop lock around migration; if someone else moves it,
7363 * that's OK. No task can be added to this CPU, so iteration is
7366 spin_unlock_irq(&rq->lock);
7367 move_task_off_dead_cpu(dead_cpu, p);
7368 spin_lock_irq(&rq->lock);
7373 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7374 static void migrate_dead_tasks(unsigned int dead_cpu)
7376 struct rq *rq = cpu_rq(dead_cpu);
7377 struct task_struct *next;
7380 if (!rq->nr_running)
7382 update_rq_clock(rq);
7383 next = pick_next_task(rq);
7386 next->sched_class->put_prev_task(rq, next);
7387 migrate_dead(dead_cpu, next);
7393 * remove the tasks which were accounted by rq from calc_load_tasks.
7395 static void calc_global_load_remove(struct rq *rq)
7397 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
7398 rq->calc_load_active = 0;
7400 #endif /* CONFIG_HOTPLUG_CPU */
7402 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7404 static struct ctl_table sd_ctl_dir[] = {
7406 .procname = "sched_domain",
7412 static struct ctl_table sd_ctl_root[] = {
7414 .procname = "kernel",
7416 .child = sd_ctl_dir,
7421 static struct ctl_table *sd_alloc_ctl_entry(int n)
7423 struct ctl_table *entry =
7424 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
7429 static void sd_free_ctl_entry(struct ctl_table **tablep)
7431 struct ctl_table *entry;
7434 * In the intermediate directories, both the child directory and
7435 * procname are dynamically allocated and could fail but the mode
7436 * will always be set. In the lowest directory the names are
7437 * static strings and all have proc handlers.
7439 for (entry = *tablep; entry->mode; entry++) {
7441 sd_free_ctl_entry(&entry->child);
7442 if (entry->proc_handler == NULL)
7443 kfree(entry->procname);
7451 set_table_entry(struct ctl_table *entry,
7452 const char *procname, void *data, int maxlen,
7453 mode_t mode, proc_handler *proc_handler)
7455 entry->procname = procname;
7457 entry->maxlen = maxlen;
7459 entry->proc_handler = proc_handler;
7462 static struct ctl_table *
7463 sd_alloc_ctl_domain_table(struct sched_domain *sd)
7465 struct ctl_table *table = sd_alloc_ctl_entry(13);
7470 set_table_entry(&table[0], "min_interval", &sd->min_interval,
7471 sizeof(long), 0644, proc_doulongvec_minmax);
7472 set_table_entry(&table[1], "max_interval", &sd->max_interval,
7473 sizeof(long), 0644, proc_doulongvec_minmax);
7474 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
7475 sizeof(int), 0644, proc_dointvec_minmax);
7476 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
7477 sizeof(int), 0644, proc_dointvec_minmax);
7478 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
7479 sizeof(int), 0644, proc_dointvec_minmax);
7480 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
7481 sizeof(int), 0644, proc_dointvec_minmax);
7482 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
7483 sizeof(int), 0644, proc_dointvec_minmax);
7484 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
7485 sizeof(int), 0644, proc_dointvec_minmax);
7486 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
7487 sizeof(int), 0644, proc_dointvec_minmax);
7488 set_table_entry(&table[9], "cache_nice_tries",
7489 &sd->cache_nice_tries,
7490 sizeof(int), 0644, proc_dointvec_minmax);
7491 set_table_entry(&table[10], "flags", &sd->flags,
7492 sizeof(int), 0644, proc_dointvec_minmax);
7493 set_table_entry(&table[11], "name", sd->name,
7494 CORENAME_MAX_SIZE, 0444, proc_dostring);
7495 /* &table[12] is terminator */
7500 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
7502 struct ctl_table *entry, *table;
7503 struct sched_domain *sd;
7504 int domain_num = 0, i;
7507 for_each_domain(cpu, sd)
7509 entry = table = sd_alloc_ctl_entry(domain_num + 1);
7514 for_each_domain(cpu, sd) {
7515 snprintf(buf, 32, "domain%d", i);
7516 entry->procname = kstrdup(buf, GFP_KERNEL);
7518 entry->child = sd_alloc_ctl_domain_table(sd);
7525 static struct ctl_table_header *sd_sysctl_header;
7526 static void register_sched_domain_sysctl(void)
7528 int i, cpu_num = num_online_cpus();
7529 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
7532 WARN_ON(sd_ctl_dir[0].child);
7533 sd_ctl_dir[0].child = entry;
7538 for_each_online_cpu(i) {
7539 snprintf(buf, 32, "cpu%d", i);
7540 entry->procname = kstrdup(buf, GFP_KERNEL);
7542 entry->child = sd_alloc_ctl_cpu_table(i);
7546 WARN_ON(sd_sysctl_header);
7547 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
7550 /* may be called multiple times per register */
7551 static void unregister_sched_domain_sysctl(void)
7553 if (sd_sysctl_header)
7554 unregister_sysctl_table(sd_sysctl_header);
7555 sd_sysctl_header = NULL;
7556 if (sd_ctl_dir[0].child)
7557 sd_free_ctl_entry(&sd_ctl_dir[0].child);
7560 static void register_sched_domain_sysctl(void)
7563 static void unregister_sched_domain_sysctl(void)
7568 static void set_rq_online(struct rq *rq)
7571 const struct sched_class *class;
7573 cpumask_set_cpu(rq->cpu, rq->rd->online);
7576 for_each_class(class) {
7577 if (class->rq_online)
7578 class->rq_online(rq);
7583 static void set_rq_offline(struct rq *rq)
7586 const struct sched_class *class;
7588 for_each_class(class) {
7589 if (class->rq_offline)
7590 class->rq_offline(rq);
7593 cpumask_clear_cpu(rq->cpu, rq->rd->online);
7599 * migration_call - callback that gets triggered when a CPU is added.
7600 * Here we can start up the necessary migration thread for the new CPU.
7602 static int __cpuinit
7603 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
7605 struct task_struct *p;
7606 int cpu = (long)hcpu;
7607 unsigned long flags;
7612 case CPU_UP_PREPARE:
7613 case CPU_UP_PREPARE_FROZEN:
7614 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
7617 kthread_bind(p, cpu);
7618 /* Must be high prio: stop_machine expects to yield to it. */
7619 rq = task_rq_lock(p, &flags);
7620 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7621 task_rq_unlock(rq, &flags);
7623 cpu_rq(cpu)->migration_thread = p;
7624 rq->calc_load_update = calc_load_update;
7628 case CPU_ONLINE_FROZEN:
7629 /* Strictly unnecessary, as first user will wake it. */
7630 wake_up_process(cpu_rq(cpu)->migration_thread);
7632 /* Update our root-domain */
7634 spin_lock_irqsave(&rq->lock, flags);
7636 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7640 spin_unlock_irqrestore(&rq->lock, flags);
7643 #ifdef CONFIG_HOTPLUG_CPU
7644 case CPU_UP_CANCELED:
7645 case CPU_UP_CANCELED_FROZEN:
7646 if (!cpu_rq(cpu)->migration_thread)
7648 /* Unbind it from offline cpu so it can run. Fall thru. */
7649 kthread_bind(cpu_rq(cpu)->migration_thread,
7650 cpumask_any(cpu_online_mask));
7651 kthread_stop(cpu_rq(cpu)->migration_thread);
7652 put_task_struct(cpu_rq(cpu)->migration_thread);
7653 cpu_rq(cpu)->migration_thread = NULL;
7657 case CPU_DEAD_FROZEN:
7658 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7659 migrate_live_tasks(cpu);
7661 kthread_stop(rq->migration_thread);
7662 put_task_struct(rq->migration_thread);
7663 rq->migration_thread = NULL;
7664 /* Idle task back to normal (off runqueue, low prio) */
7665 spin_lock_irq(&rq->lock);
7666 update_rq_clock(rq);
7667 deactivate_task(rq, rq->idle, 0);
7668 rq->idle->static_prio = MAX_PRIO;
7669 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
7670 rq->idle->sched_class = &idle_sched_class;
7671 migrate_dead_tasks(cpu);
7672 spin_unlock_irq(&rq->lock);
7674 migrate_nr_uninterruptible(rq);
7675 BUG_ON(rq->nr_running != 0);
7676 calc_global_load_remove(rq);
7678 * No need to migrate the tasks: it was best-effort if
7679 * they didn't take sched_hotcpu_mutex. Just wake up
7682 spin_lock_irq(&rq->lock);
7683 while (!list_empty(&rq->migration_queue)) {
7684 struct migration_req *req;
7686 req = list_entry(rq->migration_queue.next,
7687 struct migration_req, list);
7688 list_del_init(&req->list);
7689 spin_unlock_irq(&rq->lock);
7690 complete(&req->done);
7691 spin_lock_irq(&rq->lock);
7693 spin_unlock_irq(&rq->lock);
7697 case CPU_DYING_FROZEN:
7698 /* Update our root-domain */
7700 spin_lock_irqsave(&rq->lock, flags);
7702 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7705 spin_unlock_irqrestore(&rq->lock, flags);
7713 * Register at high priority so that task migration (migrate_all_tasks)
7714 * happens before everything else. This has to be lower priority than
7715 * the notifier in the perf_event subsystem, though.
7717 static struct notifier_block __cpuinitdata migration_notifier = {
7718 .notifier_call = migration_call,
7722 static int __init migration_init(void)
7724 void *cpu = (void *)(long)smp_processor_id();
7727 /* Start one for the boot CPU: */
7728 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
7729 BUG_ON(err == NOTIFY_BAD);
7730 migration_call(&migration_notifier, CPU_ONLINE, cpu);
7731 register_cpu_notifier(&migration_notifier);
7735 early_initcall(migration_init);
7740 #ifdef CONFIG_SCHED_DEBUG
7742 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
7743 struct cpumask *groupmask)
7745 struct sched_group *group = sd->groups;
7748 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
7749 cpumask_clear(groupmask);
7751 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
7753 if (!(sd->flags & SD_LOAD_BALANCE)) {
7754 printk("does not load-balance\n");
7756 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
7761 printk(KERN_CONT "span %s level %s\n", str, sd->name);
7763 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
7764 printk(KERN_ERR "ERROR: domain->span does not contain "
7767 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
7768 printk(KERN_ERR "ERROR: domain->groups does not contain"
7772 printk(KERN_DEBUG "%*s groups:", level + 1, "");
7776 printk(KERN_ERR "ERROR: group is NULL\n");
7780 if (!group->cpu_power) {
7781 printk(KERN_CONT "\n");
7782 printk(KERN_ERR "ERROR: domain->cpu_power not "
7787 if (!cpumask_weight(sched_group_cpus(group))) {
7788 printk(KERN_CONT "\n");
7789 printk(KERN_ERR "ERROR: empty group\n");
7793 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
7794 printk(KERN_CONT "\n");
7795 printk(KERN_ERR "ERROR: repeated CPUs\n");
7799 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
7801 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
7803 printk(KERN_CONT " %s", str);
7804 if (group->cpu_power != SCHED_LOAD_SCALE) {
7805 printk(KERN_CONT " (cpu_power = %d)",
7809 group = group->next;
7810 } while (group != sd->groups);
7811 printk(KERN_CONT "\n");
7813 if (!cpumask_equal(sched_domain_span(sd), groupmask))
7814 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
7817 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
7818 printk(KERN_ERR "ERROR: parent span is not a superset "
7819 "of domain->span\n");
7823 static void sched_domain_debug(struct sched_domain *sd, int cpu)
7825 cpumask_var_t groupmask;
7829 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7833 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7835 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
7836 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
7841 if (sched_domain_debug_one(sd, cpu, level, groupmask))
7848 free_cpumask_var(groupmask);
7850 #else /* !CONFIG_SCHED_DEBUG */
7851 # define sched_domain_debug(sd, cpu) do { } while (0)
7852 #endif /* CONFIG_SCHED_DEBUG */
7854 static int sd_degenerate(struct sched_domain *sd)
7856 if (cpumask_weight(sched_domain_span(sd)) == 1)
7859 /* Following flags need at least 2 groups */
7860 if (sd->flags & (SD_LOAD_BALANCE |
7861 SD_BALANCE_NEWIDLE |
7865 SD_SHARE_PKG_RESOURCES)) {
7866 if (sd->groups != sd->groups->next)
7870 /* Following flags don't use groups */
7871 if (sd->flags & (SD_WAKE_AFFINE))
7878 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
7880 unsigned long cflags = sd->flags, pflags = parent->flags;
7882 if (sd_degenerate(parent))
7885 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
7888 /* Flags needing groups don't count if only 1 group in parent */
7889 if (parent->groups == parent->groups->next) {
7890 pflags &= ~(SD_LOAD_BALANCE |
7891 SD_BALANCE_NEWIDLE |
7895 SD_SHARE_PKG_RESOURCES);
7896 if (nr_node_ids == 1)
7897 pflags &= ~SD_SERIALIZE;
7899 if (~cflags & pflags)
7905 static void free_rootdomain(struct root_domain *rd)
7907 cpupri_cleanup(&rd->cpupri);
7909 free_cpumask_var(rd->rto_mask);
7910 free_cpumask_var(rd->online);
7911 free_cpumask_var(rd->span);
7915 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7917 struct root_domain *old_rd = NULL;
7918 unsigned long flags;
7920 spin_lock_irqsave(&rq->lock, flags);
7925 if (cpumask_test_cpu(rq->cpu, old_rd->online))
7928 cpumask_clear_cpu(rq->cpu, old_rd->span);
7931 * If we dont want to free the old_rt yet then
7932 * set old_rd to NULL to skip the freeing later
7935 if (!atomic_dec_and_test(&old_rd->refcount))
7939 atomic_inc(&rd->refcount);
7942 cpumask_set_cpu(rq->cpu, rd->span);
7943 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
7946 spin_unlock_irqrestore(&rq->lock, flags);
7949 free_rootdomain(old_rd);
7952 static int init_rootdomain(struct root_domain *rd, bool bootmem)
7954 gfp_t gfp = GFP_KERNEL;
7956 memset(rd, 0, sizeof(*rd));
7961 if (!alloc_cpumask_var(&rd->span, gfp))
7963 if (!alloc_cpumask_var(&rd->online, gfp))
7965 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
7968 if (cpupri_init(&rd->cpupri, bootmem) != 0)
7973 free_cpumask_var(rd->rto_mask);
7975 free_cpumask_var(rd->online);
7977 free_cpumask_var(rd->span);
7982 static void init_defrootdomain(void)
7984 init_rootdomain(&def_root_domain, true);
7986 atomic_set(&def_root_domain.refcount, 1);
7989 static struct root_domain *alloc_rootdomain(void)
7991 struct root_domain *rd;
7993 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
7997 if (init_rootdomain(rd, false) != 0) {
8006 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
8007 * hold the hotplug lock.
8010 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
8012 struct rq *rq = cpu_rq(cpu);
8013 struct sched_domain *tmp;
8015 /* Remove the sched domains which do not contribute to scheduling. */
8016 for (tmp = sd; tmp; ) {
8017 struct sched_domain *parent = tmp->parent;
8021 if (sd_parent_degenerate(tmp, parent)) {
8022 tmp->parent = parent->parent;
8024 parent->parent->child = tmp;
8029 if (sd && sd_degenerate(sd)) {
8035 sched_domain_debug(sd, cpu);
8037 rq_attach_root(rq, rd);
8038 rcu_assign_pointer(rq->sd, sd);
8041 /* cpus with isolated domains */
8042 static cpumask_var_t cpu_isolated_map;
8044 /* Setup the mask of cpus configured for isolated domains */
8045 static int __init isolated_cpu_setup(char *str)
8047 cpulist_parse(str, cpu_isolated_map);
8051 __setup("isolcpus=", isolated_cpu_setup);
8054 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
8055 * to a function which identifies what group(along with sched group) a CPU
8056 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
8057 * (due to the fact that we keep track of groups covered with a struct cpumask).
8059 * init_sched_build_groups will build a circular linked list of the groups
8060 * covered by the given span, and will set each group's ->cpumask correctly,
8061 * and ->cpu_power to 0.
8064 init_sched_build_groups(const struct cpumask *span,
8065 const struct cpumask *cpu_map,
8066 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
8067 struct sched_group **sg,
8068 struct cpumask *tmpmask),
8069 struct cpumask *covered, struct cpumask *tmpmask)
8071 struct sched_group *first = NULL, *last = NULL;
8074 cpumask_clear(covered);
8076 for_each_cpu(i, span) {
8077 struct sched_group *sg;
8078 int group = group_fn(i, cpu_map, &sg, tmpmask);
8081 if (cpumask_test_cpu(i, covered))
8084 cpumask_clear(sched_group_cpus(sg));
8087 for_each_cpu(j, span) {
8088 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
8091 cpumask_set_cpu(j, covered);
8092 cpumask_set_cpu(j, sched_group_cpus(sg));
8103 #define SD_NODES_PER_DOMAIN 16
8108 * find_next_best_node - find the next node to include in a sched_domain
8109 * @node: node whose sched_domain we're building
8110 * @used_nodes: nodes already in the sched_domain
8112 * Find the next node to include in a given scheduling domain. Simply
8113 * finds the closest node not already in the @used_nodes map.
8115 * Should use nodemask_t.
8117 static int find_next_best_node(int node, nodemask_t *used_nodes)
8119 int i, n, val, min_val, best_node = 0;
8123 for (i = 0; i < nr_node_ids; i++) {
8124 /* Start at @node */
8125 n = (node + i) % nr_node_ids;
8127 if (!nr_cpus_node(n))
8130 /* Skip already used nodes */
8131 if (node_isset(n, *used_nodes))
8134 /* Simple min distance search */
8135 val = node_distance(node, n);
8137 if (val < min_val) {
8143 node_set(best_node, *used_nodes);
8148 * sched_domain_node_span - get a cpumask for a node's sched_domain
8149 * @node: node whose cpumask we're constructing
8150 * @span: resulting cpumask
8152 * Given a node, construct a good cpumask for its sched_domain to span. It
8153 * should be one that prevents unnecessary balancing, but also spreads tasks
8156 static void sched_domain_node_span(int node, struct cpumask *span)
8158 nodemask_t used_nodes;
8161 cpumask_clear(span);
8162 nodes_clear(used_nodes);
8164 cpumask_or(span, span, cpumask_of_node(node));
8165 node_set(node, used_nodes);
8167 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
8168 int next_node = find_next_best_node(node, &used_nodes);
8170 cpumask_or(span, span, cpumask_of_node(next_node));
8173 #endif /* CONFIG_NUMA */
8175 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
8178 * The cpus mask in sched_group and sched_domain hangs off the end.
8180 * ( See the the comments in include/linux/sched.h:struct sched_group
8181 * and struct sched_domain. )
8183 struct static_sched_group {
8184 struct sched_group sg;
8185 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
8188 struct static_sched_domain {
8189 struct sched_domain sd;
8190 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
8196 cpumask_var_t domainspan;
8197 cpumask_var_t covered;
8198 cpumask_var_t notcovered;
8200 cpumask_var_t nodemask;
8201 cpumask_var_t this_sibling_map;
8202 cpumask_var_t this_core_map;
8203 cpumask_var_t send_covered;
8204 cpumask_var_t tmpmask;
8205 struct sched_group **sched_group_nodes;
8206 struct root_domain *rd;
8210 sa_sched_groups = 0,
8215 sa_this_sibling_map,
8217 sa_sched_group_nodes,
8227 * SMT sched-domains:
8229 #ifdef CONFIG_SCHED_SMT
8230 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
8231 static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
8234 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
8235 struct sched_group **sg, struct cpumask *unused)
8238 *sg = &per_cpu(sched_group_cpus, cpu).sg;
8241 #endif /* CONFIG_SCHED_SMT */
8244 * multi-core sched-domains:
8246 #ifdef CONFIG_SCHED_MC
8247 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
8248 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
8249 #endif /* CONFIG_SCHED_MC */
8251 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8253 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8254 struct sched_group **sg, struct cpumask *mask)
8258 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8259 group = cpumask_first(mask);
8261 *sg = &per_cpu(sched_group_core, group).sg;
8264 #elif defined(CONFIG_SCHED_MC)
8266 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8267 struct sched_group **sg, struct cpumask *unused)
8270 *sg = &per_cpu(sched_group_core, cpu).sg;
8275 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
8276 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
8279 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
8280 struct sched_group **sg, struct cpumask *mask)
8283 #ifdef CONFIG_SCHED_MC
8284 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
8285 group = cpumask_first(mask);
8286 #elif defined(CONFIG_SCHED_SMT)
8287 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8288 group = cpumask_first(mask);
8293 *sg = &per_cpu(sched_group_phys, group).sg;
8299 * The init_sched_build_groups can't handle what we want to do with node
8300 * groups, so roll our own. Now each node has its own list of groups which
8301 * gets dynamically allocated.
8303 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
8304 static struct sched_group ***sched_group_nodes_bycpu;
8306 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
8307 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
8309 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
8310 struct sched_group **sg,
8311 struct cpumask *nodemask)
8315 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
8316 group = cpumask_first(nodemask);
8319 *sg = &per_cpu(sched_group_allnodes, group).sg;
8323 static void init_numa_sched_groups_power(struct sched_group *group_head)
8325 struct sched_group *sg = group_head;
8331 for_each_cpu(j, sched_group_cpus(sg)) {
8332 struct sched_domain *sd;
8334 sd = &per_cpu(phys_domains, j).sd;
8335 if (j != group_first_cpu(sd->groups)) {
8337 * Only add "power" once for each
8343 sg->cpu_power += sd->groups->cpu_power;
8346 } while (sg != group_head);
8349 static int build_numa_sched_groups(struct s_data *d,
8350 const struct cpumask *cpu_map, int num)
8352 struct sched_domain *sd;
8353 struct sched_group *sg, *prev;
8356 cpumask_clear(d->covered);
8357 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
8358 if (cpumask_empty(d->nodemask)) {
8359 d->sched_group_nodes[num] = NULL;
8363 sched_domain_node_span(num, d->domainspan);
8364 cpumask_and(d->domainspan, d->domainspan, cpu_map);
8366 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8369 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
8373 d->sched_group_nodes[num] = sg;
8375 for_each_cpu(j, d->nodemask) {
8376 sd = &per_cpu(node_domains, j).sd;
8381 cpumask_copy(sched_group_cpus(sg), d->nodemask);
8383 cpumask_or(d->covered, d->covered, d->nodemask);
8386 for (j = 0; j < nr_node_ids; j++) {
8387 n = (num + j) % nr_node_ids;
8388 cpumask_complement(d->notcovered, d->covered);
8389 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
8390 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
8391 if (cpumask_empty(d->tmpmask))
8393 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
8394 if (cpumask_empty(d->tmpmask))
8396 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8400 "Can not alloc domain group for node %d\n", j);
8404 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
8405 sg->next = prev->next;
8406 cpumask_or(d->covered, d->covered, d->tmpmask);
8413 #endif /* CONFIG_NUMA */
8416 /* Free memory allocated for various sched_group structures */
8417 static void free_sched_groups(const struct cpumask *cpu_map,
8418 struct cpumask *nodemask)
8422 for_each_cpu(cpu, cpu_map) {
8423 struct sched_group **sched_group_nodes
8424 = sched_group_nodes_bycpu[cpu];
8426 if (!sched_group_nodes)
8429 for (i = 0; i < nr_node_ids; i++) {
8430 struct sched_group *oldsg, *sg = sched_group_nodes[i];
8432 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8433 if (cpumask_empty(nodemask))
8443 if (oldsg != sched_group_nodes[i])
8446 kfree(sched_group_nodes);
8447 sched_group_nodes_bycpu[cpu] = NULL;
8450 #else /* !CONFIG_NUMA */
8451 static void free_sched_groups(const struct cpumask *cpu_map,
8452 struct cpumask *nodemask)
8455 #endif /* CONFIG_NUMA */
8458 * Initialize sched groups cpu_power.
8460 * cpu_power indicates the capacity of sched group, which is used while
8461 * distributing the load between different sched groups in a sched domain.
8462 * Typically cpu_power for all the groups in a sched domain will be same unless
8463 * there are asymmetries in the topology. If there are asymmetries, group
8464 * having more cpu_power will pickup more load compared to the group having
8467 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
8469 struct sched_domain *child;
8470 struct sched_group *group;
8474 WARN_ON(!sd || !sd->groups);
8476 if (cpu != group_first_cpu(sd->groups))
8481 sd->groups->cpu_power = 0;
8484 power = SCHED_LOAD_SCALE;
8485 weight = cpumask_weight(sched_domain_span(sd));
8487 * SMT siblings share the power of a single core.
8488 * Usually multiple threads get a better yield out of
8489 * that one core than a single thread would have,
8490 * reflect that in sd->smt_gain.
8492 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
8493 power *= sd->smt_gain;
8495 power >>= SCHED_LOAD_SHIFT;
8497 sd->groups->cpu_power += power;
8502 * Add cpu_power of each child group to this groups cpu_power.
8504 group = child->groups;
8506 sd->groups->cpu_power += group->cpu_power;
8507 group = group->next;
8508 } while (group != child->groups);
8512 * Initializers for schedule domains
8513 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8516 #ifdef CONFIG_SCHED_DEBUG
8517 # define SD_INIT_NAME(sd, type) sd->name = #type
8519 # define SD_INIT_NAME(sd, type) do { } while (0)
8522 #define SD_INIT(sd, type) sd_init_##type(sd)
8524 #define SD_INIT_FUNC(type) \
8525 static noinline void sd_init_##type(struct sched_domain *sd) \
8527 memset(sd, 0, sizeof(*sd)); \
8528 *sd = SD_##type##_INIT; \
8529 sd->level = SD_LV_##type; \
8530 SD_INIT_NAME(sd, type); \
8535 SD_INIT_FUNC(ALLNODES)
8538 #ifdef CONFIG_SCHED_SMT
8539 SD_INIT_FUNC(SIBLING)
8541 #ifdef CONFIG_SCHED_MC
8545 static int default_relax_domain_level = -1;
8547 static int __init setup_relax_domain_level(char *str)
8551 val = simple_strtoul(str, NULL, 0);
8552 if (val < SD_LV_MAX)
8553 default_relax_domain_level = val;
8557 __setup("relax_domain_level=", setup_relax_domain_level);
8559 static void set_domain_attribute(struct sched_domain *sd,
8560 struct sched_domain_attr *attr)
8564 if (!attr || attr->relax_domain_level < 0) {
8565 if (default_relax_domain_level < 0)
8568 request = default_relax_domain_level;
8570 request = attr->relax_domain_level;
8571 if (request < sd->level) {
8572 /* turn off idle balance on this domain */
8573 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
8575 /* turn on idle balance on this domain */
8576 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
8580 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
8581 const struct cpumask *cpu_map)
8584 case sa_sched_groups:
8585 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
8586 d->sched_group_nodes = NULL;
8588 free_rootdomain(d->rd); /* fall through */
8590 free_cpumask_var(d->tmpmask); /* fall through */
8591 case sa_send_covered:
8592 free_cpumask_var(d->send_covered); /* fall through */
8593 case sa_this_core_map:
8594 free_cpumask_var(d->this_core_map); /* fall through */
8595 case sa_this_sibling_map:
8596 free_cpumask_var(d->this_sibling_map); /* fall through */
8598 free_cpumask_var(d->nodemask); /* fall through */
8599 case sa_sched_group_nodes:
8601 kfree(d->sched_group_nodes); /* fall through */
8603 free_cpumask_var(d->notcovered); /* fall through */
8605 free_cpumask_var(d->covered); /* fall through */
8607 free_cpumask_var(d->domainspan); /* fall through */
8614 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
8615 const struct cpumask *cpu_map)
8618 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
8620 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
8621 return sa_domainspan;
8622 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
8624 /* Allocate the per-node list of sched groups */
8625 d->sched_group_nodes = kcalloc(nr_node_ids,
8626 sizeof(struct sched_group *), GFP_KERNEL);
8627 if (!d->sched_group_nodes) {
8628 printk(KERN_WARNING "Can not alloc sched group node list\n");
8629 return sa_notcovered;
8631 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
8633 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
8634 return sa_sched_group_nodes;
8635 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
8637 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
8638 return sa_this_sibling_map;
8639 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
8640 return sa_this_core_map;
8641 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
8642 return sa_send_covered;
8643 d->rd = alloc_rootdomain();
8645 printk(KERN_WARNING "Cannot alloc root domain\n");
8648 return sa_rootdomain;
8651 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
8652 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
8654 struct sched_domain *sd = NULL;
8656 struct sched_domain *parent;
8659 if (cpumask_weight(cpu_map) >
8660 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
8661 sd = &per_cpu(allnodes_domains, i).sd;
8662 SD_INIT(sd, ALLNODES);
8663 set_domain_attribute(sd, attr);
8664 cpumask_copy(sched_domain_span(sd), cpu_map);
8665 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
8670 sd = &per_cpu(node_domains, i).sd;
8672 set_domain_attribute(sd, attr);
8673 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
8674 sd->parent = parent;
8677 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
8682 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
8683 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8684 struct sched_domain *parent, int i)
8686 struct sched_domain *sd;
8687 sd = &per_cpu(phys_domains, i).sd;
8689 set_domain_attribute(sd, attr);
8690 cpumask_copy(sched_domain_span(sd), d->nodemask);
8691 sd->parent = parent;
8694 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
8698 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
8699 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8700 struct sched_domain *parent, int i)
8702 struct sched_domain *sd = parent;
8703 #ifdef CONFIG_SCHED_MC
8704 sd = &per_cpu(core_domains, i).sd;
8706 set_domain_attribute(sd, attr);
8707 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
8708 sd->parent = parent;
8710 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
8715 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
8716 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8717 struct sched_domain *parent, int i)
8719 struct sched_domain *sd = parent;
8720 #ifdef CONFIG_SCHED_SMT
8721 sd = &per_cpu(cpu_domains, i).sd;
8722 SD_INIT(sd, SIBLING);
8723 set_domain_attribute(sd, attr);
8724 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
8725 sd->parent = parent;
8727 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
8732 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
8733 const struct cpumask *cpu_map, int cpu)
8736 #ifdef CONFIG_SCHED_SMT
8737 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
8738 cpumask_and(d->this_sibling_map, cpu_map,
8739 topology_thread_cpumask(cpu));
8740 if (cpu == cpumask_first(d->this_sibling_map))
8741 init_sched_build_groups(d->this_sibling_map, cpu_map,
8743 d->send_covered, d->tmpmask);
8746 #ifdef CONFIG_SCHED_MC
8747 case SD_LV_MC: /* set up multi-core groups */
8748 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
8749 if (cpu == cpumask_first(d->this_core_map))
8750 init_sched_build_groups(d->this_core_map, cpu_map,
8752 d->send_covered, d->tmpmask);
8755 case SD_LV_CPU: /* set up physical groups */
8756 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
8757 if (!cpumask_empty(d->nodemask))
8758 init_sched_build_groups(d->nodemask, cpu_map,
8760 d->send_covered, d->tmpmask);
8763 case SD_LV_ALLNODES:
8764 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
8765 d->send_covered, d->tmpmask);
8774 * Build sched domains for a given set of cpus and attach the sched domains
8775 * to the individual cpus
8777 static int __build_sched_domains(const struct cpumask *cpu_map,
8778 struct sched_domain_attr *attr)
8780 enum s_alloc alloc_state = sa_none;
8782 struct sched_domain *sd;
8788 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
8789 if (alloc_state != sa_rootdomain)
8791 alloc_state = sa_sched_groups;
8794 * Set up domains for cpus specified by the cpu_map.
8796 for_each_cpu(i, cpu_map) {
8797 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
8800 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
8801 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
8802 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
8803 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
8806 for_each_cpu(i, cpu_map) {
8807 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
8808 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
8811 /* Set up physical groups */
8812 for (i = 0; i < nr_node_ids; i++)
8813 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
8816 /* Set up node groups */
8818 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
8820 for (i = 0; i < nr_node_ids; i++)
8821 if (build_numa_sched_groups(&d, cpu_map, i))
8825 /* Calculate CPU power for physical packages and nodes */
8826 #ifdef CONFIG_SCHED_SMT
8827 for_each_cpu(i, cpu_map) {
8828 sd = &per_cpu(cpu_domains, i).sd;
8829 init_sched_groups_power(i, sd);
8832 #ifdef CONFIG_SCHED_MC
8833 for_each_cpu(i, cpu_map) {
8834 sd = &per_cpu(core_domains, i).sd;
8835 init_sched_groups_power(i, sd);
8839 for_each_cpu(i, cpu_map) {
8840 sd = &per_cpu(phys_domains, i).sd;
8841 init_sched_groups_power(i, sd);
8845 for (i = 0; i < nr_node_ids; i++)
8846 init_numa_sched_groups_power(d.sched_group_nodes[i]);
8848 if (d.sd_allnodes) {
8849 struct sched_group *sg;
8851 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
8853 init_numa_sched_groups_power(sg);
8857 /* Attach the domains */
8858 for_each_cpu(i, cpu_map) {
8859 #ifdef CONFIG_SCHED_SMT
8860 sd = &per_cpu(cpu_domains, i).sd;
8861 #elif defined(CONFIG_SCHED_MC)
8862 sd = &per_cpu(core_domains, i).sd;
8864 sd = &per_cpu(phys_domains, i).sd;
8866 cpu_attach_domain(sd, d.rd, i);
8869 d.sched_group_nodes = NULL; /* don't free this we still need it */
8870 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
8874 __free_domain_allocs(&d, alloc_state, cpu_map);
8878 static int build_sched_domains(const struct cpumask *cpu_map)
8880 return __build_sched_domains(cpu_map, NULL);
8883 static struct cpumask *doms_cur; /* current sched domains */
8884 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
8885 static struct sched_domain_attr *dattr_cur;
8886 /* attribues of custom domains in 'doms_cur' */
8889 * Special case: If a kmalloc of a doms_cur partition (array of
8890 * cpumask) fails, then fallback to a single sched domain,
8891 * as determined by the single cpumask fallback_doms.
8893 static cpumask_var_t fallback_doms;
8896 * arch_update_cpu_topology lets virtualized architectures update the
8897 * cpu core maps. It is supposed to return 1 if the topology changed
8898 * or 0 if it stayed the same.
8900 int __attribute__((weak)) arch_update_cpu_topology(void)
8906 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
8907 * For now this just excludes isolated cpus, but could be used to
8908 * exclude other special cases in the future.
8910 static int arch_init_sched_domains(const struct cpumask *cpu_map)
8914 arch_update_cpu_topology();
8916 doms_cur = kmalloc(cpumask_size(), GFP_KERNEL);
8918 doms_cur = fallback_doms;
8919 cpumask_andnot(doms_cur, cpu_map, cpu_isolated_map);
8921 err = build_sched_domains(doms_cur);
8922 register_sched_domain_sysctl();
8927 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
8928 struct cpumask *tmpmask)
8930 free_sched_groups(cpu_map, tmpmask);
8934 * Detach sched domains from a group of cpus specified in cpu_map
8935 * These cpus will now be attached to the NULL domain
8937 static void detach_destroy_domains(const struct cpumask *cpu_map)
8939 /* Save because hotplug lock held. */
8940 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
8943 for_each_cpu(i, cpu_map)
8944 cpu_attach_domain(NULL, &def_root_domain, i);
8945 synchronize_sched();
8946 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
8949 /* handle null as "default" */
8950 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
8951 struct sched_domain_attr *new, int idx_new)
8953 struct sched_domain_attr tmp;
8960 return !memcmp(cur ? (cur + idx_cur) : &tmp,
8961 new ? (new + idx_new) : &tmp,
8962 sizeof(struct sched_domain_attr));
8966 * Partition sched domains as specified by the 'ndoms_new'
8967 * cpumasks in the array doms_new[] of cpumasks. This compares
8968 * doms_new[] to the current sched domain partitioning, doms_cur[].
8969 * It destroys each deleted domain and builds each new domain.
8971 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
8972 * The masks don't intersect (don't overlap.) We should setup one
8973 * sched domain for each mask. CPUs not in any of the cpumasks will
8974 * not be load balanced. If the same cpumask appears both in the
8975 * current 'doms_cur' domains and in the new 'doms_new', we can leave
8978 * The passed in 'doms_new' should be kmalloc'd. This routine takes
8979 * ownership of it and will kfree it when done with it. If the caller
8980 * failed the kmalloc call, then it can pass in doms_new == NULL &&
8981 * ndoms_new == 1, and partition_sched_domains() will fallback to
8982 * the single partition 'fallback_doms', it also forces the domains
8985 * If doms_new == NULL it will be replaced with cpu_online_mask.
8986 * ndoms_new == 0 is a special case for destroying existing domains,
8987 * and it will not create the default domain.
8989 * Call with hotplug lock held
8991 /* FIXME: Change to struct cpumask *doms_new[] */
8992 void partition_sched_domains(int ndoms_new, struct cpumask *doms_new,
8993 struct sched_domain_attr *dattr_new)
8998 mutex_lock(&sched_domains_mutex);
9000 /* always unregister in case we don't destroy any domains */
9001 unregister_sched_domain_sysctl();
9003 /* Let architecture update cpu core mappings. */
9004 new_topology = arch_update_cpu_topology();
9006 n = doms_new ? ndoms_new : 0;
9008 /* Destroy deleted domains */
9009 for (i = 0; i < ndoms_cur; i++) {
9010 for (j = 0; j < n && !new_topology; j++) {
9011 if (cpumask_equal(&doms_cur[i], &doms_new[j])
9012 && dattrs_equal(dattr_cur, i, dattr_new, j))
9015 /* no match - a current sched domain not in new doms_new[] */
9016 detach_destroy_domains(doms_cur + i);
9021 if (doms_new == NULL) {
9023 doms_new = fallback_doms;
9024 cpumask_andnot(&doms_new[0], cpu_online_mask, cpu_isolated_map);
9025 WARN_ON_ONCE(dattr_new);
9028 /* Build new domains */
9029 for (i = 0; i < ndoms_new; i++) {
9030 for (j = 0; j < ndoms_cur && !new_topology; j++) {
9031 if (cpumask_equal(&doms_new[i], &doms_cur[j])
9032 && dattrs_equal(dattr_new, i, dattr_cur, j))
9035 /* no match - add a new doms_new */
9036 __build_sched_domains(doms_new + i,
9037 dattr_new ? dattr_new + i : NULL);
9042 /* Remember the new sched domains */
9043 if (doms_cur != fallback_doms)
9045 kfree(dattr_cur); /* kfree(NULL) is safe */
9046 doms_cur = doms_new;
9047 dattr_cur = dattr_new;
9048 ndoms_cur = ndoms_new;
9050 register_sched_domain_sysctl();
9052 mutex_unlock(&sched_domains_mutex);
9055 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
9056 static void arch_reinit_sched_domains(void)
9060 /* Destroy domains first to force the rebuild */
9061 partition_sched_domains(0, NULL, NULL);
9063 rebuild_sched_domains();
9067 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
9069 unsigned int level = 0;
9071 if (sscanf(buf, "%u", &level) != 1)
9075 * level is always be positive so don't check for
9076 * level < POWERSAVINGS_BALANCE_NONE which is 0
9077 * What happens on 0 or 1 byte write,
9078 * need to check for count as well?
9081 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
9085 sched_smt_power_savings = level;
9087 sched_mc_power_savings = level;
9089 arch_reinit_sched_domains();
9094 #ifdef CONFIG_SCHED_MC
9095 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
9098 return sprintf(page, "%u\n", sched_mc_power_savings);
9100 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
9101 const char *buf, size_t count)
9103 return sched_power_savings_store(buf, count, 0);
9105 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
9106 sched_mc_power_savings_show,
9107 sched_mc_power_savings_store);
9110 #ifdef CONFIG_SCHED_SMT
9111 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
9114 return sprintf(page, "%u\n", sched_smt_power_savings);
9116 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
9117 const char *buf, size_t count)
9119 return sched_power_savings_store(buf, count, 1);
9121 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
9122 sched_smt_power_savings_show,
9123 sched_smt_power_savings_store);
9126 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
9130 #ifdef CONFIG_SCHED_SMT
9132 err = sysfs_create_file(&cls->kset.kobj,
9133 &attr_sched_smt_power_savings.attr);
9135 #ifdef CONFIG_SCHED_MC
9136 if (!err && mc_capable())
9137 err = sysfs_create_file(&cls->kset.kobj,
9138 &attr_sched_mc_power_savings.attr);
9142 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
9144 #ifndef CONFIG_CPUSETS
9146 * Add online and remove offline CPUs from the scheduler domains.
9147 * When cpusets are enabled they take over this function.
9149 static int update_sched_domains(struct notifier_block *nfb,
9150 unsigned long action, void *hcpu)
9154 case CPU_ONLINE_FROZEN:
9156 case CPU_DEAD_FROZEN:
9157 partition_sched_domains(1, NULL, NULL);
9166 static int update_runtime(struct notifier_block *nfb,
9167 unsigned long action, void *hcpu)
9169 int cpu = (int)(long)hcpu;
9172 case CPU_DOWN_PREPARE:
9173 case CPU_DOWN_PREPARE_FROZEN:
9174 disable_runtime(cpu_rq(cpu));
9177 case CPU_DOWN_FAILED:
9178 case CPU_DOWN_FAILED_FROZEN:
9180 case CPU_ONLINE_FROZEN:
9181 enable_runtime(cpu_rq(cpu));
9189 void __init sched_init_smp(void)
9191 cpumask_var_t non_isolated_cpus;
9193 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
9194 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
9196 #if defined(CONFIG_NUMA)
9197 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
9199 BUG_ON(sched_group_nodes_bycpu == NULL);
9202 mutex_lock(&sched_domains_mutex);
9203 arch_init_sched_domains(cpu_online_mask);
9204 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
9205 if (cpumask_empty(non_isolated_cpus))
9206 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
9207 mutex_unlock(&sched_domains_mutex);
9210 #ifndef CONFIG_CPUSETS
9211 /* XXX: Theoretical race here - CPU may be hotplugged now */
9212 hotcpu_notifier(update_sched_domains, 0);
9215 /* RT runtime code needs to handle some hotplug events */
9216 hotcpu_notifier(update_runtime, 0);
9220 /* Move init over to a non-isolated CPU */
9221 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
9223 sched_init_granularity();
9224 free_cpumask_var(non_isolated_cpus);
9226 init_sched_rt_class();
9229 void __init sched_init_smp(void)
9231 sched_init_granularity();
9233 #endif /* CONFIG_SMP */
9235 const_debug unsigned int sysctl_timer_migration = 1;
9237 int in_sched_functions(unsigned long addr)
9239 return in_lock_functions(addr) ||
9240 (addr >= (unsigned long)__sched_text_start
9241 && addr < (unsigned long)__sched_text_end);
9244 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
9246 cfs_rq->tasks_timeline = RB_ROOT;
9247 INIT_LIST_HEAD(&cfs_rq->tasks);
9248 #ifdef CONFIG_FAIR_GROUP_SCHED
9251 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9254 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
9256 struct rt_prio_array *array;
9259 array = &rt_rq->active;
9260 for (i = 0; i < MAX_RT_PRIO; i++) {
9261 INIT_LIST_HEAD(array->queue + i);
9262 __clear_bit(i, array->bitmap);
9264 /* delimiter for bitsearch: */
9265 __set_bit(MAX_RT_PRIO, array->bitmap);
9267 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9268 rt_rq->highest_prio.curr = MAX_RT_PRIO;
9270 rt_rq->highest_prio.next = MAX_RT_PRIO;
9274 rt_rq->rt_nr_migratory = 0;
9275 rt_rq->overloaded = 0;
9276 plist_head_init(&rt_rq->pushable_tasks, &rq->lock);
9280 rt_rq->rt_throttled = 0;
9281 rt_rq->rt_runtime = 0;
9282 spin_lock_init(&rt_rq->rt_runtime_lock);
9284 #ifdef CONFIG_RT_GROUP_SCHED
9285 rt_rq->rt_nr_boosted = 0;
9290 #ifdef CONFIG_FAIR_GROUP_SCHED
9291 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9292 struct sched_entity *se, int cpu, int add,
9293 struct sched_entity *parent)
9295 struct rq *rq = cpu_rq(cpu);
9296 tg->cfs_rq[cpu] = cfs_rq;
9297 init_cfs_rq(cfs_rq, rq);
9300 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
9303 /* se could be NULL for init_task_group */
9308 se->cfs_rq = &rq->cfs;
9310 se->cfs_rq = parent->my_q;
9313 se->load.weight = tg->shares;
9314 se->load.inv_weight = 0;
9315 se->parent = parent;
9319 #ifdef CONFIG_RT_GROUP_SCHED
9320 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
9321 struct sched_rt_entity *rt_se, int cpu, int add,
9322 struct sched_rt_entity *parent)
9324 struct rq *rq = cpu_rq(cpu);
9326 tg->rt_rq[cpu] = rt_rq;
9327 init_rt_rq(rt_rq, rq);
9329 rt_rq->rt_se = rt_se;
9330 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
9332 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
9334 tg->rt_se[cpu] = rt_se;
9339 rt_se->rt_rq = &rq->rt;
9341 rt_se->rt_rq = parent->my_q;
9343 rt_se->my_q = rt_rq;
9344 rt_se->parent = parent;
9345 INIT_LIST_HEAD(&rt_se->run_list);
9349 void __init sched_init(void)
9352 unsigned long alloc_size = 0, ptr;
9354 #ifdef CONFIG_FAIR_GROUP_SCHED
9355 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9357 #ifdef CONFIG_RT_GROUP_SCHED
9358 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9360 #ifdef CONFIG_USER_SCHED
9363 #ifdef CONFIG_CPUMASK_OFFSTACK
9364 alloc_size += num_possible_cpus() * cpumask_size();
9367 * As sched_init() is called before page_alloc is setup,
9368 * we use alloc_bootmem().
9371 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
9373 #ifdef CONFIG_FAIR_GROUP_SCHED
9374 init_task_group.se = (struct sched_entity **)ptr;
9375 ptr += nr_cpu_ids * sizeof(void **);
9377 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
9378 ptr += nr_cpu_ids * sizeof(void **);
9380 #ifdef CONFIG_USER_SCHED
9381 root_task_group.se = (struct sched_entity **)ptr;
9382 ptr += nr_cpu_ids * sizeof(void **);
9384 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9385 ptr += nr_cpu_ids * sizeof(void **);
9386 #endif /* CONFIG_USER_SCHED */
9387 #endif /* CONFIG_FAIR_GROUP_SCHED */
9388 #ifdef CONFIG_RT_GROUP_SCHED
9389 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
9390 ptr += nr_cpu_ids * sizeof(void **);
9392 init_task_group.rt_rq = (struct rt_rq **)ptr;
9393 ptr += nr_cpu_ids * sizeof(void **);
9395 #ifdef CONFIG_USER_SCHED
9396 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9397 ptr += nr_cpu_ids * sizeof(void **);
9399 root_task_group.rt_rq = (struct rt_rq **)ptr;
9400 ptr += nr_cpu_ids * sizeof(void **);
9401 #endif /* CONFIG_USER_SCHED */
9402 #endif /* CONFIG_RT_GROUP_SCHED */
9403 #ifdef CONFIG_CPUMASK_OFFSTACK
9404 for_each_possible_cpu(i) {
9405 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
9406 ptr += cpumask_size();
9408 #endif /* CONFIG_CPUMASK_OFFSTACK */
9412 init_defrootdomain();
9415 init_rt_bandwidth(&def_rt_bandwidth,
9416 global_rt_period(), global_rt_runtime());
9418 #ifdef CONFIG_RT_GROUP_SCHED
9419 init_rt_bandwidth(&init_task_group.rt_bandwidth,
9420 global_rt_period(), global_rt_runtime());
9421 #ifdef CONFIG_USER_SCHED
9422 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9423 global_rt_period(), RUNTIME_INF);
9424 #endif /* CONFIG_USER_SCHED */
9425 #endif /* CONFIG_RT_GROUP_SCHED */
9427 #ifdef CONFIG_GROUP_SCHED
9428 list_add(&init_task_group.list, &task_groups);
9429 INIT_LIST_HEAD(&init_task_group.children);
9431 #ifdef CONFIG_USER_SCHED
9432 INIT_LIST_HEAD(&root_task_group.children);
9433 init_task_group.parent = &root_task_group;
9434 list_add(&init_task_group.siblings, &root_task_group.children);
9435 #endif /* CONFIG_USER_SCHED */
9436 #endif /* CONFIG_GROUP_SCHED */
9438 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
9439 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
9440 __alignof__(unsigned long));
9442 for_each_possible_cpu(i) {
9446 spin_lock_init(&rq->lock);
9448 rq->calc_load_active = 0;
9449 rq->calc_load_update = jiffies + LOAD_FREQ;
9450 init_cfs_rq(&rq->cfs, rq);
9451 init_rt_rq(&rq->rt, rq);
9452 #ifdef CONFIG_FAIR_GROUP_SCHED
9453 init_task_group.shares = init_task_group_load;
9454 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9455 #ifdef CONFIG_CGROUP_SCHED
9457 * How much cpu bandwidth does init_task_group get?
9459 * In case of task-groups formed thr' the cgroup filesystem, it
9460 * gets 100% of the cpu resources in the system. This overall
9461 * system cpu resource is divided among the tasks of
9462 * init_task_group and its child task-groups in a fair manner,
9463 * based on each entity's (task or task-group's) weight
9464 * (se->load.weight).
9466 * In other words, if init_task_group has 10 tasks of weight
9467 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9468 * then A0's share of the cpu resource is:
9470 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9472 * We achieve this by letting init_task_group's tasks sit
9473 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9475 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
9476 #elif defined CONFIG_USER_SCHED
9477 root_task_group.shares = NICE_0_LOAD;
9478 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
9480 * In case of task-groups formed thr' the user id of tasks,
9481 * init_task_group represents tasks belonging to root user.
9482 * Hence it forms a sibling of all subsequent groups formed.
9483 * In this case, init_task_group gets only a fraction of overall
9484 * system cpu resource, based on the weight assigned to root
9485 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9486 * by letting tasks of init_task_group sit in a separate cfs_rq
9487 * (init_tg_cfs_rq) and having one entity represent this group of
9488 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9490 init_tg_cfs_entry(&init_task_group,
9491 &per_cpu(init_tg_cfs_rq, i),
9492 &per_cpu(init_sched_entity, i), i, 1,
9493 root_task_group.se[i]);
9496 #endif /* CONFIG_FAIR_GROUP_SCHED */
9498 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9499 #ifdef CONFIG_RT_GROUP_SCHED
9500 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
9501 #ifdef CONFIG_CGROUP_SCHED
9502 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
9503 #elif defined CONFIG_USER_SCHED
9504 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
9505 init_tg_rt_entry(&init_task_group,
9506 &per_cpu(init_rt_rq, i),
9507 &per_cpu(init_sched_rt_entity, i), i, 1,
9508 root_task_group.rt_se[i]);
9512 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
9513 rq->cpu_load[j] = 0;
9517 rq->post_schedule = 0;
9518 rq->active_balance = 0;
9519 rq->next_balance = jiffies;
9523 rq->migration_thread = NULL;
9524 INIT_LIST_HEAD(&rq->migration_queue);
9525 rq_attach_root(rq, &def_root_domain);
9528 atomic_set(&rq->nr_iowait, 0);
9531 set_load_weight(&init_task);
9533 #ifdef CONFIG_PREEMPT_NOTIFIERS
9534 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
9538 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9541 #ifdef CONFIG_RT_MUTEXES
9542 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
9546 * The boot idle thread does lazy MMU switching as well:
9548 atomic_inc(&init_mm.mm_count);
9549 enter_lazy_tlb(&init_mm, current);
9552 * Make us the idle thread. Technically, schedule() should not be
9553 * called from this thread, however somewhere below it might be,
9554 * but because we are the idle thread, we just pick up running again
9555 * when this runqueue becomes "idle".
9557 init_idle(current, smp_processor_id());
9559 calc_load_update = jiffies + LOAD_FREQ;
9562 * During early bootup we pretend to be a normal task:
9564 current->sched_class = &fair_sched_class;
9566 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9567 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
9570 zalloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
9571 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
9573 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
9578 scheduler_running = 1;
9581 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9582 static inline int preempt_count_equals(int preempt_offset)
9584 int nested = preempt_count() & ~PREEMPT_ACTIVE;
9586 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
9589 void __might_sleep(char *file, int line, int preempt_offset)
9592 static unsigned long prev_jiffy; /* ratelimiting */
9594 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
9595 system_state != SYSTEM_RUNNING || oops_in_progress)
9597 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9599 prev_jiffy = jiffies;
9602 "BUG: sleeping function called from invalid context at %s:%d\n",
9605 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9606 in_atomic(), irqs_disabled(),
9607 current->pid, current->comm);
9609 debug_show_held_locks(current);
9610 if (irqs_disabled())
9611 print_irqtrace_events(current);
9615 EXPORT_SYMBOL(__might_sleep);
9618 #ifdef CONFIG_MAGIC_SYSRQ
9619 static void normalize_task(struct rq *rq, struct task_struct *p)
9623 update_rq_clock(rq);
9624 on_rq = p->se.on_rq;
9626 deactivate_task(rq, p, 0);
9627 __setscheduler(rq, p, SCHED_NORMAL, 0);
9629 activate_task(rq, p, 0);
9630 resched_task(rq->curr);
9634 void normalize_rt_tasks(void)
9636 struct task_struct *g, *p;
9637 unsigned long flags;
9640 read_lock_irqsave(&tasklist_lock, flags);
9641 do_each_thread(g, p) {
9643 * Only normalize user tasks:
9648 p->se.exec_start = 0;
9649 #ifdef CONFIG_SCHEDSTATS
9650 p->se.wait_start = 0;
9651 p->se.sleep_start = 0;
9652 p->se.block_start = 0;
9657 * Renice negative nice level userspace
9660 if (TASK_NICE(p) < 0 && p->mm)
9661 set_user_nice(p, 0);
9665 spin_lock(&p->pi_lock);
9666 rq = __task_rq_lock(p);
9668 normalize_task(rq, p);
9670 __task_rq_unlock(rq);
9671 spin_unlock(&p->pi_lock);
9672 } while_each_thread(g, p);
9674 read_unlock_irqrestore(&tasklist_lock, flags);
9677 #endif /* CONFIG_MAGIC_SYSRQ */
9681 * These functions are only useful for the IA64 MCA handling.
9683 * They can only be called when the whole system has been
9684 * stopped - every CPU needs to be quiescent, and no scheduling
9685 * activity can take place. Using them for anything else would
9686 * be a serious bug, and as a result, they aren't even visible
9687 * under any other configuration.
9691 * curr_task - return the current task for a given cpu.
9692 * @cpu: the processor in question.
9694 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9696 struct task_struct *curr_task(int cpu)
9698 return cpu_curr(cpu);
9702 * set_curr_task - set the current task for a given cpu.
9703 * @cpu: the processor in question.
9704 * @p: the task pointer to set.
9706 * Description: This function must only be used when non-maskable interrupts
9707 * are serviced on a separate stack. It allows the architecture to switch the
9708 * notion of the current task on a cpu in a non-blocking manner. This function
9709 * must be called with all CPU's synchronized, and interrupts disabled, the
9710 * and caller must save the original value of the current task (see
9711 * curr_task() above) and restore that value before reenabling interrupts and
9712 * re-starting the system.
9714 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9716 void set_curr_task(int cpu, struct task_struct *p)
9723 #ifdef CONFIG_FAIR_GROUP_SCHED
9724 static void free_fair_sched_group(struct task_group *tg)
9728 for_each_possible_cpu(i) {
9730 kfree(tg->cfs_rq[i]);
9740 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9742 struct cfs_rq *cfs_rq;
9743 struct sched_entity *se;
9747 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9750 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9754 tg->shares = NICE_0_LOAD;
9756 for_each_possible_cpu(i) {
9759 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9760 GFP_KERNEL, cpu_to_node(i));
9764 se = kzalloc_node(sizeof(struct sched_entity),
9765 GFP_KERNEL, cpu_to_node(i));
9769 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
9778 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9780 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
9781 &cpu_rq(cpu)->leaf_cfs_rq_list);
9784 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9786 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
9788 #else /* !CONFG_FAIR_GROUP_SCHED */
9789 static inline void free_fair_sched_group(struct task_group *tg)
9794 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9799 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9803 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9806 #endif /* CONFIG_FAIR_GROUP_SCHED */
9808 #ifdef CONFIG_RT_GROUP_SCHED
9809 static void free_rt_sched_group(struct task_group *tg)
9813 destroy_rt_bandwidth(&tg->rt_bandwidth);
9815 for_each_possible_cpu(i) {
9817 kfree(tg->rt_rq[i]);
9819 kfree(tg->rt_se[i]);
9827 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9829 struct rt_rq *rt_rq;
9830 struct sched_rt_entity *rt_se;
9834 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
9837 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
9841 init_rt_bandwidth(&tg->rt_bandwidth,
9842 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
9844 for_each_possible_cpu(i) {
9847 rt_rq = kzalloc_node(sizeof(struct rt_rq),
9848 GFP_KERNEL, cpu_to_node(i));
9852 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
9853 GFP_KERNEL, cpu_to_node(i));
9857 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
9866 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9868 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
9869 &cpu_rq(cpu)->leaf_rt_rq_list);
9872 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9874 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
9876 #else /* !CONFIG_RT_GROUP_SCHED */
9877 static inline void free_rt_sched_group(struct task_group *tg)
9882 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9887 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9891 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9894 #endif /* CONFIG_RT_GROUP_SCHED */
9896 #ifdef CONFIG_GROUP_SCHED
9897 static void free_sched_group(struct task_group *tg)
9899 free_fair_sched_group(tg);
9900 free_rt_sched_group(tg);
9904 /* allocate runqueue etc for a new task group */
9905 struct task_group *sched_create_group(struct task_group *parent)
9907 struct task_group *tg;
9908 unsigned long flags;
9911 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
9913 return ERR_PTR(-ENOMEM);
9915 if (!alloc_fair_sched_group(tg, parent))
9918 if (!alloc_rt_sched_group(tg, parent))
9921 spin_lock_irqsave(&task_group_lock, flags);
9922 for_each_possible_cpu(i) {
9923 register_fair_sched_group(tg, i);
9924 register_rt_sched_group(tg, i);
9926 list_add_rcu(&tg->list, &task_groups);
9928 WARN_ON(!parent); /* root should already exist */
9930 tg->parent = parent;
9931 INIT_LIST_HEAD(&tg->children);
9932 list_add_rcu(&tg->siblings, &parent->children);
9933 spin_unlock_irqrestore(&task_group_lock, flags);
9938 free_sched_group(tg);
9939 return ERR_PTR(-ENOMEM);
9942 /* rcu callback to free various structures associated with a task group */
9943 static void free_sched_group_rcu(struct rcu_head *rhp)
9945 /* now it should be safe to free those cfs_rqs */
9946 free_sched_group(container_of(rhp, struct task_group, rcu));
9949 /* Destroy runqueue etc associated with a task group */
9950 void sched_destroy_group(struct task_group *tg)
9952 unsigned long flags;
9955 spin_lock_irqsave(&task_group_lock, flags);
9956 for_each_possible_cpu(i) {
9957 unregister_fair_sched_group(tg, i);
9958 unregister_rt_sched_group(tg, i);
9960 list_del_rcu(&tg->list);
9961 list_del_rcu(&tg->siblings);
9962 spin_unlock_irqrestore(&task_group_lock, flags);
9964 /* wait for possible concurrent references to cfs_rqs complete */
9965 call_rcu(&tg->rcu, free_sched_group_rcu);
9968 /* change task's runqueue when it moves between groups.
9969 * The caller of this function should have put the task in its new group
9970 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
9971 * reflect its new group.
9973 void sched_move_task(struct task_struct *tsk)
9976 unsigned long flags;
9979 rq = task_rq_lock(tsk, &flags);
9981 update_rq_clock(rq);
9983 running = task_current(rq, tsk);
9984 on_rq = tsk->se.on_rq;
9987 dequeue_task(rq, tsk, 0);
9988 if (unlikely(running))
9989 tsk->sched_class->put_prev_task(rq, tsk);
9991 set_task_rq(tsk, task_cpu(tsk));
9993 #ifdef CONFIG_FAIR_GROUP_SCHED
9994 if (tsk->sched_class->moved_group)
9995 tsk->sched_class->moved_group(tsk);
9998 if (unlikely(running))
9999 tsk->sched_class->set_curr_task(rq);
10001 enqueue_task(rq, tsk, 0);
10003 task_rq_unlock(rq, &flags);
10005 #endif /* CONFIG_GROUP_SCHED */
10007 #ifdef CONFIG_FAIR_GROUP_SCHED
10008 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
10010 struct cfs_rq *cfs_rq = se->cfs_rq;
10015 dequeue_entity(cfs_rq, se, 0);
10017 se->load.weight = shares;
10018 se->load.inv_weight = 0;
10021 enqueue_entity(cfs_rq, se, 0);
10024 static void set_se_shares(struct sched_entity *se, unsigned long shares)
10026 struct cfs_rq *cfs_rq = se->cfs_rq;
10027 struct rq *rq = cfs_rq->rq;
10028 unsigned long flags;
10030 spin_lock_irqsave(&rq->lock, flags);
10031 __set_se_shares(se, shares);
10032 spin_unlock_irqrestore(&rq->lock, flags);
10035 static DEFINE_MUTEX(shares_mutex);
10037 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10040 unsigned long flags;
10043 * We can't change the weight of the root cgroup.
10048 if (shares < MIN_SHARES)
10049 shares = MIN_SHARES;
10050 else if (shares > MAX_SHARES)
10051 shares = MAX_SHARES;
10053 mutex_lock(&shares_mutex);
10054 if (tg->shares == shares)
10057 spin_lock_irqsave(&task_group_lock, flags);
10058 for_each_possible_cpu(i)
10059 unregister_fair_sched_group(tg, i);
10060 list_del_rcu(&tg->siblings);
10061 spin_unlock_irqrestore(&task_group_lock, flags);
10063 /* wait for any ongoing reference to this group to finish */
10064 synchronize_sched();
10067 * Now we are free to modify the group's share on each cpu
10068 * w/o tripping rebalance_share or load_balance_fair.
10070 tg->shares = shares;
10071 for_each_possible_cpu(i) {
10073 * force a rebalance
10075 cfs_rq_set_shares(tg->cfs_rq[i], 0);
10076 set_se_shares(tg->se[i], shares);
10080 * Enable load balance activity on this group, by inserting it back on
10081 * each cpu's rq->leaf_cfs_rq_list.
10083 spin_lock_irqsave(&task_group_lock, flags);
10084 for_each_possible_cpu(i)
10085 register_fair_sched_group(tg, i);
10086 list_add_rcu(&tg->siblings, &tg->parent->children);
10087 spin_unlock_irqrestore(&task_group_lock, flags);
10089 mutex_unlock(&shares_mutex);
10093 unsigned long sched_group_shares(struct task_group *tg)
10099 #ifdef CONFIG_RT_GROUP_SCHED
10101 * Ensure that the real time constraints are schedulable.
10103 static DEFINE_MUTEX(rt_constraints_mutex);
10105 static unsigned long to_ratio(u64 period, u64 runtime)
10107 if (runtime == RUNTIME_INF)
10110 return div64_u64(runtime << 20, period);
10113 /* Must be called with tasklist_lock held */
10114 static inline int tg_has_rt_tasks(struct task_group *tg)
10116 struct task_struct *g, *p;
10118 do_each_thread(g, p) {
10119 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
10121 } while_each_thread(g, p);
10126 struct rt_schedulable_data {
10127 struct task_group *tg;
10132 static int tg_schedulable(struct task_group *tg, void *data)
10134 struct rt_schedulable_data *d = data;
10135 struct task_group *child;
10136 unsigned long total, sum = 0;
10137 u64 period, runtime;
10139 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10140 runtime = tg->rt_bandwidth.rt_runtime;
10143 period = d->rt_period;
10144 runtime = d->rt_runtime;
10147 #ifdef CONFIG_USER_SCHED
10148 if (tg == &root_task_group) {
10149 period = global_rt_period();
10150 runtime = global_rt_runtime();
10155 * Cannot have more runtime than the period.
10157 if (runtime > period && runtime != RUNTIME_INF)
10161 * Ensure we don't starve existing RT tasks.
10163 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
10166 total = to_ratio(period, runtime);
10169 * Nobody can have more than the global setting allows.
10171 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
10175 * The sum of our children's runtime should not exceed our own.
10177 list_for_each_entry_rcu(child, &tg->children, siblings) {
10178 period = ktime_to_ns(child->rt_bandwidth.rt_period);
10179 runtime = child->rt_bandwidth.rt_runtime;
10181 if (child == d->tg) {
10182 period = d->rt_period;
10183 runtime = d->rt_runtime;
10186 sum += to_ratio(period, runtime);
10195 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
10197 struct rt_schedulable_data data = {
10199 .rt_period = period,
10200 .rt_runtime = runtime,
10203 return walk_tg_tree(tg_schedulable, tg_nop, &data);
10206 static int tg_set_bandwidth(struct task_group *tg,
10207 u64 rt_period, u64 rt_runtime)
10211 mutex_lock(&rt_constraints_mutex);
10212 read_lock(&tasklist_lock);
10213 err = __rt_schedulable(tg, rt_period, rt_runtime);
10217 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10218 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
10219 tg->rt_bandwidth.rt_runtime = rt_runtime;
10221 for_each_possible_cpu(i) {
10222 struct rt_rq *rt_rq = tg->rt_rq[i];
10224 spin_lock(&rt_rq->rt_runtime_lock);
10225 rt_rq->rt_runtime = rt_runtime;
10226 spin_unlock(&rt_rq->rt_runtime_lock);
10228 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10230 read_unlock(&tasklist_lock);
10231 mutex_unlock(&rt_constraints_mutex);
10236 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
10238 u64 rt_runtime, rt_period;
10240 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10241 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
10242 if (rt_runtime_us < 0)
10243 rt_runtime = RUNTIME_INF;
10245 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10248 long sched_group_rt_runtime(struct task_group *tg)
10252 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
10255 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
10256 do_div(rt_runtime_us, NSEC_PER_USEC);
10257 return rt_runtime_us;
10260 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
10262 u64 rt_runtime, rt_period;
10264 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
10265 rt_runtime = tg->rt_bandwidth.rt_runtime;
10267 if (rt_period == 0)
10270 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10273 long sched_group_rt_period(struct task_group *tg)
10277 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
10278 do_div(rt_period_us, NSEC_PER_USEC);
10279 return rt_period_us;
10282 static int sched_rt_global_constraints(void)
10284 u64 runtime, period;
10287 if (sysctl_sched_rt_period <= 0)
10290 runtime = global_rt_runtime();
10291 period = global_rt_period();
10294 * Sanity check on the sysctl variables.
10296 if (runtime > period && runtime != RUNTIME_INF)
10299 mutex_lock(&rt_constraints_mutex);
10300 read_lock(&tasklist_lock);
10301 ret = __rt_schedulable(NULL, 0, 0);
10302 read_unlock(&tasklist_lock);
10303 mutex_unlock(&rt_constraints_mutex);
10308 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
10310 /* Don't accept realtime tasks when there is no way for them to run */
10311 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
10317 #else /* !CONFIG_RT_GROUP_SCHED */
10318 static int sched_rt_global_constraints(void)
10320 unsigned long flags;
10323 if (sysctl_sched_rt_period <= 0)
10327 * There's always some RT tasks in the root group
10328 * -- migration, kstopmachine etc..
10330 if (sysctl_sched_rt_runtime == 0)
10333 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
10334 for_each_possible_cpu(i) {
10335 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
10337 spin_lock(&rt_rq->rt_runtime_lock);
10338 rt_rq->rt_runtime = global_rt_runtime();
10339 spin_unlock(&rt_rq->rt_runtime_lock);
10341 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
10345 #endif /* CONFIG_RT_GROUP_SCHED */
10347 int sched_rt_handler(struct ctl_table *table, int write,
10348 void __user *buffer, size_t *lenp,
10352 int old_period, old_runtime;
10353 static DEFINE_MUTEX(mutex);
10355 mutex_lock(&mutex);
10356 old_period = sysctl_sched_rt_period;
10357 old_runtime = sysctl_sched_rt_runtime;
10359 ret = proc_dointvec(table, write, buffer, lenp, ppos);
10361 if (!ret && write) {
10362 ret = sched_rt_global_constraints();
10364 sysctl_sched_rt_period = old_period;
10365 sysctl_sched_rt_runtime = old_runtime;
10367 def_rt_bandwidth.rt_runtime = global_rt_runtime();
10368 def_rt_bandwidth.rt_period =
10369 ns_to_ktime(global_rt_period());
10372 mutex_unlock(&mutex);
10377 #ifdef CONFIG_CGROUP_SCHED
10379 /* return corresponding task_group object of a cgroup */
10380 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
10382 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
10383 struct task_group, css);
10386 static struct cgroup_subsys_state *
10387 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
10389 struct task_group *tg, *parent;
10391 if (!cgrp->parent) {
10392 /* This is early initialization for the top cgroup */
10393 return &init_task_group.css;
10396 parent = cgroup_tg(cgrp->parent);
10397 tg = sched_create_group(parent);
10399 return ERR_PTR(-ENOMEM);
10405 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10407 struct task_group *tg = cgroup_tg(cgrp);
10409 sched_destroy_group(tg);
10413 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
10415 #ifdef CONFIG_RT_GROUP_SCHED
10416 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
10419 /* We don't support RT-tasks being in separate groups */
10420 if (tsk->sched_class != &fair_sched_class)
10427 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10428 struct task_struct *tsk, bool threadgroup)
10430 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
10434 struct task_struct *c;
10436 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
10437 retval = cpu_cgroup_can_attach_task(cgrp, c);
10449 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10450 struct cgroup *old_cont, struct task_struct *tsk,
10453 sched_move_task(tsk);
10455 struct task_struct *c;
10457 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
10458 sched_move_task(c);
10464 #ifdef CONFIG_FAIR_GROUP_SCHED
10465 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
10468 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
10471 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
10473 struct task_group *tg = cgroup_tg(cgrp);
10475 return (u64) tg->shares;
10477 #endif /* CONFIG_FAIR_GROUP_SCHED */
10479 #ifdef CONFIG_RT_GROUP_SCHED
10480 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
10483 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
10486 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
10488 return sched_group_rt_runtime(cgroup_tg(cgrp));
10491 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
10494 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
10497 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
10499 return sched_group_rt_period(cgroup_tg(cgrp));
10501 #endif /* CONFIG_RT_GROUP_SCHED */
10503 static struct cftype cpu_files[] = {
10504 #ifdef CONFIG_FAIR_GROUP_SCHED
10507 .read_u64 = cpu_shares_read_u64,
10508 .write_u64 = cpu_shares_write_u64,
10511 #ifdef CONFIG_RT_GROUP_SCHED
10513 .name = "rt_runtime_us",
10514 .read_s64 = cpu_rt_runtime_read,
10515 .write_s64 = cpu_rt_runtime_write,
10518 .name = "rt_period_us",
10519 .read_u64 = cpu_rt_period_read_uint,
10520 .write_u64 = cpu_rt_period_write_uint,
10525 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
10527 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
10530 struct cgroup_subsys cpu_cgroup_subsys = {
10532 .create = cpu_cgroup_create,
10533 .destroy = cpu_cgroup_destroy,
10534 .can_attach = cpu_cgroup_can_attach,
10535 .attach = cpu_cgroup_attach,
10536 .populate = cpu_cgroup_populate,
10537 .subsys_id = cpu_cgroup_subsys_id,
10541 #endif /* CONFIG_CGROUP_SCHED */
10543 #ifdef CONFIG_CGROUP_CPUACCT
10546 * CPU accounting code for task groups.
10548 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10549 * (balbir@in.ibm.com).
10552 /* track cpu usage of a group of tasks and its child groups */
10554 struct cgroup_subsys_state css;
10555 /* cpuusage holds pointer to a u64-type object on every cpu */
10557 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
10558 struct cpuacct *parent;
10561 struct cgroup_subsys cpuacct_subsys;
10563 /* return cpu accounting group corresponding to this container */
10564 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
10566 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
10567 struct cpuacct, css);
10570 /* return cpu accounting group to which this task belongs */
10571 static inline struct cpuacct *task_ca(struct task_struct *tsk)
10573 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
10574 struct cpuacct, css);
10577 /* create a new cpu accounting group */
10578 static struct cgroup_subsys_state *cpuacct_create(
10579 struct cgroup_subsys *ss, struct cgroup *cgrp)
10581 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
10587 ca->cpuusage = alloc_percpu(u64);
10591 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10592 if (percpu_counter_init(&ca->cpustat[i], 0))
10593 goto out_free_counters;
10596 ca->parent = cgroup_ca(cgrp->parent);
10602 percpu_counter_destroy(&ca->cpustat[i]);
10603 free_percpu(ca->cpuusage);
10607 return ERR_PTR(-ENOMEM);
10610 /* destroy an existing cpu accounting group */
10612 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10614 struct cpuacct *ca = cgroup_ca(cgrp);
10617 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10618 percpu_counter_destroy(&ca->cpustat[i]);
10619 free_percpu(ca->cpuusage);
10623 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
10625 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10628 #ifndef CONFIG_64BIT
10630 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10632 spin_lock_irq(&cpu_rq(cpu)->lock);
10634 spin_unlock_irq(&cpu_rq(cpu)->lock);
10642 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
10644 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10646 #ifndef CONFIG_64BIT
10648 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10650 spin_lock_irq(&cpu_rq(cpu)->lock);
10652 spin_unlock_irq(&cpu_rq(cpu)->lock);
10658 /* return total cpu usage (in nanoseconds) of a group */
10659 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
10661 struct cpuacct *ca = cgroup_ca(cgrp);
10662 u64 totalcpuusage = 0;
10665 for_each_present_cpu(i)
10666 totalcpuusage += cpuacct_cpuusage_read(ca, i);
10668 return totalcpuusage;
10671 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
10674 struct cpuacct *ca = cgroup_ca(cgrp);
10683 for_each_present_cpu(i)
10684 cpuacct_cpuusage_write(ca, i, 0);
10690 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
10691 struct seq_file *m)
10693 struct cpuacct *ca = cgroup_ca(cgroup);
10697 for_each_present_cpu(i) {
10698 percpu = cpuacct_cpuusage_read(ca, i);
10699 seq_printf(m, "%llu ", (unsigned long long) percpu);
10701 seq_printf(m, "\n");
10705 static const char *cpuacct_stat_desc[] = {
10706 [CPUACCT_STAT_USER] = "user",
10707 [CPUACCT_STAT_SYSTEM] = "system",
10710 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
10711 struct cgroup_map_cb *cb)
10713 struct cpuacct *ca = cgroup_ca(cgrp);
10716 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
10717 s64 val = percpu_counter_read(&ca->cpustat[i]);
10718 val = cputime64_to_clock_t(val);
10719 cb->fill(cb, cpuacct_stat_desc[i], val);
10724 static struct cftype files[] = {
10727 .read_u64 = cpuusage_read,
10728 .write_u64 = cpuusage_write,
10731 .name = "usage_percpu",
10732 .read_seq_string = cpuacct_percpu_seq_read,
10736 .read_map = cpuacct_stats_show,
10740 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
10742 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
10746 * charge this task's execution time to its accounting group.
10748 * called with rq->lock held.
10750 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
10752 struct cpuacct *ca;
10755 if (unlikely(!cpuacct_subsys.active))
10758 cpu = task_cpu(tsk);
10764 for (; ca; ca = ca->parent) {
10765 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10766 *cpuusage += cputime;
10773 * Charge the system/user time to the task's accounting group.
10775 static void cpuacct_update_stats(struct task_struct *tsk,
10776 enum cpuacct_stat_index idx, cputime_t val)
10778 struct cpuacct *ca;
10780 if (unlikely(!cpuacct_subsys.active))
10787 percpu_counter_add(&ca->cpustat[idx], val);
10793 struct cgroup_subsys cpuacct_subsys = {
10795 .create = cpuacct_create,
10796 .destroy = cpuacct_destroy,
10797 .populate = cpuacct_populate,
10798 .subsys_id = cpuacct_subsys_id,
10800 #endif /* CONFIG_CGROUP_CPUACCT */
10804 int rcu_expedited_torture_stats(char *page)
10808 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
10810 void synchronize_sched_expedited(void)
10813 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
10815 #else /* #ifndef CONFIG_SMP */
10817 static DEFINE_PER_CPU(struct migration_req, rcu_migration_req);
10818 static DEFINE_MUTEX(rcu_sched_expedited_mutex);
10820 #define RCU_EXPEDITED_STATE_POST -2
10821 #define RCU_EXPEDITED_STATE_IDLE -1
10823 static int rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
10825 int rcu_expedited_torture_stats(char *page)
10830 cnt += sprintf(&page[cnt], "state: %d /", rcu_expedited_state);
10831 for_each_online_cpu(cpu) {
10832 cnt += sprintf(&page[cnt], " %d:%d",
10833 cpu, per_cpu(rcu_migration_req, cpu).dest_cpu);
10835 cnt += sprintf(&page[cnt], "\n");
10838 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
10840 static long synchronize_sched_expedited_count;
10843 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
10844 * approach to force grace period to end quickly. This consumes
10845 * significant time on all CPUs, and is thus not recommended for
10846 * any sort of common-case code.
10848 * Note that it is illegal to call this function while holding any
10849 * lock that is acquired by a CPU-hotplug notifier. Failing to
10850 * observe this restriction will result in deadlock.
10852 void synchronize_sched_expedited(void)
10855 unsigned long flags;
10856 bool need_full_sync = 0;
10858 struct migration_req *req;
10862 smp_mb(); /* ensure prior mod happens before capturing snap. */
10863 snap = ACCESS_ONCE(synchronize_sched_expedited_count) + 1;
10865 while (!mutex_trylock(&rcu_sched_expedited_mutex)) {
10867 if (trycount++ < 10)
10868 udelay(trycount * num_online_cpus());
10870 synchronize_sched();
10873 if (ACCESS_ONCE(synchronize_sched_expedited_count) - snap > 0) {
10874 smp_mb(); /* ensure test happens before caller kfree */
10879 rcu_expedited_state = RCU_EXPEDITED_STATE_POST;
10880 for_each_online_cpu(cpu) {
10882 req = &per_cpu(rcu_migration_req, cpu);
10883 init_completion(&req->done);
10885 req->dest_cpu = RCU_MIGRATION_NEED_QS;
10886 spin_lock_irqsave(&rq->lock, flags);
10887 list_add(&req->list, &rq->migration_queue);
10888 spin_unlock_irqrestore(&rq->lock, flags);
10889 wake_up_process(rq->migration_thread);
10891 for_each_online_cpu(cpu) {
10892 rcu_expedited_state = cpu;
10893 req = &per_cpu(rcu_migration_req, cpu);
10895 wait_for_completion(&req->done);
10896 spin_lock_irqsave(&rq->lock, flags);
10897 if (unlikely(req->dest_cpu == RCU_MIGRATION_MUST_SYNC))
10898 need_full_sync = 1;
10899 req->dest_cpu = RCU_MIGRATION_IDLE;
10900 spin_unlock_irqrestore(&rq->lock, flags);
10902 rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
10903 mutex_unlock(&rcu_sched_expedited_mutex);
10905 if (need_full_sync)
10906 synchronize_sched();
10908 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
10910 #endif /* #else #ifndef CONFIG_SMP */