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 char in_nohz_recently;
540 /* capture load from *all* tasks on this cpu: */
541 struct load_weight load;
542 unsigned long nr_load_updates;
548 #ifdef CONFIG_FAIR_GROUP_SCHED
549 /* list of leaf cfs_rq on this cpu: */
550 struct list_head leaf_cfs_rq_list;
552 #ifdef CONFIG_RT_GROUP_SCHED
553 struct list_head leaf_rt_rq_list;
557 * This is part of a global counter where only the total sum
558 * over all CPUs matters. A task can increase this counter on
559 * one CPU and if it got migrated afterwards it may decrease
560 * it on another CPU. Always updated under the runqueue lock:
562 unsigned long nr_uninterruptible;
564 struct task_struct *curr, *idle;
565 unsigned long next_balance;
566 struct mm_struct *prev_mm;
573 struct root_domain *rd;
574 struct sched_domain *sd;
576 unsigned char idle_at_tick;
577 /* For active balancing */
581 /* cpu of this runqueue: */
585 unsigned long avg_load_per_task;
587 struct task_struct *migration_thread;
588 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 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1816 set_task_rq(p, cpu);
1819 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1820 * successfuly executed on another CPU. We must ensure that updates of
1821 * per-task data have been completed by this moment.
1824 task_thread_info(p)->cpu = cpu;
1828 #include "sched_stats.h"
1829 #include "sched_idletask.c"
1830 #include "sched_fair.c"
1831 #include "sched_rt.c"
1832 #ifdef CONFIG_SCHED_DEBUG
1833 # include "sched_debug.c"
1836 #define sched_class_highest (&rt_sched_class)
1837 #define for_each_class(class) \
1838 for (class = sched_class_highest; class; class = class->next)
1840 static void inc_nr_running(struct rq *rq)
1845 static void dec_nr_running(struct rq *rq)
1850 static void set_load_weight(struct task_struct *p)
1852 if (task_has_rt_policy(p)) {
1853 p->se.load.weight = prio_to_weight[0] * 2;
1854 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1859 * SCHED_IDLE tasks get minimal weight:
1861 if (p->policy == SCHED_IDLE) {
1862 p->se.load.weight = WEIGHT_IDLEPRIO;
1863 p->se.load.inv_weight = WMULT_IDLEPRIO;
1867 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1868 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1871 static void update_avg(u64 *avg, u64 sample)
1873 s64 diff = sample - *avg;
1877 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1880 p->se.start_runtime = p->se.sum_exec_runtime;
1882 sched_info_queued(p);
1883 p->sched_class->enqueue_task(rq, p, wakeup);
1887 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1890 if (p->se.last_wakeup) {
1891 update_avg(&p->se.avg_overlap,
1892 p->se.sum_exec_runtime - p->se.last_wakeup);
1893 p->se.last_wakeup = 0;
1895 update_avg(&p->se.avg_wakeup,
1896 sysctl_sched_wakeup_granularity);
1900 sched_info_dequeued(p);
1901 p->sched_class->dequeue_task(rq, p, sleep);
1906 * __normal_prio - return the priority that is based on the static prio
1908 static inline int __normal_prio(struct task_struct *p)
1910 return p->static_prio;
1914 * Calculate the expected normal priority: i.e. priority
1915 * without taking RT-inheritance into account. Might be
1916 * boosted by interactivity modifiers. Changes upon fork,
1917 * setprio syscalls, and whenever the interactivity
1918 * estimator recalculates.
1920 static inline int normal_prio(struct task_struct *p)
1924 if (task_has_rt_policy(p))
1925 prio = MAX_RT_PRIO-1 - p->rt_priority;
1927 prio = __normal_prio(p);
1932 * Calculate the current priority, i.e. the priority
1933 * taken into account by the scheduler. This value might
1934 * be boosted by RT tasks, or might be boosted by
1935 * interactivity modifiers. Will be RT if the task got
1936 * RT-boosted. If not then it returns p->normal_prio.
1938 static int effective_prio(struct task_struct *p)
1940 p->normal_prio = normal_prio(p);
1942 * If we are RT tasks or we were boosted to RT priority,
1943 * keep the priority unchanged. Otherwise, update priority
1944 * to the normal priority:
1946 if (!rt_prio(p->prio))
1947 return p->normal_prio;
1952 * activate_task - move a task to the runqueue.
1954 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1956 if (task_contributes_to_load(p))
1957 rq->nr_uninterruptible--;
1959 enqueue_task(rq, p, wakeup);
1964 * deactivate_task - remove a task from the runqueue.
1966 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1968 if (task_contributes_to_load(p))
1969 rq->nr_uninterruptible++;
1971 dequeue_task(rq, p, sleep);
1976 * task_curr - is this task currently executing on a CPU?
1977 * @p: the task in question.
1979 inline int task_curr(const struct task_struct *p)
1981 return cpu_curr(task_cpu(p)) == p;
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 update_rq_clock(rq);
2021 set_task_cpu(p, cpu);
2022 p->cpus_allowed = cpumask_of_cpu(cpu);
2023 p->rt.nr_cpus_allowed = 1;
2024 p->flags |= PF_THREAD_BOUND;
2025 spin_unlock_irqrestore(&rq->lock, flags);
2027 EXPORT_SYMBOL(kthread_bind);
2031 * Is this task likely cache-hot:
2034 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2039 * Buddy candidates are cache hot:
2041 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2042 (&p->se == cfs_rq_of(&p->se)->next ||
2043 &p->se == cfs_rq_of(&p->se)->last))
2046 if (p->sched_class != &fair_sched_class)
2049 if (sysctl_sched_migration_cost == -1)
2051 if (sysctl_sched_migration_cost == 0)
2054 delta = now - p->se.exec_start;
2056 return delta < (s64)sysctl_sched_migration_cost;
2060 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2062 int old_cpu = task_cpu(p);
2063 struct rq *old_rq = cpu_rq(old_cpu);
2064 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
2065 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
2067 trace_sched_migrate_task(p, new_cpu);
2069 if (old_cpu != new_cpu) {
2070 p->se.nr_migrations++;
2071 #ifdef CONFIG_SCHEDSTATS
2072 if (task_hot(p, old_rq->clock, NULL))
2073 schedstat_inc(p, se.nr_forced2_migrations);
2075 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS,
2078 p->se.vruntime -= old_cfsrq->min_vruntime -
2079 new_cfsrq->min_vruntime;
2081 __set_task_cpu(p, new_cpu);
2084 struct migration_req {
2085 struct list_head list;
2087 struct task_struct *task;
2090 struct completion done;
2094 * The task's runqueue lock must be held.
2095 * Returns true if you have to wait for migration thread.
2098 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2100 struct rq *rq = task_rq(p);
2103 * If the task is not on a runqueue (and not running), then
2104 * it is sufficient to simply update the task's cpu field.
2106 if (!p->se.on_rq && !task_running(rq, p)) {
2107 update_rq_clock(rq);
2108 set_task_cpu(p, dest_cpu);
2112 init_completion(&req->done);
2114 req->dest_cpu = dest_cpu;
2115 list_add(&req->list, &rq->migration_queue);
2121 * wait_task_context_switch - wait for a thread to complete at least one
2124 * @p must not be current.
2126 void wait_task_context_switch(struct task_struct *p)
2128 unsigned long nvcsw, nivcsw, flags;
2136 * The runqueue is assigned before the actual context
2137 * switch. We need to take the runqueue lock.
2139 * We could check initially without the lock but it is
2140 * very likely that we need to take the lock in every
2143 rq = task_rq_lock(p, &flags);
2144 running = task_running(rq, p);
2145 task_rq_unlock(rq, &flags);
2147 if (likely(!running))
2150 * The switch count is incremented before the actual
2151 * context switch. We thus wait for two switches to be
2152 * sure at least one completed.
2154 if ((p->nvcsw - nvcsw) > 1)
2156 if ((p->nivcsw - nivcsw) > 1)
2164 * wait_task_inactive - wait for a thread to unschedule.
2166 * If @match_state is nonzero, it's the @p->state value just checked and
2167 * not expected to change. If it changes, i.e. @p might have woken up,
2168 * then return zero. When we succeed in waiting for @p to be off its CPU,
2169 * we return a positive number (its total switch count). If a second call
2170 * a short while later returns the same number, the caller can be sure that
2171 * @p has remained unscheduled the whole time.
2173 * The caller must ensure that the task *will* unschedule sometime soon,
2174 * else this function might spin for a *long* time. This function can't
2175 * be called with interrupts off, or it may introduce deadlock with
2176 * smp_call_function() if an IPI is sent by the same process we are
2177 * waiting to become inactive.
2179 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2181 unsigned long flags;
2188 * We do the initial early heuristics without holding
2189 * any task-queue locks at all. We'll only try to get
2190 * the runqueue lock when things look like they will
2196 * If the task is actively running on another CPU
2197 * still, just relax and busy-wait without holding
2200 * NOTE! Since we don't hold any locks, it's not
2201 * even sure that "rq" stays as the right runqueue!
2202 * But we don't care, since "task_running()" will
2203 * return false if the runqueue has changed and p
2204 * is actually now running somewhere else!
2206 while (task_running(rq, p)) {
2207 if (match_state && unlikely(p->state != match_state))
2213 * Ok, time to look more closely! We need the rq
2214 * lock now, to be *sure*. If we're wrong, we'll
2215 * just go back and repeat.
2217 rq = task_rq_lock(p, &flags);
2218 trace_sched_wait_task(rq, p);
2219 running = task_running(rq, p);
2220 on_rq = p->se.on_rq;
2222 if (!match_state || p->state == match_state)
2223 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2224 task_rq_unlock(rq, &flags);
2227 * If it changed from the expected state, bail out now.
2229 if (unlikely(!ncsw))
2233 * Was it really running after all now that we
2234 * checked with the proper locks actually held?
2236 * Oops. Go back and try again..
2238 if (unlikely(running)) {
2244 * It's not enough that it's not actively running,
2245 * it must be off the runqueue _entirely_, and not
2248 * So if it was still runnable (but just not actively
2249 * running right now), it's preempted, and we should
2250 * yield - it could be a while.
2252 if (unlikely(on_rq)) {
2253 schedule_timeout_uninterruptible(1);
2258 * Ahh, all good. It wasn't running, and it wasn't
2259 * runnable, which means that it will never become
2260 * running in the future either. We're all done!
2269 * kick_process - kick a running thread to enter/exit the kernel
2270 * @p: the to-be-kicked thread
2272 * Cause a process which is running on another CPU to enter
2273 * kernel-mode, without any delay. (to get signals handled.)
2275 * NOTE: this function doesnt have to take the runqueue lock,
2276 * because all it wants to ensure is that the remote task enters
2277 * the kernel. If the IPI races and the task has been migrated
2278 * to another CPU then no harm is done and the purpose has been
2281 void kick_process(struct task_struct *p)
2287 if ((cpu != smp_processor_id()) && task_curr(p))
2288 smp_send_reschedule(cpu);
2291 EXPORT_SYMBOL_GPL(kick_process);
2292 #endif /* CONFIG_SMP */
2295 * task_oncpu_function_call - call a function on the cpu on which a task runs
2296 * @p: the task to evaluate
2297 * @func: the function to be called
2298 * @info: the function call argument
2300 * Calls the function @func when the task is currently running. This might
2301 * be on the current CPU, which just calls the function directly
2303 void task_oncpu_function_call(struct task_struct *p,
2304 void (*func) (void *info), void *info)
2311 smp_call_function_single(cpu, func, info, 1);
2317 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2319 return p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2324 * try_to_wake_up - wake up a thread
2325 * @p: the to-be-woken-up thread
2326 * @state: the mask of task states that can be woken
2327 * @sync: do a synchronous wakeup?
2329 * Put it on the run-queue if it's not already there. The "current"
2330 * thread is always on the run-queue (except when the actual
2331 * re-schedule is in progress), and as such you're allowed to do
2332 * the simpler "current->state = TASK_RUNNING" to mark yourself
2333 * runnable without the overhead of this.
2335 * returns failure only if the task is already active.
2337 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2340 int cpu, orig_cpu, this_cpu, success = 0;
2341 unsigned long flags;
2342 struct rq *rq, *orig_rq;
2344 if (!sched_feat(SYNC_WAKEUPS))
2345 wake_flags &= ~WF_SYNC;
2347 this_cpu = get_cpu();
2350 rq = orig_rq = task_rq_lock(p, &flags);
2351 update_rq_clock(rq);
2352 if (!(p->state & state))
2362 if (unlikely(task_running(rq, p)))
2366 * In order to handle concurrent wakeups and release the rq->lock
2367 * we put the task in TASK_WAKING state.
2369 * First fix up the nr_uninterruptible count:
2371 if (task_contributes_to_load(p))
2372 rq->nr_uninterruptible--;
2373 p->state = TASK_WAKING;
2374 __task_rq_unlock(rq);
2376 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2377 if (cpu != orig_cpu)
2378 set_task_cpu(p, cpu);
2380 rq = __task_rq_lock(p);
2381 update_rq_clock(rq);
2383 WARN_ON(p->state != TASK_WAKING);
2386 #ifdef CONFIG_SCHEDSTATS
2387 schedstat_inc(rq, ttwu_count);
2388 if (cpu == this_cpu)
2389 schedstat_inc(rq, ttwu_local);
2391 struct sched_domain *sd;
2392 for_each_domain(this_cpu, sd) {
2393 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2394 schedstat_inc(sd, ttwu_wake_remote);
2399 #endif /* CONFIG_SCHEDSTATS */
2402 #endif /* CONFIG_SMP */
2403 schedstat_inc(p, se.nr_wakeups);
2404 if (wake_flags & WF_SYNC)
2405 schedstat_inc(p, se.nr_wakeups_sync);
2406 if (orig_cpu != cpu)
2407 schedstat_inc(p, se.nr_wakeups_migrate);
2408 if (cpu == this_cpu)
2409 schedstat_inc(p, se.nr_wakeups_local);
2411 schedstat_inc(p, se.nr_wakeups_remote);
2412 activate_task(rq, p, 1);
2416 * Only attribute actual wakeups done by this task.
2418 if (!in_interrupt()) {
2419 struct sched_entity *se = ¤t->se;
2420 u64 sample = se->sum_exec_runtime;
2422 if (se->last_wakeup)
2423 sample -= se->last_wakeup;
2425 sample -= se->start_runtime;
2426 update_avg(&se->avg_wakeup, sample);
2428 se->last_wakeup = se->sum_exec_runtime;
2432 trace_sched_wakeup(rq, p, success);
2433 check_preempt_curr(rq, p, wake_flags);
2435 p->state = TASK_RUNNING;
2437 if (p->sched_class->task_wake_up)
2438 p->sched_class->task_wake_up(rq, p);
2440 if (unlikely(rq->idle_stamp)) {
2441 u64 delta = rq->clock - rq->idle_stamp;
2442 u64 max = 2*sysctl_sched_migration_cost;
2447 update_avg(&rq->avg_idle, delta);
2452 task_rq_unlock(rq, &flags);
2459 * wake_up_process - Wake up a specific process
2460 * @p: The process to be woken up.
2462 * Attempt to wake up the nominated process and move it to the set of runnable
2463 * processes. Returns 1 if the process was woken up, 0 if it was already
2466 * It may be assumed that this function implies a write memory barrier before
2467 * changing the task state if and only if any tasks are woken up.
2469 int wake_up_process(struct task_struct *p)
2471 return try_to_wake_up(p, TASK_ALL, 0);
2473 EXPORT_SYMBOL(wake_up_process);
2475 int wake_up_state(struct task_struct *p, unsigned int state)
2477 return try_to_wake_up(p, state, 0);
2481 * Perform scheduler related setup for a newly forked process p.
2482 * p is forked by current.
2484 * __sched_fork() is basic setup used by init_idle() too:
2486 static void __sched_fork(struct task_struct *p)
2488 p->se.exec_start = 0;
2489 p->se.sum_exec_runtime = 0;
2490 p->se.prev_sum_exec_runtime = 0;
2491 p->se.nr_migrations = 0;
2492 p->se.last_wakeup = 0;
2493 p->se.avg_overlap = 0;
2494 p->se.start_runtime = 0;
2495 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2496 p->se.avg_running = 0;
2498 #ifdef CONFIG_SCHEDSTATS
2499 p->se.wait_start = 0;
2501 p->se.wait_count = 0;
2504 p->se.sleep_start = 0;
2505 p->se.sleep_max = 0;
2506 p->se.sum_sleep_runtime = 0;
2508 p->se.block_start = 0;
2509 p->se.block_max = 0;
2511 p->se.slice_max = 0;
2513 p->se.nr_migrations_cold = 0;
2514 p->se.nr_failed_migrations_affine = 0;
2515 p->se.nr_failed_migrations_running = 0;
2516 p->se.nr_failed_migrations_hot = 0;
2517 p->se.nr_forced_migrations = 0;
2518 p->se.nr_forced2_migrations = 0;
2520 p->se.nr_wakeups = 0;
2521 p->se.nr_wakeups_sync = 0;
2522 p->se.nr_wakeups_migrate = 0;
2523 p->se.nr_wakeups_local = 0;
2524 p->se.nr_wakeups_remote = 0;
2525 p->se.nr_wakeups_affine = 0;
2526 p->se.nr_wakeups_affine_attempts = 0;
2527 p->se.nr_wakeups_passive = 0;
2528 p->se.nr_wakeups_idle = 0;
2532 INIT_LIST_HEAD(&p->rt.run_list);
2534 INIT_LIST_HEAD(&p->se.group_node);
2536 #ifdef CONFIG_PREEMPT_NOTIFIERS
2537 INIT_HLIST_HEAD(&p->preempt_notifiers);
2541 * We mark the process as running here, but have not actually
2542 * inserted it onto the runqueue yet. This guarantees that
2543 * nobody will actually run it, and a signal or other external
2544 * event cannot wake it up and insert it on the runqueue either.
2546 p->state = TASK_RUNNING;
2550 * fork()/clone()-time setup:
2552 void sched_fork(struct task_struct *p, int clone_flags)
2554 int cpu = get_cpu();
2559 * Revert to default priority/policy on fork if requested.
2561 if (unlikely(p->sched_reset_on_fork)) {
2562 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2563 p->policy = SCHED_NORMAL;
2564 p->normal_prio = p->static_prio;
2567 if (PRIO_TO_NICE(p->static_prio) < 0) {
2568 p->static_prio = NICE_TO_PRIO(0);
2569 p->normal_prio = p->static_prio;
2574 * We don't need the reset flag anymore after the fork. It has
2575 * fulfilled its duty:
2577 p->sched_reset_on_fork = 0;
2581 * Make sure we do not leak PI boosting priority to the child.
2583 p->prio = current->normal_prio;
2585 if (!rt_prio(p->prio))
2586 p->sched_class = &fair_sched_class;
2588 if (p->sched_class->task_fork)
2589 p->sched_class->task_fork(p);
2592 cpu = select_task_rq(p, SD_BALANCE_FORK, 0);
2594 set_task_cpu(p, cpu);
2596 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2597 if (likely(sched_info_on()))
2598 memset(&p->sched_info, 0, sizeof(p->sched_info));
2600 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2603 #ifdef CONFIG_PREEMPT
2604 /* Want to start with kernel preemption disabled. */
2605 task_thread_info(p)->preempt_count = 1;
2607 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2613 * wake_up_new_task - wake up a newly created task for the first time.
2615 * This function will do some initial scheduler statistics housekeeping
2616 * that must be done for every newly created context, then puts the task
2617 * on the runqueue and wakes it.
2619 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2621 unsigned long flags;
2624 rq = task_rq_lock(p, &flags);
2625 BUG_ON(p->state != TASK_RUNNING);
2626 update_rq_clock(rq);
2627 activate_task(rq, p, 0);
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);
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 (likely(!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 * Update rq->cpu_load[] statistics. This function is usually called every
3022 * scheduler tick (TICK_NSEC).
3024 static void update_cpu_load(struct rq *this_rq)
3026 unsigned long this_load = this_rq->load.weight;
3029 this_rq->nr_load_updates++;
3031 /* Update our load: */
3032 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3033 unsigned long old_load, new_load;
3035 /* scale is effectively 1 << i now, and >> i divides by scale */
3037 old_load = this_rq->cpu_load[i];
3038 new_load = this_load;
3040 * Round up the averaging division if load is increasing. This
3041 * prevents us from getting stuck on 9 if the load is 10, for
3044 if (new_load > old_load)
3045 new_load += scale-1;
3046 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3049 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3050 this_rq->calc_load_update += LOAD_FREQ;
3051 calc_load_account_active(this_rq);
3058 * double_rq_lock - safely lock two runqueues
3060 * Note this does not disable interrupts like task_rq_lock,
3061 * you need to do so manually before calling.
3063 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
3064 __acquires(rq1->lock)
3065 __acquires(rq2->lock)
3067 BUG_ON(!irqs_disabled());
3069 spin_lock(&rq1->lock);
3070 __acquire(rq2->lock); /* Fake it out ;) */
3073 spin_lock(&rq1->lock);
3074 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
3076 spin_lock(&rq2->lock);
3077 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
3080 update_rq_clock(rq1);
3081 update_rq_clock(rq2);
3085 * double_rq_unlock - safely unlock two runqueues
3087 * Note this does not restore interrupts like task_rq_unlock,
3088 * you need to do so manually after calling.
3090 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
3091 __releases(rq1->lock)
3092 __releases(rq2->lock)
3094 spin_unlock(&rq1->lock);
3096 spin_unlock(&rq2->lock);
3098 __release(rq2->lock);
3102 * If dest_cpu is allowed for this process, migrate the task to it.
3103 * This is accomplished by forcing the cpu_allowed mask to only
3104 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3105 * the cpu_allowed mask is restored.
3107 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
3109 struct migration_req req;
3110 unsigned long flags;
3113 rq = task_rq_lock(p, &flags);
3114 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
3115 || unlikely(!cpu_active(dest_cpu)))
3118 /* force the process onto the specified CPU */
3119 if (migrate_task(p, dest_cpu, &req)) {
3120 /* Need to wait for migration thread (might exit: take ref). */
3121 struct task_struct *mt = rq->migration_thread;
3123 get_task_struct(mt);
3124 task_rq_unlock(rq, &flags);
3125 wake_up_process(mt);
3126 put_task_struct(mt);
3127 wait_for_completion(&req.done);
3132 task_rq_unlock(rq, &flags);
3136 * sched_exec - execve() is a valuable balancing opportunity, because at
3137 * this point the task has the smallest effective memory and cache footprint.
3139 void sched_exec(void)
3141 int new_cpu, this_cpu = get_cpu();
3142 new_cpu = select_task_rq(current, SD_BALANCE_EXEC, 0);
3144 if (new_cpu != this_cpu)
3145 sched_migrate_task(current, new_cpu);
3149 * pull_task - move a task from a remote runqueue to the local runqueue.
3150 * Both runqueues must be locked.
3152 static void pull_task(struct rq *src_rq, struct task_struct *p,
3153 struct rq *this_rq, int this_cpu)
3155 deactivate_task(src_rq, p, 0);
3156 set_task_cpu(p, this_cpu);
3157 activate_task(this_rq, p, 0);
3159 * Note that idle threads have a prio of MAX_PRIO, for this test
3160 * to be always true for them.
3162 check_preempt_curr(this_rq, p, 0);
3166 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3169 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3170 struct sched_domain *sd, enum cpu_idle_type idle,
3173 int tsk_cache_hot = 0;
3175 * We do not migrate tasks that are:
3176 * 1) running (obviously), or
3177 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3178 * 3) are cache-hot on their current CPU.
3180 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
3181 schedstat_inc(p, se.nr_failed_migrations_affine);
3186 if (task_running(rq, p)) {
3187 schedstat_inc(p, se.nr_failed_migrations_running);
3192 * Aggressive migration if:
3193 * 1) task is cache cold, or
3194 * 2) too many balance attempts have failed.
3197 tsk_cache_hot = task_hot(p, rq->clock, sd);
3198 if (!tsk_cache_hot ||
3199 sd->nr_balance_failed > sd->cache_nice_tries) {
3200 #ifdef CONFIG_SCHEDSTATS
3201 if (tsk_cache_hot) {
3202 schedstat_inc(sd, lb_hot_gained[idle]);
3203 schedstat_inc(p, se.nr_forced_migrations);
3209 if (tsk_cache_hot) {
3210 schedstat_inc(p, se.nr_failed_migrations_hot);
3216 static unsigned long
3217 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3218 unsigned long max_load_move, struct sched_domain *sd,
3219 enum cpu_idle_type idle, int *all_pinned,
3220 int *this_best_prio, struct rq_iterator *iterator)
3222 int loops = 0, pulled = 0, pinned = 0;
3223 struct task_struct *p;
3224 long rem_load_move = max_load_move;
3226 if (max_load_move == 0)
3232 * Start the load-balancing iterator:
3234 p = iterator->start(iterator->arg);
3236 if (!p || loops++ > sysctl_sched_nr_migrate)
3239 if ((p->se.load.weight >> 1) > rem_load_move ||
3240 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3241 p = iterator->next(iterator->arg);
3245 pull_task(busiest, p, this_rq, this_cpu);
3247 rem_load_move -= p->se.load.weight;
3249 #ifdef CONFIG_PREEMPT
3251 * NEWIDLE balancing is a source of latency, so preemptible kernels
3252 * will stop after the first task is pulled to minimize the critical
3255 if (idle == CPU_NEWLY_IDLE)
3260 * We only want to steal up to the prescribed amount of weighted load.
3262 if (rem_load_move > 0) {
3263 if (p->prio < *this_best_prio)
3264 *this_best_prio = p->prio;
3265 p = iterator->next(iterator->arg);
3270 * Right now, this is one of only two places pull_task() is called,
3271 * so we can safely collect pull_task() stats here rather than
3272 * inside pull_task().
3274 schedstat_add(sd, lb_gained[idle], pulled);
3277 *all_pinned = pinned;
3279 return max_load_move - rem_load_move;
3283 * move_tasks tries to move up to max_load_move weighted load from busiest to
3284 * this_rq, as part of a balancing operation within domain "sd".
3285 * Returns 1 if successful and 0 otherwise.
3287 * Called with both runqueues locked.
3289 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3290 unsigned long max_load_move,
3291 struct sched_domain *sd, enum cpu_idle_type idle,
3294 const struct sched_class *class = sched_class_highest;
3295 unsigned long total_load_moved = 0;
3296 int this_best_prio = this_rq->curr->prio;
3300 class->load_balance(this_rq, this_cpu, busiest,
3301 max_load_move - total_load_moved,
3302 sd, idle, all_pinned, &this_best_prio);
3303 class = class->next;
3305 #ifdef CONFIG_PREEMPT
3307 * NEWIDLE balancing is a source of latency, so preemptible
3308 * kernels will stop after the first task is pulled to minimize
3309 * the critical section.
3311 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3314 } while (class && max_load_move > total_load_moved);
3316 return total_load_moved > 0;
3320 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3321 struct sched_domain *sd, enum cpu_idle_type idle,
3322 struct rq_iterator *iterator)
3324 struct task_struct *p = iterator->start(iterator->arg);
3328 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3329 pull_task(busiest, p, this_rq, this_cpu);
3331 * Right now, this is only the second place pull_task()
3332 * is called, so we can safely collect pull_task()
3333 * stats here rather than inside pull_task().
3335 schedstat_inc(sd, lb_gained[idle]);
3339 p = iterator->next(iterator->arg);
3346 * move_one_task tries to move exactly one task from busiest to this_rq, as
3347 * part of active balancing operations within "domain".
3348 * Returns 1 if successful and 0 otherwise.
3350 * Called with both runqueues locked.
3352 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3353 struct sched_domain *sd, enum cpu_idle_type idle)
3355 const struct sched_class *class;
3357 for_each_class(class) {
3358 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3364 /********** Helpers for find_busiest_group ************************/
3366 * sd_lb_stats - Structure to store the statistics of a sched_domain
3367 * during load balancing.
3369 struct sd_lb_stats {
3370 struct sched_group *busiest; /* Busiest group in this sd */
3371 struct sched_group *this; /* Local group in this sd */
3372 unsigned long total_load; /* Total load of all groups in sd */
3373 unsigned long total_pwr; /* Total power of all groups in sd */
3374 unsigned long avg_load; /* Average load across all groups in sd */
3376 /** Statistics of this group */
3377 unsigned long this_load;
3378 unsigned long this_load_per_task;
3379 unsigned long this_nr_running;
3381 /* Statistics of the busiest group */
3382 unsigned long max_load;
3383 unsigned long busiest_load_per_task;
3384 unsigned long busiest_nr_running;
3386 int group_imb; /* Is there imbalance in this sd */
3387 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3388 int power_savings_balance; /* Is powersave balance needed for this sd */
3389 struct sched_group *group_min; /* Least loaded group in sd */
3390 struct sched_group *group_leader; /* Group which relieves group_min */
3391 unsigned long min_load_per_task; /* load_per_task in group_min */
3392 unsigned long leader_nr_running; /* Nr running of group_leader */
3393 unsigned long min_nr_running; /* Nr running of group_min */
3398 * sg_lb_stats - stats of a sched_group required for load_balancing
3400 struct sg_lb_stats {
3401 unsigned long avg_load; /*Avg load across the CPUs of the group */
3402 unsigned long group_load; /* Total load over the CPUs of the group */
3403 unsigned long sum_nr_running; /* Nr tasks running in the group */
3404 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3405 unsigned long group_capacity;
3406 int group_imb; /* Is there an imbalance in the group ? */
3410 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3411 * @group: The group whose first cpu is to be returned.
3413 static inline unsigned int group_first_cpu(struct sched_group *group)
3415 return cpumask_first(sched_group_cpus(group));
3419 * get_sd_load_idx - Obtain the load index for a given sched domain.
3420 * @sd: The sched_domain whose load_idx is to be obtained.
3421 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3423 static inline int get_sd_load_idx(struct sched_domain *sd,
3424 enum cpu_idle_type idle)
3430 load_idx = sd->busy_idx;
3433 case CPU_NEWLY_IDLE:
3434 load_idx = sd->newidle_idx;
3437 load_idx = sd->idle_idx;
3445 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3447 * init_sd_power_savings_stats - Initialize power savings statistics for
3448 * the given sched_domain, during load balancing.
3450 * @sd: Sched domain whose power-savings statistics are to be initialized.
3451 * @sds: Variable containing the statistics for sd.
3452 * @idle: Idle status of the CPU at which we're performing load-balancing.
3454 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3455 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3458 * Busy processors will not participate in power savings
3461 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3462 sds->power_savings_balance = 0;
3464 sds->power_savings_balance = 1;
3465 sds->min_nr_running = ULONG_MAX;
3466 sds->leader_nr_running = 0;
3471 * update_sd_power_savings_stats - Update the power saving stats for a
3472 * sched_domain while performing load balancing.
3474 * @group: sched_group belonging to the sched_domain under consideration.
3475 * @sds: Variable containing the statistics of the sched_domain
3476 * @local_group: Does group contain the CPU for which we're performing
3478 * @sgs: Variable containing the statistics of the group.
3480 static inline void update_sd_power_savings_stats(struct sched_group *group,
3481 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3484 if (!sds->power_savings_balance)
3488 * If the local group is idle or completely loaded
3489 * no need to do power savings balance at this domain
3491 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3492 !sds->this_nr_running))
3493 sds->power_savings_balance = 0;
3496 * If a group is already running at full capacity or idle,
3497 * don't include that group in power savings calculations
3499 if (!sds->power_savings_balance ||
3500 sgs->sum_nr_running >= sgs->group_capacity ||
3501 !sgs->sum_nr_running)
3505 * Calculate the group which has the least non-idle load.
3506 * This is the group from where we need to pick up the load
3509 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3510 (sgs->sum_nr_running == sds->min_nr_running &&
3511 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3512 sds->group_min = group;
3513 sds->min_nr_running = sgs->sum_nr_running;
3514 sds->min_load_per_task = sgs->sum_weighted_load /
3515 sgs->sum_nr_running;
3519 * Calculate the group which is almost near its
3520 * capacity but still has some space to pick up some load
3521 * from other group and save more power
3523 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
3526 if (sgs->sum_nr_running > sds->leader_nr_running ||
3527 (sgs->sum_nr_running == sds->leader_nr_running &&
3528 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3529 sds->group_leader = group;
3530 sds->leader_nr_running = sgs->sum_nr_running;
3535 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3536 * @sds: Variable containing the statistics of the sched_domain
3537 * under consideration.
3538 * @this_cpu: Cpu at which we're currently performing load-balancing.
3539 * @imbalance: Variable to store the imbalance.
3542 * Check if we have potential to perform some power-savings balance.
3543 * If yes, set the busiest group to be the least loaded group in the
3544 * sched_domain, so that it's CPUs can be put to idle.
3546 * Returns 1 if there is potential to perform power-savings balance.
3549 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3550 int this_cpu, unsigned long *imbalance)
3552 if (!sds->power_savings_balance)
3555 if (sds->this != sds->group_leader ||
3556 sds->group_leader == sds->group_min)
3559 *imbalance = sds->min_load_per_task;
3560 sds->busiest = sds->group_min;
3565 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3566 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3567 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3572 static inline void update_sd_power_savings_stats(struct sched_group *group,
3573 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3578 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3579 int this_cpu, unsigned long *imbalance)
3583 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3586 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
3588 return SCHED_LOAD_SCALE;
3591 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
3593 return default_scale_freq_power(sd, cpu);
3596 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
3598 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3599 unsigned long smt_gain = sd->smt_gain;
3606 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3608 return default_scale_smt_power(sd, cpu);
3611 unsigned long scale_rt_power(int cpu)
3613 struct rq *rq = cpu_rq(cpu);
3614 u64 total, available;
3616 sched_avg_update(rq);
3618 total = sched_avg_period() + (rq->clock - rq->age_stamp);
3619 available = total - rq->rt_avg;
3621 if (unlikely((s64)total < SCHED_LOAD_SCALE))
3622 total = SCHED_LOAD_SCALE;
3624 total >>= SCHED_LOAD_SHIFT;
3626 return div_u64(available, total);
3629 static void update_cpu_power(struct sched_domain *sd, int cpu)
3631 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3632 unsigned long power = SCHED_LOAD_SCALE;
3633 struct sched_group *sdg = sd->groups;
3635 if (sched_feat(ARCH_POWER))
3636 power *= arch_scale_freq_power(sd, cpu);
3638 power *= default_scale_freq_power(sd, cpu);
3640 power >>= SCHED_LOAD_SHIFT;
3642 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
3643 if (sched_feat(ARCH_POWER))
3644 power *= arch_scale_smt_power(sd, cpu);
3646 power *= default_scale_smt_power(sd, cpu);
3648 power >>= SCHED_LOAD_SHIFT;
3651 power *= scale_rt_power(cpu);
3652 power >>= SCHED_LOAD_SHIFT;
3657 sdg->cpu_power = power;
3660 static void update_group_power(struct sched_domain *sd, int cpu)
3662 struct sched_domain *child = sd->child;
3663 struct sched_group *group, *sdg = sd->groups;
3664 unsigned long power;
3667 update_cpu_power(sd, cpu);
3673 group = child->groups;
3675 power += group->cpu_power;
3676 group = group->next;
3677 } while (group != child->groups);
3679 sdg->cpu_power = power;
3683 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3684 * @sd: The sched_domain whose statistics are to be updated.
3685 * @group: sched_group whose statistics are to be updated.
3686 * @this_cpu: Cpu for which load balance is currently performed.
3687 * @idle: Idle status of this_cpu
3688 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3689 * @sd_idle: Idle status of the sched_domain containing group.
3690 * @local_group: Does group contain this_cpu.
3691 * @cpus: Set of cpus considered for load balancing.
3692 * @balance: Should we balance.
3693 * @sgs: variable to hold the statistics for this group.
3695 static inline void update_sg_lb_stats(struct sched_domain *sd,
3696 struct sched_group *group, int this_cpu,
3697 enum cpu_idle_type idle, int load_idx, int *sd_idle,
3698 int local_group, const struct cpumask *cpus,
3699 int *balance, struct sg_lb_stats *sgs)
3701 unsigned long load, max_cpu_load, min_cpu_load;
3703 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3704 unsigned long sum_avg_load_per_task;
3705 unsigned long avg_load_per_task;
3708 balance_cpu = group_first_cpu(group);
3709 if (balance_cpu == this_cpu)
3710 update_group_power(sd, this_cpu);
3713 /* Tally up the load of all CPUs in the group */
3714 sum_avg_load_per_task = avg_load_per_task = 0;
3716 min_cpu_load = ~0UL;
3718 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3719 struct rq *rq = cpu_rq(i);
3721 if (*sd_idle && rq->nr_running)
3724 /* Bias balancing toward cpus of our domain */
3726 if (idle_cpu(i) && !first_idle_cpu) {
3731 load = target_load(i, load_idx);
3733 load = source_load(i, load_idx);
3734 if (load > max_cpu_load)
3735 max_cpu_load = load;
3736 if (min_cpu_load > load)
3737 min_cpu_load = load;
3740 sgs->group_load += load;
3741 sgs->sum_nr_running += rq->nr_running;
3742 sgs->sum_weighted_load += weighted_cpuload(i);
3744 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3748 * First idle cpu or the first cpu(busiest) in this sched group
3749 * is eligible for doing load balancing at this and above
3750 * domains. In the newly idle case, we will allow all the cpu's
3751 * to do the newly idle load balance.
3753 if (idle != CPU_NEWLY_IDLE && local_group &&
3754 balance_cpu != this_cpu && balance) {
3759 /* Adjust by relative CPU power of the group */
3760 sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power;
3764 * Consider the group unbalanced when the imbalance is larger
3765 * than the average weight of two tasks.
3767 * APZ: with cgroup the avg task weight can vary wildly and
3768 * might not be a suitable number - should we keep a
3769 * normalized nr_running number somewhere that negates
3772 avg_load_per_task = (sum_avg_load_per_task * SCHED_LOAD_SCALE) /
3775 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3778 sgs->group_capacity =
3779 DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
3783 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3784 * @sd: sched_domain whose statistics are to be updated.
3785 * @this_cpu: Cpu for which load balance is currently performed.
3786 * @idle: Idle status of this_cpu
3787 * @sd_idle: Idle status of the sched_domain containing group.
3788 * @cpus: Set of cpus considered for load balancing.
3789 * @balance: Should we balance.
3790 * @sds: variable to hold the statistics for this sched_domain.
3792 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3793 enum cpu_idle_type idle, int *sd_idle,
3794 const struct cpumask *cpus, int *balance,
3795 struct sd_lb_stats *sds)
3797 struct sched_domain *child = sd->child;
3798 struct sched_group *group = sd->groups;
3799 struct sg_lb_stats sgs;
3800 int load_idx, prefer_sibling = 0;
3802 if (child && child->flags & SD_PREFER_SIBLING)
3805 init_sd_power_savings_stats(sd, sds, idle);
3806 load_idx = get_sd_load_idx(sd, idle);
3811 local_group = cpumask_test_cpu(this_cpu,
3812 sched_group_cpus(group));
3813 memset(&sgs, 0, sizeof(sgs));
3814 update_sg_lb_stats(sd, group, this_cpu, idle, load_idx, sd_idle,
3815 local_group, cpus, balance, &sgs);
3817 if (local_group && balance && !(*balance))
3820 sds->total_load += sgs.group_load;
3821 sds->total_pwr += group->cpu_power;
3824 * In case the child domain prefers tasks go to siblings
3825 * first, lower the group capacity to one so that we'll try
3826 * and move all the excess tasks away.
3829 sgs.group_capacity = min(sgs.group_capacity, 1UL);
3832 sds->this_load = sgs.avg_load;
3834 sds->this_nr_running = sgs.sum_nr_running;
3835 sds->this_load_per_task = sgs.sum_weighted_load;
3836 } else if (sgs.avg_load > sds->max_load &&
3837 (sgs.sum_nr_running > sgs.group_capacity ||
3839 sds->max_load = sgs.avg_load;
3840 sds->busiest = group;
3841 sds->busiest_nr_running = sgs.sum_nr_running;
3842 sds->busiest_load_per_task = sgs.sum_weighted_load;
3843 sds->group_imb = sgs.group_imb;
3846 update_sd_power_savings_stats(group, sds, local_group, &sgs);
3847 group = group->next;
3848 } while (group != sd->groups);
3852 * fix_small_imbalance - Calculate the minor imbalance that exists
3853 * amongst the groups of a sched_domain, during
3855 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3856 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3857 * @imbalance: Variable to store the imbalance.
3859 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3860 int this_cpu, unsigned long *imbalance)
3862 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3863 unsigned int imbn = 2;
3865 if (sds->this_nr_running) {
3866 sds->this_load_per_task /= sds->this_nr_running;
3867 if (sds->busiest_load_per_task >
3868 sds->this_load_per_task)
3871 sds->this_load_per_task =
3872 cpu_avg_load_per_task(this_cpu);
3874 if (sds->max_load - sds->this_load + sds->busiest_load_per_task >=
3875 sds->busiest_load_per_task * imbn) {
3876 *imbalance = sds->busiest_load_per_task;
3881 * OK, we don't have enough imbalance to justify moving tasks,
3882 * however we may be able to increase total CPU power used by
3886 pwr_now += sds->busiest->cpu_power *
3887 min(sds->busiest_load_per_task, sds->max_load);
3888 pwr_now += sds->this->cpu_power *
3889 min(sds->this_load_per_task, sds->this_load);
3890 pwr_now /= SCHED_LOAD_SCALE;
3892 /* Amount of load we'd subtract */
3893 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3894 sds->busiest->cpu_power;
3895 if (sds->max_load > tmp)
3896 pwr_move += sds->busiest->cpu_power *
3897 min(sds->busiest_load_per_task, sds->max_load - tmp);
3899 /* Amount of load we'd add */
3900 if (sds->max_load * sds->busiest->cpu_power <
3901 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
3902 tmp = (sds->max_load * sds->busiest->cpu_power) /
3903 sds->this->cpu_power;
3905 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3906 sds->this->cpu_power;
3907 pwr_move += sds->this->cpu_power *
3908 min(sds->this_load_per_task, sds->this_load + tmp);
3909 pwr_move /= SCHED_LOAD_SCALE;
3911 /* Move if we gain throughput */
3912 if (pwr_move > pwr_now)
3913 *imbalance = sds->busiest_load_per_task;
3917 * calculate_imbalance - Calculate the amount of imbalance present within the
3918 * groups of a given sched_domain during load balance.
3919 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3920 * @this_cpu: Cpu for which currently load balance is being performed.
3921 * @imbalance: The variable to store the imbalance.
3923 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3924 unsigned long *imbalance)
3926 unsigned long max_pull;
3928 * In the presence of smp nice balancing, certain scenarios can have
3929 * max load less than avg load(as we skip the groups at or below
3930 * its cpu_power, while calculating max_load..)
3932 if (sds->max_load < sds->avg_load) {
3934 return fix_small_imbalance(sds, this_cpu, imbalance);
3937 /* Don't want to pull so many tasks that a group would go idle */
3938 max_pull = min(sds->max_load - sds->avg_load,
3939 sds->max_load - sds->busiest_load_per_task);
3941 /* How much load to actually move to equalise the imbalance */
3942 *imbalance = min(max_pull * sds->busiest->cpu_power,
3943 (sds->avg_load - sds->this_load) * sds->this->cpu_power)
3947 * if *imbalance is less than the average load per runnable task
3948 * there is no gaurantee that any tasks will be moved so we'll have
3949 * a think about bumping its value to force at least one task to be
3952 if (*imbalance < sds->busiest_load_per_task)
3953 return fix_small_imbalance(sds, this_cpu, imbalance);
3956 /******* find_busiest_group() helpers end here *********************/
3959 * find_busiest_group - Returns the busiest group within the sched_domain
3960 * if there is an imbalance. If there isn't an imbalance, and
3961 * the user has opted for power-savings, it returns a group whose
3962 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3963 * such a group exists.
3965 * Also calculates the amount of weighted load which should be moved
3966 * to restore balance.
3968 * @sd: The sched_domain whose busiest group is to be returned.
3969 * @this_cpu: The cpu for which load balancing is currently being performed.
3970 * @imbalance: Variable which stores amount of weighted load which should
3971 * be moved to restore balance/put a group to idle.
3972 * @idle: The idle status of this_cpu.
3973 * @sd_idle: The idleness of sd
3974 * @cpus: The set of CPUs under consideration for load-balancing.
3975 * @balance: Pointer to a variable indicating if this_cpu
3976 * is the appropriate cpu to perform load balancing at this_level.
3978 * Returns: - the busiest group if imbalance exists.
3979 * - If no imbalance and user has opted for power-savings balance,
3980 * return the least loaded group whose CPUs can be
3981 * put to idle by rebalancing its tasks onto our group.
3983 static struct sched_group *
3984 find_busiest_group(struct sched_domain *sd, int this_cpu,
3985 unsigned long *imbalance, enum cpu_idle_type idle,
3986 int *sd_idle, const struct cpumask *cpus, int *balance)
3988 struct sd_lb_stats sds;
3990 memset(&sds, 0, sizeof(sds));
3993 * Compute the various statistics relavent for load balancing at
3996 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
3999 /* Cases where imbalance does not exist from POV of this_cpu */
4000 /* 1) this_cpu is not the appropriate cpu to perform load balancing
4002 * 2) There is no busy sibling group to pull from.
4003 * 3) This group is the busiest group.
4004 * 4) This group is more busy than the avg busieness at this
4006 * 5) The imbalance is within the specified limit.
4007 * 6) Any rebalance would lead to ping-pong
4009 if (balance && !(*balance))
4012 if (!sds.busiest || sds.busiest_nr_running == 0)
4015 if (sds.this_load >= sds.max_load)
4018 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
4020 if (sds.this_load >= sds.avg_load)
4023 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
4026 sds.busiest_load_per_task /= sds.busiest_nr_running;
4028 sds.busiest_load_per_task =
4029 min(sds.busiest_load_per_task, sds.avg_load);
4032 * We're trying to get all the cpus to the average_load, so we don't
4033 * want to push ourselves above the average load, nor do we wish to
4034 * reduce the max loaded cpu below the average load, as either of these
4035 * actions would just result in more rebalancing later, and ping-pong
4036 * tasks around. Thus we look for the minimum possible imbalance.
4037 * Negative imbalances (*we* are more loaded than anyone else) will
4038 * be counted as no imbalance for these purposes -- we can't fix that
4039 * by pulling tasks to us. Be careful of negative numbers as they'll
4040 * appear as very large values with unsigned longs.
4042 if (sds.max_load <= sds.busiest_load_per_task)
4045 /* Looks like there is an imbalance. Compute it */
4046 calculate_imbalance(&sds, this_cpu, imbalance);
4051 * There is no obvious imbalance. But check if we can do some balancing
4054 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
4062 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4065 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
4066 unsigned long imbalance, const struct cpumask *cpus)
4068 struct rq *busiest = NULL, *rq;
4069 unsigned long max_load = 0;
4072 for_each_cpu(i, sched_group_cpus(group)) {
4073 unsigned long power = power_of(i);
4074 unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
4077 if (!cpumask_test_cpu(i, cpus))
4081 wl = weighted_cpuload(i) * SCHED_LOAD_SCALE;
4084 if (capacity && rq->nr_running == 1 && wl > imbalance)
4087 if (wl > max_load) {
4097 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4098 * so long as it is large enough.
4100 #define MAX_PINNED_INTERVAL 512
4102 /* Working cpumask for load_balance and load_balance_newidle. */
4103 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4106 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4107 * tasks if there is an imbalance.
4109 static int load_balance(int this_cpu, struct rq *this_rq,
4110 struct sched_domain *sd, enum cpu_idle_type idle,
4113 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
4114 struct sched_group *group;
4115 unsigned long imbalance;
4117 unsigned long flags;
4118 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4120 cpumask_copy(cpus, cpu_active_mask);
4123 * When power savings policy is enabled for the parent domain, idle
4124 * sibling can pick up load irrespective of busy siblings. In this case,
4125 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4126 * portraying it as CPU_NOT_IDLE.
4128 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
4129 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4132 schedstat_inc(sd, lb_count[idle]);
4136 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
4143 schedstat_inc(sd, lb_nobusyg[idle]);
4147 busiest = find_busiest_queue(group, idle, imbalance, cpus);
4149 schedstat_inc(sd, lb_nobusyq[idle]);
4153 BUG_ON(busiest == this_rq);
4155 schedstat_add(sd, lb_imbalance[idle], imbalance);
4158 if (busiest->nr_running > 1) {
4160 * Attempt to move tasks. If find_busiest_group has found
4161 * an imbalance but busiest->nr_running <= 1, the group is
4162 * still unbalanced. ld_moved simply stays zero, so it is
4163 * correctly treated as an imbalance.
4165 local_irq_save(flags);
4166 double_rq_lock(this_rq, busiest);
4167 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4168 imbalance, sd, idle, &all_pinned);
4169 double_rq_unlock(this_rq, busiest);
4170 local_irq_restore(flags);
4173 * some other cpu did the load balance for us.
4175 if (ld_moved && this_cpu != smp_processor_id())
4176 resched_cpu(this_cpu);
4178 /* All tasks on this runqueue were pinned by CPU affinity */
4179 if (unlikely(all_pinned)) {
4180 cpumask_clear_cpu(cpu_of(busiest), cpus);
4181 if (!cpumask_empty(cpus))
4188 schedstat_inc(sd, lb_failed[idle]);
4189 sd->nr_balance_failed++;
4191 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
4193 spin_lock_irqsave(&busiest->lock, flags);
4195 /* don't kick the migration_thread, if the curr
4196 * task on busiest cpu can't be moved to this_cpu
4198 if (!cpumask_test_cpu(this_cpu,
4199 &busiest->curr->cpus_allowed)) {
4200 spin_unlock_irqrestore(&busiest->lock, flags);
4202 goto out_one_pinned;
4205 if (!busiest->active_balance) {
4206 busiest->active_balance = 1;
4207 busiest->push_cpu = this_cpu;
4210 spin_unlock_irqrestore(&busiest->lock, flags);
4212 wake_up_process(busiest->migration_thread);
4215 * We've kicked active balancing, reset the failure
4218 sd->nr_balance_failed = sd->cache_nice_tries+1;
4221 sd->nr_balance_failed = 0;
4223 if (likely(!active_balance)) {
4224 /* We were unbalanced, so reset the balancing interval */
4225 sd->balance_interval = sd->min_interval;
4228 * If we've begun active balancing, start to back off. This
4229 * case may not be covered by the all_pinned logic if there
4230 * is only 1 task on the busy runqueue (because we don't call
4233 if (sd->balance_interval < sd->max_interval)
4234 sd->balance_interval *= 2;
4237 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4238 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4244 schedstat_inc(sd, lb_balanced[idle]);
4246 sd->nr_balance_failed = 0;
4249 /* tune up the balancing interval */
4250 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4251 (sd->balance_interval < sd->max_interval))
4252 sd->balance_interval *= 2;
4254 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4255 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4266 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4267 * tasks if there is an imbalance.
4269 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4270 * this_rq is locked.
4273 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
4275 struct sched_group *group;
4276 struct rq *busiest = NULL;
4277 unsigned long imbalance;
4281 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4283 cpumask_copy(cpus, cpu_active_mask);
4286 * When power savings policy is enabled for the parent domain, idle
4287 * sibling can pick up load irrespective of busy siblings. In this case,
4288 * let the state of idle sibling percolate up as IDLE, instead of
4289 * portraying it as CPU_NOT_IDLE.
4291 if (sd->flags & SD_SHARE_CPUPOWER &&
4292 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4295 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
4297 update_shares_locked(this_rq, sd);
4298 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
4299 &sd_idle, cpus, NULL);
4301 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
4305 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
4307 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
4311 BUG_ON(busiest == this_rq);
4313 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
4316 if (busiest->nr_running > 1) {
4317 /* Attempt to move tasks */
4318 double_lock_balance(this_rq, busiest);
4319 /* this_rq->clock is already updated */
4320 update_rq_clock(busiest);
4321 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4322 imbalance, sd, CPU_NEWLY_IDLE,
4324 double_unlock_balance(this_rq, busiest);
4326 if (unlikely(all_pinned)) {
4327 cpumask_clear_cpu(cpu_of(busiest), cpus);
4328 if (!cpumask_empty(cpus))
4334 int active_balance = 0;
4336 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
4337 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4338 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4341 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4344 if (sd->nr_balance_failed++ < 2)
4348 * The only task running in a non-idle cpu can be moved to this
4349 * cpu in an attempt to completely freeup the other CPU
4350 * package. The same method used to move task in load_balance()
4351 * have been extended for load_balance_newidle() to speedup
4352 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4354 * The package power saving logic comes from
4355 * find_busiest_group(). If there are no imbalance, then
4356 * f_b_g() will return NULL. However when sched_mc={1,2} then
4357 * f_b_g() will select a group from which a running task may be
4358 * pulled to this cpu in order to make the other package idle.
4359 * If there is no opportunity to make a package idle and if
4360 * there are no imbalance, then f_b_g() will return NULL and no
4361 * action will be taken in load_balance_newidle().
4363 * Under normal task pull operation due to imbalance, there
4364 * will be more than one task in the source run queue and
4365 * move_tasks() will succeed. ld_moved will be true and this
4366 * active balance code will not be triggered.
4369 /* Lock busiest in correct order while this_rq is held */
4370 double_lock_balance(this_rq, busiest);
4373 * don't kick the migration_thread, if the curr
4374 * task on busiest cpu can't be moved to this_cpu
4376 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
4377 double_unlock_balance(this_rq, busiest);
4382 if (!busiest->active_balance) {
4383 busiest->active_balance = 1;
4384 busiest->push_cpu = this_cpu;
4388 double_unlock_balance(this_rq, busiest);
4390 * Should not call ttwu while holding a rq->lock
4392 spin_unlock(&this_rq->lock);
4394 wake_up_process(busiest->migration_thread);
4395 spin_lock(&this_rq->lock);
4398 sd->nr_balance_failed = 0;
4400 update_shares_locked(this_rq, sd);
4404 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
4405 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4406 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4408 sd->nr_balance_failed = 0;
4414 * idle_balance is called by schedule() if this_cpu is about to become
4415 * idle. Attempts to pull tasks from other CPUs.
4417 static void idle_balance(int this_cpu, struct rq *this_rq)
4419 struct sched_domain *sd;
4420 int pulled_task = 0;
4421 unsigned long next_balance = jiffies + HZ;
4423 this_rq->idle_stamp = this_rq->clock;
4425 if (this_rq->avg_idle < sysctl_sched_migration_cost)
4428 for_each_domain(this_cpu, sd) {
4429 unsigned long interval;
4431 if (!(sd->flags & SD_LOAD_BALANCE))
4434 if (sd->flags & SD_BALANCE_NEWIDLE)
4435 /* If we've pulled tasks over stop searching: */
4436 pulled_task = load_balance_newidle(this_cpu, this_rq,
4439 interval = msecs_to_jiffies(sd->balance_interval);
4440 if (time_after(next_balance, sd->last_balance + interval))
4441 next_balance = sd->last_balance + interval;
4443 this_rq->idle_stamp = 0;
4447 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4449 * We are going idle. next_balance may be set based on
4450 * a busy processor. So reset next_balance.
4452 this_rq->next_balance = next_balance;
4457 * active_load_balance is run by migration threads. It pushes running tasks
4458 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4459 * running on each physical CPU where possible, and avoids physical /
4460 * logical imbalances.
4462 * Called with busiest_rq locked.
4464 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
4466 int target_cpu = busiest_rq->push_cpu;
4467 struct sched_domain *sd;
4468 struct rq *target_rq;
4470 /* Is there any task to move? */
4471 if (busiest_rq->nr_running <= 1)
4474 target_rq = cpu_rq(target_cpu);
4477 * This condition is "impossible", if it occurs
4478 * we need to fix it. Originally reported by
4479 * Bjorn Helgaas on a 128-cpu setup.
4481 BUG_ON(busiest_rq == target_rq);
4483 /* move a task from busiest_rq to target_rq */
4484 double_lock_balance(busiest_rq, target_rq);
4485 update_rq_clock(busiest_rq);
4486 update_rq_clock(target_rq);
4488 /* Search for an sd spanning us and the target CPU. */
4489 for_each_domain(target_cpu, sd) {
4490 if ((sd->flags & SD_LOAD_BALANCE) &&
4491 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4496 schedstat_inc(sd, alb_count);
4498 if (move_one_task(target_rq, target_cpu, busiest_rq,
4500 schedstat_inc(sd, alb_pushed);
4502 schedstat_inc(sd, alb_failed);
4504 double_unlock_balance(busiest_rq, target_rq);
4509 atomic_t load_balancer;
4510 cpumask_var_t cpu_mask;
4511 cpumask_var_t ilb_grp_nohz_mask;
4512 } nohz ____cacheline_aligned = {
4513 .load_balancer = ATOMIC_INIT(-1),
4516 int get_nohz_load_balancer(void)
4518 return atomic_read(&nohz.load_balancer);
4521 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4523 * lowest_flag_domain - Return lowest sched_domain containing flag.
4524 * @cpu: The cpu whose lowest level of sched domain is to
4526 * @flag: The flag to check for the lowest sched_domain
4527 * for the given cpu.
4529 * Returns the lowest sched_domain of a cpu which contains the given flag.
4531 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4533 struct sched_domain *sd;
4535 for_each_domain(cpu, sd)
4536 if (sd && (sd->flags & flag))
4543 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4544 * @cpu: The cpu whose domains we're iterating over.
4545 * @sd: variable holding the value of the power_savings_sd
4547 * @flag: The flag to filter the sched_domains to be iterated.
4549 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4550 * set, starting from the lowest sched_domain to the highest.
4552 #define for_each_flag_domain(cpu, sd, flag) \
4553 for (sd = lowest_flag_domain(cpu, flag); \
4554 (sd && (sd->flags & flag)); sd = sd->parent)
4557 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4558 * @ilb_group: group to be checked for semi-idleness
4560 * Returns: 1 if the group is semi-idle. 0 otherwise.
4562 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4563 * and atleast one non-idle CPU. This helper function checks if the given
4564 * sched_group is semi-idle or not.
4566 static inline int is_semi_idle_group(struct sched_group *ilb_group)
4568 cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask,
4569 sched_group_cpus(ilb_group));
4572 * A sched_group is semi-idle when it has atleast one busy cpu
4573 * and atleast one idle cpu.
4575 if (cpumask_empty(nohz.ilb_grp_nohz_mask))
4578 if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group)))
4584 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4585 * @cpu: The cpu which is nominating a new idle_load_balancer.
4587 * Returns: Returns the id of the idle load balancer if it exists,
4588 * Else, returns >= nr_cpu_ids.
4590 * This algorithm picks the idle load balancer such that it belongs to a
4591 * semi-idle powersavings sched_domain. The idea is to try and avoid
4592 * completely idle packages/cores just for the purpose of idle load balancing
4593 * when there are other idle cpu's which are better suited for that job.
4595 static int find_new_ilb(int cpu)
4597 struct sched_domain *sd;
4598 struct sched_group *ilb_group;
4601 * Have idle load balancer selection from semi-idle packages only
4602 * when power-aware load balancing is enabled
4604 if (!(sched_smt_power_savings || sched_mc_power_savings))
4608 * Optimize for the case when we have no idle CPUs or only one
4609 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4611 if (cpumask_weight(nohz.cpu_mask) < 2)
4614 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4615 ilb_group = sd->groups;
4618 if (is_semi_idle_group(ilb_group))
4619 return cpumask_first(nohz.ilb_grp_nohz_mask);
4621 ilb_group = ilb_group->next;
4623 } while (ilb_group != sd->groups);
4627 return cpumask_first(nohz.cpu_mask);
4629 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4630 static inline int find_new_ilb(int call_cpu)
4632 return cpumask_first(nohz.cpu_mask);
4637 * This routine will try to nominate the ilb (idle load balancing)
4638 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4639 * load balancing on behalf of all those cpus. If all the cpus in the system
4640 * go into this tickless mode, then there will be no ilb owner (as there is
4641 * no need for one) and all the cpus will sleep till the next wakeup event
4644 * For the ilb owner, tick is not stopped. And this tick will be used
4645 * for idle load balancing. ilb owner will still be part of
4648 * While stopping the tick, this cpu will become the ilb owner if there
4649 * is no other owner. And will be the owner till that cpu becomes busy
4650 * or if all cpus in the system stop their ticks at which point
4651 * there is no need for ilb owner.
4653 * When the ilb owner becomes busy, it nominates another owner, during the
4654 * next busy scheduler_tick()
4656 int select_nohz_load_balancer(int stop_tick)
4658 int cpu = smp_processor_id();
4661 cpu_rq(cpu)->in_nohz_recently = 1;
4663 if (!cpu_active(cpu)) {
4664 if (atomic_read(&nohz.load_balancer) != cpu)
4668 * If we are going offline and still the leader,
4671 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4677 cpumask_set_cpu(cpu, nohz.cpu_mask);
4679 /* time for ilb owner also to sleep */
4680 if (cpumask_weight(nohz.cpu_mask) == num_active_cpus()) {
4681 if (atomic_read(&nohz.load_balancer) == cpu)
4682 atomic_set(&nohz.load_balancer, -1);
4686 if (atomic_read(&nohz.load_balancer) == -1) {
4687 /* make me the ilb owner */
4688 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4690 } else if (atomic_read(&nohz.load_balancer) == cpu) {
4693 if (!(sched_smt_power_savings ||
4694 sched_mc_power_savings))
4697 * Check to see if there is a more power-efficient
4700 new_ilb = find_new_ilb(cpu);
4701 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4702 atomic_set(&nohz.load_balancer, -1);
4703 resched_cpu(new_ilb);
4709 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4712 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4714 if (atomic_read(&nohz.load_balancer) == cpu)
4715 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4722 static DEFINE_SPINLOCK(balancing);
4725 * It checks each scheduling domain to see if it is due to be balanced,
4726 * and initiates a balancing operation if so.
4728 * Balancing parameters are set up in arch_init_sched_domains.
4730 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4733 struct rq *rq = cpu_rq(cpu);
4734 unsigned long interval;
4735 struct sched_domain *sd;
4736 /* Earliest time when we have to do rebalance again */
4737 unsigned long next_balance = jiffies + 60*HZ;
4738 int update_next_balance = 0;
4741 for_each_domain(cpu, sd) {
4742 if (!(sd->flags & SD_LOAD_BALANCE))
4745 interval = sd->balance_interval;
4746 if (idle != CPU_IDLE)
4747 interval *= sd->busy_factor;
4749 /* scale ms to jiffies */
4750 interval = msecs_to_jiffies(interval);
4751 if (unlikely(!interval))
4753 if (interval > HZ*NR_CPUS/10)
4754 interval = HZ*NR_CPUS/10;
4756 need_serialize = sd->flags & SD_SERIALIZE;
4758 if (need_serialize) {
4759 if (!spin_trylock(&balancing))
4763 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4764 if (load_balance(cpu, rq, sd, idle, &balance)) {
4766 * We've pulled tasks over so either we're no
4767 * longer idle, or one of our SMT siblings is
4770 idle = CPU_NOT_IDLE;
4772 sd->last_balance = jiffies;
4775 spin_unlock(&balancing);
4777 if (time_after(next_balance, sd->last_balance + interval)) {
4778 next_balance = sd->last_balance + interval;
4779 update_next_balance = 1;
4783 * Stop the load balance at this level. There is another
4784 * CPU in our sched group which is doing load balancing more
4792 * next_balance will be updated only when there is a need.
4793 * When the cpu is attached to null domain for ex, it will not be
4796 if (likely(update_next_balance))
4797 rq->next_balance = next_balance;
4801 * run_rebalance_domains is triggered when needed from the scheduler tick.
4802 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4803 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4805 static void run_rebalance_domains(struct softirq_action *h)
4807 int this_cpu = smp_processor_id();
4808 struct rq *this_rq = cpu_rq(this_cpu);
4809 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4810 CPU_IDLE : CPU_NOT_IDLE;
4812 rebalance_domains(this_cpu, idle);
4816 * If this cpu is the owner for idle load balancing, then do the
4817 * balancing on behalf of the other idle cpus whose ticks are
4820 if (this_rq->idle_at_tick &&
4821 atomic_read(&nohz.load_balancer) == this_cpu) {
4825 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4826 if (balance_cpu == this_cpu)
4830 * If this cpu gets work to do, stop the load balancing
4831 * work being done for other cpus. Next load
4832 * balancing owner will pick it up.
4837 rebalance_domains(balance_cpu, CPU_IDLE);
4839 rq = cpu_rq(balance_cpu);
4840 if (time_after(this_rq->next_balance, rq->next_balance))
4841 this_rq->next_balance = rq->next_balance;
4847 static inline int on_null_domain(int cpu)
4849 return !rcu_dereference(cpu_rq(cpu)->sd);
4853 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4855 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4856 * idle load balancing owner or decide to stop the periodic load balancing,
4857 * if the whole system is idle.
4859 static inline void trigger_load_balance(struct rq *rq, int cpu)
4863 * If we were in the nohz mode recently and busy at the current
4864 * scheduler tick, then check if we need to nominate new idle
4867 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4868 rq->in_nohz_recently = 0;
4870 if (atomic_read(&nohz.load_balancer) == cpu) {
4871 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4872 atomic_set(&nohz.load_balancer, -1);
4875 if (atomic_read(&nohz.load_balancer) == -1) {
4876 int ilb = find_new_ilb(cpu);
4878 if (ilb < nr_cpu_ids)
4884 * If this cpu is idle and doing idle load balancing for all the
4885 * cpus with ticks stopped, is it time for that to stop?
4887 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4888 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4894 * If this cpu is idle and the idle load balancing is done by
4895 * someone else, then no need raise the SCHED_SOFTIRQ
4897 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4898 cpumask_test_cpu(cpu, nohz.cpu_mask))
4901 /* Don't need to rebalance while attached to NULL domain */
4902 if (time_after_eq(jiffies, rq->next_balance) &&
4903 likely(!on_null_domain(cpu)))
4904 raise_softirq(SCHED_SOFTIRQ);
4907 #else /* CONFIG_SMP */
4910 * on UP we do not need to balance between CPUs:
4912 static inline void idle_balance(int cpu, struct rq *rq)
4918 DEFINE_PER_CPU(struct kernel_stat, kstat);
4920 EXPORT_PER_CPU_SYMBOL(kstat);
4923 * Return any ns on the sched_clock that have not yet been accounted in
4924 * @p in case that task is currently running.
4926 * Called with task_rq_lock() held on @rq.
4928 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
4932 if (task_current(rq, p)) {
4933 update_rq_clock(rq);
4934 ns = rq->clock - p->se.exec_start;
4942 unsigned long long task_delta_exec(struct task_struct *p)
4944 unsigned long flags;
4948 rq = task_rq_lock(p, &flags);
4949 ns = do_task_delta_exec(p, rq);
4950 task_rq_unlock(rq, &flags);
4956 * Return accounted runtime for the task.
4957 * In case the task is currently running, return the runtime plus current's
4958 * pending runtime that have not been accounted yet.
4960 unsigned long long task_sched_runtime(struct task_struct *p)
4962 unsigned long flags;
4966 rq = task_rq_lock(p, &flags);
4967 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
4968 task_rq_unlock(rq, &flags);
4974 * Return sum_exec_runtime for the thread group.
4975 * In case the task is currently running, return the sum plus current's
4976 * pending runtime that have not been accounted yet.
4978 * Note that the thread group might have other running tasks as well,
4979 * so the return value not includes other pending runtime that other
4980 * running tasks might have.
4982 unsigned long long thread_group_sched_runtime(struct task_struct *p)
4984 struct task_cputime totals;
4985 unsigned long flags;
4989 rq = task_rq_lock(p, &flags);
4990 thread_group_cputime(p, &totals);
4991 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
4992 task_rq_unlock(rq, &flags);
4998 * Account user cpu time to a process.
4999 * @p: the process that the cpu time gets accounted to
5000 * @cputime: the cpu time spent in user space since the last update
5001 * @cputime_scaled: cputime scaled by cpu frequency
5003 void account_user_time(struct task_struct *p, cputime_t cputime,
5004 cputime_t cputime_scaled)
5006 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5009 /* Add user time to process. */
5010 p->utime = cputime_add(p->utime, cputime);
5011 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5012 account_group_user_time(p, cputime);
5014 /* Add user time to cpustat. */
5015 tmp = cputime_to_cputime64(cputime);
5016 if (TASK_NICE(p) > 0)
5017 cpustat->nice = cputime64_add(cpustat->nice, tmp);
5019 cpustat->user = cputime64_add(cpustat->user, tmp);
5021 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
5022 /* Account for user time used */
5023 acct_update_integrals(p);
5027 * Account guest cpu time to a process.
5028 * @p: the process that the cpu time gets accounted to
5029 * @cputime: the cpu time spent in virtual machine since the last update
5030 * @cputime_scaled: cputime scaled by cpu frequency
5032 static void account_guest_time(struct task_struct *p, cputime_t cputime,
5033 cputime_t cputime_scaled)
5036 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5038 tmp = cputime_to_cputime64(cputime);
5040 /* Add guest time to process. */
5041 p->utime = cputime_add(p->utime, cputime);
5042 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5043 account_group_user_time(p, cputime);
5044 p->gtime = cputime_add(p->gtime, cputime);
5046 /* Add guest time to cpustat. */
5047 if (TASK_NICE(p) > 0) {
5048 cpustat->nice = cputime64_add(cpustat->nice, tmp);
5049 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
5051 cpustat->user = cputime64_add(cpustat->user, tmp);
5052 cpustat->guest = cputime64_add(cpustat->guest, tmp);
5057 * Account system cpu time to a process.
5058 * @p: the process that the cpu time gets accounted to
5059 * @hardirq_offset: the offset to subtract from hardirq_count()
5060 * @cputime: the cpu time spent in kernel space since the last update
5061 * @cputime_scaled: cputime scaled by cpu frequency
5063 void account_system_time(struct task_struct *p, int hardirq_offset,
5064 cputime_t cputime, cputime_t cputime_scaled)
5066 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5069 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
5070 account_guest_time(p, cputime, cputime_scaled);
5074 /* Add system time to process. */
5075 p->stime = cputime_add(p->stime, cputime);
5076 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
5077 account_group_system_time(p, cputime);
5079 /* Add system time to cpustat. */
5080 tmp = cputime_to_cputime64(cputime);
5081 if (hardirq_count() - hardirq_offset)
5082 cpustat->irq = cputime64_add(cpustat->irq, tmp);
5083 else if (softirq_count())
5084 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
5086 cpustat->system = cputime64_add(cpustat->system, tmp);
5088 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
5090 /* Account for system time used */
5091 acct_update_integrals(p);
5095 * Account for involuntary wait time.
5096 * @steal: the cpu time spent in involuntary wait
5098 void account_steal_time(cputime_t cputime)
5100 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5101 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5103 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
5107 * Account for idle time.
5108 * @cputime: the cpu time spent in idle wait
5110 void account_idle_time(cputime_t cputime)
5112 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5113 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5114 struct rq *rq = this_rq();
5116 if (atomic_read(&rq->nr_iowait) > 0)
5117 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
5119 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
5122 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
5125 * Account a single tick of cpu time.
5126 * @p: the process that the cpu time gets accounted to
5127 * @user_tick: indicates if the tick is a user or a system tick
5129 void account_process_tick(struct task_struct *p, int user_tick)
5131 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
5132 struct rq *rq = this_rq();
5135 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
5136 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
5137 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
5140 account_idle_time(cputime_one_jiffy);
5144 * Account multiple ticks of steal time.
5145 * @p: the process from which the cpu time has been stolen
5146 * @ticks: number of stolen ticks
5148 void account_steal_ticks(unsigned long ticks)
5150 account_steal_time(jiffies_to_cputime(ticks));
5154 * Account multiple ticks of idle time.
5155 * @ticks: number of stolen ticks
5157 void account_idle_ticks(unsigned long ticks)
5159 account_idle_time(jiffies_to_cputime(ticks));
5165 * Use precise platform statistics if available:
5167 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
5168 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5174 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5176 struct task_cputime cputime;
5178 thread_group_cputime(p, &cputime);
5180 *ut = cputime.utime;
5181 *st = cputime.stime;
5185 #ifndef nsecs_to_cputime
5186 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
5189 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5191 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
5194 * Use CFS's precise accounting:
5196 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
5201 temp = (u64)(rtime * utime);
5202 do_div(temp, total);
5203 utime = (cputime_t)temp;
5208 * Compare with previous values, to keep monotonicity:
5210 p->prev_utime = max(p->prev_utime, utime);
5211 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
5213 *ut = p->prev_utime;
5214 *st = p->prev_stime;
5218 * Must be called with siglock held.
5220 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5222 struct signal_struct *sig = p->signal;
5223 struct task_cputime cputime;
5224 cputime_t rtime, utime, total;
5226 thread_group_cputime(p, &cputime);
5228 total = cputime_add(cputime.utime, cputime.stime);
5229 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
5234 temp = (u64)(rtime * cputime.utime);
5235 do_div(temp, total);
5236 utime = (cputime_t)temp;
5240 sig->prev_utime = max(sig->prev_utime, utime);
5241 sig->prev_stime = max(sig->prev_stime,
5242 cputime_sub(rtime, sig->prev_utime));
5244 *ut = sig->prev_utime;
5245 *st = sig->prev_stime;
5250 * This function gets called by the timer code, with HZ frequency.
5251 * We call it with interrupts disabled.
5253 * It also gets called by the fork code, when changing the parent's
5256 void scheduler_tick(void)
5258 int cpu = smp_processor_id();
5259 struct rq *rq = cpu_rq(cpu);
5260 struct task_struct *curr = rq->curr;
5264 spin_lock(&rq->lock);
5265 update_rq_clock(rq);
5266 update_cpu_load(rq);
5267 curr->sched_class->task_tick(rq, curr, 0);
5268 spin_unlock(&rq->lock);
5270 perf_event_task_tick(curr, cpu);
5273 rq->idle_at_tick = idle_cpu(cpu);
5274 trigger_load_balance(rq, cpu);
5278 notrace unsigned long get_parent_ip(unsigned long addr)
5280 if (in_lock_functions(addr)) {
5281 addr = CALLER_ADDR2;
5282 if (in_lock_functions(addr))
5283 addr = CALLER_ADDR3;
5288 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5289 defined(CONFIG_PREEMPT_TRACER))
5291 void __kprobes add_preempt_count(int val)
5293 #ifdef CONFIG_DEBUG_PREEMPT
5297 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5300 preempt_count() += val;
5301 #ifdef CONFIG_DEBUG_PREEMPT
5303 * Spinlock count overflowing soon?
5305 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5308 if (preempt_count() == val)
5309 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5311 EXPORT_SYMBOL(add_preempt_count);
5313 void __kprobes sub_preempt_count(int val)
5315 #ifdef CONFIG_DEBUG_PREEMPT
5319 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5322 * Is the spinlock portion underflowing?
5324 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5325 !(preempt_count() & PREEMPT_MASK)))
5329 if (preempt_count() == val)
5330 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5331 preempt_count() -= val;
5333 EXPORT_SYMBOL(sub_preempt_count);
5338 * Print scheduling while atomic bug:
5340 static noinline void __schedule_bug(struct task_struct *prev)
5342 struct pt_regs *regs = get_irq_regs();
5344 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5345 prev->comm, prev->pid, preempt_count());
5347 debug_show_held_locks(prev);
5349 if (irqs_disabled())
5350 print_irqtrace_events(prev);
5359 * Various schedule()-time debugging checks and statistics:
5361 static inline void schedule_debug(struct task_struct *prev)
5364 * Test if we are atomic. Since do_exit() needs to call into
5365 * schedule() atomically, we ignore that path for now.
5366 * Otherwise, whine if we are scheduling when we should not be.
5368 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
5369 __schedule_bug(prev);
5371 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5373 schedstat_inc(this_rq(), sched_count);
5374 #ifdef CONFIG_SCHEDSTATS
5375 if (unlikely(prev->lock_depth >= 0)) {
5376 schedstat_inc(this_rq(), bkl_count);
5377 schedstat_inc(prev, sched_info.bkl_count);
5382 static void put_prev_task(struct rq *rq, struct task_struct *p)
5384 u64 runtime = p->se.sum_exec_runtime - p->se.prev_sum_exec_runtime;
5386 update_avg(&p->se.avg_running, runtime);
5388 if (p->state == TASK_RUNNING) {
5390 * In order to avoid avg_overlap growing stale when we are
5391 * indeed overlapping and hence not getting put to sleep, grow
5392 * the avg_overlap on preemption.
5394 * We use the average preemption runtime because that
5395 * correlates to the amount of cache footprint a task can
5398 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
5399 update_avg(&p->se.avg_overlap, runtime);
5401 update_avg(&p->se.avg_running, 0);
5403 p->sched_class->put_prev_task(rq, p);
5407 * Pick up the highest-prio task:
5409 static inline struct task_struct *
5410 pick_next_task(struct rq *rq)
5412 const struct sched_class *class;
5413 struct task_struct *p;
5416 * Optimization: we know that if all tasks are in
5417 * the fair class we can call that function directly:
5419 if (likely(rq->nr_running == rq->cfs.nr_running)) {
5420 p = fair_sched_class.pick_next_task(rq);
5425 class = sched_class_highest;
5427 p = class->pick_next_task(rq);
5431 * Will never be NULL as the idle class always
5432 * returns a non-NULL p:
5434 class = class->next;
5439 * schedule() is the main scheduler function.
5441 asmlinkage void __sched schedule(void)
5443 struct task_struct *prev, *next;
5444 unsigned long *switch_count;
5450 cpu = smp_processor_id();
5454 switch_count = &prev->nivcsw;
5456 release_kernel_lock(prev);
5457 need_resched_nonpreemptible:
5459 schedule_debug(prev);
5461 if (sched_feat(HRTICK))
5464 spin_lock_irq(&rq->lock);
5465 update_rq_clock(rq);
5466 clear_tsk_need_resched(prev);
5468 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
5469 if (unlikely(signal_pending_state(prev->state, prev)))
5470 prev->state = TASK_RUNNING;
5472 deactivate_task(rq, prev, 1);
5473 switch_count = &prev->nvcsw;
5476 pre_schedule(rq, prev);
5478 if (unlikely(!rq->nr_running))
5479 idle_balance(cpu, rq);
5481 put_prev_task(rq, prev);
5482 next = pick_next_task(rq);
5484 if (likely(prev != next)) {
5485 sched_info_switch(prev, next);
5486 perf_event_task_sched_out(prev, next, cpu);
5492 context_switch(rq, prev, next); /* unlocks the rq */
5494 * the context switch might have flipped the stack from under
5495 * us, hence refresh the local variables.
5497 cpu = smp_processor_id();
5500 spin_unlock_irq(&rq->lock);
5504 if (unlikely(reacquire_kernel_lock(current) < 0))
5505 goto need_resched_nonpreemptible;
5507 preempt_enable_no_resched();
5511 EXPORT_SYMBOL(schedule);
5513 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
5515 * Look out! "owner" is an entirely speculative pointer
5516 * access and not reliable.
5518 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5523 if (!sched_feat(OWNER_SPIN))
5526 #ifdef CONFIG_DEBUG_PAGEALLOC
5528 * Need to access the cpu field knowing that
5529 * DEBUG_PAGEALLOC could have unmapped it if
5530 * the mutex owner just released it and exited.
5532 if (probe_kernel_address(&owner->cpu, cpu))
5539 * Even if the access succeeded (likely case),
5540 * the cpu field may no longer be valid.
5542 if (cpu >= nr_cpumask_bits)
5546 * We need to validate that we can do a
5547 * get_cpu() and that we have the percpu area.
5549 if (!cpu_online(cpu))
5556 * Owner changed, break to re-assess state.
5558 if (lock->owner != owner)
5562 * Is that owner really running on that cpu?
5564 if (task_thread_info(rq->curr) != owner || need_resched())
5574 #ifdef CONFIG_PREEMPT
5576 * this is the entry point to schedule() from in-kernel preemption
5577 * off of preempt_enable. Kernel preemptions off return from interrupt
5578 * occur there and call schedule directly.
5580 asmlinkage void __sched preempt_schedule(void)
5582 struct thread_info *ti = current_thread_info();
5585 * If there is a non-zero preempt_count or interrupts are disabled,
5586 * we do not want to preempt the current task. Just return..
5588 if (likely(ti->preempt_count || irqs_disabled()))
5592 add_preempt_count(PREEMPT_ACTIVE);
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());
5603 EXPORT_SYMBOL(preempt_schedule);
5606 * this is the entry point to schedule() from kernel preemption
5607 * off of irq context.
5608 * Note, that this is called and return with irqs disabled. This will
5609 * protect us against recursive calling from irq.
5611 asmlinkage void __sched preempt_schedule_irq(void)
5613 struct thread_info *ti = current_thread_info();
5615 /* Catch callers which need to be fixed */
5616 BUG_ON(ti->preempt_count || !irqs_disabled());
5619 add_preempt_count(PREEMPT_ACTIVE);
5622 local_irq_disable();
5623 sub_preempt_count(PREEMPT_ACTIVE);
5626 * Check again in case we missed a preemption opportunity
5627 * between schedule and now.
5630 } while (need_resched());
5633 #endif /* CONFIG_PREEMPT */
5635 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
5638 return try_to_wake_up(curr->private, mode, wake_flags);
5640 EXPORT_SYMBOL(default_wake_function);
5643 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5644 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5645 * number) then we wake all the non-exclusive tasks and one exclusive task.
5647 * There are circumstances in which we can try to wake a task which has already
5648 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5649 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5651 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
5652 int nr_exclusive, int wake_flags, void *key)
5654 wait_queue_t *curr, *next;
5656 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
5657 unsigned flags = curr->flags;
5659 if (curr->func(curr, mode, wake_flags, key) &&
5660 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
5666 * __wake_up - wake up threads blocked on a waitqueue.
5668 * @mode: which threads
5669 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5670 * @key: is directly passed to the wakeup function
5672 * It may be assumed that this function implies a write memory barrier before
5673 * changing the task state if and only if any tasks are woken up.
5675 void __wake_up(wait_queue_head_t *q, unsigned int mode,
5676 int nr_exclusive, void *key)
5678 unsigned long flags;
5680 spin_lock_irqsave(&q->lock, flags);
5681 __wake_up_common(q, mode, nr_exclusive, 0, key);
5682 spin_unlock_irqrestore(&q->lock, flags);
5684 EXPORT_SYMBOL(__wake_up);
5687 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5689 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
5691 __wake_up_common(q, mode, 1, 0, NULL);
5694 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
5696 __wake_up_common(q, mode, 1, 0, key);
5700 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5702 * @mode: which threads
5703 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5704 * @key: opaque value to be passed to wakeup targets
5706 * The sync wakeup differs that the waker knows that it will schedule
5707 * away soon, so while the target thread will be woken up, it will not
5708 * be migrated to another CPU - ie. the two threads are 'synchronized'
5709 * with each other. This can prevent needless bouncing between CPUs.
5711 * On UP it can prevent extra preemption.
5713 * It may be assumed that this function implies a write memory barrier before
5714 * changing the task state if and only if any tasks are woken up.
5716 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
5717 int nr_exclusive, void *key)
5719 unsigned long flags;
5720 int wake_flags = WF_SYNC;
5725 if (unlikely(!nr_exclusive))
5728 spin_lock_irqsave(&q->lock, flags);
5729 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
5730 spin_unlock_irqrestore(&q->lock, flags);
5732 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
5735 * __wake_up_sync - see __wake_up_sync_key()
5737 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
5739 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
5741 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
5744 * complete: - signals a single thread waiting on this completion
5745 * @x: holds the state of this particular completion
5747 * This will wake up a single thread waiting on this completion. Threads will be
5748 * awakened in the same order in which they were queued.
5750 * See also complete_all(), wait_for_completion() and related routines.
5752 * It may be assumed that this function implies a write memory barrier before
5753 * changing the task state if and only if any tasks are woken up.
5755 void complete(struct completion *x)
5757 unsigned long flags;
5759 spin_lock_irqsave(&x->wait.lock, flags);
5761 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
5762 spin_unlock_irqrestore(&x->wait.lock, flags);
5764 EXPORT_SYMBOL(complete);
5767 * complete_all: - signals all threads waiting on this completion
5768 * @x: holds the state of this particular completion
5770 * This will wake up all threads waiting on this particular completion event.
5772 * It may be assumed that this function implies a write memory barrier before
5773 * changing the task state if and only if any tasks are woken up.
5775 void complete_all(struct completion *x)
5777 unsigned long flags;
5779 spin_lock_irqsave(&x->wait.lock, flags);
5780 x->done += UINT_MAX/2;
5781 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
5782 spin_unlock_irqrestore(&x->wait.lock, flags);
5784 EXPORT_SYMBOL(complete_all);
5786 static inline long __sched
5787 do_wait_for_common(struct completion *x, long timeout, int state)
5790 DECLARE_WAITQUEUE(wait, current);
5792 wait.flags |= WQ_FLAG_EXCLUSIVE;
5793 __add_wait_queue_tail(&x->wait, &wait);
5795 if (signal_pending_state(state, current)) {
5796 timeout = -ERESTARTSYS;
5799 __set_current_state(state);
5800 spin_unlock_irq(&x->wait.lock);
5801 timeout = schedule_timeout(timeout);
5802 spin_lock_irq(&x->wait.lock);
5803 } while (!x->done && timeout);
5804 __remove_wait_queue(&x->wait, &wait);
5809 return timeout ?: 1;
5813 wait_for_common(struct completion *x, long timeout, int state)
5817 spin_lock_irq(&x->wait.lock);
5818 timeout = do_wait_for_common(x, timeout, state);
5819 spin_unlock_irq(&x->wait.lock);
5824 * wait_for_completion: - waits for completion of a task
5825 * @x: holds the state of this particular completion
5827 * This waits to be signaled for completion of a specific task. It is NOT
5828 * interruptible and there is no timeout.
5830 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5831 * and interrupt capability. Also see complete().
5833 void __sched wait_for_completion(struct completion *x)
5835 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
5837 EXPORT_SYMBOL(wait_for_completion);
5840 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5841 * @x: holds the state of this particular completion
5842 * @timeout: timeout value in jiffies
5844 * This waits for either a completion of a specific task to be signaled or for a
5845 * specified timeout to expire. The timeout is in jiffies. It is not
5848 unsigned long __sched
5849 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5851 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5853 EXPORT_SYMBOL(wait_for_completion_timeout);
5856 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5857 * @x: holds the state of this particular completion
5859 * This waits for completion of a specific task to be signaled. It is
5862 int __sched wait_for_completion_interruptible(struct completion *x)
5864 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5865 if (t == -ERESTARTSYS)
5869 EXPORT_SYMBOL(wait_for_completion_interruptible);
5872 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5873 * @x: holds the state of this particular completion
5874 * @timeout: timeout value in jiffies
5876 * This waits for either a completion of a specific task to be signaled or for a
5877 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5879 unsigned long __sched
5880 wait_for_completion_interruptible_timeout(struct completion *x,
5881 unsigned long timeout)
5883 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
5885 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5888 * wait_for_completion_killable: - waits for completion of a task (killable)
5889 * @x: holds the state of this particular completion
5891 * This waits to be signaled for completion of a specific task. It can be
5892 * interrupted by a kill signal.
5894 int __sched wait_for_completion_killable(struct completion *x)
5896 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5897 if (t == -ERESTARTSYS)
5901 EXPORT_SYMBOL(wait_for_completion_killable);
5904 * try_wait_for_completion - try to decrement a completion without blocking
5905 * @x: completion structure
5907 * Returns: 0 if a decrement cannot be done without blocking
5908 * 1 if a decrement succeeded.
5910 * If a completion is being used as a counting completion,
5911 * attempt to decrement the counter without blocking. This
5912 * enables us to avoid waiting if the resource the completion
5913 * is protecting is not available.
5915 bool try_wait_for_completion(struct completion *x)
5919 spin_lock_irq(&x->wait.lock);
5924 spin_unlock_irq(&x->wait.lock);
5927 EXPORT_SYMBOL(try_wait_for_completion);
5930 * completion_done - Test to see if a completion has any waiters
5931 * @x: completion structure
5933 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5934 * 1 if there are no waiters.
5937 bool completion_done(struct completion *x)
5941 spin_lock_irq(&x->wait.lock);
5944 spin_unlock_irq(&x->wait.lock);
5947 EXPORT_SYMBOL(completion_done);
5950 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
5952 unsigned long flags;
5955 init_waitqueue_entry(&wait, current);
5957 __set_current_state(state);
5959 spin_lock_irqsave(&q->lock, flags);
5960 __add_wait_queue(q, &wait);
5961 spin_unlock(&q->lock);
5962 timeout = schedule_timeout(timeout);
5963 spin_lock_irq(&q->lock);
5964 __remove_wait_queue(q, &wait);
5965 spin_unlock_irqrestore(&q->lock, flags);
5970 void __sched interruptible_sleep_on(wait_queue_head_t *q)
5972 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5974 EXPORT_SYMBOL(interruptible_sleep_on);
5977 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
5979 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
5981 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
5983 void __sched sleep_on(wait_queue_head_t *q)
5985 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5987 EXPORT_SYMBOL(sleep_on);
5989 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
5991 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
5993 EXPORT_SYMBOL(sleep_on_timeout);
5995 #ifdef CONFIG_RT_MUTEXES
5998 * rt_mutex_setprio - set the current priority of a task
6000 * @prio: prio value (kernel-internal form)
6002 * This function changes the 'effective' priority of a task. It does
6003 * not touch ->normal_prio like __setscheduler().
6005 * Used by the rt_mutex code to implement priority inheritance logic.
6007 void rt_mutex_setprio(struct task_struct *p, int prio)
6009 unsigned long flags;
6010 int oldprio, on_rq, running;
6012 const struct sched_class *prev_class = p->sched_class;
6014 BUG_ON(prio < 0 || prio > MAX_PRIO);
6016 rq = task_rq_lock(p, &flags);
6017 update_rq_clock(rq);
6020 on_rq = p->se.on_rq;
6021 running = task_current(rq, p);
6023 dequeue_task(rq, p, 0);
6025 p->sched_class->put_prev_task(rq, p);
6028 p->sched_class = &rt_sched_class;
6030 p->sched_class = &fair_sched_class;
6035 p->sched_class->set_curr_task(rq);
6037 enqueue_task(rq, p, 0);
6039 check_class_changed(rq, p, prev_class, oldprio, running);
6041 task_rq_unlock(rq, &flags);
6046 void set_user_nice(struct task_struct *p, long nice)
6048 int old_prio, delta, on_rq;
6049 unsigned long flags;
6052 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
6055 * We have to be careful, if called from sys_setpriority(),
6056 * the task might be in the middle of scheduling on another CPU.
6058 rq = task_rq_lock(p, &flags);
6059 update_rq_clock(rq);
6061 * The RT priorities are set via sched_setscheduler(), but we still
6062 * allow the 'normal' nice value to be set - but as expected
6063 * it wont have any effect on scheduling until the task is
6064 * SCHED_FIFO/SCHED_RR:
6066 if (task_has_rt_policy(p)) {
6067 p->static_prio = NICE_TO_PRIO(nice);
6070 on_rq = p->se.on_rq;
6072 dequeue_task(rq, p, 0);
6074 p->static_prio = NICE_TO_PRIO(nice);
6077 p->prio = effective_prio(p);
6078 delta = p->prio - old_prio;
6081 enqueue_task(rq, p, 0);
6083 * If the task increased its priority or is running and
6084 * lowered its priority, then reschedule its CPU:
6086 if (delta < 0 || (delta > 0 && task_running(rq, p)))
6087 resched_task(rq->curr);
6090 task_rq_unlock(rq, &flags);
6092 EXPORT_SYMBOL(set_user_nice);
6095 * can_nice - check if a task can reduce its nice value
6099 int can_nice(const struct task_struct *p, const int nice)
6101 /* convert nice value [19,-20] to rlimit style value [1,40] */
6102 int nice_rlim = 20 - nice;
6104 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
6105 capable(CAP_SYS_NICE));
6108 #ifdef __ARCH_WANT_SYS_NICE
6111 * sys_nice - change the priority of the current process.
6112 * @increment: priority increment
6114 * sys_setpriority is a more generic, but much slower function that
6115 * does similar things.
6117 SYSCALL_DEFINE1(nice, int, increment)
6122 * Setpriority might change our priority at the same moment.
6123 * We don't have to worry. Conceptually one call occurs first
6124 * and we have a single winner.
6126 if (increment < -40)
6131 nice = TASK_NICE(current) + increment;
6137 if (increment < 0 && !can_nice(current, nice))
6140 retval = security_task_setnice(current, nice);
6144 set_user_nice(current, nice);
6151 * task_prio - return the priority value of a given task.
6152 * @p: the task in question.
6154 * This is the priority value as seen by users in /proc.
6155 * RT tasks are offset by -200. Normal tasks are centered
6156 * around 0, value goes from -16 to +15.
6158 int task_prio(const struct task_struct *p)
6160 return p->prio - MAX_RT_PRIO;
6164 * task_nice - return the nice value of a given task.
6165 * @p: the task in question.
6167 int task_nice(const struct task_struct *p)
6169 return TASK_NICE(p);
6171 EXPORT_SYMBOL(task_nice);
6174 * idle_cpu - is a given cpu idle currently?
6175 * @cpu: the processor in question.
6177 int idle_cpu(int cpu)
6179 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
6183 * idle_task - return the idle task for a given cpu.
6184 * @cpu: the processor in question.
6186 struct task_struct *idle_task(int cpu)
6188 return cpu_rq(cpu)->idle;
6192 * find_process_by_pid - find a process with a matching PID value.
6193 * @pid: the pid in question.
6195 static struct task_struct *find_process_by_pid(pid_t pid)
6197 return pid ? find_task_by_vpid(pid) : current;
6200 /* Actually do priority change: must hold rq lock. */
6202 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
6204 BUG_ON(p->se.on_rq);
6207 p->rt_priority = prio;
6208 p->normal_prio = normal_prio(p);
6209 /* we are holding p->pi_lock already */
6210 p->prio = rt_mutex_getprio(p);
6211 if (rt_prio(p->prio))
6212 p->sched_class = &rt_sched_class;
6214 p->sched_class = &fair_sched_class;
6219 * check the target process has a UID that matches the current process's
6221 static bool check_same_owner(struct task_struct *p)
6223 const struct cred *cred = current_cred(), *pcred;
6227 pcred = __task_cred(p);
6228 match = (cred->euid == pcred->euid ||
6229 cred->euid == pcred->uid);
6234 static int __sched_setscheduler(struct task_struct *p, int policy,
6235 struct sched_param *param, bool user)
6237 int retval, oldprio, oldpolicy = -1, on_rq, running;
6238 unsigned long flags;
6239 const struct sched_class *prev_class = p->sched_class;
6243 /* may grab non-irq protected spin_locks */
6244 BUG_ON(in_interrupt());
6246 /* double check policy once rq lock held */
6248 reset_on_fork = p->sched_reset_on_fork;
6249 policy = oldpolicy = p->policy;
6251 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
6252 policy &= ~SCHED_RESET_ON_FORK;
6254 if (policy != SCHED_FIFO && policy != SCHED_RR &&
6255 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
6256 policy != SCHED_IDLE)
6261 * Valid priorities for SCHED_FIFO and SCHED_RR are
6262 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6263 * SCHED_BATCH and SCHED_IDLE is 0.
6265 if (param->sched_priority < 0 ||
6266 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
6267 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
6269 if (rt_policy(policy) != (param->sched_priority != 0))
6273 * Allow unprivileged RT tasks to decrease priority:
6275 if (user && !capable(CAP_SYS_NICE)) {
6276 if (rt_policy(policy)) {
6277 unsigned long rlim_rtprio;
6279 if (!lock_task_sighand(p, &flags))
6281 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
6282 unlock_task_sighand(p, &flags);
6284 /* can't set/change the rt policy */
6285 if (policy != p->policy && !rlim_rtprio)
6288 /* can't increase priority */
6289 if (param->sched_priority > p->rt_priority &&
6290 param->sched_priority > rlim_rtprio)
6294 * Like positive nice levels, dont allow tasks to
6295 * move out of SCHED_IDLE either:
6297 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
6300 /* can't change other user's priorities */
6301 if (!check_same_owner(p))
6304 /* Normal users shall not reset the sched_reset_on_fork flag */
6305 if (p->sched_reset_on_fork && !reset_on_fork)
6310 #ifdef CONFIG_RT_GROUP_SCHED
6312 * Do not allow realtime tasks into groups that have no runtime
6315 if (rt_bandwidth_enabled() && rt_policy(policy) &&
6316 task_group(p)->rt_bandwidth.rt_runtime == 0)
6320 retval = security_task_setscheduler(p, policy, param);
6326 * make sure no PI-waiters arrive (or leave) while we are
6327 * changing the priority of the task:
6329 spin_lock_irqsave(&p->pi_lock, flags);
6331 * To be able to change p->policy safely, the apropriate
6332 * runqueue lock must be held.
6334 rq = __task_rq_lock(p);
6335 /* recheck policy now with rq lock held */
6336 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
6337 policy = oldpolicy = -1;
6338 __task_rq_unlock(rq);
6339 spin_unlock_irqrestore(&p->pi_lock, flags);
6342 update_rq_clock(rq);
6343 on_rq = p->se.on_rq;
6344 running = task_current(rq, p);
6346 deactivate_task(rq, p, 0);
6348 p->sched_class->put_prev_task(rq, p);
6350 p->sched_reset_on_fork = reset_on_fork;
6353 __setscheduler(rq, p, policy, param->sched_priority);
6356 p->sched_class->set_curr_task(rq);
6358 activate_task(rq, p, 0);
6360 check_class_changed(rq, p, prev_class, oldprio, running);
6362 __task_rq_unlock(rq);
6363 spin_unlock_irqrestore(&p->pi_lock, flags);
6365 rt_mutex_adjust_pi(p);
6371 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6372 * @p: the task in question.
6373 * @policy: new policy.
6374 * @param: structure containing the new RT priority.
6376 * NOTE that the task may be already dead.
6378 int sched_setscheduler(struct task_struct *p, int policy,
6379 struct sched_param *param)
6381 return __sched_setscheduler(p, policy, param, true);
6383 EXPORT_SYMBOL_GPL(sched_setscheduler);
6386 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6387 * @p: the task in question.
6388 * @policy: new policy.
6389 * @param: structure containing the new RT priority.
6391 * Just like sched_setscheduler, only don't bother checking if the
6392 * current context has permission. For example, this is needed in
6393 * stop_machine(): we create temporary high priority worker threads,
6394 * but our caller might not have that capability.
6396 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
6397 struct sched_param *param)
6399 return __sched_setscheduler(p, policy, param, false);
6403 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
6405 struct sched_param lparam;
6406 struct task_struct *p;
6409 if (!param || pid < 0)
6411 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
6416 p = find_process_by_pid(pid);
6418 retval = sched_setscheduler(p, policy, &lparam);
6425 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6426 * @pid: the pid in question.
6427 * @policy: new policy.
6428 * @param: structure containing the new RT priority.
6430 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
6431 struct sched_param __user *, param)
6433 /* negative values for policy are not valid */
6437 return do_sched_setscheduler(pid, policy, param);
6441 * sys_sched_setparam - set/change the RT priority of a thread
6442 * @pid: the pid in question.
6443 * @param: structure containing the new RT priority.
6445 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
6447 return do_sched_setscheduler(pid, -1, param);
6451 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6452 * @pid: the pid in question.
6454 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
6456 struct task_struct *p;
6463 read_lock(&tasklist_lock);
6464 p = find_process_by_pid(pid);
6466 retval = security_task_getscheduler(p);
6469 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
6471 read_unlock(&tasklist_lock);
6476 * sys_sched_getparam - get the RT priority of a thread
6477 * @pid: the pid in question.
6478 * @param: structure containing the RT priority.
6480 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
6482 struct sched_param lp;
6483 struct task_struct *p;
6486 if (!param || pid < 0)
6489 read_lock(&tasklist_lock);
6490 p = find_process_by_pid(pid);
6495 retval = security_task_getscheduler(p);
6499 lp.sched_priority = p->rt_priority;
6500 read_unlock(&tasklist_lock);
6503 * This one might sleep, we cannot do it with a spinlock held ...
6505 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
6510 read_unlock(&tasklist_lock);
6514 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
6516 cpumask_var_t cpus_allowed, new_mask;
6517 struct task_struct *p;
6521 read_lock(&tasklist_lock);
6523 p = find_process_by_pid(pid);
6525 read_unlock(&tasklist_lock);
6531 * It is not safe to call set_cpus_allowed with the
6532 * tasklist_lock held. We will bump the task_struct's
6533 * usage count and then drop tasklist_lock.
6536 read_unlock(&tasklist_lock);
6538 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6542 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
6544 goto out_free_cpus_allowed;
6547 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
6550 retval = security_task_setscheduler(p, 0, NULL);
6554 cpuset_cpus_allowed(p, cpus_allowed);
6555 cpumask_and(new_mask, in_mask, cpus_allowed);
6557 retval = set_cpus_allowed_ptr(p, new_mask);
6560 cpuset_cpus_allowed(p, cpus_allowed);
6561 if (!cpumask_subset(new_mask, cpus_allowed)) {
6563 * We must have raced with a concurrent cpuset
6564 * update. Just reset the cpus_allowed to the
6565 * cpuset's cpus_allowed
6567 cpumask_copy(new_mask, cpus_allowed);
6572 free_cpumask_var(new_mask);
6573 out_free_cpus_allowed:
6574 free_cpumask_var(cpus_allowed);
6581 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
6582 struct cpumask *new_mask)
6584 if (len < cpumask_size())
6585 cpumask_clear(new_mask);
6586 else if (len > cpumask_size())
6587 len = cpumask_size();
6589 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6593 * sys_sched_setaffinity - set the cpu affinity of a process
6594 * @pid: pid of the process
6595 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6596 * @user_mask_ptr: user-space pointer to the new cpu mask
6598 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6599 unsigned long __user *, user_mask_ptr)
6601 cpumask_var_t new_mask;
6604 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6607 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6609 retval = sched_setaffinity(pid, new_mask);
6610 free_cpumask_var(new_mask);
6614 long sched_getaffinity(pid_t pid, struct cpumask *mask)
6616 struct task_struct *p;
6617 unsigned long flags;
6622 read_lock(&tasklist_lock);
6625 p = find_process_by_pid(pid);
6629 retval = security_task_getscheduler(p);
6633 rq = task_rq_lock(p, &flags);
6634 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
6635 task_rq_unlock(rq, &flags);
6638 read_unlock(&tasklist_lock);
6645 * sys_sched_getaffinity - get the cpu affinity of a process
6646 * @pid: pid of the process
6647 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6648 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6650 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6651 unsigned long __user *, user_mask_ptr)
6656 if (len < cpumask_size())
6659 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6662 ret = sched_getaffinity(pid, mask);
6664 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
6667 ret = cpumask_size();
6669 free_cpumask_var(mask);
6675 * sys_sched_yield - yield the current processor to other threads.
6677 * This function yields the current CPU to other tasks. If there are no
6678 * other threads running on this CPU then this function will return.
6680 SYSCALL_DEFINE0(sched_yield)
6682 struct rq *rq = this_rq_lock();
6684 schedstat_inc(rq, yld_count);
6685 current->sched_class->yield_task(rq);
6688 * Since we are going to call schedule() anyway, there's
6689 * no need to preempt or enable interrupts:
6691 __release(rq->lock);
6692 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
6693 _raw_spin_unlock(&rq->lock);
6694 preempt_enable_no_resched();
6701 static inline int should_resched(void)
6703 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
6706 static void __cond_resched(void)
6708 add_preempt_count(PREEMPT_ACTIVE);
6710 sub_preempt_count(PREEMPT_ACTIVE);
6713 int __sched _cond_resched(void)
6715 if (should_resched()) {
6721 EXPORT_SYMBOL(_cond_resched);
6724 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
6725 * call schedule, and on return reacquire the lock.
6727 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6728 * operations here to prevent schedule() from being called twice (once via
6729 * spin_unlock(), once by hand).
6731 int __cond_resched_lock(spinlock_t *lock)
6733 int resched = should_resched();
6736 lockdep_assert_held(lock);
6738 if (spin_needbreak(lock) || resched) {
6749 EXPORT_SYMBOL(__cond_resched_lock);
6751 int __sched __cond_resched_softirq(void)
6753 BUG_ON(!in_softirq());
6755 if (should_resched()) {
6763 EXPORT_SYMBOL(__cond_resched_softirq);
6766 * yield - yield the current processor to other threads.
6768 * This is a shortcut for kernel-space yielding - it marks the
6769 * thread runnable and calls sys_sched_yield().
6771 void __sched yield(void)
6773 set_current_state(TASK_RUNNING);
6776 EXPORT_SYMBOL(yield);
6779 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6780 * that process accounting knows that this is a task in IO wait state.
6782 void __sched io_schedule(void)
6784 struct rq *rq = raw_rq();
6786 delayacct_blkio_start();
6787 atomic_inc(&rq->nr_iowait);
6788 current->in_iowait = 1;
6790 current->in_iowait = 0;
6791 atomic_dec(&rq->nr_iowait);
6792 delayacct_blkio_end();
6794 EXPORT_SYMBOL(io_schedule);
6796 long __sched io_schedule_timeout(long timeout)
6798 struct rq *rq = raw_rq();
6801 delayacct_blkio_start();
6802 atomic_inc(&rq->nr_iowait);
6803 current->in_iowait = 1;
6804 ret = schedule_timeout(timeout);
6805 current->in_iowait = 0;
6806 atomic_dec(&rq->nr_iowait);
6807 delayacct_blkio_end();
6812 * sys_sched_get_priority_max - return maximum RT priority.
6813 * @policy: scheduling class.
6815 * this syscall returns the maximum rt_priority that can be used
6816 * by a given scheduling class.
6818 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6825 ret = MAX_USER_RT_PRIO-1;
6837 * sys_sched_get_priority_min - return minimum RT priority.
6838 * @policy: scheduling class.
6840 * this syscall returns the minimum rt_priority that can be used
6841 * by a given scheduling class.
6843 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6861 * sys_sched_rr_get_interval - return the default timeslice of a process.
6862 * @pid: pid of the process.
6863 * @interval: userspace pointer to the timeslice value.
6865 * this syscall writes the default timeslice value of a given process
6866 * into the user-space timespec buffer. A value of '0' means infinity.
6868 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6869 struct timespec __user *, interval)
6871 struct task_struct *p;
6872 unsigned int time_slice;
6873 unsigned long flags;
6882 read_lock(&tasklist_lock);
6883 p = find_process_by_pid(pid);
6887 retval = security_task_getscheduler(p);
6891 rq = task_rq_lock(p, &flags);
6892 time_slice = p->sched_class->get_rr_interval(rq, p);
6893 task_rq_unlock(rq, &flags);
6895 read_unlock(&tasklist_lock);
6896 jiffies_to_timespec(time_slice, &t);
6897 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
6901 read_unlock(&tasklist_lock);
6905 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
6907 void sched_show_task(struct task_struct *p)
6909 unsigned long free = 0;
6912 state = p->state ? __ffs(p->state) + 1 : 0;
6913 printk(KERN_INFO "%-13.13s %c", p->comm,
6914 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
6915 #if BITS_PER_LONG == 32
6916 if (state == TASK_RUNNING)
6917 printk(KERN_CONT " running ");
6919 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
6921 if (state == TASK_RUNNING)
6922 printk(KERN_CONT " running task ");
6924 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
6926 #ifdef CONFIG_DEBUG_STACK_USAGE
6927 free = stack_not_used(p);
6929 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
6930 task_pid_nr(p), task_pid_nr(p->real_parent),
6931 (unsigned long)task_thread_info(p)->flags);
6933 show_stack(p, NULL);
6936 void show_state_filter(unsigned long state_filter)
6938 struct task_struct *g, *p;
6940 #if BITS_PER_LONG == 32
6942 " task PC stack pid father\n");
6945 " task PC stack pid father\n");
6947 read_lock(&tasklist_lock);
6948 do_each_thread(g, p) {
6950 * reset the NMI-timeout, listing all files on a slow
6951 * console might take alot of time:
6953 touch_nmi_watchdog();
6954 if (!state_filter || (p->state & state_filter))
6956 } while_each_thread(g, p);
6958 touch_all_softlockup_watchdogs();
6960 #ifdef CONFIG_SCHED_DEBUG
6961 sysrq_sched_debug_show();
6963 read_unlock(&tasklist_lock);
6965 * Only show locks if all tasks are dumped:
6968 debug_show_all_locks();
6971 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6973 idle->sched_class = &idle_sched_class;
6977 * init_idle - set up an idle thread for a given CPU
6978 * @idle: task in question
6979 * @cpu: cpu the idle task belongs to
6981 * NOTE: this function does not set the idle thread's NEED_RESCHED
6982 * flag, to make booting more robust.
6984 void __cpuinit init_idle(struct task_struct *idle, int cpu)
6986 struct rq *rq = cpu_rq(cpu);
6987 unsigned long flags;
6989 spin_lock_irqsave(&rq->lock, flags);
6992 idle->se.exec_start = sched_clock();
6994 idle->prio = idle->normal_prio = MAX_PRIO;
6995 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
6996 __set_task_cpu(idle, cpu);
6998 rq->curr = rq->idle = idle;
6999 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
7002 spin_unlock_irqrestore(&rq->lock, flags);
7004 /* Set the preempt count _outside_ the spinlocks! */
7005 #if defined(CONFIG_PREEMPT)
7006 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
7008 task_thread_info(idle)->preempt_count = 0;
7011 * The idle tasks have their own, simple scheduling class:
7013 idle->sched_class = &idle_sched_class;
7014 ftrace_graph_init_task(idle);
7018 * In a system that switches off the HZ timer nohz_cpu_mask
7019 * indicates which cpus entered this state. This is used
7020 * in the rcu update to wait only for active cpus. For system
7021 * which do not switch off the HZ timer nohz_cpu_mask should
7022 * always be CPU_BITS_NONE.
7024 cpumask_var_t nohz_cpu_mask;
7027 * Increase the granularity value when there are more CPUs,
7028 * because with more CPUs the 'effective latency' as visible
7029 * to users decreases. But the relationship is not linear,
7030 * so pick a second-best guess by going with the log2 of the
7033 * This idea comes from the SD scheduler of Con Kolivas:
7035 static inline void sched_init_granularity(void)
7037 unsigned int factor = 1 + ilog2(num_online_cpus());
7038 const unsigned long limit = 200000000;
7040 sysctl_sched_min_granularity *= factor;
7041 if (sysctl_sched_min_granularity > limit)
7042 sysctl_sched_min_granularity = limit;
7044 sysctl_sched_latency *= factor;
7045 if (sysctl_sched_latency > limit)
7046 sysctl_sched_latency = limit;
7048 sysctl_sched_wakeup_granularity *= factor;
7050 sysctl_sched_shares_ratelimit *= factor;
7055 * This is how migration works:
7057 * 1) we queue a struct migration_req structure in the source CPU's
7058 * runqueue and wake up that CPU's migration thread.
7059 * 2) we down() the locked semaphore => thread blocks.
7060 * 3) migration thread wakes up (implicitly it forces the migrated
7061 * thread off the CPU)
7062 * 4) it gets the migration request and checks whether the migrated
7063 * task is still in the wrong runqueue.
7064 * 5) if it's in the wrong runqueue then the migration thread removes
7065 * it and puts it into the right queue.
7066 * 6) migration thread up()s the semaphore.
7067 * 7) we wake up and the migration is done.
7071 * Change a given task's CPU affinity. Migrate the thread to a
7072 * proper CPU and schedule it away if the CPU it's executing on
7073 * is removed from the allowed bitmask.
7075 * NOTE: the caller must have a valid reference to the task, the
7076 * task must not exit() & deallocate itself prematurely. The
7077 * call is not atomic; no spinlocks may be held.
7079 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
7081 struct migration_req req;
7082 unsigned long flags;
7086 rq = task_rq_lock(p, &flags);
7087 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
7092 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
7093 !cpumask_equal(&p->cpus_allowed, new_mask))) {
7098 if (p->sched_class->set_cpus_allowed)
7099 p->sched_class->set_cpus_allowed(p, new_mask);
7101 cpumask_copy(&p->cpus_allowed, new_mask);
7102 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
7105 /* Can the task run on the task's current CPU? If so, we're done */
7106 if (cpumask_test_cpu(task_cpu(p), new_mask))
7109 if (migrate_task(p, cpumask_any_and(cpu_active_mask, new_mask), &req)) {
7110 /* Need help from migration thread: drop lock and wait. */
7111 struct task_struct *mt = rq->migration_thread;
7113 get_task_struct(mt);
7114 task_rq_unlock(rq, &flags);
7115 wake_up_process(rq->migration_thread);
7116 put_task_struct(mt);
7117 wait_for_completion(&req.done);
7118 tlb_migrate_finish(p->mm);
7122 task_rq_unlock(rq, &flags);
7126 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
7129 * Move (not current) task off this cpu, onto dest cpu. We're doing
7130 * this because either it can't run here any more (set_cpus_allowed()
7131 * away from this CPU, or CPU going down), or because we're
7132 * attempting to rebalance this task on exec (sched_exec).
7134 * So we race with normal scheduler movements, but that's OK, as long
7135 * as the task is no longer on this CPU.
7137 * Returns non-zero if task was successfully migrated.
7139 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
7141 struct rq *rq_dest, *rq_src;
7144 if (unlikely(!cpu_active(dest_cpu)))
7147 rq_src = cpu_rq(src_cpu);
7148 rq_dest = cpu_rq(dest_cpu);
7150 double_rq_lock(rq_src, rq_dest);
7151 /* Already moved. */
7152 if (task_cpu(p) != src_cpu)
7154 /* Affinity changed (again). */
7155 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7158 on_rq = p->se.on_rq;
7160 deactivate_task(rq_src, p, 0);
7162 set_task_cpu(p, dest_cpu);
7164 activate_task(rq_dest, p, 0);
7165 check_preempt_curr(rq_dest, p, 0);
7170 double_rq_unlock(rq_src, rq_dest);
7174 #define RCU_MIGRATION_IDLE 0
7175 #define RCU_MIGRATION_NEED_QS 1
7176 #define RCU_MIGRATION_GOT_QS 2
7177 #define RCU_MIGRATION_MUST_SYNC 3
7180 * migration_thread - this is a highprio system thread that performs
7181 * thread migration by bumping thread off CPU then 'pushing' onto
7184 static int migration_thread(void *data)
7187 int cpu = (long)data;
7191 BUG_ON(rq->migration_thread != current);
7193 set_current_state(TASK_INTERRUPTIBLE);
7194 while (!kthread_should_stop()) {
7195 struct migration_req *req;
7196 struct list_head *head;
7198 spin_lock_irq(&rq->lock);
7200 if (cpu_is_offline(cpu)) {
7201 spin_unlock_irq(&rq->lock);
7205 if (rq->active_balance) {
7206 active_load_balance(rq, cpu);
7207 rq->active_balance = 0;
7210 head = &rq->migration_queue;
7212 if (list_empty(head)) {
7213 spin_unlock_irq(&rq->lock);
7215 set_current_state(TASK_INTERRUPTIBLE);
7218 req = list_entry(head->next, struct migration_req, list);
7219 list_del_init(head->next);
7221 if (req->task != NULL) {
7222 spin_unlock(&rq->lock);
7223 __migrate_task(req->task, cpu, req->dest_cpu);
7224 } else if (likely(cpu == (badcpu = smp_processor_id()))) {
7225 req->dest_cpu = RCU_MIGRATION_GOT_QS;
7226 spin_unlock(&rq->lock);
7228 req->dest_cpu = RCU_MIGRATION_MUST_SYNC;
7229 spin_unlock(&rq->lock);
7230 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu, cpu);
7234 complete(&req->done);
7236 __set_current_state(TASK_RUNNING);
7241 #ifdef CONFIG_HOTPLUG_CPU
7243 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
7247 local_irq_disable();
7248 ret = __migrate_task(p, src_cpu, dest_cpu);
7254 * Figure out where task on dead CPU should go, use force if necessary.
7256 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
7259 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(dead_cpu));
7262 /* Look for allowed, online CPU in same node. */
7263 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
7264 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7267 /* Any allowed, online CPU? */
7268 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
7269 if (dest_cpu < nr_cpu_ids)
7272 /* No more Mr. Nice Guy. */
7273 if (dest_cpu >= nr_cpu_ids) {
7274 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
7275 dest_cpu = cpumask_any_and(cpu_active_mask, &p->cpus_allowed);
7278 * Don't tell them about moving exiting tasks or
7279 * kernel threads (both mm NULL), since they never
7282 if (p->mm && printk_ratelimit()) {
7283 printk(KERN_INFO "process %d (%s) no "
7284 "longer affine to cpu%d\n",
7285 task_pid_nr(p), p->comm, dead_cpu);
7290 /* It can have affinity changed while we were choosing. */
7291 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
7296 * While a dead CPU has no uninterruptible tasks queued at this point,
7297 * it might still have a nonzero ->nr_uninterruptible counter, because
7298 * for performance reasons the counter is not stricly tracking tasks to
7299 * their home CPUs. So we just add the counter to another CPU's counter,
7300 * to keep the global sum constant after CPU-down:
7302 static void migrate_nr_uninterruptible(struct rq *rq_src)
7304 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
7305 unsigned long flags;
7307 local_irq_save(flags);
7308 double_rq_lock(rq_src, rq_dest);
7309 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
7310 rq_src->nr_uninterruptible = 0;
7311 double_rq_unlock(rq_src, rq_dest);
7312 local_irq_restore(flags);
7315 /* Run through task list and migrate tasks from the dead cpu. */
7316 static void migrate_live_tasks(int src_cpu)
7318 struct task_struct *p, *t;
7320 read_lock(&tasklist_lock);
7322 do_each_thread(t, p) {
7326 if (task_cpu(p) == src_cpu)
7327 move_task_off_dead_cpu(src_cpu, p);
7328 } while_each_thread(t, p);
7330 read_unlock(&tasklist_lock);
7334 * Schedules idle task to be the next runnable task on current CPU.
7335 * It does so by boosting its priority to highest possible.
7336 * Used by CPU offline code.
7338 void sched_idle_next(void)
7340 int this_cpu = smp_processor_id();
7341 struct rq *rq = cpu_rq(this_cpu);
7342 struct task_struct *p = rq->idle;
7343 unsigned long flags;
7345 /* cpu has to be offline */
7346 BUG_ON(cpu_online(this_cpu));
7349 * Strictly not necessary since rest of the CPUs are stopped by now
7350 * and interrupts disabled on the current cpu.
7352 spin_lock_irqsave(&rq->lock, flags);
7354 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7356 update_rq_clock(rq);
7357 activate_task(rq, p, 0);
7359 spin_unlock_irqrestore(&rq->lock, flags);
7363 * Ensures that the idle task is using init_mm right before its cpu goes
7366 void idle_task_exit(void)
7368 struct mm_struct *mm = current->active_mm;
7370 BUG_ON(cpu_online(smp_processor_id()));
7373 switch_mm(mm, &init_mm, current);
7377 /* called under rq->lock with disabled interrupts */
7378 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
7380 struct rq *rq = cpu_rq(dead_cpu);
7382 /* Must be exiting, otherwise would be on tasklist. */
7383 BUG_ON(!p->exit_state);
7385 /* Cannot have done final schedule yet: would have vanished. */
7386 BUG_ON(p->state == TASK_DEAD);
7391 * Drop lock around migration; if someone else moves it,
7392 * that's OK. No task can be added to this CPU, so iteration is
7395 spin_unlock_irq(&rq->lock);
7396 move_task_off_dead_cpu(dead_cpu, p);
7397 spin_lock_irq(&rq->lock);
7402 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7403 static void migrate_dead_tasks(unsigned int dead_cpu)
7405 struct rq *rq = cpu_rq(dead_cpu);
7406 struct task_struct *next;
7409 if (!rq->nr_running)
7411 update_rq_clock(rq);
7412 next = pick_next_task(rq);
7415 next->sched_class->put_prev_task(rq, next);
7416 migrate_dead(dead_cpu, next);
7422 * remove the tasks which were accounted by rq from calc_load_tasks.
7424 static void calc_global_load_remove(struct rq *rq)
7426 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
7427 rq->calc_load_active = 0;
7429 #endif /* CONFIG_HOTPLUG_CPU */
7431 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7433 static struct ctl_table sd_ctl_dir[] = {
7435 .procname = "sched_domain",
7441 static struct ctl_table sd_ctl_root[] = {
7443 .ctl_name = CTL_KERN,
7444 .procname = "kernel",
7446 .child = sd_ctl_dir,
7451 static struct ctl_table *sd_alloc_ctl_entry(int n)
7453 struct ctl_table *entry =
7454 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
7459 static void sd_free_ctl_entry(struct ctl_table **tablep)
7461 struct ctl_table *entry;
7464 * In the intermediate directories, both the child directory and
7465 * procname are dynamically allocated and could fail but the mode
7466 * will always be set. In the lowest directory the names are
7467 * static strings and all have proc handlers.
7469 for (entry = *tablep; entry->mode; entry++) {
7471 sd_free_ctl_entry(&entry->child);
7472 if (entry->proc_handler == NULL)
7473 kfree(entry->procname);
7481 set_table_entry(struct ctl_table *entry,
7482 const char *procname, void *data, int maxlen,
7483 mode_t mode, proc_handler *proc_handler)
7485 entry->procname = procname;
7487 entry->maxlen = maxlen;
7489 entry->proc_handler = proc_handler;
7492 static struct ctl_table *
7493 sd_alloc_ctl_domain_table(struct sched_domain *sd)
7495 struct ctl_table *table = sd_alloc_ctl_entry(13);
7500 set_table_entry(&table[0], "min_interval", &sd->min_interval,
7501 sizeof(long), 0644, proc_doulongvec_minmax);
7502 set_table_entry(&table[1], "max_interval", &sd->max_interval,
7503 sizeof(long), 0644, proc_doulongvec_minmax);
7504 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
7505 sizeof(int), 0644, proc_dointvec_minmax);
7506 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
7507 sizeof(int), 0644, proc_dointvec_minmax);
7508 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
7509 sizeof(int), 0644, proc_dointvec_minmax);
7510 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
7511 sizeof(int), 0644, proc_dointvec_minmax);
7512 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
7513 sizeof(int), 0644, proc_dointvec_minmax);
7514 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
7515 sizeof(int), 0644, proc_dointvec_minmax);
7516 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
7517 sizeof(int), 0644, proc_dointvec_minmax);
7518 set_table_entry(&table[9], "cache_nice_tries",
7519 &sd->cache_nice_tries,
7520 sizeof(int), 0644, proc_dointvec_minmax);
7521 set_table_entry(&table[10], "flags", &sd->flags,
7522 sizeof(int), 0644, proc_dointvec_minmax);
7523 set_table_entry(&table[11], "name", sd->name,
7524 CORENAME_MAX_SIZE, 0444, proc_dostring);
7525 /* &table[12] is terminator */
7530 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
7532 struct ctl_table *entry, *table;
7533 struct sched_domain *sd;
7534 int domain_num = 0, i;
7537 for_each_domain(cpu, sd)
7539 entry = table = sd_alloc_ctl_entry(domain_num + 1);
7544 for_each_domain(cpu, sd) {
7545 snprintf(buf, 32, "domain%d", i);
7546 entry->procname = kstrdup(buf, GFP_KERNEL);
7548 entry->child = sd_alloc_ctl_domain_table(sd);
7555 static struct ctl_table_header *sd_sysctl_header;
7556 static void register_sched_domain_sysctl(void)
7558 int i, cpu_num = num_possible_cpus();
7559 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
7562 WARN_ON(sd_ctl_dir[0].child);
7563 sd_ctl_dir[0].child = entry;
7568 for_each_possible_cpu(i) {
7569 snprintf(buf, 32, "cpu%d", i);
7570 entry->procname = kstrdup(buf, GFP_KERNEL);
7572 entry->child = sd_alloc_ctl_cpu_table(i);
7576 WARN_ON(sd_sysctl_header);
7577 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
7580 /* may be called multiple times per register */
7581 static void unregister_sched_domain_sysctl(void)
7583 if (sd_sysctl_header)
7584 unregister_sysctl_table(sd_sysctl_header);
7585 sd_sysctl_header = NULL;
7586 if (sd_ctl_dir[0].child)
7587 sd_free_ctl_entry(&sd_ctl_dir[0].child);
7590 static void register_sched_domain_sysctl(void)
7593 static void unregister_sched_domain_sysctl(void)
7598 static void set_rq_online(struct rq *rq)
7601 const struct sched_class *class;
7603 cpumask_set_cpu(rq->cpu, rq->rd->online);
7606 for_each_class(class) {
7607 if (class->rq_online)
7608 class->rq_online(rq);
7613 static void set_rq_offline(struct rq *rq)
7616 const struct sched_class *class;
7618 for_each_class(class) {
7619 if (class->rq_offline)
7620 class->rq_offline(rq);
7623 cpumask_clear_cpu(rq->cpu, rq->rd->online);
7629 * migration_call - callback that gets triggered when a CPU is added.
7630 * Here we can start up the necessary migration thread for the new CPU.
7632 static int __cpuinit
7633 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
7635 struct task_struct *p;
7636 int cpu = (long)hcpu;
7637 unsigned long flags;
7642 case CPU_UP_PREPARE:
7643 case CPU_UP_PREPARE_FROZEN:
7644 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
7647 kthread_bind(p, cpu);
7648 /* Must be high prio: stop_machine expects to yield to it. */
7649 rq = task_rq_lock(p, &flags);
7650 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7651 task_rq_unlock(rq, &flags);
7653 cpu_rq(cpu)->migration_thread = p;
7654 rq->calc_load_update = calc_load_update;
7658 case CPU_ONLINE_FROZEN:
7659 /* Strictly unnecessary, as first user will wake it. */
7660 wake_up_process(cpu_rq(cpu)->migration_thread);
7662 /* Update our root-domain */
7664 spin_lock_irqsave(&rq->lock, flags);
7666 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7670 spin_unlock_irqrestore(&rq->lock, flags);
7673 #ifdef CONFIG_HOTPLUG_CPU
7674 case CPU_UP_CANCELED:
7675 case CPU_UP_CANCELED_FROZEN:
7676 if (!cpu_rq(cpu)->migration_thread)
7678 /* Unbind it from offline cpu so it can run. Fall thru. */
7679 kthread_bind(cpu_rq(cpu)->migration_thread,
7680 cpumask_any(cpu_online_mask));
7681 kthread_stop(cpu_rq(cpu)->migration_thread);
7682 put_task_struct(cpu_rq(cpu)->migration_thread);
7683 cpu_rq(cpu)->migration_thread = NULL;
7687 case CPU_DEAD_FROZEN:
7688 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7689 migrate_live_tasks(cpu);
7691 kthread_stop(rq->migration_thread);
7692 put_task_struct(rq->migration_thread);
7693 rq->migration_thread = NULL;
7694 /* Idle task back to normal (off runqueue, low prio) */
7695 spin_lock_irq(&rq->lock);
7696 update_rq_clock(rq);
7697 deactivate_task(rq, rq->idle, 0);
7698 rq->idle->static_prio = MAX_PRIO;
7699 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
7700 rq->idle->sched_class = &idle_sched_class;
7701 migrate_dead_tasks(cpu);
7702 spin_unlock_irq(&rq->lock);
7704 migrate_nr_uninterruptible(rq);
7705 BUG_ON(rq->nr_running != 0);
7706 calc_global_load_remove(rq);
7708 * No need to migrate the tasks: it was best-effort if
7709 * they didn't take sched_hotcpu_mutex. Just wake up
7712 spin_lock_irq(&rq->lock);
7713 while (!list_empty(&rq->migration_queue)) {
7714 struct migration_req *req;
7716 req = list_entry(rq->migration_queue.next,
7717 struct migration_req, list);
7718 list_del_init(&req->list);
7719 spin_unlock_irq(&rq->lock);
7720 complete(&req->done);
7721 spin_lock_irq(&rq->lock);
7723 spin_unlock_irq(&rq->lock);
7727 case CPU_DYING_FROZEN:
7728 /* Update our root-domain */
7730 spin_lock_irqsave(&rq->lock, flags);
7732 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7735 spin_unlock_irqrestore(&rq->lock, flags);
7743 * Register at high priority so that task migration (migrate_all_tasks)
7744 * happens before everything else. This has to be lower priority than
7745 * the notifier in the perf_event subsystem, though.
7747 static struct notifier_block __cpuinitdata migration_notifier = {
7748 .notifier_call = migration_call,
7752 static int __init migration_init(void)
7754 void *cpu = (void *)(long)smp_processor_id();
7757 /* Start one for the boot CPU: */
7758 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
7759 BUG_ON(err == NOTIFY_BAD);
7760 migration_call(&migration_notifier, CPU_ONLINE, cpu);
7761 register_cpu_notifier(&migration_notifier);
7765 early_initcall(migration_init);
7770 #ifdef CONFIG_SCHED_DEBUG
7772 static __read_mostly int sched_domain_debug_enabled;
7774 static int __init sched_domain_debug_setup(char *str)
7776 sched_domain_debug_enabled = 1;
7780 early_param("sched_debug", sched_domain_debug_setup);
7782 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
7783 struct cpumask *groupmask)
7785 struct sched_group *group = sd->groups;
7788 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
7789 cpumask_clear(groupmask);
7791 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
7793 if (!(sd->flags & SD_LOAD_BALANCE)) {
7794 printk("does not load-balance\n");
7796 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
7801 printk(KERN_CONT "span %s level %s\n", str, sd->name);
7803 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
7804 printk(KERN_ERR "ERROR: domain->span does not contain "
7807 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
7808 printk(KERN_ERR "ERROR: domain->groups does not contain"
7812 printk(KERN_DEBUG "%*s groups:", level + 1, "");
7816 printk(KERN_ERR "ERROR: group is NULL\n");
7820 if (!group->cpu_power) {
7821 printk(KERN_CONT "\n");
7822 printk(KERN_ERR "ERROR: domain->cpu_power not "
7827 if (!cpumask_weight(sched_group_cpus(group))) {
7828 printk(KERN_CONT "\n");
7829 printk(KERN_ERR "ERROR: empty group\n");
7833 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
7834 printk(KERN_CONT "\n");
7835 printk(KERN_ERR "ERROR: repeated CPUs\n");
7839 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
7841 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
7843 printk(KERN_CONT " %s", str);
7844 if (group->cpu_power != SCHED_LOAD_SCALE) {
7845 printk(KERN_CONT " (cpu_power = %d)",
7849 group = group->next;
7850 } while (group != sd->groups);
7851 printk(KERN_CONT "\n");
7853 if (!cpumask_equal(sched_domain_span(sd), groupmask))
7854 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
7857 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
7858 printk(KERN_ERR "ERROR: parent span is not a superset "
7859 "of domain->span\n");
7863 static void sched_domain_debug(struct sched_domain *sd, int cpu)
7865 cpumask_var_t groupmask;
7868 if (!sched_domain_debug_enabled)
7872 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7876 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7878 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
7879 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
7884 if (sched_domain_debug_one(sd, cpu, level, groupmask))
7891 free_cpumask_var(groupmask);
7893 #else /* !CONFIG_SCHED_DEBUG */
7894 # define sched_domain_debug(sd, cpu) do { } while (0)
7895 #endif /* CONFIG_SCHED_DEBUG */
7897 static int sd_degenerate(struct sched_domain *sd)
7899 if (cpumask_weight(sched_domain_span(sd)) == 1)
7902 /* Following flags need at least 2 groups */
7903 if (sd->flags & (SD_LOAD_BALANCE |
7904 SD_BALANCE_NEWIDLE |
7908 SD_SHARE_PKG_RESOURCES)) {
7909 if (sd->groups != sd->groups->next)
7913 /* Following flags don't use groups */
7914 if (sd->flags & (SD_WAKE_AFFINE))
7921 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
7923 unsigned long cflags = sd->flags, pflags = parent->flags;
7925 if (sd_degenerate(parent))
7928 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
7931 /* Flags needing groups don't count if only 1 group in parent */
7932 if (parent->groups == parent->groups->next) {
7933 pflags &= ~(SD_LOAD_BALANCE |
7934 SD_BALANCE_NEWIDLE |
7938 SD_SHARE_PKG_RESOURCES);
7939 if (nr_node_ids == 1)
7940 pflags &= ~SD_SERIALIZE;
7942 if (~cflags & pflags)
7948 static void free_rootdomain(struct root_domain *rd)
7950 synchronize_sched();
7952 cpupri_cleanup(&rd->cpupri);
7954 free_cpumask_var(rd->rto_mask);
7955 free_cpumask_var(rd->online);
7956 free_cpumask_var(rd->span);
7960 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7962 struct root_domain *old_rd = NULL;
7963 unsigned long flags;
7965 spin_lock_irqsave(&rq->lock, flags);
7970 if (cpumask_test_cpu(rq->cpu, old_rd->online))
7973 cpumask_clear_cpu(rq->cpu, old_rd->span);
7976 * If we dont want to free the old_rt yet then
7977 * set old_rd to NULL to skip the freeing later
7980 if (!atomic_dec_and_test(&old_rd->refcount))
7984 atomic_inc(&rd->refcount);
7987 cpumask_set_cpu(rq->cpu, rd->span);
7988 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
7991 spin_unlock_irqrestore(&rq->lock, flags);
7994 free_rootdomain(old_rd);
7997 static int init_rootdomain(struct root_domain *rd, bool bootmem)
7999 gfp_t gfp = GFP_KERNEL;
8001 memset(rd, 0, sizeof(*rd));
8006 if (!alloc_cpumask_var(&rd->span, gfp))
8008 if (!alloc_cpumask_var(&rd->online, gfp))
8010 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
8013 if (cpupri_init(&rd->cpupri, bootmem) != 0)
8018 free_cpumask_var(rd->rto_mask);
8020 free_cpumask_var(rd->online);
8022 free_cpumask_var(rd->span);
8027 static void init_defrootdomain(void)
8029 init_rootdomain(&def_root_domain, true);
8031 atomic_set(&def_root_domain.refcount, 1);
8034 static struct root_domain *alloc_rootdomain(void)
8036 struct root_domain *rd;
8038 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
8042 if (init_rootdomain(rd, false) != 0) {
8051 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
8052 * hold the hotplug lock.
8055 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
8057 struct rq *rq = cpu_rq(cpu);
8058 struct sched_domain *tmp;
8060 /* Remove the sched domains which do not contribute to scheduling. */
8061 for (tmp = sd; tmp; ) {
8062 struct sched_domain *parent = tmp->parent;
8066 if (sd_parent_degenerate(tmp, parent)) {
8067 tmp->parent = parent->parent;
8069 parent->parent->child = tmp;
8074 if (sd && sd_degenerate(sd)) {
8080 sched_domain_debug(sd, cpu);
8082 rq_attach_root(rq, rd);
8083 rcu_assign_pointer(rq->sd, sd);
8086 /* cpus with isolated domains */
8087 static cpumask_var_t cpu_isolated_map;
8089 /* Setup the mask of cpus configured for isolated domains */
8090 static int __init isolated_cpu_setup(char *str)
8092 alloc_bootmem_cpumask_var(&cpu_isolated_map);
8093 cpulist_parse(str, cpu_isolated_map);
8097 __setup("isolcpus=", isolated_cpu_setup);
8100 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
8101 * to a function which identifies what group(along with sched group) a CPU
8102 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
8103 * (due to the fact that we keep track of groups covered with a struct cpumask).
8105 * init_sched_build_groups will build a circular linked list of the groups
8106 * covered by the given span, and will set each group's ->cpumask correctly,
8107 * and ->cpu_power to 0.
8110 init_sched_build_groups(const struct cpumask *span,
8111 const struct cpumask *cpu_map,
8112 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
8113 struct sched_group **sg,
8114 struct cpumask *tmpmask),
8115 struct cpumask *covered, struct cpumask *tmpmask)
8117 struct sched_group *first = NULL, *last = NULL;
8120 cpumask_clear(covered);
8122 for_each_cpu(i, span) {
8123 struct sched_group *sg;
8124 int group = group_fn(i, cpu_map, &sg, tmpmask);
8127 if (cpumask_test_cpu(i, covered))
8130 cpumask_clear(sched_group_cpus(sg));
8133 for_each_cpu(j, span) {
8134 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
8137 cpumask_set_cpu(j, covered);
8138 cpumask_set_cpu(j, sched_group_cpus(sg));
8149 #define SD_NODES_PER_DOMAIN 16
8154 * find_next_best_node - find the next node to include in a sched_domain
8155 * @node: node whose sched_domain we're building
8156 * @used_nodes: nodes already in the sched_domain
8158 * Find the next node to include in a given scheduling domain. Simply
8159 * finds the closest node not already in the @used_nodes map.
8161 * Should use nodemask_t.
8163 static int find_next_best_node(int node, nodemask_t *used_nodes)
8165 int i, n, val, min_val, best_node = 0;
8169 for (i = 0; i < nr_node_ids; i++) {
8170 /* Start at @node */
8171 n = (node + i) % nr_node_ids;
8173 if (!nr_cpus_node(n))
8176 /* Skip already used nodes */
8177 if (node_isset(n, *used_nodes))
8180 /* Simple min distance search */
8181 val = node_distance(node, n);
8183 if (val < min_val) {
8189 node_set(best_node, *used_nodes);
8194 * sched_domain_node_span - get a cpumask for a node's sched_domain
8195 * @node: node whose cpumask we're constructing
8196 * @span: resulting cpumask
8198 * Given a node, construct a good cpumask for its sched_domain to span. It
8199 * should be one that prevents unnecessary balancing, but also spreads tasks
8202 static void sched_domain_node_span(int node, struct cpumask *span)
8204 nodemask_t used_nodes;
8207 cpumask_clear(span);
8208 nodes_clear(used_nodes);
8210 cpumask_or(span, span, cpumask_of_node(node));
8211 node_set(node, used_nodes);
8213 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
8214 int next_node = find_next_best_node(node, &used_nodes);
8216 cpumask_or(span, span, cpumask_of_node(next_node));
8219 #endif /* CONFIG_NUMA */
8221 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
8224 * The cpus mask in sched_group and sched_domain hangs off the end.
8226 * ( See the the comments in include/linux/sched.h:struct sched_group
8227 * and struct sched_domain. )
8229 struct static_sched_group {
8230 struct sched_group sg;
8231 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
8234 struct static_sched_domain {
8235 struct sched_domain sd;
8236 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
8242 cpumask_var_t domainspan;
8243 cpumask_var_t covered;
8244 cpumask_var_t notcovered;
8246 cpumask_var_t nodemask;
8247 cpumask_var_t this_sibling_map;
8248 cpumask_var_t this_core_map;
8249 cpumask_var_t send_covered;
8250 cpumask_var_t tmpmask;
8251 struct sched_group **sched_group_nodes;
8252 struct root_domain *rd;
8256 sa_sched_groups = 0,
8261 sa_this_sibling_map,
8263 sa_sched_group_nodes,
8273 * SMT sched-domains:
8275 #ifdef CONFIG_SCHED_SMT
8276 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
8277 static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
8280 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
8281 struct sched_group **sg, struct cpumask *unused)
8284 *sg = &per_cpu(sched_group_cpus, cpu).sg;
8287 #endif /* CONFIG_SCHED_SMT */
8290 * multi-core sched-domains:
8292 #ifdef CONFIG_SCHED_MC
8293 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
8294 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
8295 #endif /* CONFIG_SCHED_MC */
8297 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8299 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8300 struct sched_group **sg, struct cpumask *mask)
8304 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8305 group = cpumask_first(mask);
8307 *sg = &per_cpu(sched_group_core, group).sg;
8310 #elif defined(CONFIG_SCHED_MC)
8312 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8313 struct sched_group **sg, struct cpumask *unused)
8316 *sg = &per_cpu(sched_group_core, cpu).sg;
8321 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
8322 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
8325 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
8326 struct sched_group **sg, struct cpumask *mask)
8329 #ifdef CONFIG_SCHED_MC
8330 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
8331 group = cpumask_first(mask);
8332 #elif defined(CONFIG_SCHED_SMT)
8333 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8334 group = cpumask_first(mask);
8339 *sg = &per_cpu(sched_group_phys, group).sg;
8345 * The init_sched_build_groups can't handle what we want to do with node
8346 * groups, so roll our own. Now each node has its own list of groups which
8347 * gets dynamically allocated.
8349 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
8350 static struct sched_group ***sched_group_nodes_bycpu;
8352 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
8353 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
8355 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
8356 struct sched_group **sg,
8357 struct cpumask *nodemask)
8361 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
8362 group = cpumask_first(nodemask);
8365 *sg = &per_cpu(sched_group_allnodes, group).sg;
8369 static void init_numa_sched_groups_power(struct sched_group *group_head)
8371 struct sched_group *sg = group_head;
8377 for_each_cpu(j, sched_group_cpus(sg)) {
8378 struct sched_domain *sd;
8380 sd = &per_cpu(phys_domains, j).sd;
8381 if (j != group_first_cpu(sd->groups)) {
8383 * Only add "power" once for each
8389 sg->cpu_power += sd->groups->cpu_power;
8392 } while (sg != group_head);
8395 static int build_numa_sched_groups(struct s_data *d,
8396 const struct cpumask *cpu_map, int num)
8398 struct sched_domain *sd;
8399 struct sched_group *sg, *prev;
8402 cpumask_clear(d->covered);
8403 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
8404 if (cpumask_empty(d->nodemask)) {
8405 d->sched_group_nodes[num] = NULL;
8409 sched_domain_node_span(num, d->domainspan);
8410 cpumask_and(d->domainspan, d->domainspan, cpu_map);
8412 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8415 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
8419 d->sched_group_nodes[num] = sg;
8421 for_each_cpu(j, d->nodemask) {
8422 sd = &per_cpu(node_domains, j).sd;
8427 cpumask_copy(sched_group_cpus(sg), d->nodemask);
8429 cpumask_or(d->covered, d->covered, d->nodemask);
8432 for (j = 0; j < nr_node_ids; j++) {
8433 n = (num + j) % nr_node_ids;
8434 cpumask_complement(d->notcovered, d->covered);
8435 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
8436 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
8437 if (cpumask_empty(d->tmpmask))
8439 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
8440 if (cpumask_empty(d->tmpmask))
8442 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8446 "Can not alloc domain group for node %d\n", j);
8450 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
8451 sg->next = prev->next;
8452 cpumask_or(d->covered, d->covered, d->tmpmask);
8459 #endif /* CONFIG_NUMA */
8462 /* Free memory allocated for various sched_group structures */
8463 static void free_sched_groups(const struct cpumask *cpu_map,
8464 struct cpumask *nodemask)
8468 for_each_cpu(cpu, cpu_map) {
8469 struct sched_group **sched_group_nodes
8470 = sched_group_nodes_bycpu[cpu];
8472 if (!sched_group_nodes)
8475 for (i = 0; i < nr_node_ids; i++) {
8476 struct sched_group *oldsg, *sg = sched_group_nodes[i];
8478 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8479 if (cpumask_empty(nodemask))
8489 if (oldsg != sched_group_nodes[i])
8492 kfree(sched_group_nodes);
8493 sched_group_nodes_bycpu[cpu] = NULL;
8496 #else /* !CONFIG_NUMA */
8497 static void free_sched_groups(const struct cpumask *cpu_map,
8498 struct cpumask *nodemask)
8501 #endif /* CONFIG_NUMA */
8504 * Initialize sched groups cpu_power.
8506 * cpu_power indicates the capacity of sched group, which is used while
8507 * distributing the load between different sched groups in a sched domain.
8508 * Typically cpu_power for all the groups in a sched domain will be same unless
8509 * there are asymmetries in the topology. If there are asymmetries, group
8510 * having more cpu_power will pickup more load compared to the group having
8513 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
8515 struct sched_domain *child;
8516 struct sched_group *group;
8520 WARN_ON(!sd || !sd->groups);
8522 if (cpu != group_first_cpu(sd->groups))
8527 sd->groups->cpu_power = 0;
8530 power = SCHED_LOAD_SCALE;
8531 weight = cpumask_weight(sched_domain_span(sd));
8533 * SMT siblings share the power of a single core.
8534 * Usually multiple threads get a better yield out of
8535 * that one core than a single thread would have,
8536 * reflect that in sd->smt_gain.
8538 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
8539 power *= sd->smt_gain;
8541 power >>= SCHED_LOAD_SHIFT;
8543 sd->groups->cpu_power += power;
8548 * Add cpu_power of each child group to this groups cpu_power.
8550 group = child->groups;
8552 sd->groups->cpu_power += group->cpu_power;
8553 group = group->next;
8554 } while (group != child->groups);
8558 * Initializers for schedule domains
8559 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8562 #ifdef CONFIG_SCHED_DEBUG
8563 # define SD_INIT_NAME(sd, type) sd->name = #type
8565 # define SD_INIT_NAME(sd, type) do { } while (0)
8568 #define SD_INIT(sd, type) sd_init_##type(sd)
8570 #define SD_INIT_FUNC(type) \
8571 static noinline void sd_init_##type(struct sched_domain *sd) \
8573 memset(sd, 0, sizeof(*sd)); \
8574 *sd = SD_##type##_INIT; \
8575 sd->level = SD_LV_##type; \
8576 SD_INIT_NAME(sd, type); \
8581 SD_INIT_FUNC(ALLNODES)
8584 #ifdef CONFIG_SCHED_SMT
8585 SD_INIT_FUNC(SIBLING)
8587 #ifdef CONFIG_SCHED_MC
8591 static int default_relax_domain_level = -1;
8593 static int __init setup_relax_domain_level(char *str)
8597 val = simple_strtoul(str, NULL, 0);
8598 if (val < SD_LV_MAX)
8599 default_relax_domain_level = val;
8603 __setup("relax_domain_level=", setup_relax_domain_level);
8605 static void set_domain_attribute(struct sched_domain *sd,
8606 struct sched_domain_attr *attr)
8610 if (!attr || attr->relax_domain_level < 0) {
8611 if (default_relax_domain_level < 0)
8614 request = default_relax_domain_level;
8616 request = attr->relax_domain_level;
8617 if (request < sd->level) {
8618 /* turn off idle balance on this domain */
8619 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
8621 /* turn on idle balance on this domain */
8622 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
8626 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
8627 const struct cpumask *cpu_map)
8630 case sa_sched_groups:
8631 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
8632 d->sched_group_nodes = NULL;
8634 free_rootdomain(d->rd); /* fall through */
8636 free_cpumask_var(d->tmpmask); /* fall through */
8637 case sa_send_covered:
8638 free_cpumask_var(d->send_covered); /* fall through */
8639 case sa_this_core_map:
8640 free_cpumask_var(d->this_core_map); /* fall through */
8641 case sa_this_sibling_map:
8642 free_cpumask_var(d->this_sibling_map); /* fall through */
8644 free_cpumask_var(d->nodemask); /* fall through */
8645 case sa_sched_group_nodes:
8647 kfree(d->sched_group_nodes); /* fall through */
8649 free_cpumask_var(d->notcovered); /* fall through */
8651 free_cpumask_var(d->covered); /* fall through */
8653 free_cpumask_var(d->domainspan); /* fall through */
8660 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
8661 const struct cpumask *cpu_map)
8664 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
8666 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
8667 return sa_domainspan;
8668 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
8670 /* Allocate the per-node list of sched groups */
8671 d->sched_group_nodes = kcalloc(nr_node_ids,
8672 sizeof(struct sched_group *), GFP_KERNEL);
8673 if (!d->sched_group_nodes) {
8674 printk(KERN_WARNING "Can not alloc sched group node list\n");
8675 return sa_notcovered;
8677 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
8679 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
8680 return sa_sched_group_nodes;
8681 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
8683 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
8684 return sa_this_sibling_map;
8685 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
8686 return sa_this_core_map;
8687 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
8688 return sa_send_covered;
8689 d->rd = alloc_rootdomain();
8691 printk(KERN_WARNING "Cannot alloc root domain\n");
8694 return sa_rootdomain;
8697 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
8698 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
8700 struct sched_domain *sd = NULL;
8702 struct sched_domain *parent;
8705 if (cpumask_weight(cpu_map) >
8706 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
8707 sd = &per_cpu(allnodes_domains, i).sd;
8708 SD_INIT(sd, ALLNODES);
8709 set_domain_attribute(sd, attr);
8710 cpumask_copy(sched_domain_span(sd), cpu_map);
8711 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
8716 sd = &per_cpu(node_domains, i).sd;
8718 set_domain_attribute(sd, attr);
8719 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
8720 sd->parent = parent;
8723 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
8728 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
8729 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8730 struct sched_domain *parent, int i)
8732 struct sched_domain *sd;
8733 sd = &per_cpu(phys_domains, i).sd;
8735 set_domain_attribute(sd, attr);
8736 cpumask_copy(sched_domain_span(sd), d->nodemask);
8737 sd->parent = parent;
8740 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
8744 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
8745 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8746 struct sched_domain *parent, int i)
8748 struct sched_domain *sd = parent;
8749 #ifdef CONFIG_SCHED_MC
8750 sd = &per_cpu(core_domains, i).sd;
8752 set_domain_attribute(sd, attr);
8753 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
8754 sd->parent = parent;
8756 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
8761 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
8762 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8763 struct sched_domain *parent, int i)
8765 struct sched_domain *sd = parent;
8766 #ifdef CONFIG_SCHED_SMT
8767 sd = &per_cpu(cpu_domains, i).sd;
8768 SD_INIT(sd, SIBLING);
8769 set_domain_attribute(sd, attr);
8770 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
8771 sd->parent = parent;
8773 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
8778 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
8779 const struct cpumask *cpu_map, int cpu)
8782 #ifdef CONFIG_SCHED_SMT
8783 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
8784 cpumask_and(d->this_sibling_map, cpu_map,
8785 topology_thread_cpumask(cpu));
8786 if (cpu == cpumask_first(d->this_sibling_map))
8787 init_sched_build_groups(d->this_sibling_map, cpu_map,
8789 d->send_covered, d->tmpmask);
8792 #ifdef CONFIG_SCHED_MC
8793 case SD_LV_MC: /* set up multi-core groups */
8794 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
8795 if (cpu == cpumask_first(d->this_core_map))
8796 init_sched_build_groups(d->this_core_map, cpu_map,
8798 d->send_covered, d->tmpmask);
8801 case SD_LV_CPU: /* set up physical groups */
8802 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
8803 if (!cpumask_empty(d->nodemask))
8804 init_sched_build_groups(d->nodemask, cpu_map,
8806 d->send_covered, d->tmpmask);
8809 case SD_LV_ALLNODES:
8810 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
8811 d->send_covered, d->tmpmask);
8820 * Build sched domains for a given set of cpus and attach the sched domains
8821 * to the individual cpus
8823 static int __build_sched_domains(const struct cpumask *cpu_map,
8824 struct sched_domain_attr *attr)
8826 enum s_alloc alloc_state = sa_none;
8828 struct sched_domain *sd;
8834 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
8835 if (alloc_state != sa_rootdomain)
8837 alloc_state = sa_sched_groups;
8840 * Set up domains for cpus specified by the cpu_map.
8842 for_each_cpu(i, cpu_map) {
8843 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
8846 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
8847 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
8848 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
8849 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
8852 for_each_cpu(i, cpu_map) {
8853 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
8854 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
8857 /* Set up physical groups */
8858 for (i = 0; i < nr_node_ids; i++)
8859 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
8862 /* Set up node groups */
8864 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
8866 for (i = 0; i < nr_node_ids; i++)
8867 if (build_numa_sched_groups(&d, cpu_map, i))
8871 /* Calculate CPU power for physical packages and nodes */
8872 #ifdef CONFIG_SCHED_SMT
8873 for_each_cpu(i, cpu_map) {
8874 sd = &per_cpu(cpu_domains, i).sd;
8875 init_sched_groups_power(i, sd);
8878 #ifdef CONFIG_SCHED_MC
8879 for_each_cpu(i, cpu_map) {
8880 sd = &per_cpu(core_domains, i).sd;
8881 init_sched_groups_power(i, sd);
8885 for_each_cpu(i, cpu_map) {
8886 sd = &per_cpu(phys_domains, i).sd;
8887 init_sched_groups_power(i, sd);
8891 for (i = 0; i < nr_node_ids; i++)
8892 init_numa_sched_groups_power(d.sched_group_nodes[i]);
8894 if (d.sd_allnodes) {
8895 struct sched_group *sg;
8897 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
8899 init_numa_sched_groups_power(sg);
8903 /* Attach the domains */
8904 for_each_cpu(i, cpu_map) {
8905 #ifdef CONFIG_SCHED_SMT
8906 sd = &per_cpu(cpu_domains, i).sd;
8907 #elif defined(CONFIG_SCHED_MC)
8908 sd = &per_cpu(core_domains, i).sd;
8910 sd = &per_cpu(phys_domains, i).sd;
8912 cpu_attach_domain(sd, d.rd, i);
8915 d.sched_group_nodes = NULL; /* don't free this we still need it */
8916 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
8920 __free_domain_allocs(&d, alloc_state, cpu_map);
8924 static int build_sched_domains(const struct cpumask *cpu_map)
8926 return __build_sched_domains(cpu_map, NULL);
8929 static cpumask_var_t *doms_cur; /* current sched domains */
8930 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
8931 static struct sched_domain_attr *dattr_cur;
8932 /* attribues of custom domains in 'doms_cur' */
8935 * Special case: If a kmalloc of a doms_cur partition (array of
8936 * cpumask) fails, then fallback to a single sched domain,
8937 * as determined by the single cpumask fallback_doms.
8939 static cpumask_var_t fallback_doms;
8942 * arch_update_cpu_topology lets virtualized architectures update the
8943 * cpu core maps. It is supposed to return 1 if the topology changed
8944 * or 0 if it stayed the same.
8946 int __attribute__((weak)) arch_update_cpu_topology(void)
8951 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
8954 cpumask_var_t *doms;
8956 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
8959 for (i = 0; i < ndoms; i++) {
8960 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
8961 free_sched_domains(doms, i);
8968 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
8971 for (i = 0; i < ndoms; i++)
8972 free_cpumask_var(doms[i]);
8977 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
8978 * For now this just excludes isolated cpus, but could be used to
8979 * exclude other special cases in the future.
8981 static int arch_init_sched_domains(const struct cpumask *cpu_map)
8985 arch_update_cpu_topology();
8987 doms_cur = alloc_sched_domains(ndoms_cur);
8989 doms_cur = &fallback_doms;
8990 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
8992 err = build_sched_domains(doms_cur[0]);
8993 register_sched_domain_sysctl();
8998 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
8999 struct cpumask *tmpmask)
9001 free_sched_groups(cpu_map, tmpmask);
9005 * Detach sched domains from a group of cpus specified in cpu_map
9006 * These cpus will now be attached to the NULL domain
9008 static void detach_destroy_domains(const struct cpumask *cpu_map)
9010 /* Save because hotplug lock held. */
9011 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
9014 for_each_cpu(i, cpu_map)
9015 cpu_attach_domain(NULL, &def_root_domain, i);
9016 synchronize_sched();
9017 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
9020 /* handle null as "default" */
9021 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
9022 struct sched_domain_attr *new, int idx_new)
9024 struct sched_domain_attr tmp;
9031 return !memcmp(cur ? (cur + idx_cur) : &tmp,
9032 new ? (new + idx_new) : &tmp,
9033 sizeof(struct sched_domain_attr));
9037 * Partition sched domains as specified by the 'ndoms_new'
9038 * cpumasks in the array doms_new[] of cpumasks. This compares
9039 * doms_new[] to the current sched domain partitioning, doms_cur[].
9040 * It destroys each deleted domain and builds each new domain.
9042 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
9043 * The masks don't intersect (don't overlap.) We should setup one
9044 * sched domain for each mask. CPUs not in any of the cpumasks will
9045 * not be load balanced. If the same cpumask appears both in the
9046 * current 'doms_cur' domains and in the new 'doms_new', we can leave
9049 * The passed in 'doms_new' should be allocated using
9050 * alloc_sched_domains. This routine takes ownership of it and will
9051 * free_sched_domains it when done with it. If the caller failed the
9052 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
9053 * and partition_sched_domains() will fallback to the single partition
9054 * 'fallback_doms', it also forces the domains to be rebuilt.
9056 * If doms_new == NULL it will be replaced with cpu_online_mask.
9057 * ndoms_new == 0 is a special case for destroying existing domains,
9058 * and it will not create the default domain.
9060 * Call with hotplug lock held
9062 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
9063 struct sched_domain_attr *dattr_new)
9068 mutex_lock(&sched_domains_mutex);
9070 /* always unregister in case we don't destroy any domains */
9071 unregister_sched_domain_sysctl();
9073 /* Let architecture update cpu core mappings. */
9074 new_topology = arch_update_cpu_topology();
9076 n = doms_new ? ndoms_new : 0;
9078 /* Destroy deleted domains */
9079 for (i = 0; i < ndoms_cur; i++) {
9080 for (j = 0; j < n && !new_topology; j++) {
9081 if (cpumask_equal(doms_cur[i], doms_new[j])
9082 && dattrs_equal(dattr_cur, i, dattr_new, j))
9085 /* no match - a current sched domain not in new doms_new[] */
9086 detach_destroy_domains(doms_cur[i]);
9091 if (doms_new == NULL) {
9093 doms_new = &fallback_doms;
9094 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
9095 WARN_ON_ONCE(dattr_new);
9098 /* Build new domains */
9099 for (i = 0; i < ndoms_new; i++) {
9100 for (j = 0; j < ndoms_cur && !new_topology; j++) {
9101 if (cpumask_equal(doms_new[i], doms_cur[j])
9102 && dattrs_equal(dattr_new, i, dattr_cur, j))
9105 /* no match - add a new doms_new */
9106 __build_sched_domains(doms_new[i],
9107 dattr_new ? dattr_new + i : NULL);
9112 /* Remember the new sched domains */
9113 if (doms_cur != &fallback_doms)
9114 free_sched_domains(doms_cur, ndoms_cur);
9115 kfree(dattr_cur); /* kfree(NULL) is safe */
9116 doms_cur = doms_new;
9117 dattr_cur = dattr_new;
9118 ndoms_cur = ndoms_new;
9120 register_sched_domain_sysctl();
9122 mutex_unlock(&sched_domains_mutex);
9125 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
9126 static void arch_reinit_sched_domains(void)
9130 /* Destroy domains first to force the rebuild */
9131 partition_sched_domains(0, NULL, NULL);
9133 rebuild_sched_domains();
9137 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
9139 unsigned int level = 0;
9141 if (sscanf(buf, "%u", &level) != 1)
9145 * level is always be positive so don't check for
9146 * level < POWERSAVINGS_BALANCE_NONE which is 0
9147 * What happens on 0 or 1 byte write,
9148 * need to check for count as well?
9151 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
9155 sched_smt_power_savings = level;
9157 sched_mc_power_savings = level;
9159 arch_reinit_sched_domains();
9164 #ifdef CONFIG_SCHED_MC
9165 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
9168 return sprintf(page, "%u\n", sched_mc_power_savings);
9170 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
9171 const char *buf, size_t count)
9173 return sched_power_savings_store(buf, count, 0);
9175 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
9176 sched_mc_power_savings_show,
9177 sched_mc_power_savings_store);
9180 #ifdef CONFIG_SCHED_SMT
9181 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
9184 return sprintf(page, "%u\n", sched_smt_power_savings);
9186 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
9187 const char *buf, size_t count)
9189 return sched_power_savings_store(buf, count, 1);
9191 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
9192 sched_smt_power_savings_show,
9193 sched_smt_power_savings_store);
9196 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
9200 #ifdef CONFIG_SCHED_SMT
9202 err = sysfs_create_file(&cls->kset.kobj,
9203 &attr_sched_smt_power_savings.attr);
9205 #ifdef CONFIG_SCHED_MC
9206 if (!err && mc_capable())
9207 err = sysfs_create_file(&cls->kset.kobj,
9208 &attr_sched_mc_power_savings.attr);
9212 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
9214 #ifndef CONFIG_CPUSETS
9216 * Add online and remove offline CPUs from the scheduler domains.
9217 * When cpusets are enabled they take over this function.
9219 static int update_sched_domains(struct notifier_block *nfb,
9220 unsigned long action, void *hcpu)
9224 case CPU_ONLINE_FROZEN:
9225 case CPU_DOWN_PREPARE:
9226 case CPU_DOWN_PREPARE_FROZEN:
9227 case CPU_DOWN_FAILED:
9228 case CPU_DOWN_FAILED_FROZEN:
9229 partition_sched_domains(1, NULL, NULL);
9238 static int update_runtime(struct notifier_block *nfb,
9239 unsigned long action, void *hcpu)
9241 int cpu = (int)(long)hcpu;
9244 case CPU_DOWN_PREPARE:
9245 case CPU_DOWN_PREPARE_FROZEN:
9246 disable_runtime(cpu_rq(cpu));
9249 case CPU_DOWN_FAILED:
9250 case CPU_DOWN_FAILED_FROZEN:
9252 case CPU_ONLINE_FROZEN:
9253 enable_runtime(cpu_rq(cpu));
9261 void __init sched_init_smp(void)
9263 cpumask_var_t non_isolated_cpus;
9265 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
9266 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
9268 #if defined(CONFIG_NUMA)
9269 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
9271 BUG_ON(sched_group_nodes_bycpu == NULL);
9274 mutex_lock(&sched_domains_mutex);
9275 arch_init_sched_domains(cpu_active_mask);
9276 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
9277 if (cpumask_empty(non_isolated_cpus))
9278 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
9279 mutex_unlock(&sched_domains_mutex);
9282 #ifndef CONFIG_CPUSETS
9283 /* XXX: Theoretical race here - CPU may be hotplugged now */
9284 hotcpu_notifier(update_sched_domains, 0);
9287 /* RT runtime code needs to handle some hotplug events */
9288 hotcpu_notifier(update_runtime, 0);
9292 /* Move init over to a non-isolated CPU */
9293 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
9295 sched_init_granularity();
9296 free_cpumask_var(non_isolated_cpus);
9298 init_sched_rt_class();
9301 void __init sched_init_smp(void)
9303 sched_init_granularity();
9305 #endif /* CONFIG_SMP */
9307 const_debug unsigned int sysctl_timer_migration = 1;
9309 int in_sched_functions(unsigned long addr)
9311 return in_lock_functions(addr) ||
9312 (addr >= (unsigned long)__sched_text_start
9313 && addr < (unsigned long)__sched_text_end);
9316 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
9318 cfs_rq->tasks_timeline = RB_ROOT;
9319 INIT_LIST_HEAD(&cfs_rq->tasks);
9320 #ifdef CONFIG_FAIR_GROUP_SCHED
9323 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9326 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
9328 struct rt_prio_array *array;
9331 array = &rt_rq->active;
9332 for (i = 0; i < MAX_RT_PRIO; i++) {
9333 INIT_LIST_HEAD(array->queue + i);
9334 __clear_bit(i, array->bitmap);
9336 /* delimiter for bitsearch: */
9337 __set_bit(MAX_RT_PRIO, array->bitmap);
9339 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9340 rt_rq->highest_prio.curr = MAX_RT_PRIO;
9342 rt_rq->highest_prio.next = MAX_RT_PRIO;
9346 rt_rq->rt_nr_migratory = 0;
9347 rt_rq->overloaded = 0;
9348 plist_head_init(&rt_rq->pushable_tasks, &rq->lock);
9352 rt_rq->rt_throttled = 0;
9353 rt_rq->rt_runtime = 0;
9354 spin_lock_init(&rt_rq->rt_runtime_lock);
9356 #ifdef CONFIG_RT_GROUP_SCHED
9357 rt_rq->rt_nr_boosted = 0;
9362 #ifdef CONFIG_FAIR_GROUP_SCHED
9363 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9364 struct sched_entity *se, int cpu, int add,
9365 struct sched_entity *parent)
9367 struct rq *rq = cpu_rq(cpu);
9368 tg->cfs_rq[cpu] = cfs_rq;
9369 init_cfs_rq(cfs_rq, rq);
9372 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
9375 /* se could be NULL for init_task_group */
9380 se->cfs_rq = &rq->cfs;
9382 se->cfs_rq = parent->my_q;
9385 se->load.weight = tg->shares;
9386 se->load.inv_weight = 0;
9387 se->parent = parent;
9391 #ifdef CONFIG_RT_GROUP_SCHED
9392 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
9393 struct sched_rt_entity *rt_se, int cpu, int add,
9394 struct sched_rt_entity *parent)
9396 struct rq *rq = cpu_rq(cpu);
9398 tg->rt_rq[cpu] = rt_rq;
9399 init_rt_rq(rt_rq, rq);
9401 rt_rq->rt_se = rt_se;
9402 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
9404 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
9406 tg->rt_se[cpu] = rt_se;
9411 rt_se->rt_rq = &rq->rt;
9413 rt_se->rt_rq = parent->my_q;
9415 rt_se->my_q = rt_rq;
9416 rt_se->parent = parent;
9417 INIT_LIST_HEAD(&rt_se->run_list);
9421 void __init sched_init(void)
9424 unsigned long alloc_size = 0, ptr;
9426 #ifdef CONFIG_FAIR_GROUP_SCHED
9427 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9429 #ifdef CONFIG_RT_GROUP_SCHED
9430 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9432 #ifdef CONFIG_USER_SCHED
9435 #ifdef CONFIG_CPUMASK_OFFSTACK
9436 alloc_size += num_possible_cpus() * cpumask_size();
9439 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
9441 #ifdef CONFIG_FAIR_GROUP_SCHED
9442 init_task_group.se = (struct sched_entity **)ptr;
9443 ptr += nr_cpu_ids * sizeof(void **);
9445 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
9446 ptr += nr_cpu_ids * sizeof(void **);
9448 #ifdef CONFIG_USER_SCHED
9449 root_task_group.se = (struct sched_entity **)ptr;
9450 ptr += nr_cpu_ids * sizeof(void **);
9452 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9453 ptr += nr_cpu_ids * sizeof(void **);
9454 #endif /* CONFIG_USER_SCHED */
9455 #endif /* CONFIG_FAIR_GROUP_SCHED */
9456 #ifdef CONFIG_RT_GROUP_SCHED
9457 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
9458 ptr += nr_cpu_ids * sizeof(void **);
9460 init_task_group.rt_rq = (struct rt_rq **)ptr;
9461 ptr += nr_cpu_ids * sizeof(void **);
9463 #ifdef CONFIG_USER_SCHED
9464 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9465 ptr += nr_cpu_ids * sizeof(void **);
9467 root_task_group.rt_rq = (struct rt_rq **)ptr;
9468 ptr += nr_cpu_ids * sizeof(void **);
9469 #endif /* CONFIG_USER_SCHED */
9470 #endif /* CONFIG_RT_GROUP_SCHED */
9471 #ifdef CONFIG_CPUMASK_OFFSTACK
9472 for_each_possible_cpu(i) {
9473 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
9474 ptr += cpumask_size();
9476 #endif /* CONFIG_CPUMASK_OFFSTACK */
9480 init_defrootdomain();
9483 init_rt_bandwidth(&def_rt_bandwidth,
9484 global_rt_period(), global_rt_runtime());
9486 #ifdef CONFIG_RT_GROUP_SCHED
9487 init_rt_bandwidth(&init_task_group.rt_bandwidth,
9488 global_rt_period(), global_rt_runtime());
9489 #ifdef CONFIG_USER_SCHED
9490 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9491 global_rt_period(), RUNTIME_INF);
9492 #endif /* CONFIG_USER_SCHED */
9493 #endif /* CONFIG_RT_GROUP_SCHED */
9495 #ifdef CONFIG_GROUP_SCHED
9496 list_add(&init_task_group.list, &task_groups);
9497 INIT_LIST_HEAD(&init_task_group.children);
9499 #ifdef CONFIG_USER_SCHED
9500 INIT_LIST_HEAD(&root_task_group.children);
9501 init_task_group.parent = &root_task_group;
9502 list_add(&init_task_group.siblings, &root_task_group.children);
9503 #endif /* CONFIG_USER_SCHED */
9504 #endif /* CONFIG_GROUP_SCHED */
9506 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
9507 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
9508 __alignof__(unsigned long));
9510 for_each_possible_cpu(i) {
9514 spin_lock_init(&rq->lock);
9516 rq->calc_load_active = 0;
9517 rq->calc_load_update = jiffies + LOAD_FREQ;
9518 init_cfs_rq(&rq->cfs, rq);
9519 init_rt_rq(&rq->rt, rq);
9520 #ifdef CONFIG_FAIR_GROUP_SCHED
9521 init_task_group.shares = init_task_group_load;
9522 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9523 #ifdef CONFIG_CGROUP_SCHED
9525 * How much cpu bandwidth does init_task_group get?
9527 * In case of task-groups formed thr' the cgroup filesystem, it
9528 * gets 100% of the cpu resources in the system. This overall
9529 * system cpu resource is divided among the tasks of
9530 * init_task_group and its child task-groups in a fair manner,
9531 * based on each entity's (task or task-group's) weight
9532 * (se->load.weight).
9534 * In other words, if init_task_group has 10 tasks of weight
9535 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9536 * then A0's share of the cpu resource is:
9538 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9540 * We achieve this by letting init_task_group's tasks sit
9541 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9543 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
9544 #elif defined CONFIG_USER_SCHED
9545 root_task_group.shares = NICE_0_LOAD;
9546 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
9548 * In case of task-groups formed thr' the user id of tasks,
9549 * init_task_group represents tasks belonging to root user.
9550 * Hence it forms a sibling of all subsequent groups formed.
9551 * In this case, init_task_group gets only a fraction of overall
9552 * system cpu resource, based on the weight assigned to root
9553 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9554 * by letting tasks of init_task_group sit in a separate cfs_rq
9555 * (init_tg_cfs_rq) and having one entity represent this group of
9556 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9558 init_tg_cfs_entry(&init_task_group,
9559 &per_cpu(init_tg_cfs_rq, i),
9560 &per_cpu(init_sched_entity, i), i, 1,
9561 root_task_group.se[i]);
9564 #endif /* CONFIG_FAIR_GROUP_SCHED */
9566 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9567 #ifdef CONFIG_RT_GROUP_SCHED
9568 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
9569 #ifdef CONFIG_CGROUP_SCHED
9570 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
9571 #elif defined CONFIG_USER_SCHED
9572 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
9573 init_tg_rt_entry(&init_task_group,
9574 &per_cpu(init_rt_rq, i),
9575 &per_cpu(init_sched_rt_entity, i), i, 1,
9576 root_task_group.rt_se[i]);
9580 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
9581 rq->cpu_load[j] = 0;
9585 rq->post_schedule = 0;
9586 rq->active_balance = 0;
9587 rq->next_balance = jiffies;
9591 rq->migration_thread = NULL;
9593 rq->avg_idle = 2*sysctl_sched_migration_cost;
9594 INIT_LIST_HEAD(&rq->migration_queue);
9595 rq_attach_root(rq, &def_root_domain);
9598 atomic_set(&rq->nr_iowait, 0);
9601 set_load_weight(&init_task);
9603 #ifdef CONFIG_PREEMPT_NOTIFIERS
9604 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
9608 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9611 #ifdef CONFIG_RT_MUTEXES
9612 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
9616 * The boot idle thread does lazy MMU switching as well:
9618 atomic_inc(&init_mm.mm_count);
9619 enter_lazy_tlb(&init_mm, current);
9622 * Make us the idle thread. Technically, schedule() should not be
9623 * called from this thread, however somewhere below it might be,
9624 * but because we are the idle thread, we just pick up running again
9625 * when this runqueue becomes "idle".
9627 init_idle(current, smp_processor_id());
9629 calc_load_update = jiffies + LOAD_FREQ;
9632 * During early bootup we pretend to be a normal task:
9634 current->sched_class = &fair_sched_class;
9636 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9637 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
9640 zalloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
9641 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
9643 /* May be allocated at isolcpus cmdline parse time */
9644 if (cpu_isolated_map == NULL)
9645 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
9650 scheduler_running = 1;
9653 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9654 static inline int preempt_count_equals(int preempt_offset)
9656 int nested = preempt_count() & ~PREEMPT_ACTIVE;
9658 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
9661 void __might_sleep(char *file, int line, int preempt_offset)
9664 static unsigned long prev_jiffy; /* ratelimiting */
9666 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
9667 system_state != SYSTEM_RUNNING || oops_in_progress)
9669 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9671 prev_jiffy = jiffies;
9674 "BUG: sleeping function called from invalid context at %s:%d\n",
9677 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9678 in_atomic(), irqs_disabled(),
9679 current->pid, current->comm);
9681 debug_show_held_locks(current);
9682 if (irqs_disabled())
9683 print_irqtrace_events(current);
9687 EXPORT_SYMBOL(__might_sleep);
9690 #ifdef CONFIG_MAGIC_SYSRQ
9691 static void normalize_task(struct rq *rq, struct task_struct *p)
9695 update_rq_clock(rq);
9696 on_rq = p->se.on_rq;
9698 deactivate_task(rq, p, 0);
9699 __setscheduler(rq, p, SCHED_NORMAL, 0);
9701 activate_task(rq, p, 0);
9702 resched_task(rq->curr);
9706 void normalize_rt_tasks(void)
9708 struct task_struct *g, *p;
9709 unsigned long flags;
9712 read_lock_irqsave(&tasklist_lock, flags);
9713 do_each_thread(g, p) {
9715 * Only normalize user tasks:
9720 p->se.exec_start = 0;
9721 #ifdef CONFIG_SCHEDSTATS
9722 p->se.wait_start = 0;
9723 p->se.sleep_start = 0;
9724 p->se.block_start = 0;
9729 * Renice negative nice level userspace
9732 if (TASK_NICE(p) < 0 && p->mm)
9733 set_user_nice(p, 0);
9737 spin_lock(&p->pi_lock);
9738 rq = __task_rq_lock(p);
9740 normalize_task(rq, p);
9742 __task_rq_unlock(rq);
9743 spin_unlock(&p->pi_lock);
9744 } while_each_thread(g, p);
9746 read_unlock_irqrestore(&tasklist_lock, flags);
9749 #endif /* CONFIG_MAGIC_SYSRQ */
9753 * These functions are only useful for the IA64 MCA handling.
9755 * They can only be called when the whole system has been
9756 * stopped - every CPU needs to be quiescent, and no scheduling
9757 * activity can take place. Using them for anything else would
9758 * be a serious bug, and as a result, they aren't even visible
9759 * under any other configuration.
9763 * curr_task - return the current task for a given cpu.
9764 * @cpu: the processor in question.
9766 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9768 struct task_struct *curr_task(int cpu)
9770 return cpu_curr(cpu);
9774 * set_curr_task - set the current task for a given cpu.
9775 * @cpu: the processor in question.
9776 * @p: the task pointer to set.
9778 * Description: This function must only be used when non-maskable interrupts
9779 * are serviced on a separate stack. It allows the architecture to switch the
9780 * notion of the current task on a cpu in a non-blocking manner. This function
9781 * must be called with all CPU's synchronized, and interrupts disabled, the
9782 * and caller must save the original value of the current task (see
9783 * curr_task() above) and restore that value before reenabling interrupts and
9784 * re-starting the system.
9786 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9788 void set_curr_task(int cpu, struct task_struct *p)
9795 #ifdef CONFIG_FAIR_GROUP_SCHED
9796 static void free_fair_sched_group(struct task_group *tg)
9800 for_each_possible_cpu(i) {
9802 kfree(tg->cfs_rq[i]);
9812 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9814 struct cfs_rq *cfs_rq;
9815 struct sched_entity *se;
9819 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9822 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9826 tg->shares = NICE_0_LOAD;
9828 for_each_possible_cpu(i) {
9831 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9832 GFP_KERNEL, cpu_to_node(i));
9836 se = kzalloc_node(sizeof(struct sched_entity),
9837 GFP_KERNEL, cpu_to_node(i));
9841 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
9850 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9852 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
9853 &cpu_rq(cpu)->leaf_cfs_rq_list);
9856 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9858 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
9860 #else /* !CONFG_FAIR_GROUP_SCHED */
9861 static inline void free_fair_sched_group(struct task_group *tg)
9866 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9871 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9875 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9878 #endif /* CONFIG_FAIR_GROUP_SCHED */
9880 #ifdef CONFIG_RT_GROUP_SCHED
9881 static void free_rt_sched_group(struct task_group *tg)
9885 destroy_rt_bandwidth(&tg->rt_bandwidth);
9887 for_each_possible_cpu(i) {
9889 kfree(tg->rt_rq[i]);
9891 kfree(tg->rt_se[i]);
9899 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9901 struct rt_rq *rt_rq;
9902 struct sched_rt_entity *rt_se;
9906 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
9909 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
9913 init_rt_bandwidth(&tg->rt_bandwidth,
9914 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
9916 for_each_possible_cpu(i) {
9919 rt_rq = kzalloc_node(sizeof(struct rt_rq),
9920 GFP_KERNEL, cpu_to_node(i));
9924 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
9925 GFP_KERNEL, cpu_to_node(i));
9929 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
9938 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9940 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
9941 &cpu_rq(cpu)->leaf_rt_rq_list);
9944 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9946 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
9948 #else /* !CONFIG_RT_GROUP_SCHED */
9949 static inline void free_rt_sched_group(struct task_group *tg)
9954 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9959 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9963 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9966 #endif /* CONFIG_RT_GROUP_SCHED */
9968 #ifdef CONFIG_GROUP_SCHED
9969 static void free_sched_group(struct task_group *tg)
9971 free_fair_sched_group(tg);
9972 free_rt_sched_group(tg);
9976 /* allocate runqueue etc for a new task group */
9977 struct task_group *sched_create_group(struct task_group *parent)
9979 struct task_group *tg;
9980 unsigned long flags;
9983 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
9985 return ERR_PTR(-ENOMEM);
9987 if (!alloc_fair_sched_group(tg, parent))
9990 if (!alloc_rt_sched_group(tg, parent))
9993 spin_lock_irqsave(&task_group_lock, flags);
9994 for_each_possible_cpu(i) {
9995 register_fair_sched_group(tg, i);
9996 register_rt_sched_group(tg, i);
9998 list_add_rcu(&tg->list, &task_groups);
10000 WARN_ON(!parent); /* root should already exist */
10002 tg->parent = parent;
10003 INIT_LIST_HEAD(&tg->children);
10004 list_add_rcu(&tg->siblings, &parent->children);
10005 spin_unlock_irqrestore(&task_group_lock, flags);
10010 free_sched_group(tg);
10011 return ERR_PTR(-ENOMEM);
10014 /* rcu callback to free various structures associated with a task group */
10015 static void free_sched_group_rcu(struct rcu_head *rhp)
10017 /* now it should be safe to free those cfs_rqs */
10018 free_sched_group(container_of(rhp, struct task_group, rcu));
10021 /* Destroy runqueue etc associated with a task group */
10022 void sched_destroy_group(struct task_group *tg)
10024 unsigned long flags;
10027 spin_lock_irqsave(&task_group_lock, flags);
10028 for_each_possible_cpu(i) {
10029 unregister_fair_sched_group(tg, i);
10030 unregister_rt_sched_group(tg, i);
10032 list_del_rcu(&tg->list);
10033 list_del_rcu(&tg->siblings);
10034 spin_unlock_irqrestore(&task_group_lock, flags);
10036 /* wait for possible concurrent references to cfs_rqs complete */
10037 call_rcu(&tg->rcu, free_sched_group_rcu);
10040 /* change task's runqueue when it moves between groups.
10041 * The caller of this function should have put the task in its new group
10042 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
10043 * reflect its new group.
10045 void sched_move_task(struct task_struct *tsk)
10047 int on_rq, running;
10048 unsigned long flags;
10051 rq = task_rq_lock(tsk, &flags);
10053 update_rq_clock(rq);
10055 running = task_current(rq, tsk);
10056 on_rq = tsk->se.on_rq;
10059 dequeue_task(rq, tsk, 0);
10060 if (unlikely(running))
10061 tsk->sched_class->put_prev_task(rq, tsk);
10063 set_task_rq(tsk, task_cpu(tsk));
10065 #ifdef CONFIG_FAIR_GROUP_SCHED
10066 if (tsk->sched_class->moved_group)
10067 tsk->sched_class->moved_group(tsk);
10070 if (unlikely(running))
10071 tsk->sched_class->set_curr_task(rq);
10073 enqueue_task(rq, tsk, 0);
10075 task_rq_unlock(rq, &flags);
10077 #endif /* CONFIG_GROUP_SCHED */
10079 #ifdef CONFIG_FAIR_GROUP_SCHED
10080 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
10082 struct cfs_rq *cfs_rq = se->cfs_rq;
10087 dequeue_entity(cfs_rq, se, 0);
10089 se->load.weight = shares;
10090 se->load.inv_weight = 0;
10093 enqueue_entity(cfs_rq, se, 0);
10096 static void set_se_shares(struct sched_entity *se, unsigned long shares)
10098 struct cfs_rq *cfs_rq = se->cfs_rq;
10099 struct rq *rq = cfs_rq->rq;
10100 unsigned long flags;
10102 spin_lock_irqsave(&rq->lock, flags);
10103 __set_se_shares(se, shares);
10104 spin_unlock_irqrestore(&rq->lock, flags);
10107 static DEFINE_MUTEX(shares_mutex);
10109 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10112 unsigned long flags;
10115 * We can't change the weight of the root cgroup.
10120 if (shares < MIN_SHARES)
10121 shares = MIN_SHARES;
10122 else if (shares > MAX_SHARES)
10123 shares = MAX_SHARES;
10125 mutex_lock(&shares_mutex);
10126 if (tg->shares == shares)
10129 spin_lock_irqsave(&task_group_lock, flags);
10130 for_each_possible_cpu(i)
10131 unregister_fair_sched_group(tg, i);
10132 list_del_rcu(&tg->siblings);
10133 spin_unlock_irqrestore(&task_group_lock, flags);
10135 /* wait for any ongoing reference to this group to finish */
10136 synchronize_sched();
10139 * Now we are free to modify the group's share on each cpu
10140 * w/o tripping rebalance_share or load_balance_fair.
10142 tg->shares = shares;
10143 for_each_possible_cpu(i) {
10145 * force a rebalance
10147 cfs_rq_set_shares(tg->cfs_rq[i], 0);
10148 set_se_shares(tg->se[i], shares);
10152 * Enable load balance activity on this group, by inserting it back on
10153 * each cpu's rq->leaf_cfs_rq_list.
10155 spin_lock_irqsave(&task_group_lock, flags);
10156 for_each_possible_cpu(i)
10157 register_fair_sched_group(tg, i);
10158 list_add_rcu(&tg->siblings, &tg->parent->children);
10159 spin_unlock_irqrestore(&task_group_lock, flags);
10161 mutex_unlock(&shares_mutex);
10165 unsigned long sched_group_shares(struct task_group *tg)
10171 #ifdef CONFIG_RT_GROUP_SCHED
10173 * Ensure that the real time constraints are schedulable.
10175 static DEFINE_MUTEX(rt_constraints_mutex);
10177 static unsigned long to_ratio(u64 period, u64 runtime)
10179 if (runtime == RUNTIME_INF)
10182 return div64_u64(runtime << 20, period);
10185 /* Must be called with tasklist_lock held */
10186 static inline int tg_has_rt_tasks(struct task_group *tg)
10188 struct task_struct *g, *p;
10190 do_each_thread(g, p) {
10191 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
10193 } while_each_thread(g, p);
10198 struct rt_schedulable_data {
10199 struct task_group *tg;
10204 static int tg_schedulable(struct task_group *tg, void *data)
10206 struct rt_schedulable_data *d = data;
10207 struct task_group *child;
10208 unsigned long total, sum = 0;
10209 u64 period, runtime;
10211 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10212 runtime = tg->rt_bandwidth.rt_runtime;
10215 period = d->rt_period;
10216 runtime = d->rt_runtime;
10219 #ifdef CONFIG_USER_SCHED
10220 if (tg == &root_task_group) {
10221 period = global_rt_period();
10222 runtime = global_rt_runtime();
10227 * Cannot have more runtime than the period.
10229 if (runtime > period && runtime != RUNTIME_INF)
10233 * Ensure we don't starve existing RT tasks.
10235 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
10238 total = to_ratio(period, runtime);
10241 * Nobody can have more than the global setting allows.
10243 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
10247 * The sum of our children's runtime should not exceed our own.
10249 list_for_each_entry_rcu(child, &tg->children, siblings) {
10250 period = ktime_to_ns(child->rt_bandwidth.rt_period);
10251 runtime = child->rt_bandwidth.rt_runtime;
10253 if (child == d->tg) {
10254 period = d->rt_period;
10255 runtime = d->rt_runtime;
10258 sum += to_ratio(period, runtime);
10267 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
10269 struct rt_schedulable_data data = {
10271 .rt_period = period,
10272 .rt_runtime = runtime,
10275 return walk_tg_tree(tg_schedulable, tg_nop, &data);
10278 static int tg_set_bandwidth(struct task_group *tg,
10279 u64 rt_period, u64 rt_runtime)
10283 mutex_lock(&rt_constraints_mutex);
10284 read_lock(&tasklist_lock);
10285 err = __rt_schedulable(tg, rt_period, rt_runtime);
10289 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10290 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
10291 tg->rt_bandwidth.rt_runtime = rt_runtime;
10293 for_each_possible_cpu(i) {
10294 struct rt_rq *rt_rq = tg->rt_rq[i];
10296 spin_lock(&rt_rq->rt_runtime_lock);
10297 rt_rq->rt_runtime = rt_runtime;
10298 spin_unlock(&rt_rq->rt_runtime_lock);
10300 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10302 read_unlock(&tasklist_lock);
10303 mutex_unlock(&rt_constraints_mutex);
10308 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
10310 u64 rt_runtime, rt_period;
10312 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10313 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
10314 if (rt_runtime_us < 0)
10315 rt_runtime = RUNTIME_INF;
10317 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10320 long sched_group_rt_runtime(struct task_group *tg)
10324 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
10327 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
10328 do_div(rt_runtime_us, NSEC_PER_USEC);
10329 return rt_runtime_us;
10332 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
10334 u64 rt_runtime, rt_period;
10336 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
10337 rt_runtime = tg->rt_bandwidth.rt_runtime;
10339 if (rt_period == 0)
10342 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10345 long sched_group_rt_period(struct task_group *tg)
10349 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
10350 do_div(rt_period_us, NSEC_PER_USEC);
10351 return rt_period_us;
10354 static int sched_rt_global_constraints(void)
10356 u64 runtime, period;
10359 if (sysctl_sched_rt_period <= 0)
10362 runtime = global_rt_runtime();
10363 period = global_rt_period();
10366 * Sanity check on the sysctl variables.
10368 if (runtime > period && runtime != RUNTIME_INF)
10371 mutex_lock(&rt_constraints_mutex);
10372 read_lock(&tasklist_lock);
10373 ret = __rt_schedulable(NULL, 0, 0);
10374 read_unlock(&tasklist_lock);
10375 mutex_unlock(&rt_constraints_mutex);
10380 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
10382 /* Don't accept realtime tasks when there is no way for them to run */
10383 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
10389 #else /* !CONFIG_RT_GROUP_SCHED */
10390 static int sched_rt_global_constraints(void)
10392 unsigned long flags;
10395 if (sysctl_sched_rt_period <= 0)
10399 * There's always some RT tasks in the root group
10400 * -- migration, kstopmachine etc..
10402 if (sysctl_sched_rt_runtime == 0)
10405 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
10406 for_each_possible_cpu(i) {
10407 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
10409 spin_lock(&rt_rq->rt_runtime_lock);
10410 rt_rq->rt_runtime = global_rt_runtime();
10411 spin_unlock(&rt_rq->rt_runtime_lock);
10413 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
10417 #endif /* CONFIG_RT_GROUP_SCHED */
10419 int sched_rt_handler(struct ctl_table *table, int write,
10420 void __user *buffer, size_t *lenp,
10424 int old_period, old_runtime;
10425 static DEFINE_MUTEX(mutex);
10427 mutex_lock(&mutex);
10428 old_period = sysctl_sched_rt_period;
10429 old_runtime = sysctl_sched_rt_runtime;
10431 ret = proc_dointvec(table, write, buffer, lenp, ppos);
10433 if (!ret && write) {
10434 ret = sched_rt_global_constraints();
10436 sysctl_sched_rt_period = old_period;
10437 sysctl_sched_rt_runtime = old_runtime;
10439 def_rt_bandwidth.rt_runtime = global_rt_runtime();
10440 def_rt_bandwidth.rt_period =
10441 ns_to_ktime(global_rt_period());
10444 mutex_unlock(&mutex);
10449 #ifdef CONFIG_CGROUP_SCHED
10451 /* return corresponding task_group object of a cgroup */
10452 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
10454 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
10455 struct task_group, css);
10458 static struct cgroup_subsys_state *
10459 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
10461 struct task_group *tg, *parent;
10463 if (!cgrp->parent) {
10464 /* This is early initialization for the top cgroup */
10465 return &init_task_group.css;
10468 parent = cgroup_tg(cgrp->parent);
10469 tg = sched_create_group(parent);
10471 return ERR_PTR(-ENOMEM);
10477 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10479 struct task_group *tg = cgroup_tg(cgrp);
10481 sched_destroy_group(tg);
10485 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
10487 #ifdef CONFIG_RT_GROUP_SCHED
10488 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
10491 /* We don't support RT-tasks being in separate groups */
10492 if (tsk->sched_class != &fair_sched_class)
10499 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10500 struct task_struct *tsk, bool threadgroup)
10502 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
10506 struct task_struct *c;
10508 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
10509 retval = cpu_cgroup_can_attach_task(cgrp, c);
10521 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10522 struct cgroup *old_cont, struct task_struct *tsk,
10525 sched_move_task(tsk);
10527 struct task_struct *c;
10529 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
10530 sched_move_task(c);
10536 #ifdef CONFIG_FAIR_GROUP_SCHED
10537 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
10540 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
10543 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
10545 struct task_group *tg = cgroup_tg(cgrp);
10547 return (u64) tg->shares;
10549 #endif /* CONFIG_FAIR_GROUP_SCHED */
10551 #ifdef CONFIG_RT_GROUP_SCHED
10552 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
10555 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
10558 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
10560 return sched_group_rt_runtime(cgroup_tg(cgrp));
10563 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
10566 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
10569 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
10571 return sched_group_rt_period(cgroup_tg(cgrp));
10573 #endif /* CONFIG_RT_GROUP_SCHED */
10575 static struct cftype cpu_files[] = {
10576 #ifdef CONFIG_FAIR_GROUP_SCHED
10579 .read_u64 = cpu_shares_read_u64,
10580 .write_u64 = cpu_shares_write_u64,
10583 #ifdef CONFIG_RT_GROUP_SCHED
10585 .name = "rt_runtime_us",
10586 .read_s64 = cpu_rt_runtime_read,
10587 .write_s64 = cpu_rt_runtime_write,
10590 .name = "rt_period_us",
10591 .read_u64 = cpu_rt_period_read_uint,
10592 .write_u64 = cpu_rt_period_write_uint,
10597 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
10599 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
10602 struct cgroup_subsys cpu_cgroup_subsys = {
10604 .create = cpu_cgroup_create,
10605 .destroy = cpu_cgroup_destroy,
10606 .can_attach = cpu_cgroup_can_attach,
10607 .attach = cpu_cgroup_attach,
10608 .populate = cpu_cgroup_populate,
10609 .subsys_id = cpu_cgroup_subsys_id,
10613 #endif /* CONFIG_CGROUP_SCHED */
10615 #ifdef CONFIG_CGROUP_CPUACCT
10618 * CPU accounting code for task groups.
10620 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10621 * (balbir@in.ibm.com).
10624 /* track cpu usage of a group of tasks and its child groups */
10626 struct cgroup_subsys_state css;
10627 /* cpuusage holds pointer to a u64-type object on every cpu */
10629 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
10630 struct cpuacct *parent;
10633 struct cgroup_subsys cpuacct_subsys;
10635 /* return cpu accounting group corresponding to this container */
10636 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
10638 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
10639 struct cpuacct, css);
10642 /* return cpu accounting group to which this task belongs */
10643 static inline struct cpuacct *task_ca(struct task_struct *tsk)
10645 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
10646 struct cpuacct, css);
10649 /* create a new cpu accounting group */
10650 static struct cgroup_subsys_state *cpuacct_create(
10651 struct cgroup_subsys *ss, struct cgroup *cgrp)
10653 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
10659 ca->cpuusage = alloc_percpu(u64);
10663 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10664 if (percpu_counter_init(&ca->cpustat[i], 0))
10665 goto out_free_counters;
10668 ca->parent = cgroup_ca(cgrp->parent);
10674 percpu_counter_destroy(&ca->cpustat[i]);
10675 free_percpu(ca->cpuusage);
10679 return ERR_PTR(-ENOMEM);
10682 /* destroy an existing cpu accounting group */
10684 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10686 struct cpuacct *ca = cgroup_ca(cgrp);
10689 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10690 percpu_counter_destroy(&ca->cpustat[i]);
10691 free_percpu(ca->cpuusage);
10695 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
10697 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10700 #ifndef CONFIG_64BIT
10702 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10704 spin_lock_irq(&cpu_rq(cpu)->lock);
10706 spin_unlock_irq(&cpu_rq(cpu)->lock);
10714 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
10716 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10718 #ifndef CONFIG_64BIT
10720 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10722 spin_lock_irq(&cpu_rq(cpu)->lock);
10724 spin_unlock_irq(&cpu_rq(cpu)->lock);
10730 /* return total cpu usage (in nanoseconds) of a group */
10731 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
10733 struct cpuacct *ca = cgroup_ca(cgrp);
10734 u64 totalcpuusage = 0;
10737 for_each_present_cpu(i)
10738 totalcpuusage += cpuacct_cpuusage_read(ca, i);
10740 return totalcpuusage;
10743 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
10746 struct cpuacct *ca = cgroup_ca(cgrp);
10755 for_each_present_cpu(i)
10756 cpuacct_cpuusage_write(ca, i, 0);
10762 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
10763 struct seq_file *m)
10765 struct cpuacct *ca = cgroup_ca(cgroup);
10769 for_each_present_cpu(i) {
10770 percpu = cpuacct_cpuusage_read(ca, i);
10771 seq_printf(m, "%llu ", (unsigned long long) percpu);
10773 seq_printf(m, "\n");
10777 static const char *cpuacct_stat_desc[] = {
10778 [CPUACCT_STAT_USER] = "user",
10779 [CPUACCT_STAT_SYSTEM] = "system",
10782 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
10783 struct cgroup_map_cb *cb)
10785 struct cpuacct *ca = cgroup_ca(cgrp);
10788 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
10789 s64 val = percpu_counter_read(&ca->cpustat[i]);
10790 val = cputime64_to_clock_t(val);
10791 cb->fill(cb, cpuacct_stat_desc[i], val);
10796 static struct cftype files[] = {
10799 .read_u64 = cpuusage_read,
10800 .write_u64 = cpuusage_write,
10803 .name = "usage_percpu",
10804 .read_seq_string = cpuacct_percpu_seq_read,
10808 .read_map = cpuacct_stats_show,
10812 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
10814 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
10818 * charge this task's execution time to its accounting group.
10820 * called with rq->lock held.
10822 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
10824 struct cpuacct *ca;
10827 if (unlikely(!cpuacct_subsys.active))
10830 cpu = task_cpu(tsk);
10836 for (; ca; ca = ca->parent) {
10837 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10838 *cpuusage += cputime;
10845 * Charge the system/user time to the task's accounting group.
10847 static void cpuacct_update_stats(struct task_struct *tsk,
10848 enum cpuacct_stat_index idx, cputime_t val)
10850 struct cpuacct *ca;
10852 if (unlikely(!cpuacct_subsys.active))
10859 percpu_counter_add(&ca->cpustat[idx], val);
10865 struct cgroup_subsys cpuacct_subsys = {
10867 .create = cpuacct_create,
10868 .destroy = cpuacct_destroy,
10869 .populate = cpuacct_populate,
10870 .subsys_id = cpuacct_subsys_id,
10872 #endif /* CONFIG_CGROUP_CPUACCT */
10876 int rcu_expedited_torture_stats(char *page)
10880 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
10882 void synchronize_sched_expedited(void)
10885 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
10887 #else /* #ifndef CONFIG_SMP */
10889 static DEFINE_PER_CPU(struct migration_req, rcu_migration_req);
10890 static DEFINE_MUTEX(rcu_sched_expedited_mutex);
10892 #define RCU_EXPEDITED_STATE_POST -2
10893 #define RCU_EXPEDITED_STATE_IDLE -1
10895 static int rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
10897 int rcu_expedited_torture_stats(char *page)
10902 cnt += sprintf(&page[cnt], "state: %d /", rcu_expedited_state);
10903 for_each_online_cpu(cpu) {
10904 cnt += sprintf(&page[cnt], " %d:%d",
10905 cpu, per_cpu(rcu_migration_req, cpu).dest_cpu);
10907 cnt += sprintf(&page[cnt], "\n");
10910 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
10912 static long synchronize_sched_expedited_count;
10915 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
10916 * approach to force grace period to end quickly. This consumes
10917 * significant time on all CPUs, and is thus not recommended for
10918 * any sort of common-case code.
10920 * Note that it is illegal to call this function while holding any
10921 * lock that is acquired by a CPU-hotplug notifier. Failing to
10922 * observe this restriction will result in deadlock.
10924 void synchronize_sched_expedited(void)
10927 unsigned long flags;
10928 bool need_full_sync = 0;
10930 struct migration_req *req;
10934 smp_mb(); /* ensure prior mod happens before capturing snap. */
10935 snap = ACCESS_ONCE(synchronize_sched_expedited_count) + 1;
10937 while (!mutex_trylock(&rcu_sched_expedited_mutex)) {
10939 if (trycount++ < 10)
10940 udelay(trycount * num_online_cpus());
10942 synchronize_sched();
10945 if (ACCESS_ONCE(synchronize_sched_expedited_count) - snap > 0) {
10946 smp_mb(); /* ensure test happens before caller kfree */
10951 rcu_expedited_state = RCU_EXPEDITED_STATE_POST;
10952 for_each_online_cpu(cpu) {
10954 req = &per_cpu(rcu_migration_req, cpu);
10955 init_completion(&req->done);
10957 req->dest_cpu = RCU_MIGRATION_NEED_QS;
10958 spin_lock_irqsave(&rq->lock, flags);
10959 list_add(&req->list, &rq->migration_queue);
10960 spin_unlock_irqrestore(&rq->lock, flags);
10961 wake_up_process(rq->migration_thread);
10963 for_each_online_cpu(cpu) {
10964 rcu_expedited_state = cpu;
10965 req = &per_cpu(rcu_migration_req, cpu);
10967 wait_for_completion(&req->done);
10968 spin_lock_irqsave(&rq->lock, flags);
10969 if (unlikely(req->dest_cpu == RCU_MIGRATION_MUST_SYNC))
10970 need_full_sync = 1;
10971 req->dest_cpu = RCU_MIGRATION_IDLE;
10972 spin_unlock_irqrestore(&rq->lock, flags);
10974 rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
10975 synchronize_sched_expedited_count++;
10976 mutex_unlock(&rcu_sched_expedited_mutex);
10978 if (need_full_sync)
10979 synchronize_sched();
10981 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
10983 #endif /* #else #ifndef CONFIG_SMP */