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);
313 static int root_task_group_empty(void)
315 return list_empty(&root_task_group.children);
319 #ifdef CONFIG_FAIR_GROUP_SCHED
320 #ifdef CONFIG_USER_SCHED
321 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
322 #else /* !CONFIG_USER_SCHED */
323 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
324 #endif /* CONFIG_USER_SCHED */
327 * A weight of 0 or 1 can cause arithmetics problems.
328 * A weight of a cfs_rq is the sum of weights of which entities
329 * are queued on this cfs_rq, so a weight of a entity should not be
330 * too large, so as the shares value of a task group.
331 * (The default weight is 1024 - so there's no practical
332 * limitation from this.)
335 #define MAX_SHARES (1UL << 18)
337 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
340 /* Default task group.
341 * Every task in system belong to this group at bootup.
343 struct task_group init_task_group;
345 /* return group to which a task belongs */
346 static inline struct task_group *task_group(struct task_struct *p)
348 struct task_group *tg;
350 #ifdef CONFIG_USER_SCHED
352 tg = __task_cred(p)->user->tg;
354 #elif defined(CONFIG_CGROUP_SCHED)
355 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
356 struct task_group, css);
358 tg = &init_task_group;
363 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
364 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
366 #ifdef CONFIG_FAIR_GROUP_SCHED
367 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
368 p->se.parent = task_group(p)->se[cpu];
371 #ifdef CONFIG_RT_GROUP_SCHED
372 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
373 p->rt.parent = task_group(p)->rt_se[cpu];
379 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
380 static inline struct task_group *task_group(struct task_struct *p)
385 #endif /* CONFIG_GROUP_SCHED */
387 /* CFS-related fields in a runqueue */
389 struct load_weight load;
390 unsigned long nr_running;
395 struct rb_root tasks_timeline;
396 struct rb_node *rb_leftmost;
398 struct list_head tasks;
399 struct list_head *balance_iterator;
402 * 'curr' points to currently running entity on this cfs_rq.
403 * It is set to NULL otherwise (i.e when none are currently running).
405 struct sched_entity *curr, *next, *last;
407 unsigned int nr_spread_over;
409 #ifdef CONFIG_FAIR_GROUP_SCHED
410 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
413 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
414 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
415 * (like users, containers etc.)
417 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
418 * list is used during load balance.
420 struct list_head leaf_cfs_rq_list;
421 struct task_group *tg; /* group that "owns" this runqueue */
425 * the part of load.weight contributed by tasks
427 unsigned long task_weight;
430 * h_load = weight * f(tg)
432 * Where f(tg) is the recursive weight fraction assigned to
435 unsigned long h_load;
438 * this cpu's part of tg->shares
440 unsigned long shares;
443 * load.weight at the time we set shares
445 unsigned long rq_weight;
450 /* Real-Time classes' related field in a runqueue: */
452 struct rt_prio_array active;
453 unsigned long rt_nr_running;
454 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
456 int curr; /* highest queued rt task prio */
458 int next; /* next highest */
463 unsigned long rt_nr_migratory;
464 unsigned long rt_nr_total;
466 struct plist_head pushable_tasks;
471 /* Nests inside the rq lock: */
472 spinlock_t rt_runtime_lock;
474 #ifdef CONFIG_RT_GROUP_SCHED
475 unsigned long rt_nr_boosted;
478 struct list_head leaf_rt_rq_list;
479 struct task_group *tg;
480 struct sched_rt_entity *rt_se;
487 * We add the notion of a root-domain which will be used to define per-domain
488 * variables. Each exclusive cpuset essentially defines an island domain by
489 * fully partitioning the member cpus from any other cpuset. Whenever a new
490 * exclusive cpuset is created, we also create and attach a new root-domain
497 cpumask_var_t online;
500 * The "RT overload" flag: it gets set if a CPU has more than
501 * one runnable RT task.
503 cpumask_var_t rto_mask;
506 struct cpupri cpupri;
511 * By default the system creates a single root-domain with all cpus as
512 * members (mimicking the global state we have today).
514 static struct root_domain def_root_domain;
519 * This is the main, per-CPU runqueue data structure.
521 * Locking rule: those places that want to lock multiple runqueues
522 * (such as the load balancing or the thread migration code), lock
523 * acquire operations must be ordered by ascending &runqueue.
530 * nr_running and cpu_load should be in the same cacheline because
531 * remote CPUs use both these fields when doing load calculation.
533 unsigned long nr_running;
534 #define CPU_LOAD_IDX_MAX 5
535 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
537 unsigned long last_tick_seen;
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;
544 u64 nr_migrations_in;
549 #ifdef CONFIG_FAIR_GROUP_SCHED
550 /* list of leaf cfs_rq on this cpu: */
551 struct list_head leaf_cfs_rq_list;
553 #ifdef CONFIG_RT_GROUP_SCHED
554 struct list_head leaf_rt_rq_list;
558 * This is part of a global counter where only the total sum
559 * over all CPUs matters. A task can increase this counter on
560 * one CPU and if it got migrated afterwards it may decrease
561 * it on another CPU. Always updated under the runqueue lock:
563 unsigned long nr_uninterruptible;
565 struct task_struct *curr, *idle;
566 unsigned long next_balance;
567 struct mm_struct *prev_mm;
574 struct root_domain *rd;
575 struct sched_domain *sd;
577 unsigned char idle_at_tick;
578 /* For active balancing */
582 /* cpu of this runqueue: */
586 unsigned long avg_load_per_task;
588 struct task_struct *migration_thread;
589 struct list_head migration_queue;
595 /* calc_load related fields */
596 unsigned long calc_load_update;
597 long calc_load_active;
599 #ifdef CONFIG_SCHED_HRTICK
601 int hrtick_csd_pending;
602 struct call_single_data hrtick_csd;
604 struct hrtimer hrtick_timer;
607 #ifdef CONFIG_SCHEDSTATS
609 struct sched_info rq_sched_info;
610 unsigned long long rq_cpu_time;
611 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
613 /* sys_sched_yield() stats */
614 unsigned int yld_count;
616 /* schedule() stats */
617 unsigned int sched_switch;
618 unsigned int sched_count;
619 unsigned int sched_goidle;
621 /* try_to_wake_up() stats */
622 unsigned int ttwu_count;
623 unsigned int ttwu_local;
626 unsigned int bkl_count;
630 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
633 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
635 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
638 static inline int cpu_of(struct rq *rq)
648 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
649 * See detach_destroy_domains: synchronize_sched for details.
651 * The domain tree of any CPU may only be accessed from within
652 * preempt-disabled sections.
654 #define for_each_domain(cpu, __sd) \
655 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
657 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
658 #define this_rq() (&__get_cpu_var(runqueues))
659 #define task_rq(p) cpu_rq(task_cpu(p))
660 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
661 #define raw_rq() (&__raw_get_cpu_var(runqueues))
663 inline void update_rq_clock(struct rq *rq)
665 rq->clock = sched_clock_cpu(cpu_of(rq));
669 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
671 #ifdef CONFIG_SCHED_DEBUG
672 # define const_debug __read_mostly
674 # define const_debug static const
679 * @cpu: the processor in question.
681 * Returns true if the current cpu runqueue is locked.
682 * This interface allows printk to be called with the runqueue lock
683 * held and know whether or not it is OK to wake up the klogd.
685 int runqueue_is_locked(int cpu)
687 return spin_is_locked(&cpu_rq(cpu)->lock);
691 * Debugging: various feature bits
694 #define SCHED_FEAT(name, enabled) \
695 __SCHED_FEAT_##name ,
698 #include "sched_features.h"
703 #define SCHED_FEAT(name, enabled) \
704 (1UL << __SCHED_FEAT_##name) * enabled |
706 const_debug unsigned int sysctl_sched_features =
707 #include "sched_features.h"
712 #ifdef CONFIG_SCHED_DEBUG
713 #define SCHED_FEAT(name, enabled) \
716 static __read_mostly char *sched_feat_names[] = {
717 #include "sched_features.h"
723 static int sched_feat_show(struct seq_file *m, void *v)
727 for (i = 0; sched_feat_names[i]; i++) {
728 if (!(sysctl_sched_features & (1UL << i)))
730 seq_printf(m, "%s ", sched_feat_names[i]);
738 sched_feat_write(struct file *filp, const char __user *ubuf,
739 size_t cnt, loff_t *ppos)
749 if (copy_from_user(&buf, ubuf, cnt))
754 if (strncmp(buf, "NO_", 3) == 0) {
759 for (i = 0; sched_feat_names[i]; i++) {
760 int len = strlen(sched_feat_names[i]);
762 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
764 sysctl_sched_features &= ~(1UL << i);
766 sysctl_sched_features |= (1UL << i);
771 if (!sched_feat_names[i])
779 static int sched_feat_open(struct inode *inode, struct file *filp)
781 return single_open(filp, sched_feat_show, NULL);
784 static const struct file_operations sched_feat_fops = {
785 .open = sched_feat_open,
786 .write = sched_feat_write,
789 .release = single_release,
792 static __init int sched_init_debug(void)
794 debugfs_create_file("sched_features", 0644, NULL, NULL,
799 late_initcall(sched_init_debug);
803 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
806 * Number of tasks to iterate in a single balance run.
807 * Limited because this is done with IRQs disabled.
809 const_debug unsigned int sysctl_sched_nr_migrate = 32;
812 * ratelimit for updating the group shares.
815 unsigned int sysctl_sched_shares_ratelimit = 250000;
818 * Inject some fuzzyness into changing the per-cpu group shares
819 * this avoids remote rq-locks at the expense of fairness.
822 unsigned int sysctl_sched_shares_thresh = 4;
825 * period over which we average the RT time consumption, measured
830 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
833 * period over which we measure -rt task cpu usage in us.
836 unsigned int sysctl_sched_rt_period = 1000000;
838 static __read_mostly int scheduler_running;
841 * part of the period that we allow rt tasks to run in us.
844 int sysctl_sched_rt_runtime = 950000;
846 static inline u64 global_rt_period(void)
848 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
851 static inline u64 global_rt_runtime(void)
853 if (sysctl_sched_rt_runtime < 0)
856 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
859 #ifndef prepare_arch_switch
860 # define prepare_arch_switch(next) do { } while (0)
862 #ifndef finish_arch_switch
863 # define finish_arch_switch(prev) do { } while (0)
866 static inline int task_current(struct rq *rq, struct task_struct *p)
868 return rq->curr == p;
871 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
872 static inline int task_running(struct rq *rq, struct task_struct *p)
874 return task_current(rq, p);
877 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
881 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
883 #ifdef CONFIG_DEBUG_SPINLOCK
884 /* this is a valid case when another task releases the spinlock */
885 rq->lock.owner = current;
888 * If we are tracking spinlock dependencies then we have to
889 * fix up the runqueue lock - which gets 'carried over' from
892 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
894 spin_unlock_irq(&rq->lock);
897 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
898 static inline int task_running(struct rq *rq, struct task_struct *p)
903 return task_current(rq, p);
907 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
911 * We can optimise this out completely for !SMP, because the
912 * SMP rebalancing from interrupt is the only thing that cares
917 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
918 spin_unlock_irq(&rq->lock);
920 spin_unlock(&rq->lock);
924 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
928 * After ->oncpu is cleared, the task can be moved to a different CPU.
929 * We must ensure this doesn't happen until the switch is completely
935 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
939 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
942 * __task_rq_lock - lock the runqueue a given task resides on.
943 * Must be called interrupts disabled.
945 static inline struct rq *__task_rq_lock(struct task_struct *p)
949 struct rq *rq = task_rq(p);
950 spin_lock(&rq->lock);
951 if (likely(rq == task_rq(p)))
953 spin_unlock(&rq->lock);
958 * task_rq_lock - lock the runqueue a given task resides on and disable
959 * interrupts. Note the ordering: we can safely lookup the task_rq without
960 * explicitly disabling preemption.
962 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
968 local_irq_save(*flags);
970 spin_lock(&rq->lock);
971 if (likely(rq == task_rq(p)))
973 spin_unlock_irqrestore(&rq->lock, *flags);
977 void task_rq_unlock_wait(struct task_struct *p)
979 struct rq *rq = task_rq(p);
981 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
982 spin_unlock_wait(&rq->lock);
985 static void __task_rq_unlock(struct rq *rq)
988 spin_unlock(&rq->lock);
991 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
994 spin_unlock_irqrestore(&rq->lock, *flags);
998 * this_rq_lock - lock this runqueue and disable interrupts.
1000 static struct rq *this_rq_lock(void)
1001 __acquires(rq->lock)
1005 local_irq_disable();
1007 spin_lock(&rq->lock);
1012 #ifdef CONFIG_SCHED_HRTICK
1014 * Use HR-timers to deliver accurate preemption points.
1016 * Its all a bit involved since we cannot program an hrt while holding the
1017 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1020 * When we get rescheduled we reprogram the hrtick_timer outside of the
1026 * - enabled by features
1027 * - hrtimer is actually high res
1029 static inline int hrtick_enabled(struct rq *rq)
1031 if (!sched_feat(HRTICK))
1033 if (!cpu_active(cpu_of(rq)))
1035 return hrtimer_is_hres_active(&rq->hrtick_timer);
1038 static void hrtick_clear(struct rq *rq)
1040 if (hrtimer_active(&rq->hrtick_timer))
1041 hrtimer_cancel(&rq->hrtick_timer);
1045 * High-resolution timer tick.
1046 * Runs from hardirq context with interrupts disabled.
1048 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1050 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1052 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1054 spin_lock(&rq->lock);
1055 update_rq_clock(rq);
1056 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1057 spin_unlock(&rq->lock);
1059 return HRTIMER_NORESTART;
1064 * called from hardirq (IPI) context
1066 static void __hrtick_start(void *arg)
1068 struct rq *rq = arg;
1070 spin_lock(&rq->lock);
1071 hrtimer_restart(&rq->hrtick_timer);
1072 rq->hrtick_csd_pending = 0;
1073 spin_unlock(&rq->lock);
1077 * Called to set the hrtick timer state.
1079 * called with rq->lock held and irqs disabled
1081 static void hrtick_start(struct rq *rq, u64 delay)
1083 struct hrtimer *timer = &rq->hrtick_timer;
1084 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1086 hrtimer_set_expires(timer, time);
1088 if (rq == this_rq()) {
1089 hrtimer_restart(timer);
1090 } else if (!rq->hrtick_csd_pending) {
1091 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1092 rq->hrtick_csd_pending = 1;
1097 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1099 int cpu = (int)(long)hcpu;
1102 case CPU_UP_CANCELED:
1103 case CPU_UP_CANCELED_FROZEN:
1104 case CPU_DOWN_PREPARE:
1105 case CPU_DOWN_PREPARE_FROZEN:
1107 case CPU_DEAD_FROZEN:
1108 hrtick_clear(cpu_rq(cpu));
1115 static __init void init_hrtick(void)
1117 hotcpu_notifier(hotplug_hrtick, 0);
1121 * Called to set the hrtick timer state.
1123 * called with rq->lock held and irqs disabled
1125 static void hrtick_start(struct rq *rq, u64 delay)
1127 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1128 HRTIMER_MODE_REL_PINNED, 0);
1131 static inline void init_hrtick(void)
1134 #endif /* CONFIG_SMP */
1136 static void init_rq_hrtick(struct rq *rq)
1139 rq->hrtick_csd_pending = 0;
1141 rq->hrtick_csd.flags = 0;
1142 rq->hrtick_csd.func = __hrtick_start;
1143 rq->hrtick_csd.info = rq;
1146 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1147 rq->hrtick_timer.function = hrtick;
1149 #else /* CONFIG_SCHED_HRTICK */
1150 static inline void hrtick_clear(struct rq *rq)
1154 static inline void init_rq_hrtick(struct rq *rq)
1158 static inline void init_hrtick(void)
1161 #endif /* CONFIG_SCHED_HRTICK */
1164 * resched_task - mark a task 'to be rescheduled now'.
1166 * On UP this means the setting of the need_resched flag, on SMP it
1167 * might also involve a cross-CPU call to trigger the scheduler on
1172 #ifndef tsk_is_polling
1173 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1176 static void resched_task(struct task_struct *p)
1180 assert_spin_locked(&task_rq(p)->lock);
1182 if (test_tsk_need_resched(p))
1185 set_tsk_need_resched(p);
1188 if (cpu == smp_processor_id())
1191 /* NEED_RESCHED must be visible before we test polling */
1193 if (!tsk_is_polling(p))
1194 smp_send_reschedule(cpu);
1197 static void resched_cpu(int cpu)
1199 struct rq *rq = cpu_rq(cpu);
1200 unsigned long flags;
1202 if (!spin_trylock_irqsave(&rq->lock, flags))
1204 resched_task(cpu_curr(cpu));
1205 spin_unlock_irqrestore(&rq->lock, flags);
1210 * When add_timer_on() enqueues a timer into the timer wheel of an
1211 * idle CPU then this timer might expire before the next timer event
1212 * which is scheduled to wake up that CPU. In case of a completely
1213 * idle system the next event might even be infinite time into the
1214 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1215 * leaves the inner idle loop so the newly added timer is taken into
1216 * account when the CPU goes back to idle and evaluates the timer
1217 * wheel for the next timer event.
1219 void wake_up_idle_cpu(int cpu)
1221 struct rq *rq = cpu_rq(cpu);
1223 if (cpu == smp_processor_id())
1227 * This is safe, as this function is called with the timer
1228 * wheel base lock of (cpu) held. When the CPU is on the way
1229 * to idle and has not yet set rq->curr to idle then it will
1230 * be serialized on the timer wheel base lock and take the new
1231 * timer into account automatically.
1233 if (rq->curr != rq->idle)
1237 * We can set TIF_RESCHED on the idle task of the other CPU
1238 * lockless. The worst case is that the other CPU runs the
1239 * idle task through an additional NOOP schedule()
1241 set_tsk_need_resched(rq->idle);
1243 /* NEED_RESCHED must be visible before we test polling */
1245 if (!tsk_is_polling(rq->idle))
1246 smp_send_reschedule(cpu);
1248 #endif /* CONFIG_NO_HZ */
1250 static u64 sched_avg_period(void)
1252 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1255 static void sched_avg_update(struct rq *rq)
1257 s64 period = sched_avg_period();
1259 while ((s64)(rq->clock - rq->age_stamp) > period) {
1260 rq->age_stamp += period;
1265 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1267 rq->rt_avg += rt_delta;
1268 sched_avg_update(rq);
1271 #else /* !CONFIG_SMP */
1272 static void resched_task(struct task_struct *p)
1274 assert_spin_locked(&task_rq(p)->lock);
1275 set_tsk_need_resched(p);
1278 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1281 #endif /* CONFIG_SMP */
1283 #if BITS_PER_LONG == 32
1284 # define WMULT_CONST (~0UL)
1286 # define WMULT_CONST (1UL << 32)
1289 #define WMULT_SHIFT 32
1292 * Shift right and round:
1294 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1297 * delta *= weight / lw
1299 static unsigned long
1300 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1301 struct load_weight *lw)
1305 if (!lw->inv_weight) {
1306 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1309 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1313 tmp = (u64)delta_exec * weight;
1315 * Check whether we'd overflow the 64-bit multiplication:
1317 if (unlikely(tmp > WMULT_CONST))
1318 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1321 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1323 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1326 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1332 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1339 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1340 * of tasks with abnormal "nice" values across CPUs the contribution that
1341 * each task makes to its run queue's load is weighted according to its
1342 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1343 * scaled version of the new time slice allocation that they receive on time
1347 #define WEIGHT_IDLEPRIO 3
1348 #define WMULT_IDLEPRIO 1431655765
1351 * Nice levels are multiplicative, with a gentle 10% change for every
1352 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1353 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1354 * that remained on nice 0.
1356 * The "10% effect" is relative and cumulative: from _any_ nice level,
1357 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1358 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1359 * If a task goes up by ~10% and another task goes down by ~10% then
1360 * the relative distance between them is ~25%.)
1362 static const int prio_to_weight[40] = {
1363 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1364 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1365 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1366 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1367 /* 0 */ 1024, 820, 655, 526, 423,
1368 /* 5 */ 335, 272, 215, 172, 137,
1369 /* 10 */ 110, 87, 70, 56, 45,
1370 /* 15 */ 36, 29, 23, 18, 15,
1374 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1376 * In cases where the weight does not change often, we can use the
1377 * precalculated inverse to speed up arithmetics by turning divisions
1378 * into multiplications:
1380 static const u32 prio_to_wmult[40] = {
1381 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1382 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1383 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1384 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1385 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1386 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1387 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1388 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1391 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1394 * runqueue iterator, to support SMP load-balancing between different
1395 * scheduling classes, without having to expose their internal data
1396 * structures to the load-balancing proper:
1398 struct rq_iterator {
1400 struct task_struct *(*start)(void *);
1401 struct task_struct *(*next)(void *);
1405 static unsigned long
1406 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1407 unsigned long max_load_move, struct sched_domain *sd,
1408 enum cpu_idle_type idle, int *all_pinned,
1409 int *this_best_prio, struct rq_iterator *iterator);
1412 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1413 struct sched_domain *sd, enum cpu_idle_type idle,
1414 struct rq_iterator *iterator);
1417 /* Time spent by the tasks of the cpu accounting group executing in ... */
1418 enum cpuacct_stat_index {
1419 CPUACCT_STAT_USER, /* ... user mode */
1420 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1422 CPUACCT_STAT_NSTATS,
1425 #ifdef CONFIG_CGROUP_CPUACCT
1426 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1427 static void cpuacct_update_stats(struct task_struct *tsk,
1428 enum cpuacct_stat_index idx, cputime_t val);
1430 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1431 static inline void cpuacct_update_stats(struct task_struct *tsk,
1432 enum cpuacct_stat_index idx, cputime_t val) {}
1435 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1437 update_load_add(&rq->load, load);
1440 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1442 update_load_sub(&rq->load, load);
1445 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1446 typedef int (*tg_visitor)(struct task_group *, void *);
1449 * Iterate the full tree, calling @down when first entering a node and @up when
1450 * leaving it for the final time.
1452 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1454 struct task_group *parent, *child;
1458 parent = &root_task_group;
1460 ret = (*down)(parent, data);
1463 list_for_each_entry_rcu(child, &parent->children, siblings) {
1470 ret = (*up)(parent, data);
1475 parent = parent->parent;
1484 static int tg_nop(struct task_group *tg, void *data)
1491 /* Used instead of source_load when we know the type == 0 */
1492 static unsigned long weighted_cpuload(const int cpu)
1494 return cpu_rq(cpu)->load.weight;
1498 * Return a low guess at the load of a migration-source cpu weighted
1499 * according to the scheduling class and "nice" value.
1501 * We want to under-estimate the load of migration sources, to
1502 * balance conservatively.
1504 static unsigned long source_load(int cpu, int type)
1506 struct rq *rq = cpu_rq(cpu);
1507 unsigned long total = weighted_cpuload(cpu);
1509 if (type == 0 || !sched_feat(LB_BIAS))
1512 return min(rq->cpu_load[type-1], total);
1516 * Return a high guess at the load of a migration-target cpu weighted
1517 * according to the scheduling class and "nice" value.
1519 static unsigned long target_load(int cpu, int type)
1521 struct rq *rq = cpu_rq(cpu);
1522 unsigned long total = weighted_cpuload(cpu);
1524 if (type == 0 || !sched_feat(LB_BIAS))
1527 return max(rq->cpu_load[type-1], total);
1530 static struct sched_group *group_of(int cpu)
1532 struct sched_domain *sd = rcu_dereference(cpu_rq(cpu)->sd);
1540 static unsigned long power_of(int cpu)
1542 struct sched_group *group = group_of(cpu);
1545 return SCHED_LOAD_SCALE;
1547 return group->cpu_power;
1550 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1552 static unsigned long cpu_avg_load_per_task(int cpu)
1554 struct rq *rq = cpu_rq(cpu);
1555 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1558 rq->avg_load_per_task = rq->load.weight / nr_running;
1560 rq->avg_load_per_task = 0;
1562 return rq->avg_load_per_task;
1565 #ifdef CONFIG_FAIR_GROUP_SCHED
1567 struct update_shares_data {
1568 unsigned long rq_weight[NR_CPUS];
1571 static DEFINE_PER_CPU(struct update_shares_data, update_shares_data);
1573 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1576 * Calculate and set the cpu's group shares.
1578 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1579 unsigned long sd_shares,
1580 unsigned long sd_rq_weight,
1581 struct update_shares_data *usd)
1583 unsigned long shares, rq_weight;
1586 rq_weight = usd->rq_weight[cpu];
1589 rq_weight = NICE_0_LOAD;
1593 * \Sum_j shares_j * rq_weight_i
1594 * shares_i = -----------------------------
1595 * \Sum_j rq_weight_j
1597 shares = (sd_shares * rq_weight) / sd_rq_weight;
1598 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1600 if (abs(shares - tg->se[cpu]->load.weight) >
1601 sysctl_sched_shares_thresh) {
1602 struct rq *rq = cpu_rq(cpu);
1603 unsigned long flags;
1605 spin_lock_irqsave(&rq->lock, flags);
1606 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1607 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1608 __set_se_shares(tg->se[cpu], shares);
1609 spin_unlock_irqrestore(&rq->lock, flags);
1614 * Re-compute the task group their per cpu shares over the given domain.
1615 * This needs to be done in a bottom-up fashion because the rq weight of a
1616 * parent group depends on the shares of its child groups.
1618 static int tg_shares_up(struct task_group *tg, void *data)
1620 unsigned long weight, rq_weight = 0, shares = 0;
1621 struct update_shares_data *usd;
1622 struct sched_domain *sd = data;
1623 unsigned long flags;
1629 local_irq_save(flags);
1630 usd = &__get_cpu_var(update_shares_data);
1632 for_each_cpu(i, sched_domain_span(sd)) {
1633 weight = tg->cfs_rq[i]->load.weight;
1634 usd->rq_weight[i] = weight;
1637 * If there are currently no tasks on the cpu pretend there
1638 * is one of average load so that when a new task gets to
1639 * run here it will not get delayed by group starvation.
1642 weight = NICE_0_LOAD;
1644 rq_weight += weight;
1645 shares += tg->cfs_rq[i]->shares;
1648 if ((!shares && rq_weight) || shares > tg->shares)
1649 shares = tg->shares;
1651 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1652 shares = tg->shares;
1654 for_each_cpu(i, sched_domain_span(sd))
1655 update_group_shares_cpu(tg, i, shares, rq_weight, usd);
1657 local_irq_restore(flags);
1663 * Compute the cpu's hierarchical load factor for each task group.
1664 * This needs to be done in a top-down fashion because the load of a child
1665 * group is a fraction of its parents load.
1667 static int tg_load_down(struct task_group *tg, void *data)
1670 long cpu = (long)data;
1673 load = cpu_rq(cpu)->load.weight;
1675 load = tg->parent->cfs_rq[cpu]->h_load;
1676 load *= tg->cfs_rq[cpu]->shares;
1677 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1680 tg->cfs_rq[cpu]->h_load = load;
1685 static void update_shares(struct sched_domain *sd)
1690 if (root_task_group_empty())
1693 now = cpu_clock(raw_smp_processor_id());
1694 elapsed = now - sd->last_update;
1696 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1697 sd->last_update = now;
1698 walk_tg_tree(tg_nop, tg_shares_up, sd);
1702 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1704 if (root_task_group_empty())
1707 spin_unlock(&rq->lock);
1709 spin_lock(&rq->lock);
1712 static void update_h_load(long cpu)
1714 if (root_task_group_empty())
1717 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1722 static inline void update_shares(struct sched_domain *sd)
1726 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1732 #ifdef CONFIG_PREEMPT
1734 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1737 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1738 * way at the expense of forcing extra atomic operations in all
1739 * invocations. This assures that the double_lock is acquired using the
1740 * same underlying policy as the spinlock_t on this architecture, which
1741 * reduces latency compared to the unfair variant below. However, it
1742 * also adds more overhead and therefore may reduce throughput.
1744 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1745 __releases(this_rq->lock)
1746 __acquires(busiest->lock)
1747 __acquires(this_rq->lock)
1749 spin_unlock(&this_rq->lock);
1750 double_rq_lock(this_rq, busiest);
1757 * Unfair double_lock_balance: Optimizes throughput at the expense of
1758 * latency by eliminating extra atomic operations when the locks are
1759 * already in proper order on entry. This favors lower cpu-ids and will
1760 * grant the double lock to lower cpus over higher ids under contention,
1761 * regardless of entry order into the function.
1763 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1764 __releases(this_rq->lock)
1765 __acquires(busiest->lock)
1766 __acquires(this_rq->lock)
1770 if (unlikely(!spin_trylock(&busiest->lock))) {
1771 if (busiest < this_rq) {
1772 spin_unlock(&this_rq->lock);
1773 spin_lock(&busiest->lock);
1774 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
1777 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
1782 #endif /* CONFIG_PREEMPT */
1785 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1787 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1789 if (unlikely(!irqs_disabled())) {
1790 /* printk() doesn't work good under rq->lock */
1791 spin_unlock(&this_rq->lock);
1795 return _double_lock_balance(this_rq, busiest);
1798 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1799 __releases(busiest->lock)
1801 spin_unlock(&busiest->lock);
1802 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1806 #ifdef CONFIG_FAIR_GROUP_SCHED
1807 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1810 cfs_rq->shares = shares;
1815 static void calc_load_account_active(struct rq *this_rq);
1817 #include "sched_stats.h"
1818 #include "sched_idletask.c"
1819 #include "sched_fair.c"
1820 #include "sched_rt.c"
1821 #ifdef CONFIG_SCHED_DEBUG
1822 # include "sched_debug.c"
1825 #define sched_class_highest (&rt_sched_class)
1826 #define for_each_class(class) \
1827 for (class = sched_class_highest; class; class = class->next)
1829 static void inc_nr_running(struct rq *rq)
1834 static void dec_nr_running(struct rq *rq)
1839 static void set_load_weight(struct task_struct *p)
1841 if (task_has_rt_policy(p)) {
1842 p->se.load.weight = prio_to_weight[0] * 2;
1843 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1848 * SCHED_IDLE tasks get minimal weight:
1850 if (p->policy == SCHED_IDLE) {
1851 p->se.load.weight = WEIGHT_IDLEPRIO;
1852 p->se.load.inv_weight = WMULT_IDLEPRIO;
1856 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1857 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1860 static void update_avg(u64 *avg, u64 sample)
1862 s64 diff = sample - *avg;
1866 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1869 p->se.start_runtime = p->se.sum_exec_runtime;
1871 sched_info_queued(p);
1872 p->sched_class->enqueue_task(rq, p, wakeup);
1876 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1879 if (p->se.last_wakeup) {
1880 update_avg(&p->se.avg_overlap,
1881 p->se.sum_exec_runtime - p->se.last_wakeup);
1882 p->se.last_wakeup = 0;
1884 update_avg(&p->se.avg_wakeup,
1885 sysctl_sched_wakeup_granularity);
1889 sched_info_dequeued(p);
1890 p->sched_class->dequeue_task(rq, p, sleep);
1895 * __normal_prio - return the priority that is based on the static prio
1897 static inline int __normal_prio(struct task_struct *p)
1899 return p->static_prio;
1903 * Calculate the expected normal priority: i.e. priority
1904 * without taking RT-inheritance into account. Might be
1905 * boosted by interactivity modifiers. Changes upon fork,
1906 * setprio syscalls, and whenever the interactivity
1907 * estimator recalculates.
1909 static inline int normal_prio(struct task_struct *p)
1913 if (task_has_rt_policy(p))
1914 prio = MAX_RT_PRIO-1 - p->rt_priority;
1916 prio = __normal_prio(p);
1921 * Calculate the current priority, i.e. the priority
1922 * taken into account by the scheduler. This value might
1923 * be boosted by RT tasks, or might be boosted by
1924 * interactivity modifiers. Will be RT if the task got
1925 * RT-boosted. If not then it returns p->normal_prio.
1927 static int effective_prio(struct task_struct *p)
1929 p->normal_prio = normal_prio(p);
1931 * If we are RT tasks or we were boosted to RT priority,
1932 * keep the priority unchanged. Otherwise, update priority
1933 * to the normal priority:
1935 if (!rt_prio(p->prio))
1936 return p->normal_prio;
1941 * activate_task - move a task to the runqueue.
1943 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1945 if (task_contributes_to_load(p))
1946 rq->nr_uninterruptible--;
1948 enqueue_task(rq, p, wakeup);
1953 * deactivate_task - remove a task from the runqueue.
1955 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1957 if (task_contributes_to_load(p))
1958 rq->nr_uninterruptible++;
1960 dequeue_task(rq, p, sleep);
1965 * task_curr - is this task currently executing on a CPU?
1966 * @p: the task in question.
1968 inline int task_curr(const struct task_struct *p)
1970 return cpu_curr(task_cpu(p)) == p;
1973 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1975 set_task_rq(p, cpu);
1978 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1979 * successfuly executed on another CPU. We must ensure that updates of
1980 * per-task data have been completed by this moment.
1983 task_thread_info(p)->cpu = cpu;
1987 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1988 const struct sched_class *prev_class,
1989 int oldprio, int running)
1991 if (prev_class != p->sched_class) {
1992 if (prev_class->switched_from)
1993 prev_class->switched_from(rq, p, running);
1994 p->sched_class->switched_to(rq, p, running);
1996 p->sched_class->prio_changed(rq, p, oldprio, running);
2000 * kthread_bind - bind a just-created kthread to a cpu.
2001 * @k: thread created by kthread_create().
2002 * @cpu: cpu (might not be online, must be possible) for @k to run on.
2004 * Description: This function is equivalent to set_cpus_allowed(),
2005 * except that @cpu doesn't need to be online, and the thread must be
2006 * stopped (i.e., just returned from kthread_create()).
2008 * Function lives here instead of kthread.c because it messes with
2009 * scheduler internals which require locking.
2011 void kthread_bind(struct task_struct *p, unsigned int cpu)
2013 struct rq *rq = cpu_rq(cpu);
2014 unsigned long flags;
2016 /* Must have done schedule() in kthread() before we set_task_cpu */
2017 if (!wait_task_inactive(p, TASK_UNINTERRUPTIBLE)) {
2022 spin_lock_irqsave(&rq->lock, flags);
2023 set_task_cpu(p, cpu);
2024 p->cpus_allowed = cpumask_of_cpu(cpu);
2025 p->rt.nr_cpus_allowed = 1;
2026 p->flags |= PF_THREAD_BOUND;
2027 spin_unlock_irqrestore(&rq->lock, flags);
2029 EXPORT_SYMBOL(kthread_bind);
2033 * Is this task likely cache-hot:
2036 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2041 * Buddy candidates are cache hot:
2043 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2044 (&p->se == cfs_rq_of(&p->se)->next ||
2045 &p->se == cfs_rq_of(&p->se)->last))
2048 if (p->sched_class != &fair_sched_class)
2051 if (sysctl_sched_migration_cost == -1)
2053 if (sysctl_sched_migration_cost == 0)
2056 delta = now - p->se.exec_start;
2058 return delta < (s64)sysctl_sched_migration_cost;
2062 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2064 int old_cpu = task_cpu(p);
2065 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
2066 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
2067 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
2070 clock_offset = old_rq->clock - new_rq->clock;
2072 trace_sched_migrate_task(p, new_cpu);
2074 #ifdef CONFIG_SCHEDSTATS
2075 if (p->se.wait_start)
2076 p->se.wait_start -= clock_offset;
2077 if (p->se.sleep_start)
2078 p->se.sleep_start -= clock_offset;
2079 if (p->se.block_start)
2080 p->se.block_start -= clock_offset;
2082 if (old_cpu != new_cpu) {
2083 p->se.nr_migrations++;
2084 new_rq->nr_migrations_in++;
2085 #ifdef CONFIG_SCHEDSTATS
2086 if (task_hot(p, old_rq->clock, NULL))
2087 schedstat_inc(p, se.nr_forced2_migrations);
2089 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS,
2092 p->se.vruntime -= old_cfsrq->min_vruntime -
2093 new_cfsrq->min_vruntime;
2095 __set_task_cpu(p, new_cpu);
2098 struct migration_req {
2099 struct list_head list;
2101 struct task_struct *task;
2104 struct completion done;
2108 * The task's runqueue lock must be held.
2109 * Returns true if you have to wait for migration thread.
2112 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2114 struct rq *rq = task_rq(p);
2117 * If the task is not on a runqueue (and not running), then
2118 * it is sufficient to simply update the task's cpu field.
2120 if (!p->se.on_rq && !task_running(rq, p)) {
2121 set_task_cpu(p, dest_cpu);
2125 init_completion(&req->done);
2127 req->dest_cpu = dest_cpu;
2128 list_add(&req->list, &rq->migration_queue);
2134 * wait_task_context_switch - wait for a thread to complete at least one
2137 * @p must not be current.
2139 void wait_task_context_switch(struct task_struct *p)
2141 unsigned long nvcsw, nivcsw, flags;
2149 * The runqueue is assigned before the actual context
2150 * switch. We need to take the runqueue lock.
2152 * We could check initially without the lock but it is
2153 * very likely that we need to take the lock in every
2156 rq = task_rq_lock(p, &flags);
2157 running = task_running(rq, p);
2158 task_rq_unlock(rq, &flags);
2160 if (likely(!running))
2163 * The switch count is incremented before the actual
2164 * context switch. We thus wait for two switches to be
2165 * sure at least one completed.
2167 if ((p->nvcsw - nvcsw) > 1)
2169 if ((p->nivcsw - nivcsw) > 1)
2177 * wait_task_inactive - wait for a thread to unschedule.
2179 * If @match_state is nonzero, it's the @p->state value just checked and
2180 * not expected to change. If it changes, i.e. @p might have woken up,
2181 * then return zero. When we succeed in waiting for @p to be off its CPU,
2182 * we return a positive number (its total switch count). If a second call
2183 * a short while later returns the same number, the caller can be sure that
2184 * @p has remained unscheduled the whole time.
2186 * The caller must ensure that the task *will* unschedule sometime soon,
2187 * else this function might spin for a *long* time. This function can't
2188 * be called with interrupts off, or it may introduce deadlock with
2189 * smp_call_function() if an IPI is sent by the same process we are
2190 * waiting to become inactive.
2192 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2194 unsigned long flags;
2201 * We do the initial early heuristics without holding
2202 * any task-queue locks at all. We'll only try to get
2203 * the runqueue lock when things look like they will
2209 * If the task is actively running on another CPU
2210 * still, just relax and busy-wait without holding
2213 * NOTE! Since we don't hold any locks, it's not
2214 * even sure that "rq" stays as the right runqueue!
2215 * But we don't care, since "task_running()" will
2216 * return false if the runqueue has changed and p
2217 * is actually now running somewhere else!
2219 while (task_running(rq, p)) {
2220 if (match_state && unlikely(p->state != match_state))
2226 * Ok, time to look more closely! We need the rq
2227 * lock now, to be *sure*. If we're wrong, we'll
2228 * just go back and repeat.
2230 rq = task_rq_lock(p, &flags);
2231 trace_sched_wait_task(rq, p);
2232 running = task_running(rq, p);
2233 on_rq = p->se.on_rq;
2235 if (!match_state || p->state == match_state)
2236 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2237 task_rq_unlock(rq, &flags);
2240 * If it changed from the expected state, bail out now.
2242 if (unlikely(!ncsw))
2246 * Was it really running after all now that we
2247 * checked with the proper locks actually held?
2249 * Oops. Go back and try again..
2251 if (unlikely(running)) {
2257 * It's not enough that it's not actively running,
2258 * it must be off the runqueue _entirely_, and not
2261 * So if it was still runnable (but just not actively
2262 * running right now), it's preempted, and we should
2263 * yield - it could be a while.
2265 if (unlikely(on_rq)) {
2266 schedule_timeout_uninterruptible(1);
2271 * Ahh, all good. It wasn't running, and it wasn't
2272 * runnable, which means that it will never become
2273 * running in the future either. We're all done!
2282 * kick_process - kick a running thread to enter/exit the kernel
2283 * @p: the to-be-kicked thread
2285 * Cause a process which is running on another CPU to enter
2286 * kernel-mode, without any delay. (to get signals handled.)
2288 * NOTE: this function doesnt have to take the runqueue lock,
2289 * because all it wants to ensure is that the remote task enters
2290 * the kernel. If the IPI races and the task has been migrated
2291 * to another CPU then no harm is done and the purpose has been
2294 void kick_process(struct task_struct *p)
2300 if ((cpu != smp_processor_id()) && task_curr(p))
2301 smp_send_reschedule(cpu);
2304 EXPORT_SYMBOL_GPL(kick_process);
2305 #endif /* CONFIG_SMP */
2308 * task_oncpu_function_call - call a function on the cpu on which a task runs
2309 * @p: the task to evaluate
2310 * @func: the function to be called
2311 * @info: the function call argument
2313 * Calls the function @func when the task is currently running. This might
2314 * be on the current CPU, which just calls the function directly
2316 void task_oncpu_function_call(struct task_struct *p,
2317 void (*func) (void *info), void *info)
2324 smp_call_function_single(cpu, func, info, 1);
2329 * try_to_wake_up - wake up a thread
2330 * @p: the to-be-woken-up thread
2331 * @state: the mask of task states that can be woken
2332 * @sync: do a synchronous wakeup?
2334 * Put it on the run-queue if it's not already there. The "current"
2335 * thread is always on the run-queue (except when the actual
2336 * re-schedule is in progress), and as such you're allowed to do
2337 * the simpler "current->state = TASK_RUNNING" to mark yourself
2338 * runnable without the overhead of this.
2340 * returns failure only if the task is already active.
2342 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2345 int cpu, orig_cpu, this_cpu, success = 0;
2346 unsigned long flags;
2347 struct rq *rq, *orig_rq;
2349 if (!sched_feat(SYNC_WAKEUPS))
2350 wake_flags &= ~WF_SYNC;
2352 this_cpu = get_cpu();
2355 rq = orig_rq = task_rq_lock(p, &flags);
2356 update_rq_clock(rq);
2357 if (!(p->state & state))
2367 if (unlikely(task_running(rq, p)))
2371 * In order to handle concurrent wakeups and release the rq->lock
2372 * we put the task in TASK_WAKING state.
2374 * First fix up the nr_uninterruptible count:
2376 if (task_contributes_to_load(p))
2377 rq->nr_uninterruptible--;
2378 p->state = TASK_WAKING;
2379 task_rq_unlock(rq, &flags);
2381 cpu = p->sched_class->select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2382 if (cpu != orig_cpu)
2383 set_task_cpu(p, cpu);
2385 rq = task_rq_lock(p, &flags);
2388 update_rq_clock(rq);
2390 WARN_ON(p->state != TASK_WAKING);
2393 #ifdef CONFIG_SCHEDSTATS
2394 schedstat_inc(rq, ttwu_count);
2395 if (cpu == this_cpu)
2396 schedstat_inc(rq, ttwu_local);
2398 struct sched_domain *sd;
2399 for_each_domain(this_cpu, sd) {
2400 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2401 schedstat_inc(sd, ttwu_wake_remote);
2406 #endif /* CONFIG_SCHEDSTATS */
2409 #endif /* CONFIG_SMP */
2410 schedstat_inc(p, se.nr_wakeups);
2411 if (wake_flags & WF_SYNC)
2412 schedstat_inc(p, se.nr_wakeups_sync);
2413 if (orig_cpu != cpu)
2414 schedstat_inc(p, se.nr_wakeups_migrate);
2415 if (cpu == this_cpu)
2416 schedstat_inc(p, se.nr_wakeups_local);
2418 schedstat_inc(p, se.nr_wakeups_remote);
2419 activate_task(rq, p, 1);
2423 * Only attribute actual wakeups done by this task.
2425 if (!in_interrupt()) {
2426 struct sched_entity *se = ¤t->se;
2427 u64 sample = se->sum_exec_runtime;
2429 if (se->last_wakeup)
2430 sample -= se->last_wakeup;
2432 sample -= se->start_runtime;
2433 update_avg(&se->avg_wakeup, sample);
2435 se->last_wakeup = se->sum_exec_runtime;
2439 trace_sched_wakeup(rq, p, success);
2440 check_preempt_curr(rq, p, wake_flags);
2442 p->state = TASK_RUNNING;
2444 if (p->sched_class->task_wake_up)
2445 p->sched_class->task_wake_up(rq, p);
2448 task_rq_unlock(rq, &flags);
2455 * wake_up_process - Wake up a specific process
2456 * @p: The process to be woken up.
2458 * Attempt to wake up the nominated process and move it to the set of runnable
2459 * processes. Returns 1 if the process was woken up, 0 if it was already
2462 * It may be assumed that this function implies a write memory barrier before
2463 * changing the task state if and only if any tasks are woken up.
2465 int wake_up_process(struct task_struct *p)
2467 return try_to_wake_up(p, TASK_ALL, 0);
2469 EXPORT_SYMBOL(wake_up_process);
2471 int wake_up_state(struct task_struct *p, unsigned int state)
2473 return try_to_wake_up(p, state, 0);
2477 * Perform scheduler related setup for a newly forked process p.
2478 * p is forked by current.
2480 * __sched_fork() is basic setup used by init_idle() too:
2482 static void __sched_fork(struct task_struct *p)
2484 p->se.exec_start = 0;
2485 p->se.sum_exec_runtime = 0;
2486 p->se.prev_sum_exec_runtime = 0;
2487 p->se.nr_migrations = 0;
2488 p->se.last_wakeup = 0;
2489 p->se.avg_overlap = 0;
2490 p->se.start_runtime = 0;
2491 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2492 p->se.avg_running = 0;
2494 #ifdef CONFIG_SCHEDSTATS
2495 p->se.wait_start = 0;
2497 p->se.wait_count = 0;
2500 p->se.sleep_start = 0;
2501 p->se.sleep_max = 0;
2502 p->se.sum_sleep_runtime = 0;
2504 p->se.block_start = 0;
2505 p->se.block_max = 0;
2507 p->se.slice_max = 0;
2509 p->se.nr_migrations_cold = 0;
2510 p->se.nr_failed_migrations_affine = 0;
2511 p->se.nr_failed_migrations_running = 0;
2512 p->se.nr_failed_migrations_hot = 0;
2513 p->se.nr_forced_migrations = 0;
2514 p->se.nr_forced2_migrations = 0;
2516 p->se.nr_wakeups = 0;
2517 p->se.nr_wakeups_sync = 0;
2518 p->se.nr_wakeups_migrate = 0;
2519 p->se.nr_wakeups_local = 0;
2520 p->se.nr_wakeups_remote = 0;
2521 p->se.nr_wakeups_affine = 0;
2522 p->se.nr_wakeups_affine_attempts = 0;
2523 p->se.nr_wakeups_passive = 0;
2524 p->se.nr_wakeups_idle = 0;
2528 INIT_LIST_HEAD(&p->rt.run_list);
2530 INIT_LIST_HEAD(&p->se.group_node);
2532 #ifdef CONFIG_PREEMPT_NOTIFIERS
2533 INIT_HLIST_HEAD(&p->preempt_notifiers);
2537 * We mark the process as running here, but have not actually
2538 * inserted it onto the runqueue yet. This guarantees that
2539 * nobody will actually run it, and a signal or other external
2540 * event cannot wake it up and insert it on the runqueue either.
2542 p->state = TASK_RUNNING;
2546 * fork()/clone()-time setup:
2548 void sched_fork(struct task_struct *p, int clone_flags)
2550 int cpu = get_cpu();
2555 * Revert to default priority/policy on fork if requested.
2557 if (unlikely(p->sched_reset_on_fork)) {
2558 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2559 p->policy = SCHED_NORMAL;
2560 p->normal_prio = p->static_prio;
2563 if (PRIO_TO_NICE(p->static_prio) < 0) {
2564 p->static_prio = NICE_TO_PRIO(0);
2565 p->normal_prio = p->static_prio;
2570 * We don't need the reset flag anymore after the fork. It has
2571 * fulfilled its duty:
2573 p->sched_reset_on_fork = 0;
2577 * Make sure we do not leak PI boosting priority to the child.
2579 p->prio = current->normal_prio;
2581 if (!rt_prio(p->prio))
2582 p->sched_class = &fair_sched_class;
2585 cpu = p->sched_class->select_task_rq(p, SD_BALANCE_FORK, 0);
2587 set_task_cpu(p, cpu);
2589 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2590 if (likely(sched_info_on()))
2591 memset(&p->sched_info, 0, sizeof(p->sched_info));
2593 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2596 #ifdef CONFIG_PREEMPT
2597 /* Want to start with kernel preemption disabled. */
2598 task_thread_info(p)->preempt_count = 1;
2600 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2606 * wake_up_new_task - wake up a newly created task for the first time.
2608 * This function will do some initial scheduler statistics housekeeping
2609 * that must be done for every newly created context, then puts the task
2610 * on the runqueue and wakes it.
2612 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2614 unsigned long flags;
2617 rq = task_rq_lock(p, &flags);
2618 BUG_ON(p->state != TASK_RUNNING);
2619 update_rq_clock(rq);
2621 if (!p->sched_class->task_new || !current->se.on_rq) {
2622 activate_task(rq, p, 0);
2625 * Let the scheduling class do new task startup
2626 * management (if any):
2628 p->sched_class->task_new(rq, p);
2631 trace_sched_wakeup_new(rq, p, 1);
2632 check_preempt_curr(rq, p, WF_FORK);
2634 if (p->sched_class->task_wake_up)
2635 p->sched_class->task_wake_up(rq, p);
2637 task_rq_unlock(rq, &flags);
2640 #ifdef CONFIG_PREEMPT_NOTIFIERS
2643 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2644 * @notifier: notifier struct to register
2646 void preempt_notifier_register(struct preempt_notifier *notifier)
2648 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2650 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2653 * preempt_notifier_unregister - no longer interested in preemption notifications
2654 * @notifier: notifier struct to unregister
2656 * This is safe to call from within a preemption notifier.
2658 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2660 hlist_del(¬ifier->link);
2662 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2664 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2666 struct preempt_notifier *notifier;
2667 struct hlist_node *node;
2669 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2670 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2674 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2675 struct task_struct *next)
2677 struct preempt_notifier *notifier;
2678 struct hlist_node *node;
2680 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2681 notifier->ops->sched_out(notifier, next);
2684 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2686 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2691 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2692 struct task_struct *next)
2696 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2699 * prepare_task_switch - prepare to switch tasks
2700 * @rq: the runqueue preparing to switch
2701 * @prev: the current task that is being switched out
2702 * @next: the task we are going to switch to.
2704 * This is called with the rq lock held and interrupts off. It must
2705 * be paired with a subsequent finish_task_switch after the context
2708 * prepare_task_switch sets up locking and calls architecture specific
2712 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2713 struct task_struct *next)
2715 fire_sched_out_preempt_notifiers(prev, next);
2716 prepare_lock_switch(rq, next);
2717 prepare_arch_switch(next);
2721 * finish_task_switch - clean up after a task-switch
2722 * @rq: runqueue associated with task-switch
2723 * @prev: the thread we just switched away from.
2725 * finish_task_switch must be called after the context switch, paired
2726 * with a prepare_task_switch call before the context switch.
2727 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2728 * and do any other architecture-specific cleanup actions.
2730 * Note that we may have delayed dropping an mm in context_switch(). If
2731 * so, we finish that here outside of the runqueue lock. (Doing it
2732 * with the lock held can cause deadlocks; see schedule() for
2735 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2736 __releases(rq->lock)
2738 struct mm_struct *mm = rq->prev_mm;
2744 * A task struct has one reference for the use as "current".
2745 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2746 * schedule one last time. The schedule call will never return, and
2747 * the scheduled task must drop that reference.
2748 * The test for TASK_DEAD must occur while the runqueue locks are
2749 * still held, otherwise prev could be scheduled on another cpu, die
2750 * there before we look at prev->state, and then the reference would
2752 * Manfred Spraul <manfred@colorfullife.com>
2754 prev_state = prev->state;
2755 finish_arch_switch(prev);
2756 perf_event_task_sched_in(current, cpu_of(rq));
2757 finish_lock_switch(rq, prev);
2759 fire_sched_in_preempt_notifiers(current);
2762 if (unlikely(prev_state == TASK_DEAD)) {
2764 * Remove function-return probe instances associated with this
2765 * task and put them back on the free list.
2767 kprobe_flush_task(prev);
2768 put_task_struct(prev);
2774 /* assumes rq->lock is held */
2775 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2777 if (prev->sched_class->pre_schedule)
2778 prev->sched_class->pre_schedule(rq, prev);
2781 /* rq->lock is NOT held, but preemption is disabled */
2782 static inline void post_schedule(struct rq *rq)
2784 if (rq->post_schedule) {
2785 unsigned long flags;
2787 spin_lock_irqsave(&rq->lock, flags);
2788 if (rq->curr->sched_class->post_schedule)
2789 rq->curr->sched_class->post_schedule(rq);
2790 spin_unlock_irqrestore(&rq->lock, flags);
2792 rq->post_schedule = 0;
2798 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2802 static inline void post_schedule(struct rq *rq)
2809 * schedule_tail - first thing a freshly forked thread must call.
2810 * @prev: the thread we just switched away from.
2812 asmlinkage void schedule_tail(struct task_struct *prev)
2813 __releases(rq->lock)
2815 struct rq *rq = this_rq();
2817 finish_task_switch(rq, prev);
2820 * FIXME: do we need to worry about rq being invalidated by the
2825 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2826 /* In this case, finish_task_switch does not reenable preemption */
2829 if (current->set_child_tid)
2830 put_user(task_pid_vnr(current), current->set_child_tid);
2834 * context_switch - switch to the new MM and the new
2835 * thread's register state.
2838 context_switch(struct rq *rq, struct task_struct *prev,
2839 struct task_struct *next)
2841 struct mm_struct *mm, *oldmm;
2843 prepare_task_switch(rq, prev, next);
2844 trace_sched_switch(rq, prev, next);
2846 oldmm = prev->active_mm;
2848 * For paravirt, this is coupled with an exit in switch_to to
2849 * combine the page table reload and the switch backend into
2852 arch_start_context_switch(prev);
2854 if (unlikely(!mm)) {
2855 next->active_mm = oldmm;
2856 atomic_inc(&oldmm->mm_count);
2857 enter_lazy_tlb(oldmm, next);
2859 switch_mm(oldmm, mm, next);
2861 if (unlikely(!prev->mm)) {
2862 prev->active_mm = NULL;
2863 rq->prev_mm = oldmm;
2866 * Since the runqueue lock will be released by the next
2867 * task (which is an invalid locking op but in the case
2868 * of the scheduler it's an obvious special-case), so we
2869 * do an early lockdep release here:
2871 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2872 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2875 /* Here we just switch the register state and the stack. */
2876 switch_to(prev, next, prev);
2880 * this_rq must be evaluated again because prev may have moved
2881 * CPUs since it called schedule(), thus the 'rq' on its stack
2882 * frame will be invalid.
2884 finish_task_switch(this_rq(), prev);
2888 * nr_running, nr_uninterruptible and nr_context_switches:
2890 * externally visible scheduler statistics: current number of runnable
2891 * threads, current number of uninterruptible-sleeping threads, total
2892 * number of context switches performed since bootup.
2894 unsigned long nr_running(void)
2896 unsigned long i, sum = 0;
2898 for_each_online_cpu(i)
2899 sum += cpu_rq(i)->nr_running;
2904 unsigned long nr_uninterruptible(void)
2906 unsigned long i, sum = 0;
2908 for_each_possible_cpu(i)
2909 sum += cpu_rq(i)->nr_uninterruptible;
2912 * Since we read the counters lockless, it might be slightly
2913 * inaccurate. Do not allow it to go below zero though:
2915 if (unlikely((long)sum < 0))
2921 unsigned long long nr_context_switches(void)
2924 unsigned long long sum = 0;
2926 for_each_possible_cpu(i)
2927 sum += cpu_rq(i)->nr_switches;
2932 unsigned long nr_iowait(void)
2934 unsigned long i, sum = 0;
2936 for_each_possible_cpu(i)
2937 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2942 unsigned long nr_iowait_cpu(void)
2944 struct rq *this = this_rq();
2945 return atomic_read(&this->nr_iowait);
2948 unsigned long this_cpu_load(void)
2950 struct rq *this = this_rq();
2951 return this->cpu_load[0];
2955 /* Variables and functions for calc_load */
2956 static atomic_long_t calc_load_tasks;
2957 static unsigned long calc_load_update;
2958 unsigned long avenrun[3];
2959 EXPORT_SYMBOL(avenrun);
2962 * get_avenrun - get the load average array
2963 * @loads: pointer to dest load array
2964 * @offset: offset to add
2965 * @shift: shift count to shift the result left
2967 * These values are estimates at best, so no need for locking.
2969 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2971 loads[0] = (avenrun[0] + offset) << shift;
2972 loads[1] = (avenrun[1] + offset) << shift;
2973 loads[2] = (avenrun[2] + offset) << shift;
2976 static unsigned long
2977 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2980 load += active * (FIXED_1 - exp);
2981 return load >> FSHIFT;
2985 * calc_load - update the avenrun load estimates 10 ticks after the
2986 * CPUs have updated calc_load_tasks.
2988 void calc_global_load(void)
2990 unsigned long upd = calc_load_update + 10;
2993 if (time_before(jiffies, upd))
2996 active = atomic_long_read(&calc_load_tasks);
2997 active = active > 0 ? active * FIXED_1 : 0;
2999 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3000 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3001 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3003 calc_load_update += LOAD_FREQ;
3007 * Either called from update_cpu_load() or from a cpu going idle
3009 static void calc_load_account_active(struct rq *this_rq)
3011 long nr_active, delta;
3013 nr_active = this_rq->nr_running;
3014 nr_active += (long) this_rq->nr_uninterruptible;
3016 if (nr_active != this_rq->calc_load_active) {
3017 delta = nr_active - this_rq->calc_load_active;
3018 this_rq->calc_load_active = nr_active;
3019 atomic_long_add(delta, &calc_load_tasks);
3024 * Externally visible per-cpu scheduler statistics:
3025 * cpu_nr_migrations(cpu) - number of migrations into that cpu
3027 u64 cpu_nr_migrations(int cpu)
3029 return cpu_rq(cpu)->nr_migrations_in;
3033 * Update rq->cpu_load[] statistics. This function is usually called every
3034 * scheduler tick (TICK_NSEC).
3036 static void update_cpu_load(struct rq *this_rq)
3038 unsigned long this_load = this_rq->load.weight;
3041 this_rq->nr_load_updates++;
3043 /* Update our load: */
3044 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3045 unsigned long old_load, new_load;
3047 /* scale is effectively 1 << i now, and >> i divides by scale */
3049 old_load = this_rq->cpu_load[i];
3050 new_load = this_load;
3052 * Round up the averaging division if load is increasing. This
3053 * prevents us from getting stuck on 9 if the load is 10, for
3056 if (new_load > old_load)
3057 new_load += scale-1;
3058 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3061 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3062 this_rq->calc_load_update += LOAD_FREQ;
3063 calc_load_account_active(this_rq);
3070 * double_rq_lock - safely lock two runqueues
3072 * Note this does not disable interrupts like task_rq_lock,
3073 * you need to do so manually before calling.
3075 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
3076 __acquires(rq1->lock)
3077 __acquires(rq2->lock)
3079 BUG_ON(!irqs_disabled());
3081 spin_lock(&rq1->lock);
3082 __acquire(rq2->lock); /* Fake it out ;) */
3085 spin_lock(&rq1->lock);
3086 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
3088 spin_lock(&rq2->lock);
3089 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
3092 update_rq_clock(rq1);
3093 update_rq_clock(rq2);
3097 * double_rq_unlock - safely unlock two runqueues
3099 * Note this does not restore interrupts like task_rq_unlock,
3100 * you need to do so manually after calling.
3102 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
3103 __releases(rq1->lock)
3104 __releases(rq2->lock)
3106 spin_unlock(&rq1->lock);
3108 spin_unlock(&rq2->lock);
3110 __release(rq2->lock);
3114 * If dest_cpu is allowed for this process, migrate the task to it.
3115 * This is accomplished by forcing the cpu_allowed mask to only
3116 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3117 * the cpu_allowed mask is restored.
3119 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
3121 struct migration_req req;
3122 unsigned long flags;
3125 rq = task_rq_lock(p, &flags);
3126 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
3127 || unlikely(!cpu_active(dest_cpu)))
3130 /* force the process onto the specified CPU */
3131 if (migrate_task(p, dest_cpu, &req)) {
3132 /* Need to wait for migration thread (might exit: take ref). */
3133 struct task_struct *mt = rq->migration_thread;
3135 get_task_struct(mt);
3136 task_rq_unlock(rq, &flags);
3137 wake_up_process(mt);
3138 put_task_struct(mt);
3139 wait_for_completion(&req.done);
3144 task_rq_unlock(rq, &flags);
3148 * sched_exec - execve() is a valuable balancing opportunity, because at
3149 * this point the task has the smallest effective memory and cache footprint.
3151 void sched_exec(void)
3153 int new_cpu, this_cpu = get_cpu();
3154 new_cpu = current->sched_class->select_task_rq(current, SD_BALANCE_EXEC, 0);
3156 if (new_cpu != this_cpu)
3157 sched_migrate_task(current, new_cpu);
3161 * pull_task - move a task from a remote runqueue to the local runqueue.
3162 * Both runqueues must be locked.
3164 static void pull_task(struct rq *src_rq, struct task_struct *p,
3165 struct rq *this_rq, int this_cpu)
3167 deactivate_task(src_rq, p, 0);
3168 set_task_cpu(p, this_cpu);
3169 activate_task(this_rq, p, 0);
3171 * Note that idle threads have a prio of MAX_PRIO, for this test
3172 * to be always true for them.
3174 check_preempt_curr(this_rq, p, 0);
3178 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3181 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3182 struct sched_domain *sd, enum cpu_idle_type idle,
3185 int tsk_cache_hot = 0;
3187 * We do not migrate tasks that are:
3188 * 1) running (obviously), or
3189 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3190 * 3) are cache-hot on their current CPU.
3192 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
3193 schedstat_inc(p, se.nr_failed_migrations_affine);
3198 if (task_running(rq, p)) {
3199 schedstat_inc(p, se.nr_failed_migrations_running);
3204 * Aggressive migration if:
3205 * 1) task is cache cold, or
3206 * 2) too many balance attempts have failed.
3209 tsk_cache_hot = task_hot(p, rq->clock, sd);
3210 if (!tsk_cache_hot ||
3211 sd->nr_balance_failed > sd->cache_nice_tries) {
3212 #ifdef CONFIG_SCHEDSTATS
3213 if (tsk_cache_hot) {
3214 schedstat_inc(sd, lb_hot_gained[idle]);
3215 schedstat_inc(p, se.nr_forced_migrations);
3221 if (tsk_cache_hot) {
3222 schedstat_inc(p, se.nr_failed_migrations_hot);
3228 static unsigned long
3229 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3230 unsigned long max_load_move, struct sched_domain *sd,
3231 enum cpu_idle_type idle, int *all_pinned,
3232 int *this_best_prio, struct rq_iterator *iterator)
3234 int loops = 0, pulled = 0, pinned = 0;
3235 struct task_struct *p;
3236 long rem_load_move = max_load_move;
3238 if (max_load_move == 0)
3244 * Start the load-balancing iterator:
3246 p = iterator->start(iterator->arg);
3248 if (!p || loops++ > sysctl_sched_nr_migrate)
3251 if ((p->se.load.weight >> 1) > rem_load_move ||
3252 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3253 p = iterator->next(iterator->arg);
3257 pull_task(busiest, p, this_rq, this_cpu);
3259 rem_load_move -= p->se.load.weight;
3261 #ifdef CONFIG_PREEMPT
3263 * NEWIDLE balancing is a source of latency, so preemptible kernels
3264 * will stop after the first task is pulled to minimize the critical
3267 if (idle == CPU_NEWLY_IDLE)
3272 * We only want to steal up to the prescribed amount of weighted load.
3274 if (rem_load_move > 0) {
3275 if (p->prio < *this_best_prio)
3276 *this_best_prio = p->prio;
3277 p = iterator->next(iterator->arg);
3282 * Right now, this is one of only two places pull_task() is called,
3283 * so we can safely collect pull_task() stats here rather than
3284 * inside pull_task().
3286 schedstat_add(sd, lb_gained[idle], pulled);
3289 *all_pinned = pinned;
3291 return max_load_move - rem_load_move;
3295 * move_tasks tries to move up to max_load_move weighted load from busiest to
3296 * this_rq, as part of a balancing operation within domain "sd".
3297 * Returns 1 if successful and 0 otherwise.
3299 * Called with both runqueues locked.
3301 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3302 unsigned long max_load_move,
3303 struct sched_domain *sd, enum cpu_idle_type idle,
3306 const struct sched_class *class = sched_class_highest;
3307 unsigned long total_load_moved = 0;
3308 int this_best_prio = this_rq->curr->prio;
3312 class->load_balance(this_rq, this_cpu, busiest,
3313 max_load_move - total_load_moved,
3314 sd, idle, all_pinned, &this_best_prio);
3315 class = class->next;
3317 #ifdef CONFIG_PREEMPT
3319 * NEWIDLE balancing is a source of latency, so preemptible
3320 * kernels will stop after the first task is pulled to minimize
3321 * the critical section.
3323 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3326 } while (class && max_load_move > total_load_moved);
3328 return total_load_moved > 0;
3332 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3333 struct sched_domain *sd, enum cpu_idle_type idle,
3334 struct rq_iterator *iterator)
3336 struct task_struct *p = iterator->start(iterator->arg);
3340 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3341 pull_task(busiest, p, this_rq, this_cpu);
3343 * Right now, this is only the second place pull_task()
3344 * is called, so we can safely collect pull_task()
3345 * stats here rather than inside pull_task().
3347 schedstat_inc(sd, lb_gained[idle]);
3351 p = iterator->next(iterator->arg);
3358 * move_one_task tries to move exactly one task from busiest to this_rq, as
3359 * part of active balancing operations within "domain".
3360 * Returns 1 if successful and 0 otherwise.
3362 * Called with both runqueues locked.
3364 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3365 struct sched_domain *sd, enum cpu_idle_type idle)
3367 const struct sched_class *class;
3369 for_each_class(class) {
3370 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3376 /********** Helpers for find_busiest_group ************************/
3378 * sd_lb_stats - Structure to store the statistics of a sched_domain
3379 * during load balancing.
3381 struct sd_lb_stats {
3382 struct sched_group *busiest; /* Busiest group in this sd */
3383 struct sched_group *this; /* Local group in this sd */
3384 unsigned long total_load; /* Total load of all groups in sd */
3385 unsigned long total_pwr; /* Total power of all groups in sd */
3386 unsigned long avg_load; /* Average load across all groups in sd */
3388 /** Statistics of this group */
3389 unsigned long this_load;
3390 unsigned long this_load_per_task;
3391 unsigned long this_nr_running;
3393 /* Statistics of the busiest group */
3394 unsigned long max_load;
3395 unsigned long busiest_load_per_task;
3396 unsigned long busiest_nr_running;
3398 int group_imb; /* Is there imbalance in this sd */
3399 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3400 int power_savings_balance; /* Is powersave balance needed for this sd */
3401 struct sched_group *group_min; /* Least loaded group in sd */
3402 struct sched_group *group_leader; /* Group which relieves group_min */
3403 unsigned long min_load_per_task; /* load_per_task in group_min */
3404 unsigned long leader_nr_running; /* Nr running of group_leader */
3405 unsigned long min_nr_running; /* Nr running of group_min */
3410 * sg_lb_stats - stats of a sched_group required for load_balancing
3412 struct sg_lb_stats {
3413 unsigned long avg_load; /*Avg load across the CPUs of the group */
3414 unsigned long group_load; /* Total load over the CPUs of the group */
3415 unsigned long sum_nr_running; /* Nr tasks running in the group */
3416 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3417 unsigned long group_capacity;
3418 int group_imb; /* Is there an imbalance in the group ? */
3422 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3423 * @group: The group whose first cpu is to be returned.
3425 static inline unsigned int group_first_cpu(struct sched_group *group)
3427 return cpumask_first(sched_group_cpus(group));
3431 * get_sd_load_idx - Obtain the load index for a given sched domain.
3432 * @sd: The sched_domain whose load_idx is to be obtained.
3433 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3435 static inline int get_sd_load_idx(struct sched_domain *sd,
3436 enum cpu_idle_type idle)
3442 load_idx = sd->busy_idx;
3445 case CPU_NEWLY_IDLE:
3446 load_idx = sd->newidle_idx;
3449 load_idx = sd->idle_idx;
3457 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3459 * init_sd_power_savings_stats - Initialize power savings statistics for
3460 * the given sched_domain, during load balancing.
3462 * @sd: Sched domain whose power-savings statistics are to be initialized.
3463 * @sds: Variable containing the statistics for sd.
3464 * @idle: Idle status of the CPU at which we're performing load-balancing.
3466 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3467 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3470 * Busy processors will not participate in power savings
3473 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3474 sds->power_savings_balance = 0;
3476 sds->power_savings_balance = 1;
3477 sds->min_nr_running = ULONG_MAX;
3478 sds->leader_nr_running = 0;
3483 * update_sd_power_savings_stats - Update the power saving stats for a
3484 * sched_domain while performing load balancing.
3486 * @group: sched_group belonging to the sched_domain under consideration.
3487 * @sds: Variable containing the statistics of the sched_domain
3488 * @local_group: Does group contain the CPU for which we're performing
3490 * @sgs: Variable containing the statistics of the group.
3492 static inline void update_sd_power_savings_stats(struct sched_group *group,
3493 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3496 if (!sds->power_savings_balance)
3500 * If the local group is idle or completely loaded
3501 * no need to do power savings balance at this domain
3503 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3504 !sds->this_nr_running))
3505 sds->power_savings_balance = 0;
3508 * If a group is already running at full capacity or idle,
3509 * don't include that group in power savings calculations
3511 if (!sds->power_savings_balance ||
3512 sgs->sum_nr_running >= sgs->group_capacity ||
3513 !sgs->sum_nr_running)
3517 * Calculate the group which has the least non-idle load.
3518 * This is the group from where we need to pick up the load
3521 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3522 (sgs->sum_nr_running == sds->min_nr_running &&
3523 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3524 sds->group_min = group;
3525 sds->min_nr_running = sgs->sum_nr_running;
3526 sds->min_load_per_task = sgs->sum_weighted_load /
3527 sgs->sum_nr_running;
3531 * Calculate the group which is almost near its
3532 * capacity but still has some space to pick up some load
3533 * from other group and save more power
3535 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
3538 if (sgs->sum_nr_running > sds->leader_nr_running ||
3539 (sgs->sum_nr_running == sds->leader_nr_running &&
3540 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3541 sds->group_leader = group;
3542 sds->leader_nr_running = sgs->sum_nr_running;
3547 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3548 * @sds: Variable containing the statistics of the sched_domain
3549 * under consideration.
3550 * @this_cpu: Cpu at which we're currently performing load-balancing.
3551 * @imbalance: Variable to store the imbalance.
3554 * Check if we have potential to perform some power-savings balance.
3555 * If yes, set the busiest group to be the least loaded group in the
3556 * sched_domain, so that it's CPUs can be put to idle.
3558 * Returns 1 if there is potential to perform power-savings balance.
3561 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3562 int this_cpu, unsigned long *imbalance)
3564 if (!sds->power_savings_balance)
3567 if (sds->this != sds->group_leader ||
3568 sds->group_leader == sds->group_min)
3571 *imbalance = sds->min_load_per_task;
3572 sds->busiest = sds->group_min;
3577 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3578 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3579 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3584 static inline void update_sd_power_savings_stats(struct sched_group *group,
3585 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3590 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3591 int this_cpu, unsigned long *imbalance)
3595 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3598 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
3600 return SCHED_LOAD_SCALE;
3603 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
3605 return default_scale_freq_power(sd, cpu);
3608 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
3610 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3611 unsigned long smt_gain = sd->smt_gain;
3618 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3620 return default_scale_smt_power(sd, cpu);
3623 unsigned long scale_rt_power(int cpu)
3625 struct rq *rq = cpu_rq(cpu);
3626 u64 total, available;
3628 sched_avg_update(rq);
3630 total = sched_avg_period() + (rq->clock - rq->age_stamp);
3631 available = total - rq->rt_avg;
3633 if (unlikely((s64)total < SCHED_LOAD_SCALE))
3634 total = SCHED_LOAD_SCALE;
3636 total >>= SCHED_LOAD_SHIFT;
3638 return div_u64(available, total);
3641 static void update_cpu_power(struct sched_domain *sd, int cpu)
3643 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3644 unsigned long power = SCHED_LOAD_SCALE;
3645 struct sched_group *sdg = sd->groups;
3647 if (sched_feat(ARCH_POWER))
3648 power *= arch_scale_freq_power(sd, cpu);
3650 power *= default_scale_freq_power(sd, cpu);
3652 power >>= SCHED_LOAD_SHIFT;
3654 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
3655 if (sched_feat(ARCH_POWER))
3656 power *= arch_scale_smt_power(sd, cpu);
3658 power *= default_scale_smt_power(sd, cpu);
3660 power >>= SCHED_LOAD_SHIFT;
3663 power *= scale_rt_power(cpu);
3664 power >>= SCHED_LOAD_SHIFT;
3669 sdg->cpu_power = power;
3672 static void update_group_power(struct sched_domain *sd, int cpu)
3674 struct sched_domain *child = sd->child;
3675 struct sched_group *group, *sdg = sd->groups;
3676 unsigned long power;
3679 update_cpu_power(sd, cpu);
3685 group = child->groups;
3687 power += group->cpu_power;
3688 group = group->next;
3689 } while (group != child->groups);
3691 sdg->cpu_power = power;
3695 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3696 * @sd: The sched_domain whose statistics are to be updated.
3697 * @group: sched_group whose statistics are to be updated.
3698 * @this_cpu: Cpu for which load balance is currently performed.
3699 * @idle: Idle status of this_cpu
3700 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3701 * @sd_idle: Idle status of the sched_domain containing group.
3702 * @local_group: Does group contain this_cpu.
3703 * @cpus: Set of cpus considered for load balancing.
3704 * @balance: Should we balance.
3705 * @sgs: variable to hold the statistics for this group.
3707 static inline void update_sg_lb_stats(struct sched_domain *sd,
3708 struct sched_group *group, int this_cpu,
3709 enum cpu_idle_type idle, int load_idx, int *sd_idle,
3710 int local_group, const struct cpumask *cpus,
3711 int *balance, struct sg_lb_stats *sgs)
3713 unsigned long load, max_cpu_load, min_cpu_load;
3715 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3716 unsigned long sum_avg_load_per_task;
3717 unsigned long avg_load_per_task;
3720 balance_cpu = group_first_cpu(group);
3721 if (balance_cpu == this_cpu)
3722 update_group_power(sd, this_cpu);
3725 /* Tally up the load of all CPUs in the group */
3726 sum_avg_load_per_task = avg_load_per_task = 0;
3728 min_cpu_load = ~0UL;
3730 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3731 struct rq *rq = cpu_rq(i);
3733 if (*sd_idle && rq->nr_running)
3736 /* Bias balancing toward cpus of our domain */
3738 if (idle_cpu(i) && !first_idle_cpu) {
3743 load = target_load(i, load_idx);
3745 load = source_load(i, load_idx);
3746 if (load > max_cpu_load)
3747 max_cpu_load = load;
3748 if (min_cpu_load > load)
3749 min_cpu_load = load;
3752 sgs->group_load += load;
3753 sgs->sum_nr_running += rq->nr_running;
3754 sgs->sum_weighted_load += weighted_cpuload(i);
3756 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3760 * First idle cpu or the first cpu(busiest) in this sched group
3761 * is eligible for doing load balancing at this and above
3762 * domains. In the newly idle case, we will allow all the cpu's
3763 * to do the newly idle load balance.
3765 if (idle != CPU_NEWLY_IDLE && local_group &&
3766 balance_cpu != this_cpu && balance) {
3771 /* Adjust by relative CPU power of the group */
3772 sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power;
3776 * Consider the group unbalanced when the imbalance is larger
3777 * than the average weight of two tasks.
3779 * APZ: with cgroup the avg task weight can vary wildly and
3780 * might not be a suitable number - should we keep a
3781 * normalized nr_running number somewhere that negates
3784 avg_load_per_task = (sum_avg_load_per_task * SCHED_LOAD_SCALE) /
3787 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3790 sgs->group_capacity =
3791 DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
3795 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3796 * @sd: sched_domain whose statistics are to be updated.
3797 * @this_cpu: Cpu for which load balance is currently performed.
3798 * @idle: Idle status of this_cpu
3799 * @sd_idle: Idle status of the sched_domain containing group.
3800 * @cpus: Set of cpus considered for load balancing.
3801 * @balance: Should we balance.
3802 * @sds: variable to hold the statistics for this sched_domain.
3804 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3805 enum cpu_idle_type idle, int *sd_idle,
3806 const struct cpumask *cpus, int *balance,
3807 struct sd_lb_stats *sds)
3809 struct sched_domain *child = sd->child;
3810 struct sched_group *group = sd->groups;
3811 struct sg_lb_stats sgs;
3812 int load_idx, prefer_sibling = 0;
3814 if (child && child->flags & SD_PREFER_SIBLING)
3817 init_sd_power_savings_stats(sd, sds, idle);
3818 load_idx = get_sd_load_idx(sd, idle);
3823 local_group = cpumask_test_cpu(this_cpu,
3824 sched_group_cpus(group));
3825 memset(&sgs, 0, sizeof(sgs));
3826 update_sg_lb_stats(sd, group, this_cpu, idle, load_idx, sd_idle,
3827 local_group, cpus, balance, &sgs);
3829 if (local_group && balance && !(*balance))
3832 sds->total_load += sgs.group_load;
3833 sds->total_pwr += group->cpu_power;
3836 * In case the child domain prefers tasks go to siblings
3837 * first, lower the group capacity to one so that we'll try
3838 * and move all the excess tasks away.
3841 sgs.group_capacity = min(sgs.group_capacity, 1UL);
3844 sds->this_load = sgs.avg_load;
3846 sds->this_nr_running = sgs.sum_nr_running;
3847 sds->this_load_per_task = sgs.sum_weighted_load;
3848 } else if (sgs.avg_load > sds->max_load &&
3849 (sgs.sum_nr_running > sgs.group_capacity ||
3851 sds->max_load = sgs.avg_load;
3852 sds->busiest = group;
3853 sds->busiest_nr_running = sgs.sum_nr_running;
3854 sds->busiest_load_per_task = sgs.sum_weighted_load;
3855 sds->group_imb = sgs.group_imb;
3858 update_sd_power_savings_stats(group, sds, local_group, &sgs);
3859 group = group->next;
3860 } while (group != sd->groups);
3864 * fix_small_imbalance - Calculate the minor imbalance that exists
3865 * amongst the groups of a sched_domain, during
3867 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3868 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3869 * @imbalance: Variable to store the imbalance.
3871 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3872 int this_cpu, unsigned long *imbalance)
3874 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3875 unsigned int imbn = 2;
3877 if (sds->this_nr_running) {
3878 sds->this_load_per_task /= sds->this_nr_running;
3879 if (sds->busiest_load_per_task >
3880 sds->this_load_per_task)
3883 sds->this_load_per_task =
3884 cpu_avg_load_per_task(this_cpu);
3886 if (sds->max_load - sds->this_load + sds->busiest_load_per_task >=
3887 sds->busiest_load_per_task * imbn) {
3888 *imbalance = sds->busiest_load_per_task;
3893 * OK, we don't have enough imbalance to justify moving tasks,
3894 * however we may be able to increase total CPU power used by
3898 pwr_now += sds->busiest->cpu_power *
3899 min(sds->busiest_load_per_task, sds->max_load);
3900 pwr_now += sds->this->cpu_power *
3901 min(sds->this_load_per_task, sds->this_load);
3902 pwr_now /= SCHED_LOAD_SCALE;
3904 /* Amount of load we'd subtract */
3905 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3906 sds->busiest->cpu_power;
3907 if (sds->max_load > tmp)
3908 pwr_move += sds->busiest->cpu_power *
3909 min(sds->busiest_load_per_task, sds->max_load - tmp);
3911 /* Amount of load we'd add */
3912 if (sds->max_load * sds->busiest->cpu_power <
3913 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
3914 tmp = (sds->max_load * sds->busiest->cpu_power) /
3915 sds->this->cpu_power;
3917 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3918 sds->this->cpu_power;
3919 pwr_move += sds->this->cpu_power *
3920 min(sds->this_load_per_task, sds->this_load + tmp);
3921 pwr_move /= SCHED_LOAD_SCALE;
3923 /* Move if we gain throughput */
3924 if (pwr_move > pwr_now)
3925 *imbalance = sds->busiest_load_per_task;
3929 * calculate_imbalance - Calculate the amount of imbalance present within the
3930 * groups of a given sched_domain during load balance.
3931 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3932 * @this_cpu: Cpu for which currently load balance is being performed.
3933 * @imbalance: The variable to store the imbalance.
3935 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3936 unsigned long *imbalance)
3938 unsigned long max_pull;
3940 * In the presence of smp nice balancing, certain scenarios can have
3941 * max load less than avg load(as we skip the groups at or below
3942 * its cpu_power, while calculating max_load..)
3944 if (sds->max_load < sds->avg_load) {
3946 return fix_small_imbalance(sds, this_cpu, imbalance);
3949 /* Don't want to pull so many tasks that a group would go idle */
3950 max_pull = min(sds->max_load - sds->avg_load,
3951 sds->max_load - sds->busiest_load_per_task);
3953 /* How much load to actually move to equalise the imbalance */
3954 *imbalance = min(max_pull * sds->busiest->cpu_power,
3955 (sds->avg_load - sds->this_load) * sds->this->cpu_power)
3959 * if *imbalance is less than the average load per runnable task
3960 * there is no gaurantee that any tasks will be moved so we'll have
3961 * a think about bumping its value to force at least one task to be
3964 if (*imbalance < sds->busiest_load_per_task)
3965 return fix_small_imbalance(sds, this_cpu, imbalance);
3968 /******* find_busiest_group() helpers end here *********************/
3971 * find_busiest_group - Returns the busiest group within the sched_domain
3972 * if there is an imbalance. If there isn't an imbalance, and
3973 * the user has opted for power-savings, it returns a group whose
3974 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3975 * such a group exists.
3977 * Also calculates the amount of weighted load which should be moved
3978 * to restore balance.
3980 * @sd: The sched_domain whose busiest group is to be returned.
3981 * @this_cpu: The cpu for which load balancing is currently being performed.
3982 * @imbalance: Variable which stores amount of weighted load which should
3983 * be moved to restore balance/put a group to idle.
3984 * @idle: The idle status of this_cpu.
3985 * @sd_idle: The idleness of sd
3986 * @cpus: The set of CPUs under consideration for load-balancing.
3987 * @balance: Pointer to a variable indicating if this_cpu
3988 * is the appropriate cpu to perform load balancing at this_level.
3990 * Returns: - the busiest group if imbalance exists.
3991 * - If no imbalance and user has opted for power-savings balance,
3992 * return the least loaded group whose CPUs can be
3993 * put to idle by rebalancing its tasks onto our group.
3995 static struct sched_group *
3996 find_busiest_group(struct sched_domain *sd, int this_cpu,
3997 unsigned long *imbalance, enum cpu_idle_type idle,
3998 int *sd_idle, const struct cpumask *cpus, int *balance)
4000 struct sd_lb_stats sds;
4002 memset(&sds, 0, sizeof(sds));
4005 * Compute the various statistics relavent for load balancing at
4008 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
4011 /* Cases where imbalance does not exist from POV of this_cpu */
4012 /* 1) this_cpu is not the appropriate cpu to perform load balancing
4014 * 2) There is no busy sibling group to pull from.
4015 * 3) This group is the busiest group.
4016 * 4) This group is more busy than the avg busieness at this
4018 * 5) The imbalance is within the specified limit.
4019 * 6) Any rebalance would lead to ping-pong
4021 if (balance && !(*balance))
4024 if (!sds.busiest || sds.busiest_nr_running == 0)
4027 if (sds.this_load >= sds.max_load)
4030 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
4032 if (sds.this_load >= sds.avg_load)
4035 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
4038 sds.busiest_load_per_task /= sds.busiest_nr_running;
4040 sds.busiest_load_per_task =
4041 min(sds.busiest_load_per_task, sds.avg_load);
4044 * We're trying to get all the cpus to the average_load, so we don't
4045 * want to push ourselves above the average load, nor do we wish to
4046 * reduce the max loaded cpu below the average load, as either of these
4047 * actions would just result in more rebalancing later, and ping-pong
4048 * tasks around. Thus we look for the minimum possible imbalance.
4049 * Negative imbalances (*we* are more loaded than anyone else) will
4050 * be counted as no imbalance for these purposes -- we can't fix that
4051 * by pulling tasks to us. Be careful of negative numbers as they'll
4052 * appear as very large values with unsigned longs.
4054 if (sds.max_load <= sds.busiest_load_per_task)
4057 /* Looks like there is an imbalance. Compute it */
4058 calculate_imbalance(&sds, this_cpu, imbalance);
4063 * There is no obvious imbalance. But check if we can do some balancing
4066 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
4074 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4077 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
4078 unsigned long imbalance, const struct cpumask *cpus)
4080 struct rq *busiest = NULL, *rq;
4081 unsigned long max_load = 0;
4084 for_each_cpu(i, sched_group_cpus(group)) {
4085 unsigned long power = power_of(i);
4086 unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
4089 if (!cpumask_test_cpu(i, cpus))
4093 wl = weighted_cpuload(i) * SCHED_LOAD_SCALE;
4096 if (capacity && rq->nr_running == 1 && wl > imbalance)
4099 if (wl > max_load) {
4109 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4110 * so long as it is large enough.
4112 #define MAX_PINNED_INTERVAL 512
4114 /* Working cpumask for load_balance and load_balance_newidle. */
4115 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4118 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4119 * tasks if there is an imbalance.
4121 static int load_balance(int this_cpu, struct rq *this_rq,
4122 struct sched_domain *sd, enum cpu_idle_type idle,
4125 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
4126 struct sched_group *group;
4127 unsigned long imbalance;
4129 unsigned long flags;
4130 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4132 cpumask_setall(cpus);
4135 * When power savings policy is enabled for the parent domain, idle
4136 * sibling can pick up load irrespective of busy siblings. In this case,
4137 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4138 * portraying it as CPU_NOT_IDLE.
4140 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
4141 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4144 schedstat_inc(sd, lb_count[idle]);
4148 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
4155 schedstat_inc(sd, lb_nobusyg[idle]);
4159 busiest = find_busiest_queue(group, idle, imbalance, cpus);
4161 schedstat_inc(sd, lb_nobusyq[idle]);
4165 BUG_ON(busiest == this_rq);
4167 schedstat_add(sd, lb_imbalance[idle], imbalance);
4170 if (busiest->nr_running > 1) {
4172 * Attempt to move tasks. If find_busiest_group has found
4173 * an imbalance but busiest->nr_running <= 1, the group is
4174 * still unbalanced. ld_moved simply stays zero, so it is
4175 * correctly treated as an imbalance.
4177 local_irq_save(flags);
4178 double_rq_lock(this_rq, busiest);
4179 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4180 imbalance, sd, idle, &all_pinned);
4181 double_rq_unlock(this_rq, busiest);
4182 local_irq_restore(flags);
4185 * some other cpu did the load balance for us.
4187 if (ld_moved && this_cpu != smp_processor_id())
4188 resched_cpu(this_cpu);
4190 /* All tasks on this runqueue were pinned by CPU affinity */
4191 if (unlikely(all_pinned)) {
4192 cpumask_clear_cpu(cpu_of(busiest), cpus);
4193 if (!cpumask_empty(cpus))
4200 schedstat_inc(sd, lb_failed[idle]);
4201 sd->nr_balance_failed++;
4203 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
4205 spin_lock_irqsave(&busiest->lock, flags);
4207 /* don't kick the migration_thread, if the curr
4208 * task on busiest cpu can't be moved to this_cpu
4210 if (!cpumask_test_cpu(this_cpu,
4211 &busiest->curr->cpus_allowed)) {
4212 spin_unlock_irqrestore(&busiest->lock, flags);
4214 goto out_one_pinned;
4217 if (!busiest->active_balance) {
4218 busiest->active_balance = 1;
4219 busiest->push_cpu = this_cpu;
4222 spin_unlock_irqrestore(&busiest->lock, flags);
4224 wake_up_process(busiest->migration_thread);
4227 * We've kicked active balancing, reset the failure
4230 sd->nr_balance_failed = sd->cache_nice_tries+1;
4233 sd->nr_balance_failed = 0;
4235 if (likely(!active_balance)) {
4236 /* We were unbalanced, so reset the balancing interval */
4237 sd->balance_interval = sd->min_interval;
4240 * If we've begun active balancing, start to back off. This
4241 * case may not be covered by the all_pinned logic if there
4242 * is only 1 task on the busy runqueue (because we don't call
4245 if (sd->balance_interval < sd->max_interval)
4246 sd->balance_interval *= 2;
4249 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4250 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4256 schedstat_inc(sd, lb_balanced[idle]);
4258 sd->nr_balance_failed = 0;
4261 /* tune up the balancing interval */
4262 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4263 (sd->balance_interval < sd->max_interval))
4264 sd->balance_interval *= 2;
4266 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4267 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4278 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4279 * tasks if there is an imbalance.
4281 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4282 * this_rq is locked.
4285 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
4287 struct sched_group *group;
4288 struct rq *busiest = NULL;
4289 unsigned long imbalance;
4293 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4295 cpumask_setall(cpus);
4298 * When power savings policy is enabled for the parent domain, idle
4299 * sibling can pick up load irrespective of busy siblings. In this case,
4300 * let the state of idle sibling percolate up as IDLE, instead of
4301 * portraying it as CPU_NOT_IDLE.
4303 if (sd->flags & SD_SHARE_CPUPOWER &&
4304 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4307 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
4309 update_shares_locked(this_rq, sd);
4310 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
4311 &sd_idle, cpus, NULL);
4313 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
4317 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
4319 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
4323 BUG_ON(busiest == this_rq);
4325 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
4328 if (busiest->nr_running > 1) {
4329 /* Attempt to move tasks */
4330 double_lock_balance(this_rq, busiest);
4331 /* this_rq->clock is already updated */
4332 update_rq_clock(busiest);
4333 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4334 imbalance, sd, CPU_NEWLY_IDLE,
4336 double_unlock_balance(this_rq, busiest);
4338 if (unlikely(all_pinned)) {
4339 cpumask_clear_cpu(cpu_of(busiest), cpus);
4340 if (!cpumask_empty(cpus))
4346 int active_balance = 0;
4348 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
4349 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4350 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4353 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4356 if (sd->nr_balance_failed++ < 2)
4360 * The only task running in a non-idle cpu can be moved to this
4361 * cpu in an attempt to completely freeup the other CPU
4362 * package. The same method used to move task in load_balance()
4363 * have been extended for load_balance_newidle() to speedup
4364 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4366 * The package power saving logic comes from
4367 * find_busiest_group(). If there are no imbalance, then
4368 * f_b_g() will return NULL. However when sched_mc={1,2} then
4369 * f_b_g() will select a group from which a running task may be
4370 * pulled to this cpu in order to make the other package idle.
4371 * If there is no opportunity to make a package idle and if
4372 * there are no imbalance, then f_b_g() will return NULL and no
4373 * action will be taken in load_balance_newidle().
4375 * Under normal task pull operation due to imbalance, there
4376 * will be more than one task in the source run queue and
4377 * move_tasks() will succeed. ld_moved will be true and this
4378 * active balance code will not be triggered.
4381 /* Lock busiest in correct order while this_rq is held */
4382 double_lock_balance(this_rq, busiest);
4385 * don't kick the migration_thread, if the curr
4386 * task on busiest cpu can't be moved to this_cpu
4388 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
4389 double_unlock_balance(this_rq, busiest);
4394 if (!busiest->active_balance) {
4395 busiest->active_balance = 1;
4396 busiest->push_cpu = this_cpu;
4400 double_unlock_balance(this_rq, busiest);
4402 * Should not call ttwu while holding a rq->lock
4404 spin_unlock(&this_rq->lock);
4406 wake_up_process(busiest->migration_thread);
4407 spin_lock(&this_rq->lock);
4410 sd->nr_balance_failed = 0;
4412 update_shares_locked(this_rq, sd);
4416 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
4417 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4418 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4420 sd->nr_balance_failed = 0;
4426 * idle_balance is called by schedule() if this_cpu is about to become
4427 * idle. Attempts to pull tasks from other CPUs.
4429 static void idle_balance(int this_cpu, struct rq *this_rq)
4431 struct sched_domain *sd;
4432 int pulled_task = 0;
4433 unsigned long next_balance = jiffies + HZ;
4435 for_each_domain(this_cpu, sd) {
4436 unsigned long interval;
4438 if (!(sd->flags & SD_LOAD_BALANCE))
4441 if (sd->flags & SD_BALANCE_NEWIDLE)
4442 /* If we've pulled tasks over stop searching: */
4443 pulled_task = load_balance_newidle(this_cpu, this_rq,
4446 interval = msecs_to_jiffies(sd->balance_interval);
4447 if (time_after(next_balance, sd->last_balance + interval))
4448 next_balance = sd->last_balance + interval;
4452 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4454 * We are going idle. next_balance may be set based on
4455 * a busy processor. So reset next_balance.
4457 this_rq->next_balance = next_balance;
4462 * active_load_balance is run by migration threads. It pushes running tasks
4463 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4464 * running on each physical CPU where possible, and avoids physical /
4465 * logical imbalances.
4467 * Called with busiest_rq locked.
4469 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
4471 int target_cpu = busiest_rq->push_cpu;
4472 struct sched_domain *sd;
4473 struct rq *target_rq;
4475 /* Is there any task to move? */
4476 if (busiest_rq->nr_running <= 1)
4479 target_rq = cpu_rq(target_cpu);
4482 * This condition is "impossible", if it occurs
4483 * we need to fix it. Originally reported by
4484 * Bjorn Helgaas on a 128-cpu setup.
4486 BUG_ON(busiest_rq == target_rq);
4488 /* move a task from busiest_rq to target_rq */
4489 double_lock_balance(busiest_rq, target_rq);
4490 update_rq_clock(busiest_rq);
4491 update_rq_clock(target_rq);
4493 /* Search for an sd spanning us and the target CPU. */
4494 for_each_domain(target_cpu, sd) {
4495 if ((sd->flags & SD_LOAD_BALANCE) &&
4496 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4501 schedstat_inc(sd, alb_count);
4503 if (move_one_task(target_rq, target_cpu, busiest_rq,
4505 schedstat_inc(sd, alb_pushed);
4507 schedstat_inc(sd, alb_failed);
4509 double_unlock_balance(busiest_rq, target_rq);
4514 atomic_t load_balancer;
4515 cpumask_var_t cpu_mask;
4516 cpumask_var_t ilb_grp_nohz_mask;
4517 } nohz ____cacheline_aligned = {
4518 .load_balancer = ATOMIC_INIT(-1),
4521 int get_nohz_load_balancer(void)
4523 return atomic_read(&nohz.load_balancer);
4526 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4528 * lowest_flag_domain - Return lowest sched_domain containing flag.
4529 * @cpu: The cpu whose lowest level of sched domain is to
4531 * @flag: The flag to check for the lowest sched_domain
4532 * for the given cpu.
4534 * Returns the lowest sched_domain of a cpu which contains the given flag.
4536 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4538 struct sched_domain *sd;
4540 for_each_domain(cpu, sd)
4541 if (sd && (sd->flags & flag))
4548 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4549 * @cpu: The cpu whose domains we're iterating over.
4550 * @sd: variable holding the value of the power_savings_sd
4552 * @flag: The flag to filter the sched_domains to be iterated.
4554 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4555 * set, starting from the lowest sched_domain to the highest.
4557 #define for_each_flag_domain(cpu, sd, flag) \
4558 for (sd = lowest_flag_domain(cpu, flag); \
4559 (sd && (sd->flags & flag)); sd = sd->parent)
4562 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4563 * @ilb_group: group to be checked for semi-idleness
4565 * Returns: 1 if the group is semi-idle. 0 otherwise.
4567 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4568 * and atleast one non-idle CPU. This helper function checks if the given
4569 * sched_group is semi-idle or not.
4571 static inline int is_semi_idle_group(struct sched_group *ilb_group)
4573 cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask,
4574 sched_group_cpus(ilb_group));
4577 * A sched_group is semi-idle when it has atleast one busy cpu
4578 * and atleast one idle cpu.
4580 if (cpumask_empty(nohz.ilb_grp_nohz_mask))
4583 if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group)))
4589 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4590 * @cpu: The cpu which is nominating a new idle_load_balancer.
4592 * Returns: Returns the id of the idle load balancer if it exists,
4593 * Else, returns >= nr_cpu_ids.
4595 * This algorithm picks the idle load balancer such that it belongs to a
4596 * semi-idle powersavings sched_domain. The idea is to try and avoid
4597 * completely idle packages/cores just for the purpose of idle load balancing
4598 * when there are other idle cpu's which are better suited for that job.
4600 static int find_new_ilb(int cpu)
4602 struct sched_domain *sd;
4603 struct sched_group *ilb_group;
4606 * Have idle load balancer selection from semi-idle packages only
4607 * when power-aware load balancing is enabled
4609 if (!(sched_smt_power_savings || sched_mc_power_savings))
4613 * Optimize for the case when we have no idle CPUs or only one
4614 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4616 if (cpumask_weight(nohz.cpu_mask) < 2)
4619 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4620 ilb_group = sd->groups;
4623 if (is_semi_idle_group(ilb_group))
4624 return cpumask_first(nohz.ilb_grp_nohz_mask);
4626 ilb_group = ilb_group->next;
4628 } while (ilb_group != sd->groups);
4632 return cpumask_first(nohz.cpu_mask);
4634 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4635 static inline int find_new_ilb(int call_cpu)
4637 return cpumask_first(nohz.cpu_mask);
4642 * This routine will try to nominate the ilb (idle load balancing)
4643 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4644 * load balancing on behalf of all those cpus. If all the cpus in the system
4645 * go into this tickless mode, then there will be no ilb owner (as there is
4646 * no need for one) and all the cpus will sleep till the next wakeup event
4649 * For the ilb owner, tick is not stopped. And this tick will be used
4650 * for idle load balancing. ilb owner will still be part of
4653 * While stopping the tick, this cpu will become the ilb owner if there
4654 * is no other owner. And will be the owner till that cpu becomes busy
4655 * or if all cpus in the system stop their ticks at which point
4656 * there is no need for ilb owner.
4658 * When the ilb owner becomes busy, it nominates another owner, during the
4659 * next busy scheduler_tick()
4661 int select_nohz_load_balancer(int stop_tick)
4663 int cpu = smp_processor_id();
4666 cpu_rq(cpu)->in_nohz_recently = 1;
4668 if (!cpu_active(cpu)) {
4669 if (atomic_read(&nohz.load_balancer) != cpu)
4673 * If we are going offline and still the leader,
4676 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4682 cpumask_set_cpu(cpu, nohz.cpu_mask);
4684 /* time for ilb owner also to sleep */
4685 if (cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4686 if (atomic_read(&nohz.load_balancer) == cpu)
4687 atomic_set(&nohz.load_balancer, -1);
4691 if (atomic_read(&nohz.load_balancer) == -1) {
4692 /* make me the ilb owner */
4693 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4695 } else if (atomic_read(&nohz.load_balancer) == cpu) {
4698 if (!(sched_smt_power_savings ||
4699 sched_mc_power_savings))
4702 * Check to see if there is a more power-efficient
4705 new_ilb = find_new_ilb(cpu);
4706 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4707 atomic_set(&nohz.load_balancer, -1);
4708 resched_cpu(new_ilb);
4714 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4717 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4719 if (atomic_read(&nohz.load_balancer) == cpu)
4720 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4727 static DEFINE_SPINLOCK(balancing);
4730 * It checks each scheduling domain to see if it is due to be balanced,
4731 * and initiates a balancing operation if so.
4733 * Balancing parameters are set up in arch_init_sched_domains.
4735 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4738 struct rq *rq = cpu_rq(cpu);
4739 unsigned long interval;
4740 struct sched_domain *sd;
4741 /* Earliest time when we have to do rebalance again */
4742 unsigned long next_balance = jiffies + 60*HZ;
4743 int update_next_balance = 0;
4746 for_each_domain(cpu, sd) {
4747 if (!(sd->flags & SD_LOAD_BALANCE))
4750 interval = sd->balance_interval;
4751 if (idle != CPU_IDLE)
4752 interval *= sd->busy_factor;
4754 /* scale ms to jiffies */
4755 interval = msecs_to_jiffies(interval);
4756 if (unlikely(!interval))
4758 if (interval > HZ*NR_CPUS/10)
4759 interval = HZ*NR_CPUS/10;
4761 need_serialize = sd->flags & SD_SERIALIZE;
4763 if (need_serialize) {
4764 if (!spin_trylock(&balancing))
4768 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4769 if (load_balance(cpu, rq, sd, idle, &balance)) {
4771 * We've pulled tasks over so either we're no
4772 * longer idle, or one of our SMT siblings is
4775 idle = CPU_NOT_IDLE;
4777 sd->last_balance = jiffies;
4780 spin_unlock(&balancing);
4782 if (time_after(next_balance, sd->last_balance + interval)) {
4783 next_balance = sd->last_balance + interval;
4784 update_next_balance = 1;
4788 * Stop the load balance at this level. There is another
4789 * CPU in our sched group which is doing load balancing more
4797 * next_balance will be updated only when there is a need.
4798 * When the cpu is attached to null domain for ex, it will not be
4801 if (likely(update_next_balance))
4802 rq->next_balance = next_balance;
4806 * run_rebalance_domains is triggered when needed from the scheduler tick.
4807 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4808 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4810 static void run_rebalance_domains(struct softirq_action *h)
4812 int this_cpu = smp_processor_id();
4813 struct rq *this_rq = cpu_rq(this_cpu);
4814 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4815 CPU_IDLE : CPU_NOT_IDLE;
4817 rebalance_domains(this_cpu, idle);
4821 * If this cpu is the owner for idle load balancing, then do the
4822 * balancing on behalf of the other idle cpus whose ticks are
4825 if (this_rq->idle_at_tick &&
4826 atomic_read(&nohz.load_balancer) == this_cpu) {
4830 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4831 if (balance_cpu == this_cpu)
4835 * If this cpu gets work to do, stop the load balancing
4836 * work being done for other cpus. Next load
4837 * balancing owner will pick it up.
4842 rebalance_domains(balance_cpu, CPU_IDLE);
4844 rq = cpu_rq(balance_cpu);
4845 if (time_after(this_rq->next_balance, rq->next_balance))
4846 this_rq->next_balance = rq->next_balance;
4852 static inline int on_null_domain(int cpu)
4854 return !rcu_dereference(cpu_rq(cpu)->sd);
4858 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4860 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4861 * idle load balancing owner or decide to stop the periodic load balancing,
4862 * if the whole system is idle.
4864 static inline void trigger_load_balance(struct rq *rq, int cpu)
4868 * If we were in the nohz mode recently and busy at the current
4869 * scheduler tick, then check if we need to nominate new idle
4872 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4873 rq->in_nohz_recently = 0;
4875 if (atomic_read(&nohz.load_balancer) == cpu) {
4876 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4877 atomic_set(&nohz.load_balancer, -1);
4880 if (atomic_read(&nohz.load_balancer) == -1) {
4881 int ilb = find_new_ilb(cpu);
4883 if (ilb < nr_cpu_ids)
4889 * If this cpu is idle and doing idle load balancing for all the
4890 * cpus with ticks stopped, is it time for that to stop?
4892 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4893 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4899 * If this cpu is idle and the idle load balancing is done by
4900 * someone else, then no need raise the SCHED_SOFTIRQ
4902 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4903 cpumask_test_cpu(cpu, nohz.cpu_mask))
4906 /* Don't need to rebalance while attached to NULL domain */
4907 if (time_after_eq(jiffies, rq->next_balance) &&
4908 likely(!on_null_domain(cpu)))
4909 raise_softirq(SCHED_SOFTIRQ);
4912 #else /* CONFIG_SMP */
4915 * on UP we do not need to balance between CPUs:
4917 static inline void idle_balance(int cpu, struct rq *rq)
4923 DEFINE_PER_CPU(struct kernel_stat, kstat);
4925 EXPORT_PER_CPU_SYMBOL(kstat);
4928 * Return any ns on the sched_clock that have not yet been accounted in
4929 * @p in case that task is currently running.
4931 * Called with task_rq_lock() held on @rq.
4933 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
4937 if (task_current(rq, p)) {
4938 update_rq_clock(rq);
4939 ns = rq->clock - p->se.exec_start;
4947 unsigned long long task_delta_exec(struct task_struct *p)
4949 unsigned long flags;
4953 rq = task_rq_lock(p, &flags);
4954 ns = do_task_delta_exec(p, rq);
4955 task_rq_unlock(rq, &flags);
4961 * Return accounted runtime for the task.
4962 * In case the task is currently running, return the runtime plus current's
4963 * pending runtime that have not been accounted yet.
4965 unsigned long long task_sched_runtime(struct task_struct *p)
4967 unsigned long flags;
4971 rq = task_rq_lock(p, &flags);
4972 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
4973 task_rq_unlock(rq, &flags);
4979 * Return sum_exec_runtime for the thread group.
4980 * In case the task is currently running, return the sum plus current's
4981 * pending runtime that have not been accounted yet.
4983 * Note that the thread group might have other running tasks as well,
4984 * so the return value not includes other pending runtime that other
4985 * running tasks might have.
4987 unsigned long long thread_group_sched_runtime(struct task_struct *p)
4989 struct task_cputime totals;
4990 unsigned long flags;
4994 rq = task_rq_lock(p, &flags);
4995 thread_group_cputime(p, &totals);
4996 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
4997 task_rq_unlock(rq, &flags);
5003 * Account user cpu time to a process.
5004 * @p: the process that the cpu time gets accounted to
5005 * @cputime: the cpu time spent in user space since the last update
5006 * @cputime_scaled: cputime scaled by cpu frequency
5008 void account_user_time(struct task_struct *p, cputime_t cputime,
5009 cputime_t cputime_scaled)
5011 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5014 /* Add user time to process. */
5015 p->utime = cputime_add(p->utime, cputime);
5016 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5017 account_group_user_time(p, cputime);
5019 /* Add user time to cpustat. */
5020 tmp = cputime_to_cputime64(cputime);
5021 if (TASK_NICE(p) > 0)
5022 cpustat->nice = cputime64_add(cpustat->nice, tmp);
5024 cpustat->user = cputime64_add(cpustat->user, tmp);
5026 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
5027 /* Account for user time used */
5028 acct_update_integrals(p);
5032 * Account guest cpu time to a process.
5033 * @p: the process that the cpu time gets accounted to
5034 * @cputime: the cpu time spent in virtual machine since the last update
5035 * @cputime_scaled: cputime scaled by cpu frequency
5037 static void account_guest_time(struct task_struct *p, cputime_t cputime,
5038 cputime_t cputime_scaled)
5041 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5043 tmp = cputime_to_cputime64(cputime);
5045 /* Add guest time to process. */
5046 p->utime = cputime_add(p->utime, cputime);
5047 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5048 account_group_user_time(p, cputime);
5049 p->gtime = cputime_add(p->gtime, cputime);
5051 /* Add guest time to cpustat. */
5052 cpustat->user = cputime64_add(cpustat->user, tmp);
5053 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 cputime_t task_utime(struct task_struct *p)
5173 cputime_t task_stime(struct task_struct *p)
5178 cputime_t task_utime(struct task_struct *p)
5180 clock_t utime = cputime_to_clock_t(p->utime),
5181 total = utime + cputime_to_clock_t(p->stime);
5185 * Use CFS's precise accounting:
5187 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
5191 do_div(temp, total);
5193 utime = (clock_t)temp;
5195 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
5196 return p->prev_utime;
5199 cputime_t task_stime(struct task_struct *p)
5204 * Use CFS's precise accounting. (we subtract utime from
5205 * the total, to make sure the total observed by userspace
5206 * grows monotonically - apps rely on that):
5208 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
5209 cputime_to_clock_t(task_utime(p));
5212 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
5214 return p->prev_stime;
5218 inline cputime_t task_gtime(struct task_struct *p)
5224 * This function gets called by the timer code, with HZ frequency.
5225 * We call it with interrupts disabled.
5227 * It also gets called by the fork code, when changing the parent's
5230 void scheduler_tick(void)
5232 int cpu = smp_processor_id();
5233 struct rq *rq = cpu_rq(cpu);
5234 struct task_struct *curr = rq->curr;
5238 spin_lock(&rq->lock);
5239 update_rq_clock(rq);
5240 update_cpu_load(rq);
5241 curr->sched_class->task_tick(rq, curr, 0);
5242 spin_unlock(&rq->lock);
5244 perf_event_task_tick(curr, cpu);
5247 rq->idle_at_tick = idle_cpu(cpu);
5248 trigger_load_balance(rq, cpu);
5252 notrace unsigned long get_parent_ip(unsigned long addr)
5254 if (in_lock_functions(addr)) {
5255 addr = CALLER_ADDR2;
5256 if (in_lock_functions(addr))
5257 addr = CALLER_ADDR3;
5262 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5263 defined(CONFIG_PREEMPT_TRACER))
5265 void __kprobes add_preempt_count(int val)
5267 #ifdef CONFIG_DEBUG_PREEMPT
5271 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5274 preempt_count() += val;
5275 #ifdef CONFIG_DEBUG_PREEMPT
5277 * Spinlock count overflowing soon?
5279 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5282 if (preempt_count() == val)
5283 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5285 EXPORT_SYMBOL(add_preempt_count);
5287 void __kprobes sub_preempt_count(int val)
5289 #ifdef CONFIG_DEBUG_PREEMPT
5293 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5296 * Is the spinlock portion underflowing?
5298 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5299 !(preempt_count() & PREEMPT_MASK)))
5303 if (preempt_count() == val)
5304 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5305 preempt_count() -= val;
5307 EXPORT_SYMBOL(sub_preempt_count);
5312 * Print scheduling while atomic bug:
5314 static noinline void __schedule_bug(struct task_struct *prev)
5316 struct pt_regs *regs = get_irq_regs();
5318 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5319 prev->comm, prev->pid, preempt_count());
5321 debug_show_held_locks(prev);
5323 if (irqs_disabled())
5324 print_irqtrace_events(prev);
5333 * Various schedule()-time debugging checks and statistics:
5335 static inline void schedule_debug(struct task_struct *prev)
5338 * Test if we are atomic. Since do_exit() needs to call into
5339 * schedule() atomically, we ignore that path for now.
5340 * Otherwise, whine if we are scheduling when we should not be.
5342 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
5343 __schedule_bug(prev);
5345 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5347 schedstat_inc(this_rq(), sched_count);
5348 #ifdef CONFIG_SCHEDSTATS
5349 if (unlikely(prev->lock_depth >= 0)) {
5350 schedstat_inc(this_rq(), bkl_count);
5351 schedstat_inc(prev, sched_info.bkl_count);
5356 static void put_prev_task(struct rq *rq, struct task_struct *p)
5358 u64 runtime = p->se.sum_exec_runtime - p->se.prev_sum_exec_runtime;
5360 update_avg(&p->se.avg_running, runtime);
5362 if (p->state == TASK_RUNNING) {
5364 * In order to avoid avg_overlap growing stale when we are
5365 * indeed overlapping and hence not getting put to sleep, grow
5366 * the avg_overlap on preemption.
5368 * We use the average preemption runtime because that
5369 * correlates to the amount of cache footprint a task can
5372 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
5373 update_avg(&p->se.avg_overlap, runtime);
5375 update_avg(&p->se.avg_running, 0);
5377 p->sched_class->put_prev_task(rq, p);
5381 * Pick up the highest-prio task:
5383 static inline struct task_struct *
5384 pick_next_task(struct rq *rq)
5386 const struct sched_class *class;
5387 struct task_struct *p;
5390 * Optimization: we know that if all tasks are in
5391 * the fair class we can call that function directly:
5393 if (likely(rq->nr_running == rq->cfs.nr_running)) {
5394 p = fair_sched_class.pick_next_task(rq);
5399 class = sched_class_highest;
5401 p = class->pick_next_task(rq);
5405 * Will never be NULL as the idle class always
5406 * returns a non-NULL p:
5408 class = class->next;
5413 * schedule() is the main scheduler function.
5415 asmlinkage void __sched schedule(void)
5417 struct task_struct *prev, *next;
5418 unsigned long *switch_count;
5424 cpu = smp_processor_id();
5428 switch_count = &prev->nivcsw;
5430 release_kernel_lock(prev);
5431 need_resched_nonpreemptible:
5433 schedule_debug(prev);
5435 if (sched_feat(HRTICK))
5438 spin_lock_irq(&rq->lock);
5439 update_rq_clock(rq);
5440 clear_tsk_need_resched(prev);
5442 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
5443 if (unlikely(signal_pending_state(prev->state, prev)))
5444 prev->state = TASK_RUNNING;
5446 deactivate_task(rq, prev, 1);
5447 switch_count = &prev->nvcsw;
5450 pre_schedule(rq, prev);
5452 if (unlikely(!rq->nr_running))
5453 idle_balance(cpu, rq);
5455 put_prev_task(rq, prev);
5456 next = pick_next_task(rq);
5458 if (likely(prev != next)) {
5459 sched_info_switch(prev, next);
5460 perf_event_task_sched_out(prev, next, cpu);
5466 context_switch(rq, prev, next); /* unlocks the rq */
5468 * the context switch might have flipped the stack from under
5469 * us, hence refresh the local variables.
5471 cpu = smp_processor_id();
5474 spin_unlock_irq(&rq->lock);
5478 if (unlikely(reacquire_kernel_lock(current) < 0))
5479 goto need_resched_nonpreemptible;
5481 preempt_enable_no_resched();
5485 EXPORT_SYMBOL(schedule);
5489 * Look out! "owner" is an entirely speculative pointer
5490 * access and not reliable.
5492 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5497 if (!sched_feat(OWNER_SPIN))
5500 #ifdef CONFIG_DEBUG_PAGEALLOC
5502 * Need to access the cpu field knowing that
5503 * DEBUG_PAGEALLOC could have unmapped it if
5504 * the mutex owner just released it and exited.
5506 if (probe_kernel_address(&owner->cpu, cpu))
5513 * Even if the access succeeded (likely case),
5514 * the cpu field may no longer be valid.
5516 if (cpu >= nr_cpumask_bits)
5520 * We need to validate that we can do a
5521 * get_cpu() and that we have the percpu area.
5523 if (!cpu_online(cpu))
5530 * Owner changed, break to re-assess state.
5532 if (lock->owner != owner)
5536 * Is that owner really running on that cpu?
5538 if (task_thread_info(rq->curr) != owner || need_resched())
5548 #ifdef CONFIG_PREEMPT
5550 * this is the entry point to schedule() from in-kernel preemption
5551 * off of preempt_enable. Kernel preemptions off return from interrupt
5552 * occur there and call schedule directly.
5554 asmlinkage void __sched preempt_schedule(void)
5556 struct thread_info *ti = current_thread_info();
5559 * If there is a non-zero preempt_count or interrupts are disabled,
5560 * we do not want to preempt the current task. Just return..
5562 if (likely(ti->preempt_count || irqs_disabled()))
5566 add_preempt_count(PREEMPT_ACTIVE);
5568 sub_preempt_count(PREEMPT_ACTIVE);
5571 * Check again in case we missed a preemption opportunity
5572 * between schedule and now.
5575 } while (need_resched());
5577 EXPORT_SYMBOL(preempt_schedule);
5580 * this is the entry point to schedule() from kernel preemption
5581 * off of irq context.
5582 * Note, that this is called and return with irqs disabled. This will
5583 * protect us against recursive calling from irq.
5585 asmlinkage void __sched preempt_schedule_irq(void)
5587 struct thread_info *ti = current_thread_info();
5589 /* Catch callers which need to be fixed */
5590 BUG_ON(ti->preempt_count || !irqs_disabled());
5593 add_preempt_count(PREEMPT_ACTIVE);
5596 local_irq_disable();
5597 sub_preempt_count(PREEMPT_ACTIVE);
5600 * Check again in case we missed a preemption opportunity
5601 * between schedule and now.
5604 } while (need_resched());
5607 #endif /* CONFIG_PREEMPT */
5609 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
5612 return try_to_wake_up(curr->private, mode, wake_flags);
5614 EXPORT_SYMBOL(default_wake_function);
5617 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5618 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5619 * number) then we wake all the non-exclusive tasks and one exclusive task.
5621 * There are circumstances in which we can try to wake a task which has already
5622 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5623 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5625 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
5626 int nr_exclusive, int wake_flags, void *key)
5628 wait_queue_t *curr, *next;
5630 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
5631 unsigned flags = curr->flags;
5633 if (curr->func(curr, mode, wake_flags, key) &&
5634 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
5640 * __wake_up - wake up threads blocked on a waitqueue.
5642 * @mode: which threads
5643 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5644 * @key: is directly passed to the wakeup function
5646 * It may be assumed that this function implies a write memory barrier before
5647 * changing the task state if and only if any tasks are woken up.
5649 void __wake_up(wait_queue_head_t *q, unsigned int mode,
5650 int nr_exclusive, void *key)
5652 unsigned long flags;
5654 spin_lock_irqsave(&q->lock, flags);
5655 __wake_up_common(q, mode, nr_exclusive, 0, key);
5656 spin_unlock_irqrestore(&q->lock, flags);
5658 EXPORT_SYMBOL(__wake_up);
5661 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5663 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
5665 __wake_up_common(q, mode, 1, 0, NULL);
5668 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
5670 __wake_up_common(q, mode, 1, 0, key);
5674 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5676 * @mode: which threads
5677 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5678 * @key: opaque value to be passed to wakeup targets
5680 * The sync wakeup differs that the waker knows that it will schedule
5681 * away soon, so while the target thread will be woken up, it will not
5682 * be migrated to another CPU - ie. the two threads are 'synchronized'
5683 * with each other. This can prevent needless bouncing between CPUs.
5685 * On UP it can prevent extra preemption.
5687 * It may be assumed that this function implies a write memory barrier before
5688 * changing the task state if and only if any tasks are woken up.
5690 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
5691 int nr_exclusive, void *key)
5693 unsigned long flags;
5694 int wake_flags = WF_SYNC;
5699 if (unlikely(!nr_exclusive))
5702 spin_lock_irqsave(&q->lock, flags);
5703 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
5704 spin_unlock_irqrestore(&q->lock, flags);
5706 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
5709 * __wake_up_sync - see __wake_up_sync_key()
5711 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
5713 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
5715 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
5718 * complete: - signals a single thread waiting on this completion
5719 * @x: holds the state of this particular completion
5721 * This will wake up a single thread waiting on this completion. Threads will be
5722 * awakened in the same order in which they were queued.
5724 * See also complete_all(), wait_for_completion() and related routines.
5726 * It may be assumed that this function implies a write memory barrier before
5727 * changing the task state if and only if any tasks are woken up.
5729 void complete(struct completion *x)
5731 unsigned long flags;
5733 spin_lock_irqsave(&x->wait.lock, flags);
5735 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
5736 spin_unlock_irqrestore(&x->wait.lock, flags);
5738 EXPORT_SYMBOL(complete);
5741 * complete_all: - signals all threads waiting on this completion
5742 * @x: holds the state of this particular completion
5744 * This will wake up all threads waiting on this particular completion event.
5746 * It may be assumed that this function implies a write memory barrier before
5747 * changing the task state if and only if any tasks are woken up.
5749 void complete_all(struct completion *x)
5751 unsigned long flags;
5753 spin_lock_irqsave(&x->wait.lock, flags);
5754 x->done += UINT_MAX/2;
5755 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
5756 spin_unlock_irqrestore(&x->wait.lock, flags);
5758 EXPORT_SYMBOL(complete_all);
5760 static inline long __sched
5761 do_wait_for_common(struct completion *x, long timeout, int state)
5764 DECLARE_WAITQUEUE(wait, current);
5766 wait.flags |= WQ_FLAG_EXCLUSIVE;
5767 __add_wait_queue_tail(&x->wait, &wait);
5769 if (signal_pending_state(state, current)) {
5770 timeout = -ERESTARTSYS;
5773 __set_current_state(state);
5774 spin_unlock_irq(&x->wait.lock);
5775 timeout = schedule_timeout(timeout);
5776 spin_lock_irq(&x->wait.lock);
5777 } while (!x->done && timeout);
5778 __remove_wait_queue(&x->wait, &wait);
5783 return timeout ?: 1;
5787 wait_for_common(struct completion *x, long timeout, int state)
5791 spin_lock_irq(&x->wait.lock);
5792 timeout = do_wait_for_common(x, timeout, state);
5793 spin_unlock_irq(&x->wait.lock);
5798 * wait_for_completion: - waits for completion of a task
5799 * @x: holds the state of this particular completion
5801 * This waits to be signaled for completion of a specific task. It is NOT
5802 * interruptible and there is no timeout.
5804 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5805 * and interrupt capability. Also see complete().
5807 void __sched wait_for_completion(struct completion *x)
5809 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
5811 EXPORT_SYMBOL(wait_for_completion);
5814 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5815 * @x: holds the state of this particular completion
5816 * @timeout: timeout value in jiffies
5818 * This waits for either a completion of a specific task to be signaled or for a
5819 * specified timeout to expire. The timeout is in jiffies. It is not
5822 unsigned long __sched
5823 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5825 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5827 EXPORT_SYMBOL(wait_for_completion_timeout);
5830 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5831 * @x: holds the state of this particular completion
5833 * This waits for completion of a specific task to be signaled. It is
5836 int __sched wait_for_completion_interruptible(struct completion *x)
5838 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5839 if (t == -ERESTARTSYS)
5843 EXPORT_SYMBOL(wait_for_completion_interruptible);
5846 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5847 * @x: holds the state of this particular completion
5848 * @timeout: timeout value in jiffies
5850 * This waits for either a completion of a specific task to be signaled or for a
5851 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5853 unsigned long __sched
5854 wait_for_completion_interruptible_timeout(struct completion *x,
5855 unsigned long timeout)
5857 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
5859 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5862 * wait_for_completion_killable: - waits for completion of a task (killable)
5863 * @x: holds the state of this particular completion
5865 * This waits to be signaled for completion of a specific task. It can be
5866 * interrupted by a kill signal.
5868 int __sched wait_for_completion_killable(struct completion *x)
5870 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5871 if (t == -ERESTARTSYS)
5875 EXPORT_SYMBOL(wait_for_completion_killable);
5878 * try_wait_for_completion - try to decrement a completion without blocking
5879 * @x: completion structure
5881 * Returns: 0 if a decrement cannot be done without blocking
5882 * 1 if a decrement succeeded.
5884 * If a completion is being used as a counting completion,
5885 * attempt to decrement the counter without blocking. This
5886 * enables us to avoid waiting if the resource the completion
5887 * is protecting is not available.
5889 bool try_wait_for_completion(struct completion *x)
5893 spin_lock_irq(&x->wait.lock);
5898 spin_unlock_irq(&x->wait.lock);
5901 EXPORT_SYMBOL(try_wait_for_completion);
5904 * completion_done - Test to see if a completion has any waiters
5905 * @x: completion structure
5907 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5908 * 1 if there are no waiters.
5911 bool completion_done(struct completion *x)
5915 spin_lock_irq(&x->wait.lock);
5918 spin_unlock_irq(&x->wait.lock);
5921 EXPORT_SYMBOL(completion_done);
5924 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
5926 unsigned long flags;
5929 init_waitqueue_entry(&wait, current);
5931 __set_current_state(state);
5933 spin_lock_irqsave(&q->lock, flags);
5934 __add_wait_queue(q, &wait);
5935 spin_unlock(&q->lock);
5936 timeout = schedule_timeout(timeout);
5937 spin_lock_irq(&q->lock);
5938 __remove_wait_queue(q, &wait);
5939 spin_unlock_irqrestore(&q->lock, flags);
5944 void __sched interruptible_sleep_on(wait_queue_head_t *q)
5946 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5948 EXPORT_SYMBOL(interruptible_sleep_on);
5951 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
5953 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
5955 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
5957 void __sched sleep_on(wait_queue_head_t *q)
5959 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5961 EXPORT_SYMBOL(sleep_on);
5963 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
5965 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
5967 EXPORT_SYMBOL(sleep_on_timeout);
5969 #ifdef CONFIG_RT_MUTEXES
5972 * rt_mutex_setprio - set the current priority of a task
5974 * @prio: prio value (kernel-internal form)
5976 * This function changes the 'effective' priority of a task. It does
5977 * not touch ->normal_prio like __setscheduler().
5979 * Used by the rt_mutex code to implement priority inheritance logic.
5981 void rt_mutex_setprio(struct task_struct *p, int prio)
5983 unsigned long flags;
5984 int oldprio, on_rq, running;
5986 const struct sched_class *prev_class = p->sched_class;
5988 BUG_ON(prio < 0 || prio > MAX_PRIO);
5990 rq = task_rq_lock(p, &flags);
5991 update_rq_clock(rq);
5994 on_rq = p->se.on_rq;
5995 running = task_current(rq, p);
5997 dequeue_task(rq, p, 0);
5999 p->sched_class->put_prev_task(rq, p);
6002 p->sched_class = &rt_sched_class;
6004 p->sched_class = &fair_sched_class;
6009 p->sched_class->set_curr_task(rq);
6011 enqueue_task(rq, p, 0);
6013 check_class_changed(rq, p, prev_class, oldprio, running);
6015 task_rq_unlock(rq, &flags);
6020 void set_user_nice(struct task_struct *p, long nice)
6022 int old_prio, delta, on_rq;
6023 unsigned long flags;
6026 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
6029 * We have to be careful, if called from sys_setpriority(),
6030 * the task might be in the middle of scheduling on another CPU.
6032 rq = task_rq_lock(p, &flags);
6033 update_rq_clock(rq);
6035 * The RT priorities are set via sched_setscheduler(), but we still
6036 * allow the 'normal' nice value to be set - but as expected
6037 * it wont have any effect on scheduling until the task is
6038 * SCHED_FIFO/SCHED_RR:
6040 if (task_has_rt_policy(p)) {
6041 p->static_prio = NICE_TO_PRIO(nice);
6044 on_rq = p->se.on_rq;
6046 dequeue_task(rq, p, 0);
6048 p->static_prio = NICE_TO_PRIO(nice);
6051 p->prio = effective_prio(p);
6052 delta = p->prio - old_prio;
6055 enqueue_task(rq, p, 0);
6057 * If the task increased its priority or is running and
6058 * lowered its priority, then reschedule its CPU:
6060 if (delta < 0 || (delta > 0 && task_running(rq, p)))
6061 resched_task(rq->curr);
6064 task_rq_unlock(rq, &flags);
6066 EXPORT_SYMBOL(set_user_nice);
6069 * can_nice - check if a task can reduce its nice value
6073 int can_nice(const struct task_struct *p, const int nice)
6075 /* convert nice value [19,-20] to rlimit style value [1,40] */
6076 int nice_rlim = 20 - nice;
6078 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
6079 capable(CAP_SYS_NICE));
6082 #ifdef __ARCH_WANT_SYS_NICE
6085 * sys_nice - change the priority of the current process.
6086 * @increment: priority increment
6088 * sys_setpriority is a more generic, but much slower function that
6089 * does similar things.
6091 SYSCALL_DEFINE1(nice, int, increment)
6096 * Setpriority might change our priority at the same moment.
6097 * We don't have to worry. Conceptually one call occurs first
6098 * and we have a single winner.
6100 if (increment < -40)
6105 nice = TASK_NICE(current) + increment;
6111 if (increment < 0 && !can_nice(current, nice))
6114 retval = security_task_setnice(current, nice);
6118 set_user_nice(current, nice);
6125 * task_prio - return the priority value of a given task.
6126 * @p: the task in question.
6128 * This is the priority value as seen by users in /proc.
6129 * RT tasks are offset by -200. Normal tasks are centered
6130 * around 0, value goes from -16 to +15.
6132 int task_prio(const struct task_struct *p)
6134 return p->prio - MAX_RT_PRIO;
6138 * task_nice - return the nice value of a given task.
6139 * @p: the task in question.
6141 int task_nice(const struct task_struct *p)
6143 return TASK_NICE(p);
6145 EXPORT_SYMBOL(task_nice);
6148 * idle_cpu - is a given cpu idle currently?
6149 * @cpu: the processor in question.
6151 int idle_cpu(int cpu)
6153 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
6157 * idle_task - return the idle task for a given cpu.
6158 * @cpu: the processor in question.
6160 struct task_struct *idle_task(int cpu)
6162 return cpu_rq(cpu)->idle;
6166 * find_process_by_pid - find a process with a matching PID value.
6167 * @pid: the pid in question.
6169 static struct task_struct *find_process_by_pid(pid_t pid)
6171 return pid ? find_task_by_vpid(pid) : current;
6174 /* Actually do priority change: must hold rq lock. */
6176 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
6178 BUG_ON(p->se.on_rq);
6181 switch (p->policy) {
6185 p->sched_class = &fair_sched_class;
6189 p->sched_class = &rt_sched_class;
6193 p->rt_priority = prio;
6194 p->normal_prio = normal_prio(p);
6195 /* we are holding p->pi_lock already */
6196 p->prio = rt_mutex_getprio(p);
6201 * check the target process has a UID that matches the current process's
6203 static bool check_same_owner(struct task_struct *p)
6205 const struct cred *cred = current_cred(), *pcred;
6209 pcred = __task_cred(p);
6210 match = (cred->euid == pcred->euid ||
6211 cred->euid == pcred->uid);
6216 static int __sched_setscheduler(struct task_struct *p, int policy,
6217 struct sched_param *param, bool user)
6219 int retval, oldprio, oldpolicy = -1, on_rq, running;
6220 unsigned long flags;
6221 const struct sched_class *prev_class = p->sched_class;
6225 /* may grab non-irq protected spin_locks */
6226 BUG_ON(in_interrupt());
6228 /* double check policy once rq lock held */
6230 reset_on_fork = p->sched_reset_on_fork;
6231 policy = oldpolicy = p->policy;
6233 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
6234 policy &= ~SCHED_RESET_ON_FORK;
6236 if (policy != SCHED_FIFO && policy != SCHED_RR &&
6237 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
6238 policy != SCHED_IDLE)
6243 * Valid priorities for SCHED_FIFO and SCHED_RR are
6244 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6245 * SCHED_BATCH and SCHED_IDLE is 0.
6247 if (param->sched_priority < 0 ||
6248 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
6249 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
6251 if (rt_policy(policy) != (param->sched_priority != 0))
6255 * Allow unprivileged RT tasks to decrease priority:
6257 if (user && !capable(CAP_SYS_NICE)) {
6258 if (rt_policy(policy)) {
6259 unsigned long rlim_rtprio;
6261 if (!lock_task_sighand(p, &flags))
6263 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
6264 unlock_task_sighand(p, &flags);
6266 /* can't set/change the rt policy */
6267 if (policy != p->policy && !rlim_rtprio)
6270 /* can't increase priority */
6271 if (param->sched_priority > p->rt_priority &&
6272 param->sched_priority > rlim_rtprio)
6276 * Like positive nice levels, dont allow tasks to
6277 * move out of SCHED_IDLE either:
6279 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
6282 /* can't change other user's priorities */
6283 if (!check_same_owner(p))
6286 /* Normal users shall not reset the sched_reset_on_fork flag */
6287 if (p->sched_reset_on_fork && !reset_on_fork)
6292 #ifdef CONFIG_RT_GROUP_SCHED
6294 * Do not allow realtime tasks into groups that have no runtime
6297 if (rt_bandwidth_enabled() && rt_policy(policy) &&
6298 task_group(p)->rt_bandwidth.rt_runtime == 0)
6302 retval = security_task_setscheduler(p, policy, param);
6308 * make sure no PI-waiters arrive (or leave) while we are
6309 * changing the priority of the task:
6311 spin_lock_irqsave(&p->pi_lock, flags);
6313 * To be able to change p->policy safely, the apropriate
6314 * runqueue lock must be held.
6316 rq = __task_rq_lock(p);
6317 /* recheck policy now with rq lock held */
6318 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
6319 policy = oldpolicy = -1;
6320 __task_rq_unlock(rq);
6321 spin_unlock_irqrestore(&p->pi_lock, flags);
6324 update_rq_clock(rq);
6325 on_rq = p->se.on_rq;
6326 running = task_current(rq, p);
6328 deactivate_task(rq, p, 0);
6330 p->sched_class->put_prev_task(rq, p);
6332 p->sched_reset_on_fork = reset_on_fork;
6335 __setscheduler(rq, p, policy, param->sched_priority);
6338 p->sched_class->set_curr_task(rq);
6340 activate_task(rq, p, 0);
6342 check_class_changed(rq, p, prev_class, oldprio, running);
6344 __task_rq_unlock(rq);
6345 spin_unlock_irqrestore(&p->pi_lock, flags);
6347 rt_mutex_adjust_pi(p);
6353 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6354 * @p: the task in question.
6355 * @policy: new policy.
6356 * @param: structure containing the new RT priority.
6358 * NOTE that the task may be already dead.
6360 int sched_setscheduler(struct task_struct *p, int policy,
6361 struct sched_param *param)
6363 return __sched_setscheduler(p, policy, param, true);
6365 EXPORT_SYMBOL_GPL(sched_setscheduler);
6368 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6369 * @p: the task in question.
6370 * @policy: new policy.
6371 * @param: structure containing the new RT priority.
6373 * Just like sched_setscheduler, only don't bother checking if the
6374 * current context has permission. For example, this is needed in
6375 * stop_machine(): we create temporary high priority worker threads,
6376 * but our caller might not have that capability.
6378 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
6379 struct sched_param *param)
6381 return __sched_setscheduler(p, policy, param, false);
6385 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
6387 struct sched_param lparam;
6388 struct task_struct *p;
6391 if (!param || pid < 0)
6393 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
6398 p = find_process_by_pid(pid);
6400 retval = sched_setscheduler(p, policy, &lparam);
6407 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6408 * @pid: the pid in question.
6409 * @policy: new policy.
6410 * @param: structure containing the new RT priority.
6412 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
6413 struct sched_param __user *, param)
6415 /* negative values for policy are not valid */
6419 return do_sched_setscheduler(pid, policy, param);
6423 * sys_sched_setparam - set/change the RT priority of a thread
6424 * @pid: the pid in question.
6425 * @param: structure containing the new RT priority.
6427 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
6429 return do_sched_setscheduler(pid, -1, param);
6433 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6434 * @pid: the pid in question.
6436 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
6438 struct task_struct *p;
6445 read_lock(&tasklist_lock);
6446 p = find_process_by_pid(pid);
6448 retval = security_task_getscheduler(p);
6451 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
6453 read_unlock(&tasklist_lock);
6458 * sys_sched_getparam - get the RT priority of a thread
6459 * @pid: the pid in question.
6460 * @param: structure containing the RT priority.
6462 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
6464 struct sched_param lp;
6465 struct task_struct *p;
6468 if (!param || pid < 0)
6471 read_lock(&tasklist_lock);
6472 p = find_process_by_pid(pid);
6477 retval = security_task_getscheduler(p);
6481 lp.sched_priority = p->rt_priority;
6482 read_unlock(&tasklist_lock);
6485 * This one might sleep, we cannot do it with a spinlock held ...
6487 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
6492 read_unlock(&tasklist_lock);
6496 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
6498 cpumask_var_t cpus_allowed, new_mask;
6499 struct task_struct *p;
6503 read_lock(&tasklist_lock);
6505 p = find_process_by_pid(pid);
6507 read_unlock(&tasklist_lock);
6513 * It is not safe to call set_cpus_allowed with the
6514 * tasklist_lock held. We will bump the task_struct's
6515 * usage count and then drop tasklist_lock.
6518 read_unlock(&tasklist_lock);
6520 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6524 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
6526 goto out_free_cpus_allowed;
6529 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
6532 retval = security_task_setscheduler(p, 0, NULL);
6536 cpuset_cpus_allowed(p, cpus_allowed);
6537 cpumask_and(new_mask, in_mask, cpus_allowed);
6539 retval = set_cpus_allowed_ptr(p, new_mask);
6542 cpuset_cpus_allowed(p, cpus_allowed);
6543 if (!cpumask_subset(new_mask, cpus_allowed)) {
6545 * We must have raced with a concurrent cpuset
6546 * update. Just reset the cpus_allowed to the
6547 * cpuset's cpus_allowed
6549 cpumask_copy(new_mask, cpus_allowed);
6554 free_cpumask_var(new_mask);
6555 out_free_cpus_allowed:
6556 free_cpumask_var(cpus_allowed);
6563 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
6564 struct cpumask *new_mask)
6566 if (len < cpumask_size())
6567 cpumask_clear(new_mask);
6568 else if (len > cpumask_size())
6569 len = cpumask_size();
6571 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6575 * sys_sched_setaffinity - set the cpu affinity of a process
6576 * @pid: pid of the process
6577 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6578 * @user_mask_ptr: user-space pointer to the new cpu mask
6580 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6581 unsigned long __user *, user_mask_ptr)
6583 cpumask_var_t new_mask;
6586 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6589 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6591 retval = sched_setaffinity(pid, new_mask);
6592 free_cpumask_var(new_mask);
6596 long sched_getaffinity(pid_t pid, struct cpumask *mask)
6598 struct task_struct *p;
6602 read_lock(&tasklist_lock);
6605 p = find_process_by_pid(pid);
6609 retval = security_task_getscheduler(p);
6613 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
6616 read_unlock(&tasklist_lock);
6623 * sys_sched_getaffinity - get the cpu affinity of a process
6624 * @pid: pid of the process
6625 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6626 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6628 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6629 unsigned long __user *, user_mask_ptr)
6634 if (len < cpumask_size())
6637 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6640 ret = sched_getaffinity(pid, mask);
6642 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
6645 ret = cpumask_size();
6647 free_cpumask_var(mask);
6653 * sys_sched_yield - yield the current processor to other threads.
6655 * This function yields the current CPU to other tasks. If there are no
6656 * other threads running on this CPU then this function will return.
6658 SYSCALL_DEFINE0(sched_yield)
6660 struct rq *rq = this_rq_lock();
6662 schedstat_inc(rq, yld_count);
6663 current->sched_class->yield_task(rq);
6666 * Since we are going to call schedule() anyway, there's
6667 * no need to preempt or enable interrupts:
6669 __release(rq->lock);
6670 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
6671 _raw_spin_unlock(&rq->lock);
6672 preempt_enable_no_resched();
6679 static inline int should_resched(void)
6681 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
6684 static void __cond_resched(void)
6686 add_preempt_count(PREEMPT_ACTIVE);
6688 sub_preempt_count(PREEMPT_ACTIVE);
6691 int __sched _cond_resched(void)
6693 if (should_resched()) {
6699 EXPORT_SYMBOL(_cond_resched);
6702 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
6703 * call schedule, and on return reacquire the lock.
6705 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6706 * operations here to prevent schedule() from being called twice (once via
6707 * spin_unlock(), once by hand).
6709 int __cond_resched_lock(spinlock_t *lock)
6711 int resched = should_resched();
6714 lockdep_assert_held(lock);
6716 if (spin_needbreak(lock) || resched) {
6727 EXPORT_SYMBOL(__cond_resched_lock);
6729 int __sched __cond_resched_softirq(void)
6731 BUG_ON(!in_softirq());
6733 if (should_resched()) {
6741 EXPORT_SYMBOL(__cond_resched_softirq);
6744 * yield - yield the current processor to other threads.
6746 * This is a shortcut for kernel-space yielding - it marks the
6747 * thread runnable and calls sys_sched_yield().
6749 void __sched yield(void)
6751 set_current_state(TASK_RUNNING);
6754 EXPORT_SYMBOL(yield);
6757 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6758 * that process accounting knows that this is a task in IO wait state.
6760 * But don't do that if it is a deliberate, throttling IO wait (this task
6761 * has set its backing_dev_info: the queue against which it should throttle)
6763 void __sched io_schedule(void)
6765 struct rq *rq = raw_rq();
6767 delayacct_blkio_start();
6768 atomic_inc(&rq->nr_iowait);
6769 current->in_iowait = 1;
6771 current->in_iowait = 0;
6772 atomic_dec(&rq->nr_iowait);
6773 delayacct_blkio_end();
6775 EXPORT_SYMBOL(io_schedule);
6777 long __sched io_schedule_timeout(long timeout)
6779 struct rq *rq = raw_rq();
6782 delayacct_blkio_start();
6783 atomic_inc(&rq->nr_iowait);
6784 current->in_iowait = 1;
6785 ret = schedule_timeout(timeout);
6786 current->in_iowait = 0;
6787 atomic_dec(&rq->nr_iowait);
6788 delayacct_blkio_end();
6793 * sys_sched_get_priority_max - return maximum RT priority.
6794 * @policy: scheduling class.
6796 * this syscall returns the maximum rt_priority that can be used
6797 * by a given scheduling class.
6799 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6806 ret = MAX_USER_RT_PRIO-1;
6818 * sys_sched_get_priority_min - return minimum RT priority.
6819 * @policy: scheduling class.
6821 * this syscall returns the minimum rt_priority that can be used
6822 * by a given scheduling class.
6824 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6842 * sys_sched_rr_get_interval - return the default timeslice of a process.
6843 * @pid: pid of the process.
6844 * @interval: userspace pointer to the timeslice value.
6846 * this syscall writes the default timeslice value of a given process
6847 * into the user-space timespec buffer. A value of '0' means infinity.
6849 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6850 struct timespec __user *, interval)
6852 struct task_struct *p;
6853 unsigned int time_slice;
6861 read_lock(&tasklist_lock);
6862 p = find_process_by_pid(pid);
6866 retval = security_task_getscheduler(p);
6870 time_slice = p->sched_class->get_rr_interval(p);
6872 read_unlock(&tasklist_lock);
6873 jiffies_to_timespec(time_slice, &t);
6874 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
6878 read_unlock(&tasklist_lock);
6882 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
6884 void sched_show_task(struct task_struct *p)
6886 unsigned long free = 0;
6889 state = p->state ? __ffs(p->state) + 1 : 0;
6890 printk(KERN_INFO "%-13.13s %c", p->comm,
6891 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
6892 #if BITS_PER_LONG == 32
6893 if (state == TASK_RUNNING)
6894 printk(KERN_CONT " running ");
6896 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
6898 if (state == TASK_RUNNING)
6899 printk(KERN_CONT " running task ");
6901 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
6903 #ifdef CONFIG_DEBUG_STACK_USAGE
6904 free = stack_not_used(p);
6906 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
6907 task_pid_nr(p), task_pid_nr(p->real_parent),
6908 (unsigned long)task_thread_info(p)->flags);
6910 show_stack(p, NULL);
6913 void show_state_filter(unsigned long state_filter)
6915 struct task_struct *g, *p;
6917 #if BITS_PER_LONG == 32
6919 " task PC stack pid father\n");
6922 " task PC stack pid father\n");
6924 read_lock(&tasklist_lock);
6925 do_each_thread(g, p) {
6927 * reset the NMI-timeout, listing all files on a slow
6928 * console might take alot of time:
6930 touch_nmi_watchdog();
6931 if (!state_filter || (p->state & state_filter))
6933 } while_each_thread(g, p);
6935 touch_all_softlockup_watchdogs();
6937 #ifdef CONFIG_SCHED_DEBUG
6938 sysrq_sched_debug_show();
6940 read_unlock(&tasklist_lock);
6942 * Only show locks if all tasks are dumped:
6944 if (state_filter == -1)
6945 debug_show_all_locks();
6948 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6950 idle->sched_class = &idle_sched_class;
6954 * init_idle - set up an idle thread for a given CPU
6955 * @idle: task in question
6956 * @cpu: cpu the idle task belongs to
6958 * NOTE: this function does not set the idle thread's NEED_RESCHED
6959 * flag, to make booting more robust.
6961 void __cpuinit init_idle(struct task_struct *idle, int cpu)
6963 struct rq *rq = cpu_rq(cpu);
6964 unsigned long flags;
6966 spin_lock_irqsave(&rq->lock, flags);
6969 idle->se.exec_start = sched_clock();
6971 idle->prio = idle->normal_prio = MAX_PRIO;
6972 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
6973 __set_task_cpu(idle, cpu);
6975 rq->curr = rq->idle = idle;
6976 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6979 spin_unlock_irqrestore(&rq->lock, flags);
6981 /* Set the preempt count _outside_ the spinlocks! */
6982 #if defined(CONFIG_PREEMPT)
6983 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
6985 task_thread_info(idle)->preempt_count = 0;
6988 * The idle tasks have their own, simple scheduling class:
6990 idle->sched_class = &idle_sched_class;
6991 ftrace_graph_init_task(idle);
6995 * In a system that switches off the HZ timer nohz_cpu_mask
6996 * indicates which cpus entered this state. This is used
6997 * in the rcu update to wait only for active cpus. For system
6998 * which do not switch off the HZ timer nohz_cpu_mask should
6999 * always be CPU_BITS_NONE.
7001 cpumask_var_t nohz_cpu_mask;
7004 * Increase the granularity value when there are more CPUs,
7005 * because with more CPUs the 'effective latency' as visible
7006 * to users decreases. But the relationship is not linear,
7007 * so pick a second-best guess by going with the log2 of the
7010 * This idea comes from the SD scheduler of Con Kolivas:
7012 static inline void sched_init_granularity(void)
7014 unsigned int factor = 1 + ilog2(num_online_cpus());
7015 const unsigned long limit = 200000000;
7017 sysctl_sched_min_granularity *= factor;
7018 if (sysctl_sched_min_granularity > limit)
7019 sysctl_sched_min_granularity = limit;
7021 sysctl_sched_latency *= factor;
7022 if (sysctl_sched_latency > limit)
7023 sysctl_sched_latency = limit;
7025 sysctl_sched_wakeup_granularity *= factor;
7027 sysctl_sched_shares_ratelimit *= factor;
7032 * This is how migration works:
7034 * 1) we queue a struct migration_req structure in the source CPU's
7035 * runqueue and wake up that CPU's migration thread.
7036 * 2) we down() the locked semaphore => thread blocks.
7037 * 3) migration thread wakes up (implicitly it forces the migrated
7038 * thread off the CPU)
7039 * 4) it gets the migration request and checks whether the migrated
7040 * task is still in the wrong runqueue.
7041 * 5) if it's in the wrong runqueue then the migration thread removes
7042 * it and puts it into the right queue.
7043 * 6) migration thread up()s the semaphore.
7044 * 7) we wake up and the migration is done.
7048 * Change a given task's CPU affinity. Migrate the thread to a
7049 * proper CPU and schedule it away if the CPU it's executing on
7050 * is removed from the allowed bitmask.
7052 * NOTE: the caller must have a valid reference to the task, the
7053 * task must not exit() & deallocate itself prematurely. The
7054 * call is not atomic; no spinlocks may be held.
7056 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
7058 struct migration_req req;
7059 unsigned long flags;
7063 rq = task_rq_lock(p, &flags);
7064 if (!cpumask_intersects(new_mask, cpu_online_mask)) {
7069 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
7070 !cpumask_equal(&p->cpus_allowed, new_mask))) {
7075 if (p->sched_class->set_cpus_allowed)
7076 p->sched_class->set_cpus_allowed(p, new_mask);
7078 cpumask_copy(&p->cpus_allowed, new_mask);
7079 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
7082 /* Can the task run on the task's current CPU? If so, we're done */
7083 if (cpumask_test_cpu(task_cpu(p), new_mask))
7086 if (migrate_task(p, cpumask_any_and(cpu_online_mask, new_mask), &req)) {
7087 /* Need help from migration thread: drop lock and wait. */
7088 struct task_struct *mt = rq->migration_thread;
7090 get_task_struct(mt);
7091 task_rq_unlock(rq, &flags);
7092 wake_up_process(rq->migration_thread);
7093 put_task_struct(mt);
7094 wait_for_completion(&req.done);
7095 tlb_migrate_finish(p->mm);
7099 task_rq_unlock(rq, &flags);
7103 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
7106 * Move (not current) task off this cpu, onto dest cpu. We're doing
7107 * this because either it can't run here any more (set_cpus_allowed()
7108 * away from this CPU, or CPU going down), or because we're
7109 * attempting to rebalance this task on exec (sched_exec).
7111 * So we race with normal scheduler movements, but that's OK, as long
7112 * as the task is no longer on this CPU.
7114 * Returns non-zero if task was successfully migrated.
7116 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
7118 struct rq *rq_dest, *rq_src;
7121 if (unlikely(!cpu_active(dest_cpu)))
7124 rq_src = cpu_rq(src_cpu);
7125 rq_dest = cpu_rq(dest_cpu);
7127 double_rq_lock(rq_src, rq_dest);
7128 /* Already moved. */
7129 if (task_cpu(p) != src_cpu)
7131 /* Affinity changed (again). */
7132 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7135 on_rq = p->se.on_rq;
7137 deactivate_task(rq_src, p, 0);
7139 set_task_cpu(p, dest_cpu);
7141 activate_task(rq_dest, p, 0);
7142 check_preempt_curr(rq_dest, p, 0);
7147 double_rq_unlock(rq_src, rq_dest);
7151 #define RCU_MIGRATION_IDLE 0
7152 #define RCU_MIGRATION_NEED_QS 1
7153 #define RCU_MIGRATION_GOT_QS 2
7154 #define RCU_MIGRATION_MUST_SYNC 3
7157 * migration_thread - this is a highprio system thread that performs
7158 * thread migration by bumping thread off CPU then 'pushing' onto
7161 static int migration_thread(void *data)
7164 int cpu = (long)data;
7168 BUG_ON(rq->migration_thread != current);
7170 set_current_state(TASK_INTERRUPTIBLE);
7171 while (!kthread_should_stop()) {
7172 struct migration_req *req;
7173 struct list_head *head;
7175 spin_lock_irq(&rq->lock);
7177 if (cpu_is_offline(cpu)) {
7178 spin_unlock_irq(&rq->lock);
7182 if (rq->active_balance) {
7183 active_load_balance(rq, cpu);
7184 rq->active_balance = 0;
7187 head = &rq->migration_queue;
7189 if (list_empty(head)) {
7190 spin_unlock_irq(&rq->lock);
7192 set_current_state(TASK_INTERRUPTIBLE);
7195 req = list_entry(head->next, struct migration_req, list);
7196 list_del_init(head->next);
7198 if (req->task != NULL) {
7199 spin_unlock(&rq->lock);
7200 __migrate_task(req->task, cpu, req->dest_cpu);
7201 } else if (likely(cpu == (badcpu = smp_processor_id()))) {
7202 req->dest_cpu = RCU_MIGRATION_GOT_QS;
7203 spin_unlock(&rq->lock);
7205 req->dest_cpu = RCU_MIGRATION_MUST_SYNC;
7206 spin_unlock(&rq->lock);
7207 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu, cpu);
7211 complete(&req->done);
7213 __set_current_state(TASK_RUNNING);
7218 #ifdef CONFIG_HOTPLUG_CPU
7220 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
7224 local_irq_disable();
7225 ret = __migrate_task(p, src_cpu, dest_cpu);
7231 * Figure out where task on dead CPU should go, use force if necessary.
7233 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
7236 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(dead_cpu));
7239 /* Look for allowed, online CPU in same node. */
7240 for_each_cpu_and(dest_cpu, nodemask, cpu_online_mask)
7241 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7244 /* Any allowed, online CPU? */
7245 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_online_mask);
7246 if (dest_cpu < nr_cpu_ids)
7249 /* No more Mr. Nice Guy. */
7250 if (dest_cpu >= nr_cpu_ids) {
7251 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
7252 dest_cpu = cpumask_any_and(cpu_online_mask, &p->cpus_allowed);
7255 * Don't tell them about moving exiting tasks or
7256 * kernel threads (both mm NULL), since they never
7259 if (p->mm && printk_ratelimit()) {
7260 printk(KERN_INFO "process %d (%s) no "
7261 "longer affine to cpu%d\n",
7262 task_pid_nr(p), p->comm, dead_cpu);
7267 /* It can have affinity changed while we were choosing. */
7268 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
7273 * While a dead CPU has no uninterruptible tasks queued at this point,
7274 * it might still have a nonzero ->nr_uninterruptible counter, because
7275 * for performance reasons the counter is not stricly tracking tasks to
7276 * their home CPUs. So we just add the counter to another CPU's counter,
7277 * to keep the global sum constant after CPU-down:
7279 static void migrate_nr_uninterruptible(struct rq *rq_src)
7281 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_online_mask));
7282 unsigned long flags;
7284 local_irq_save(flags);
7285 double_rq_lock(rq_src, rq_dest);
7286 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
7287 rq_src->nr_uninterruptible = 0;
7288 double_rq_unlock(rq_src, rq_dest);
7289 local_irq_restore(flags);
7292 /* Run through task list and migrate tasks from the dead cpu. */
7293 static void migrate_live_tasks(int src_cpu)
7295 struct task_struct *p, *t;
7297 read_lock(&tasklist_lock);
7299 do_each_thread(t, p) {
7303 if (task_cpu(p) == src_cpu)
7304 move_task_off_dead_cpu(src_cpu, p);
7305 } while_each_thread(t, p);
7307 read_unlock(&tasklist_lock);
7311 * Schedules idle task to be the next runnable task on current CPU.
7312 * It does so by boosting its priority to highest possible.
7313 * Used by CPU offline code.
7315 void sched_idle_next(void)
7317 int this_cpu = smp_processor_id();
7318 struct rq *rq = cpu_rq(this_cpu);
7319 struct task_struct *p = rq->idle;
7320 unsigned long flags;
7322 /* cpu has to be offline */
7323 BUG_ON(cpu_online(this_cpu));
7326 * Strictly not necessary since rest of the CPUs are stopped by now
7327 * and interrupts disabled on the current cpu.
7329 spin_lock_irqsave(&rq->lock, flags);
7331 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7333 update_rq_clock(rq);
7334 activate_task(rq, p, 0);
7336 spin_unlock_irqrestore(&rq->lock, flags);
7340 * Ensures that the idle task is using init_mm right before its cpu goes
7343 void idle_task_exit(void)
7345 struct mm_struct *mm = current->active_mm;
7347 BUG_ON(cpu_online(smp_processor_id()));
7350 switch_mm(mm, &init_mm, current);
7354 /* called under rq->lock with disabled interrupts */
7355 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
7357 struct rq *rq = cpu_rq(dead_cpu);
7359 /* Must be exiting, otherwise would be on tasklist. */
7360 BUG_ON(!p->exit_state);
7362 /* Cannot have done final schedule yet: would have vanished. */
7363 BUG_ON(p->state == TASK_DEAD);
7368 * Drop lock around migration; if someone else moves it,
7369 * that's OK. No task can be added to this CPU, so iteration is
7372 spin_unlock_irq(&rq->lock);
7373 move_task_off_dead_cpu(dead_cpu, p);
7374 spin_lock_irq(&rq->lock);
7379 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7380 static void migrate_dead_tasks(unsigned int dead_cpu)
7382 struct rq *rq = cpu_rq(dead_cpu);
7383 struct task_struct *next;
7386 if (!rq->nr_running)
7388 update_rq_clock(rq);
7389 next = pick_next_task(rq);
7392 next->sched_class->put_prev_task(rq, next);
7393 migrate_dead(dead_cpu, next);
7399 * remove the tasks which were accounted by rq from calc_load_tasks.
7401 static void calc_global_load_remove(struct rq *rq)
7403 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
7404 rq->calc_load_active = 0;
7406 #endif /* CONFIG_HOTPLUG_CPU */
7408 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7410 static struct ctl_table sd_ctl_dir[] = {
7412 .procname = "sched_domain",
7418 static struct ctl_table sd_ctl_root[] = {
7420 .ctl_name = CTL_KERN,
7421 .procname = "kernel",
7423 .child = sd_ctl_dir,
7428 static struct ctl_table *sd_alloc_ctl_entry(int n)
7430 struct ctl_table *entry =
7431 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
7436 static void sd_free_ctl_entry(struct ctl_table **tablep)
7438 struct ctl_table *entry;
7441 * In the intermediate directories, both the child directory and
7442 * procname are dynamically allocated and could fail but the mode
7443 * will always be set. In the lowest directory the names are
7444 * static strings and all have proc handlers.
7446 for (entry = *tablep; entry->mode; entry++) {
7448 sd_free_ctl_entry(&entry->child);
7449 if (entry->proc_handler == NULL)
7450 kfree(entry->procname);
7458 set_table_entry(struct ctl_table *entry,
7459 const char *procname, void *data, int maxlen,
7460 mode_t mode, proc_handler *proc_handler)
7462 entry->procname = procname;
7464 entry->maxlen = maxlen;
7466 entry->proc_handler = proc_handler;
7469 static struct ctl_table *
7470 sd_alloc_ctl_domain_table(struct sched_domain *sd)
7472 struct ctl_table *table = sd_alloc_ctl_entry(13);
7477 set_table_entry(&table[0], "min_interval", &sd->min_interval,
7478 sizeof(long), 0644, proc_doulongvec_minmax);
7479 set_table_entry(&table[1], "max_interval", &sd->max_interval,
7480 sizeof(long), 0644, proc_doulongvec_minmax);
7481 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
7482 sizeof(int), 0644, proc_dointvec_minmax);
7483 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
7484 sizeof(int), 0644, proc_dointvec_minmax);
7485 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
7486 sizeof(int), 0644, proc_dointvec_minmax);
7487 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
7488 sizeof(int), 0644, proc_dointvec_minmax);
7489 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
7490 sizeof(int), 0644, proc_dointvec_minmax);
7491 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
7492 sizeof(int), 0644, proc_dointvec_minmax);
7493 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
7494 sizeof(int), 0644, proc_dointvec_minmax);
7495 set_table_entry(&table[9], "cache_nice_tries",
7496 &sd->cache_nice_tries,
7497 sizeof(int), 0644, proc_dointvec_minmax);
7498 set_table_entry(&table[10], "flags", &sd->flags,
7499 sizeof(int), 0644, proc_dointvec_minmax);
7500 set_table_entry(&table[11], "name", sd->name,
7501 CORENAME_MAX_SIZE, 0444, proc_dostring);
7502 /* &table[12] is terminator */
7507 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
7509 struct ctl_table *entry, *table;
7510 struct sched_domain *sd;
7511 int domain_num = 0, i;
7514 for_each_domain(cpu, sd)
7516 entry = table = sd_alloc_ctl_entry(domain_num + 1);
7521 for_each_domain(cpu, sd) {
7522 snprintf(buf, 32, "domain%d", i);
7523 entry->procname = kstrdup(buf, GFP_KERNEL);
7525 entry->child = sd_alloc_ctl_domain_table(sd);
7532 static struct ctl_table_header *sd_sysctl_header;
7533 static void register_sched_domain_sysctl(void)
7535 int i, cpu_num = num_online_cpus();
7536 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
7539 WARN_ON(sd_ctl_dir[0].child);
7540 sd_ctl_dir[0].child = entry;
7545 for_each_online_cpu(i) {
7546 snprintf(buf, 32, "cpu%d", i);
7547 entry->procname = kstrdup(buf, GFP_KERNEL);
7549 entry->child = sd_alloc_ctl_cpu_table(i);
7553 WARN_ON(sd_sysctl_header);
7554 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
7557 /* may be called multiple times per register */
7558 static void unregister_sched_domain_sysctl(void)
7560 if (sd_sysctl_header)
7561 unregister_sysctl_table(sd_sysctl_header);
7562 sd_sysctl_header = NULL;
7563 if (sd_ctl_dir[0].child)
7564 sd_free_ctl_entry(&sd_ctl_dir[0].child);
7567 static void register_sched_domain_sysctl(void)
7570 static void unregister_sched_domain_sysctl(void)
7575 static void set_rq_online(struct rq *rq)
7578 const struct sched_class *class;
7580 cpumask_set_cpu(rq->cpu, rq->rd->online);
7583 for_each_class(class) {
7584 if (class->rq_online)
7585 class->rq_online(rq);
7590 static void set_rq_offline(struct rq *rq)
7593 const struct sched_class *class;
7595 for_each_class(class) {
7596 if (class->rq_offline)
7597 class->rq_offline(rq);
7600 cpumask_clear_cpu(rq->cpu, rq->rd->online);
7606 * migration_call - callback that gets triggered when a CPU is added.
7607 * Here we can start up the necessary migration thread for the new CPU.
7609 static int __cpuinit
7610 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
7612 struct task_struct *p;
7613 int cpu = (long)hcpu;
7614 unsigned long flags;
7619 case CPU_UP_PREPARE:
7620 case CPU_UP_PREPARE_FROZEN:
7621 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
7624 kthread_bind(p, cpu);
7625 /* Must be high prio: stop_machine expects to yield to it. */
7626 rq = task_rq_lock(p, &flags);
7627 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7628 task_rq_unlock(rq, &flags);
7630 cpu_rq(cpu)->migration_thread = p;
7631 rq->calc_load_update = calc_load_update;
7635 case CPU_ONLINE_FROZEN:
7636 /* Strictly unnecessary, as first user will wake it. */
7637 wake_up_process(cpu_rq(cpu)->migration_thread);
7639 /* Update our root-domain */
7641 spin_lock_irqsave(&rq->lock, flags);
7643 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7647 spin_unlock_irqrestore(&rq->lock, flags);
7650 #ifdef CONFIG_HOTPLUG_CPU
7651 case CPU_UP_CANCELED:
7652 case CPU_UP_CANCELED_FROZEN:
7653 if (!cpu_rq(cpu)->migration_thread)
7655 /* Unbind it from offline cpu so it can run. Fall thru. */
7656 kthread_bind(cpu_rq(cpu)->migration_thread,
7657 cpumask_any(cpu_online_mask));
7658 kthread_stop(cpu_rq(cpu)->migration_thread);
7659 put_task_struct(cpu_rq(cpu)->migration_thread);
7660 cpu_rq(cpu)->migration_thread = NULL;
7664 case CPU_DEAD_FROZEN:
7665 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7666 migrate_live_tasks(cpu);
7668 kthread_stop(rq->migration_thread);
7669 put_task_struct(rq->migration_thread);
7670 rq->migration_thread = NULL;
7671 /* Idle task back to normal (off runqueue, low prio) */
7672 spin_lock_irq(&rq->lock);
7673 update_rq_clock(rq);
7674 deactivate_task(rq, rq->idle, 0);
7675 rq->idle->static_prio = MAX_PRIO;
7676 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
7677 rq->idle->sched_class = &idle_sched_class;
7678 migrate_dead_tasks(cpu);
7679 spin_unlock_irq(&rq->lock);
7681 migrate_nr_uninterruptible(rq);
7682 BUG_ON(rq->nr_running != 0);
7683 calc_global_load_remove(rq);
7685 * No need to migrate the tasks: it was best-effort if
7686 * they didn't take sched_hotcpu_mutex. Just wake up
7689 spin_lock_irq(&rq->lock);
7690 while (!list_empty(&rq->migration_queue)) {
7691 struct migration_req *req;
7693 req = list_entry(rq->migration_queue.next,
7694 struct migration_req, list);
7695 list_del_init(&req->list);
7696 spin_unlock_irq(&rq->lock);
7697 complete(&req->done);
7698 spin_lock_irq(&rq->lock);
7700 spin_unlock_irq(&rq->lock);
7704 case CPU_DYING_FROZEN:
7705 /* Update our root-domain */
7707 spin_lock_irqsave(&rq->lock, flags);
7709 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7712 spin_unlock_irqrestore(&rq->lock, flags);
7720 * Register at high priority so that task migration (migrate_all_tasks)
7721 * happens before everything else. This has to be lower priority than
7722 * the notifier in the perf_event subsystem, though.
7724 static struct notifier_block __cpuinitdata migration_notifier = {
7725 .notifier_call = migration_call,
7729 static int __init migration_init(void)
7731 void *cpu = (void *)(long)smp_processor_id();
7734 /* Start one for the boot CPU: */
7735 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
7736 BUG_ON(err == NOTIFY_BAD);
7737 migration_call(&migration_notifier, CPU_ONLINE, cpu);
7738 register_cpu_notifier(&migration_notifier);
7742 early_initcall(migration_init);
7747 #ifdef CONFIG_SCHED_DEBUG
7749 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
7750 struct cpumask *groupmask)
7752 struct sched_group *group = sd->groups;
7755 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
7756 cpumask_clear(groupmask);
7758 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
7760 if (!(sd->flags & SD_LOAD_BALANCE)) {
7761 printk("does not load-balance\n");
7763 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
7768 printk(KERN_CONT "span %s level %s\n", str, sd->name);
7770 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
7771 printk(KERN_ERR "ERROR: domain->span does not contain "
7774 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
7775 printk(KERN_ERR "ERROR: domain->groups does not contain"
7779 printk(KERN_DEBUG "%*s groups:", level + 1, "");
7783 printk(KERN_ERR "ERROR: group is NULL\n");
7787 if (!group->cpu_power) {
7788 printk(KERN_CONT "\n");
7789 printk(KERN_ERR "ERROR: domain->cpu_power not "
7794 if (!cpumask_weight(sched_group_cpus(group))) {
7795 printk(KERN_CONT "\n");
7796 printk(KERN_ERR "ERROR: empty group\n");
7800 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
7801 printk(KERN_CONT "\n");
7802 printk(KERN_ERR "ERROR: repeated CPUs\n");
7806 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
7808 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
7810 printk(KERN_CONT " %s", str);
7811 if (group->cpu_power != SCHED_LOAD_SCALE) {
7812 printk(KERN_CONT " (cpu_power = %d)",
7816 group = group->next;
7817 } while (group != sd->groups);
7818 printk(KERN_CONT "\n");
7820 if (!cpumask_equal(sched_domain_span(sd), groupmask))
7821 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
7824 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
7825 printk(KERN_ERR "ERROR: parent span is not a superset "
7826 "of domain->span\n");
7830 static void sched_domain_debug(struct sched_domain *sd, int cpu)
7832 cpumask_var_t groupmask;
7836 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7840 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7842 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
7843 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
7848 if (sched_domain_debug_one(sd, cpu, level, groupmask))
7855 free_cpumask_var(groupmask);
7857 #else /* !CONFIG_SCHED_DEBUG */
7858 # define sched_domain_debug(sd, cpu) do { } while (0)
7859 #endif /* CONFIG_SCHED_DEBUG */
7861 static int sd_degenerate(struct sched_domain *sd)
7863 if (cpumask_weight(sched_domain_span(sd)) == 1)
7866 /* Following flags need at least 2 groups */
7867 if (sd->flags & (SD_LOAD_BALANCE |
7868 SD_BALANCE_NEWIDLE |
7872 SD_SHARE_PKG_RESOURCES)) {
7873 if (sd->groups != sd->groups->next)
7877 /* Following flags don't use groups */
7878 if (sd->flags & (SD_WAKE_AFFINE))
7885 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
7887 unsigned long cflags = sd->flags, pflags = parent->flags;
7889 if (sd_degenerate(parent))
7892 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
7895 /* Flags needing groups don't count if only 1 group in parent */
7896 if (parent->groups == parent->groups->next) {
7897 pflags &= ~(SD_LOAD_BALANCE |
7898 SD_BALANCE_NEWIDLE |
7902 SD_SHARE_PKG_RESOURCES);
7903 if (nr_node_ids == 1)
7904 pflags &= ~SD_SERIALIZE;
7906 if (~cflags & pflags)
7912 static void free_rootdomain(struct root_domain *rd)
7914 cpupri_cleanup(&rd->cpupri);
7916 free_cpumask_var(rd->rto_mask);
7917 free_cpumask_var(rd->online);
7918 free_cpumask_var(rd->span);
7922 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7924 struct root_domain *old_rd = NULL;
7925 unsigned long flags;
7927 spin_lock_irqsave(&rq->lock, flags);
7932 if (cpumask_test_cpu(rq->cpu, old_rd->online))
7935 cpumask_clear_cpu(rq->cpu, old_rd->span);
7938 * If we dont want to free the old_rt yet then
7939 * set old_rd to NULL to skip the freeing later
7942 if (!atomic_dec_and_test(&old_rd->refcount))
7946 atomic_inc(&rd->refcount);
7949 cpumask_set_cpu(rq->cpu, rd->span);
7950 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
7953 spin_unlock_irqrestore(&rq->lock, flags);
7956 free_rootdomain(old_rd);
7959 static int init_rootdomain(struct root_domain *rd, bool bootmem)
7961 gfp_t gfp = GFP_KERNEL;
7963 memset(rd, 0, sizeof(*rd));
7968 if (!alloc_cpumask_var(&rd->span, gfp))
7970 if (!alloc_cpumask_var(&rd->online, gfp))
7972 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
7975 if (cpupri_init(&rd->cpupri, bootmem) != 0)
7980 free_cpumask_var(rd->rto_mask);
7982 free_cpumask_var(rd->online);
7984 free_cpumask_var(rd->span);
7989 static void init_defrootdomain(void)
7991 init_rootdomain(&def_root_domain, true);
7993 atomic_set(&def_root_domain.refcount, 1);
7996 static struct root_domain *alloc_rootdomain(void)
7998 struct root_domain *rd;
8000 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
8004 if (init_rootdomain(rd, false) != 0) {
8013 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
8014 * hold the hotplug lock.
8017 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
8019 struct rq *rq = cpu_rq(cpu);
8020 struct sched_domain *tmp;
8022 /* Remove the sched domains which do not contribute to scheduling. */
8023 for (tmp = sd; tmp; ) {
8024 struct sched_domain *parent = tmp->parent;
8028 if (sd_parent_degenerate(tmp, parent)) {
8029 tmp->parent = parent->parent;
8031 parent->parent->child = tmp;
8036 if (sd && sd_degenerate(sd)) {
8042 sched_domain_debug(sd, cpu);
8044 rq_attach_root(rq, rd);
8045 rcu_assign_pointer(rq->sd, sd);
8048 /* cpus with isolated domains */
8049 static cpumask_var_t cpu_isolated_map;
8051 /* Setup the mask of cpus configured for isolated domains */
8052 static int __init isolated_cpu_setup(char *str)
8054 cpulist_parse(str, cpu_isolated_map);
8058 __setup("isolcpus=", isolated_cpu_setup);
8061 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
8062 * to a function which identifies what group(along with sched group) a CPU
8063 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
8064 * (due to the fact that we keep track of groups covered with a struct cpumask).
8066 * init_sched_build_groups will build a circular linked list of the groups
8067 * covered by the given span, and will set each group's ->cpumask correctly,
8068 * and ->cpu_power to 0.
8071 init_sched_build_groups(const struct cpumask *span,
8072 const struct cpumask *cpu_map,
8073 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
8074 struct sched_group **sg,
8075 struct cpumask *tmpmask),
8076 struct cpumask *covered, struct cpumask *tmpmask)
8078 struct sched_group *first = NULL, *last = NULL;
8081 cpumask_clear(covered);
8083 for_each_cpu(i, span) {
8084 struct sched_group *sg;
8085 int group = group_fn(i, cpu_map, &sg, tmpmask);
8088 if (cpumask_test_cpu(i, covered))
8091 cpumask_clear(sched_group_cpus(sg));
8094 for_each_cpu(j, span) {
8095 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
8098 cpumask_set_cpu(j, covered);
8099 cpumask_set_cpu(j, sched_group_cpus(sg));
8110 #define SD_NODES_PER_DOMAIN 16
8115 * find_next_best_node - find the next node to include in a sched_domain
8116 * @node: node whose sched_domain we're building
8117 * @used_nodes: nodes already in the sched_domain
8119 * Find the next node to include in a given scheduling domain. Simply
8120 * finds the closest node not already in the @used_nodes map.
8122 * Should use nodemask_t.
8124 static int find_next_best_node(int node, nodemask_t *used_nodes)
8126 int i, n, val, min_val, best_node = 0;
8130 for (i = 0; i < nr_node_ids; i++) {
8131 /* Start at @node */
8132 n = (node + i) % nr_node_ids;
8134 if (!nr_cpus_node(n))
8137 /* Skip already used nodes */
8138 if (node_isset(n, *used_nodes))
8141 /* Simple min distance search */
8142 val = node_distance(node, n);
8144 if (val < min_val) {
8150 node_set(best_node, *used_nodes);
8155 * sched_domain_node_span - get a cpumask for a node's sched_domain
8156 * @node: node whose cpumask we're constructing
8157 * @span: resulting cpumask
8159 * Given a node, construct a good cpumask for its sched_domain to span. It
8160 * should be one that prevents unnecessary balancing, but also spreads tasks
8163 static void sched_domain_node_span(int node, struct cpumask *span)
8165 nodemask_t used_nodes;
8168 cpumask_clear(span);
8169 nodes_clear(used_nodes);
8171 cpumask_or(span, span, cpumask_of_node(node));
8172 node_set(node, used_nodes);
8174 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
8175 int next_node = find_next_best_node(node, &used_nodes);
8177 cpumask_or(span, span, cpumask_of_node(next_node));
8180 #endif /* CONFIG_NUMA */
8182 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
8185 * The cpus mask in sched_group and sched_domain hangs off the end.
8187 * ( See the the comments in include/linux/sched.h:struct sched_group
8188 * and struct sched_domain. )
8190 struct static_sched_group {
8191 struct sched_group sg;
8192 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
8195 struct static_sched_domain {
8196 struct sched_domain sd;
8197 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
8203 cpumask_var_t domainspan;
8204 cpumask_var_t covered;
8205 cpumask_var_t notcovered;
8207 cpumask_var_t nodemask;
8208 cpumask_var_t this_sibling_map;
8209 cpumask_var_t this_core_map;
8210 cpumask_var_t send_covered;
8211 cpumask_var_t tmpmask;
8212 struct sched_group **sched_group_nodes;
8213 struct root_domain *rd;
8217 sa_sched_groups = 0,
8222 sa_this_sibling_map,
8224 sa_sched_group_nodes,
8234 * SMT sched-domains:
8236 #ifdef CONFIG_SCHED_SMT
8237 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
8238 static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
8241 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
8242 struct sched_group **sg, struct cpumask *unused)
8245 *sg = &per_cpu(sched_group_cpus, cpu).sg;
8248 #endif /* CONFIG_SCHED_SMT */
8251 * multi-core sched-domains:
8253 #ifdef CONFIG_SCHED_MC
8254 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
8255 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
8256 #endif /* CONFIG_SCHED_MC */
8258 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8260 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8261 struct sched_group **sg, struct cpumask *mask)
8265 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8266 group = cpumask_first(mask);
8268 *sg = &per_cpu(sched_group_core, group).sg;
8271 #elif defined(CONFIG_SCHED_MC)
8273 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8274 struct sched_group **sg, struct cpumask *unused)
8277 *sg = &per_cpu(sched_group_core, cpu).sg;
8282 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
8283 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
8286 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
8287 struct sched_group **sg, struct cpumask *mask)
8290 #ifdef CONFIG_SCHED_MC
8291 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
8292 group = cpumask_first(mask);
8293 #elif defined(CONFIG_SCHED_SMT)
8294 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8295 group = cpumask_first(mask);
8300 *sg = &per_cpu(sched_group_phys, group).sg;
8306 * The init_sched_build_groups can't handle what we want to do with node
8307 * groups, so roll our own. Now each node has its own list of groups which
8308 * gets dynamically allocated.
8310 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
8311 static struct sched_group ***sched_group_nodes_bycpu;
8313 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
8314 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
8316 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
8317 struct sched_group **sg,
8318 struct cpumask *nodemask)
8322 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
8323 group = cpumask_first(nodemask);
8326 *sg = &per_cpu(sched_group_allnodes, group).sg;
8330 static void init_numa_sched_groups_power(struct sched_group *group_head)
8332 struct sched_group *sg = group_head;
8338 for_each_cpu(j, sched_group_cpus(sg)) {
8339 struct sched_domain *sd;
8341 sd = &per_cpu(phys_domains, j).sd;
8342 if (j != group_first_cpu(sd->groups)) {
8344 * Only add "power" once for each
8350 sg->cpu_power += sd->groups->cpu_power;
8353 } while (sg != group_head);
8356 static int build_numa_sched_groups(struct s_data *d,
8357 const struct cpumask *cpu_map, int num)
8359 struct sched_domain *sd;
8360 struct sched_group *sg, *prev;
8363 cpumask_clear(d->covered);
8364 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
8365 if (cpumask_empty(d->nodemask)) {
8366 d->sched_group_nodes[num] = NULL;
8370 sched_domain_node_span(num, d->domainspan);
8371 cpumask_and(d->domainspan, d->domainspan, cpu_map);
8373 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8376 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
8380 d->sched_group_nodes[num] = sg;
8382 for_each_cpu(j, d->nodemask) {
8383 sd = &per_cpu(node_domains, j).sd;
8388 cpumask_copy(sched_group_cpus(sg), d->nodemask);
8390 cpumask_or(d->covered, d->covered, d->nodemask);
8393 for (j = 0; j < nr_node_ids; j++) {
8394 n = (num + j) % nr_node_ids;
8395 cpumask_complement(d->notcovered, d->covered);
8396 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
8397 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
8398 if (cpumask_empty(d->tmpmask))
8400 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
8401 if (cpumask_empty(d->tmpmask))
8403 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8407 "Can not alloc domain group for node %d\n", j);
8411 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
8412 sg->next = prev->next;
8413 cpumask_or(d->covered, d->covered, d->tmpmask);
8420 #endif /* CONFIG_NUMA */
8423 /* Free memory allocated for various sched_group structures */
8424 static void free_sched_groups(const struct cpumask *cpu_map,
8425 struct cpumask *nodemask)
8429 for_each_cpu(cpu, cpu_map) {
8430 struct sched_group **sched_group_nodes
8431 = sched_group_nodes_bycpu[cpu];
8433 if (!sched_group_nodes)
8436 for (i = 0; i < nr_node_ids; i++) {
8437 struct sched_group *oldsg, *sg = sched_group_nodes[i];
8439 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8440 if (cpumask_empty(nodemask))
8450 if (oldsg != sched_group_nodes[i])
8453 kfree(sched_group_nodes);
8454 sched_group_nodes_bycpu[cpu] = NULL;
8457 #else /* !CONFIG_NUMA */
8458 static void free_sched_groups(const struct cpumask *cpu_map,
8459 struct cpumask *nodemask)
8462 #endif /* CONFIG_NUMA */
8465 * Initialize sched groups cpu_power.
8467 * cpu_power indicates the capacity of sched group, which is used while
8468 * distributing the load between different sched groups in a sched domain.
8469 * Typically cpu_power for all the groups in a sched domain will be same unless
8470 * there are asymmetries in the topology. If there are asymmetries, group
8471 * having more cpu_power will pickup more load compared to the group having
8474 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
8476 struct sched_domain *child;
8477 struct sched_group *group;
8481 WARN_ON(!sd || !sd->groups);
8483 if (cpu != group_first_cpu(sd->groups))
8488 sd->groups->cpu_power = 0;
8491 power = SCHED_LOAD_SCALE;
8492 weight = cpumask_weight(sched_domain_span(sd));
8494 * SMT siblings share the power of a single core.
8495 * Usually multiple threads get a better yield out of
8496 * that one core than a single thread would have,
8497 * reflect that in sd->smt_gain.
8499 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
8500 power *= sd->smt_gain;
8502 power >>= SCHED_LOAD_SHIFT;
8504 sd->groups->cpu_power += power;
8509 * Add cpu_power of each child group to this groups cpu_power.
8511 group = child->groups;
8513 sd->groups->cpu_power += group->cpu_power;
8514 group = group->next;
8515 } while (group != child->groups);
8519 * Initializers for schedule domains
8520 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8523 #ifdef CONFIG_SCHED_DEBUG
8524 # define SD_INIT_NAME(sd, type) sd->name = #type
8526 # define SD_INIT_NAME(sd, type) do { } while (0)
8529 #define SD_INIT(sd, type) sd_init_##type(sd)
8531 #define SD_INIT_FUNC(type) \
8532 static noinline void sd_init_##type(struct sched_domain *sd) \
8534 memset(sd, 0, sizeof(*sd)); \
8535 *sd = SD_##type##_INIT; \
8536 sd->level = SD_LV_##type; \
8537 SD_INIT_NAME(sd, type); \
8542 SD_INIT_FUNC(ALLNODES)
8545 #ifdef CONFIG_SCHED_SMT
8546 SD_INIT_FUNC(SIBLING)
8548 #ifdef CONFIG_SCHED_MC
8552 static int default_relax_domain_level = -1;
8554 static int __init setup_relax_domain_level(char *str)
8558 val = simple_strtoul(str, NULL, 0);
8559 if (val < SD_LV_MAX)
8560 default_relax_domain_level = val;
8564 __setup("relax_domain_level=", setup_relax_domain_level);
8566 static void set_domain_attribute(struct sched_domain *sd,
8567 struct sched_domain_attr *attr)
8571 if (!attr || attr->relax_domain_level < 0) {
8572 if (default_relax_domain_level < 0)
8575 request = default_relax_domain_level;
8577 request = attr->relax_domain_level;
8578 if (request < sd->level) {
8579 /* turn off idle balance on this domain */
8580 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
8582 /* turn on idle balance on this domain */
8583 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
8587 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
8588 const struct cpumask *cpu_map)
8591 case sa_sched_groups:
8592 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
8593 d->sched_group_nodes = NULL;
8595 free_rootdomain(d->rd); /* fall through */
8597 free_cpumask_var(d->tmpmask); /* fall through */
8598 case sa_send_covered:
8599 free_cpumask_var(d->send_covered); /* fall through */
8600 case sa_this_core_map:
8601 free_cpumask_var(d->this_core_map); /* fall through */
8602 case sa_this_sibling_map:
8603 free_cpumask_var(d->this_sibling_map); /* fall through */
8605 free_cpumask_var(d->nodemask); /* fall through */
8606 case sa_sched_group_nodes:
8608 kfree(d->sched_group_nodes); /* fall through */
8610 free_cpumask_var(d->notcovered); /* fall through */
8612 free_cpumask_var(d->covered); /* fall through */
8614 free_cpumask_var(d->domainspan); /* fall through */
8621 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
8622 const struct cpumask *cpu_map)
8625 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
8627 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
8628 return sa_domainspan;
8629 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
8631 /* Allocate the per-node list of sched groups */
8632 d->sched_group_nodes = kcalloc(nr_node_ids,
8633 sizeof(struct sched_group *), GFP_KERNEL);
8634 if (!d->sched_group_nodes) {
8635 printk(KERN_WARNING "Can not alloc sched group node list\n");
8636 return sa_notcovered;
8638 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
8640 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
8641 return sa_sched_group_nodes;
8642 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
8644 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
8645 return sa_this_sibling_map;
8646 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
8647 return sa_this_core_map;
8648 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
8649 return sa_send_covered;
8650 d->rd = alloc_rootdomain();
8652 printk(KERN_WARNING "Cannot alloc root domain\n");
8655 return sa_rootdomain;
8658 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
8659 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
8661 struct sched_domain *sd = NULL;
8663 struct sched_domain *parent;
8666 if (cpumask_weight(cpu_map) >
8667 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
8668 sd = &per_cpu(allnodes_domains, i).sd;
8669 SD_INIT(sd, ALLNODES);
8670 set_domain_attribute(sd, attr);
8671 cpumask_copy(sched_domain_span(sd), cpu_map);
8672 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
8677 sd = &per_cpu(node_domains, i).sd;
8679 set_domain_attribute(sd, attr);
8680 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
8681 sd->parent = parent;
8684 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
8689 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
8690 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8691 struct sched_domain *parent, int i)
8693 struct sched_domain *sd;
8694 sd = &per_cpu(phys_domains, i).sd;
8696 set_domain_attribute(sd, attr);
8697 cpumask_copy(sched_domain_span(sd), d->nodemask);
8698 sd->parent = parent;
8701 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
8705 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
8706 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8707 struct sched_domain *parent, int i)
8709 struct sched_domain *sd = parent;
8710 #ifdef CONFIG_SCHED_MC
8711 sd = &per_cpu(core_domains, i).sd;
8713 set_domain_attribute(sd, attr);
8714 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
8715 sd->parent = parent;
8717 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
8722 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
8723 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8724 struct sched_domain *parent, int i)
8726 struct sched_domain *sd = parent;
8727 #ifdef CONFIG_SCHED_SMT
8728 sd = &per_cpu(cpu_domains, i).sd;
8729 SD_INIT(sd, SIBLING);
8730 set_domain_attribute(sd, attr);
8731 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
8732 sd->parent = parent;
8734 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
8739 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
8740 const struct cpumask *cpu_map, int cpu)
8743 #ifdef CONFIG_SCHED_SMT
8744 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
8745 cpumask_and(d->this_sibling_map, cpu_map,
8746 topology_thread_cpumask(cpu));
8747 if (cpu == cpumask_first(d->this_sibling_map))
8748 init_sched_build_groups(d->this_sibling_map, cpu_map,
8750 d->send_covered, d->tmpmask);
8753 #ifdef CONFIG_SCHED_MC
8754 case SD_LV_MC: /* set up multi-core groups */
8755 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
8756 if (cpu == cpumask_first(d->this_core_map))
8757 init_sched_build_groups(d->this_core_map, cpu_map,
8759 d->send_covered, d->tmpmask);
8762 case SD_LV_CPU: /* set up physical groups */
8763 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
8764 if (!cpumask_empty(d->nodemask))
8765 init_sched_build_groups(d->nodemask, cpu_map,
8767 d->send_covered, d->tmpmask);
8770 case SD_LV_ALLNODES:
8771 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
8772 d->send_covered, d->tmpmask);
8781 * Build sched domains for a given set of cpus and attach the sched domains
8782 * to the individual cpus
8784 static int __build_sched_domains(const struct cpumask *cpu_map,
8785 struct sched_domain_attr *attr)
8787 enum s_alloc alloc_state = sa_none;
8789 struct sched_domain *sd;
8795 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
8796 if (alloc_state != sa_rootdomain)
8798 alloc_state = sa_sched_groups;
8801 * Set up domains for cpus specified by the cpu_map.
8803 for_each_cpu(i, cpu_map) {
8804 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
8807 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
8808 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
8809 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
8810 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
8813 for_each_cpu(i, cpu_map) {
8814 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
8815 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
8818 /* Set up physical groups */
8819 for (i = 0; i < nr_node_ids; i++)
8820 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
8823 /* Set up node groups */
8825 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
8827 for (i = 0; i < nr_node_ids; i++)
8828 if (build_numa_sched_groups(&d, cpu_map, i))
8832 /* Calculate CPU power for physical packages and nodes */
8833 #ifdef CONFIG_SCHED_SMT
8834 for_each_cpu(i, cpu_map) {
8835 sd = &per_cpu(cpu_domains, i).sd;
8836 init_sched_groups_power(i, sd);
8839 #ifdef CONFIG_SCHED_MC
8840 for_each_cpu(i, cpu_map) {
8841 sd = &per_cpu(core_domains, i).sd;
8842 init_sched_groups_power(i, sd);
8846 for_each_cpu(i, cpu_map) {
8847 sd = &per_cpu(phys_domains, i).sd;
8848 init_sched_groups_power(i, sd);
8852 for (i = 0; i < nr_node_ids; i++)
8853 init_numa_sched_groups_power(d.sched_group_nodes[i]);
8855 if (d.sd_allnodes) {
8856 struct sched_group *sg;
8858 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
8860 init_numa_sched_groups_power(sg);
8864 /* Attach the domains */
8865 for_each_cpu(i, cpu_map) {
8866 #ifdef CONFIG_SCHED_SMT
8867 sd = &per_cpu(cpu_domains, i).sd;
8868 #elif defined(CONFIG_SCHED_MC)
8869 sd = &per_cpu(core_domains, i).sd;
8871 sd = &per_cpu(phys_domains, i).sd;
8873 cpu_attach_domain(sd, d.rd, i);
8876 d.sched_group_nodes = NULL; /* don't free this we still need it */
8877 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
8881 __free_domain_allocs(&d, alloc_state, cpu_map);
8885 static int build_sched_domains(const struct cpumask *cpu_map)
8887 return __build_sched_domains(cpu_map, NULL);
8890 static struct cpumask *doms_cur; /* current sched domains */
8891 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
8892 static struct sched_domain_attr *dattr_cur;
8893 /* attribues of custom domains in 'doms_cur' */
8896 * Special case: If a kmalloc of a doms_cur partition (array of
8897 * cpumask) fails, then fallback to a single sched domain,
8898 * as determined by the single cpumask fallback_doms.
8900 static cpumask_var_t fallback_doms;
8903 * arch_update_cpu_topology lets virtualized architectures update the
8904 * cpu core maps. It is supposed to return 1 if the topology changed
8905 * or 0 if it stayed the same.
8907 int __attribute__((weak)) arch_update_cpu_topology(void)
8913 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
8914 * For now this just excludes isolated cpus, but could be used to
8915 * exclude other special cases in the future.
8917 static int arch_init_sched_domains(const struct cpumask *cpu_map)
8921 arch_update_cpu_topology();
8923 doms_cur = kmalloc(cpumask_size(), GFP_KERNEL);
8925 doms_cur = fallback_doms;
8926 cpumask_andnot(doms_cur, cpu_map, cpu_isolated_map);
8928 err = build_sched_domains(doms_cur);
8929 register_sched_domain_sysctl();
8934 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
8935 struct cpumask *tmpmask)
8937 free_sched_groups(cpu_map, tmpmask);
8941 * Detach sched domains from a group of cpus specified in cpu_map
8942 * These cpus will now be attached to the NULL domain
8944 static void detach_destroy_domains(const struct cpumask *cpu_map)
8946 /* Save because hotplug lock held. */
8947 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
8950 for_each_cpu(i, cpu_map)
8951 cpu_attach_domain(NULL, &def_root_domain, i);
8952 synchronize_sched();
8953 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
8956 /* handle null as "default" */
8957 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
8958 struct sched_domain_attr *new, int idx_new)
8960 struct sched_domain_attr tmp;
8967 return !memcmp(cur ? (cur + idx_cur) : &tmp,
8968 new ? (new + idx_new) : &tmp,
8969 sizeof(struct sched_domain_attr));
8973 * Partition sched domains as specified by the 'ndoms_new'
8974 * cpumasks in the array doms_new[] of cpumasks. This compares
8975 * doms_new[] to the current sched domain partitioning, doms_cur[].
8976 * It destroys each deleted domain and builds each new domain.
8978 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
8979 * The masks don't intersect (don't overlap.) We should setup one
8980 * sched domain for each mask. CPUs not in any of the cpumasks will
8981 * not be load balanced. If the same cpumask appears both in the
8982 * current 'doms_cur' domains and in the new 'doms_new', we can leave
8985 * The passed in 'doms_new' should be kmalloc'd. This routine takes
8986 * ownership of it and will kfree it when done with it. If the caller
8987 * failed the kmalloc call, then it can pass in doms_new == NULL &&
8988 * ndoms_new == 1, and partition_sched_domains() will fallback to
8989 * the single partition 'fallback_doms', it also forces the domains
8992 * If doms_new == NULL it will be replaced with cpu_online_mask.
8993 * ndoms_new == 0 is a special case for destroying existing domains,
8994 * and it will not create the default domain.
8996 * Call with hotplug lock held
8998 /* FIXME: Change to struct cpumask *doms_new[] */
8999 void partition_sched_domains(int ndoms_new, struct cpumask *doms_new,
9000 struct sched_domain_attr *dattr_new)
9005 mutex_lock(&sched_domains_mutex);
9007 /* always unregister in case we don't destroy any domains */
9008 unregister_sched_domain_sysctl();
9010 /* Let architecture update cpu core mappings. */
9011 new_topology = arch_update_cpu_topology();
9013 n = doms_new ? ndoms_new : 0;
9015 /* Destroy deleted domains */
9016 for (i = 0; i < ndoms_cur; i++) {
9017 for (j = 0; j < n && !new_topology; j++) {
9018 if (cpumask_equal(&doms_cur[i], &doms_new[j])
9019 && dattrs_equal(dattr_cur, i, dattr_new, j))
9022 /* no match - a current sched domain not in new doms_new[] */
9023 detach_destroy_domains(doms_cur + i);
9028 if (doms_new == NULL) {
9030 doms_new = fallback_doms;
9031 cpumask_andnot(&doms_new[0], cpu_online_mask, cpu_isolated_map);
9032 WARN_ON_ONCE(dattr_new);
9035 /* Build new domains */
9036 for (i = 0; i < ndoms_new; i++) {
9037 for (j = 0; j < ndoms_cur && !new_topology; j++) {
9038 if (cpumask_equal(&doms_new[i], &doms_cur[j])
9039 && dattrs_equal(dattr_new, i, dattr_cur, j))
9042 /* no match - add a new doms_new */
9043 __build_sched_domains(doms_new + i,
9044 dattr_new ? dattr_new + i : NULL);
9049 /* Remember the new sched domains */
9050 if (doms_cur != fallback_doms)
9052 kfree(dattr_cur); /* kfree(NULL) is safe */
9053 doms_cur = doms_new;
9054 dattr_cur = dattr_new;
9055 ndoms_cur = ndoms_new;
9057 register_sched_domain_sysctl();
9059 mutex_unlock(&sched_domains_mutex);
9062 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
9063 static void arch_reinit_sched_domains(void)
9067 /* Destroy domains first to force the rebuild */
9068 partition_sched_domains(0, NULL, NULL);
9070 rebuild_sched_domains();
9074 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
9076 unsigned int level = 0;
9078 if (sscanf(buf, "%u", &level) != 1)
9082 * level is always be positive so don't check for
9083 * level < POWERSAVINGS_BALANCE_NONE which is 0
9084 * What happens on 0 or 1 byte write,
9085 * need to check for count as well?
9088 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
9092 sched_smt_power_savings = level;
9094 sched_mc_power_savings = level;
9096 arch_reinit_sched_domains();
9101 #ifdef CONFIG_SCHED_MC
9102 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
9105 return sprintf(page, "%u\n", sched_mc_power_savings);
9107 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
9108 const char *buf, size_t count)
9110 return sched_power_savings_store(buf, count, 0);
9112 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
9113 sched_mc_power_savings_show,
9114 sched_mc_power_savings_store);
9117 #ifdef CONFIG_SCHED_SMT
9118 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
9121 return sprintf(page, "%u\n", sched_smt_power_savings);
9123 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
9124 const char *buf, size_t count)
9126 return sched_power_savings_store(buf, count, 1);
9128 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
9129 sched_smt_power_savings_show,
9130 sched_smt_power_savings_store);
9133 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
9137 #ifdef CONFIG_SCHED_SMT
9139 err = sysfs_create_file(&cls->kset.kobj,
9140 &attr_sched_smt_power_savings.attr);
9142 #ifdef CONFIG_SCHED_MC
9143 if (!err && mc_capable())
9144 err = sysfs_create_file(&cls->kset.kobj,
9145 &attr_sched_mc_power_savings.attr);
9149 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
9151 #ifndef CONFIG_CPUSETS
9153 * Add online and remove offline CPUs from the scheduler domains.
9154 * When cpusets are enabled they take over this function.
9156 static int update_sched_domains(struct notifier_block *nfb,
9157 unsigned long action, void *hcpu)
9161 case CPU_ONLINE_FROZEN:
9163 case CPU_DEAD_FROZEN:
9164 partition_sched_domains(1, NULL, NULL);
9173 static int update_runtime(struct notifier_block *nfb,
9174 unsigned long action, void *hcpu)
9176 int cpu = (int)(long)hcpu;
9179 case CPU_DOWN_PREPARE:
9180 case CPU_DOWN_PREPARE_FROZEN:
9181 disable_runtime(cpu_rq(cpu));
9184 case CPU_DOWN_FAILED:
9185 case CPU_DOWN_FAILED_FROZEN:
9187 case CPU_ONLINE_FROZEN:
9188 enable_runtime(cpu_rq(cpu));
9196 void __init sched_init_smp(void)
9198 cpumask_var_t non_isolated_cpus;
9200 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
9201 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
9203 #if defined(CONFIG_NUMA)
9204 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
9206 BUG_ON(sched_group_nodes_bycpu == NULL);
9209 mutex_lock(&sched_domains_mutex);
9210 arch_init_sched_domains(cpu_online_mask);
9211 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
9212 if (cpumask_empty(non_isolated_cpus))
9213 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
9214 mutex_unlock(&sched_domains_mutex);
9217 #ifndef CONFIG_CPUSETS
9218 /* XXX: Theoretical race here - CPU may be hotplugged now */
9219 hotcpu_notifier(update_sched_domains, 0);
9222 /* RT runtime code needs to handle some hotplug events */
9223 hotcpu_notifier(update_runtime, 0);
9227 /* Move init over to a non-isolated CPU */
9228 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
9230 sched_init_granularity();
9231 free_cpumask_var(non_isolated_cpus);
9233 init_sched_rt_class();
9236 void __init sched_init_smp(void)
9238 sched_init_granularity();
9240 #endif /* CONFIG_SMP */
9242 const_debug unsigned int sysctl_timer_migration = 1;
9244 int in_sched_functions(unsigned long addr)
9246 return in_lock_functions(addr) ||
9247 (addr >= (unsigned long)__sched_text_start
9248 && addr < (unsigned long)__sched_text_end);
9251 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
9253 cfs_rq->tasks_timeline = RB_ROOT;
9254 INIT_LIST_HEAD(&cfs_rq->tasks);
9255 #ifdef CONFIG_FAIR_GROUP_SCHED
9258 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9261 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
9263 struct rt_prio_array *array;
9266 array = &rt_rq->active;
9267 for (i = 0; i < MAX_RT_PRIO; i++) {
9268 INIT_LIST_HEAD(array->queue + i);
9269 __clear_bit(i, array->bitmap);
9271 /* delimiter for bitsearch: */
9272 __set_bit(MAX_RT_PRIO, array->bitmap);
9274 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9275 rt_rq->highest_prio.curr = MAX_RT_PRIO;
9277 rt_rq->highest_prio.next = MAX_RT_PRIO;
9281 rt_rq->rt_nr_migratory = 0;
9282 rt_rq->overloaded = 0;
9283 plist_head_init(&rt_rq->pushable_tasks, &rq->lock);
9287 rt_rq->rt_throttled = 0;
9288 rt_rq->rt_runtime = 0;
9289 spin_lock_init(&rt_rq->rt_runtime_lock);
9291 #ifdef CONFIG_RT_GROUP_SCHED
9292 rt_rq->rt_nr_boosted = 0;
9297 #ifdef CONFIG_FAIR_GROUP_SCHED
9298 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9299 struct sched_entity *se, int cpu, int add,
9300 struct sched_entity *parent)
9302 struct rq *rq = cpu_rq(cpu);
9303 tg->cfs_rq[cpu] = cfs_rq;
9304 init_cfs_rq(cfs_rq, rq);
9307 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
9310 /* se could be NULL for init_task_group */
9315 se->cfs_rq = &rq->cfs;
9317 se->cfs_rq = parent->my_q;
9320 se->load.weight = tg->shares;
9321 se->load.inv_weight = 0;
9322 se->parent = parent;
9326 #ifdef CONFIG_RT_GROUP_SCHED
9327 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
9328 struct sched_rt_entity *rt_se, int cpu, int add,
9329 struct sched_rt_entity *parent)
9331 struct rq *rq = cpu_rq(cpu);
9333 tg->rt_rq[cpu] = rt_rq;
9334 init_rt_rq(rt_rq, rq);
9336 rt_rq->rt_se = rt_se;
9337 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
9339 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
9341 tg->rt_se[cpu] = rt_se;
9346 rt_se->rt_rq = &rq->rt;
9348 rt_se->rt_rq = parent->my_q;
9350 rt_se->my_q = rt_rq;
9351 rt_se->parent = parent;
9352 INIT_LIST_HEAD(&rt_se->run_list);
9356 void __init sched_init(void)
9359 unsigned long alloc_size = 0, ptr;
9361 #ifdef CONFIG_FAIR_GROUP_SCHED
9362 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9364 #ifdef CONFIG_RT_GROUP_SCHED
9365 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9367 #ifdef CONFIG_USER_SCHED
9370 #ifdef CONFIG_CPUMASK_OFFSTACK
9371 alloc_size += num_possible_cpus() * cpumask_size();
9374 * As sched_init() is called before page_alloc is setup,
9375 * we use alloc_bootmem().
9378 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
9380 #ifdef CONFIG_FAIR_GROUP_SCHED
9381 init_task_group.se = (struct sched_entity **)ptr;
9382 ptr += nr_cpu_ids * sizeof(void **);
9384 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
9385 ptr += nr_cpu_ids * sizeof(void **);
9387 #ifdef CONFIG_USER_SCHED
9388 root_task_group.se = (struct sched_entity **)ptr;
9389 ptr += nr_cpu_ids * sizeof(void **);
9391 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9392 ptr += nr_cpu_ids * sizeof(void **);
9393 #endif /* CONFIG_USER_SCHED */
9394 #endif /* CONFIG_FAIR_GROUP_SCHED */
9395 #ifdef CONFIG_RT_GROUP_SCHED
9396 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
9397 ptr += nr_cpu_ids * sizeof(void **);
9399 init_task_group.rt_rq = (struct rt_rq **)ptr;
9400 ptr += nr_cpu_ids * sizeof(void **);
9402 #ifdef CONFIG_USER_SCHED
9403 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9404 ptr += nr_cpu_ids * sizeof(void **);
9406 root_task_group.rt_rq = (struct rt_rq **)ptr;
9407 ptr += nr_cpu_ids * sizeof(void **);
9408 #endif /* CONFIG_USER_SCHED */
9409 #endif /* CONFIG_RT_GROUP_SCHED */
9410 #ifdef CONFIG_CPUMASK_OFFSTACK
9411 for_each_possible_cpu(i) {
9412 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
9413 ptr += cpumask_size();
9415 #endif /* CONFIG_CPUMASK_OFFSTACK */
9419 init_defrootdomain();
9422 init_rt_bandwidth(&def_rt_bandwidth,
9423 global_rt_period(), global_rt_runtime());
9425 #ifdef CONFIG_RT_GROUP_SCHED
9426 init_rt_bandwidth(&init_task_group.rt_bandwidth,
9427 global_rt_period(), global_rt_runtime());
9428 #ifdef CONFIG_USER_SCHED
9429 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9430 global_rt_period(), RUNTIME_INF);
9431 #endif /* CONFIG_USER_SCHED */
9432 #endif /* CONFIG_RT_GROUP_SCHED */
9434 #ifdef CONFIG_GROUP_SCHED
9435 list_add(&init_task_group.list, &task_groups);
9436 INIT_LIST_HEAD(&init_task_group.children);
9438 #ifdef CONFIG_USER_SCHED
9439 INIT_LIST_HEAD(&root_task_group.children);
9440 init_task_group.parent = &root_task_group;
9441 list_add(&init_task_group.siblings, &root_task_group.children);
9442 #endif /* CONFIG_USER_SCHED */
9443 #endif /* CONFIG_GROUP_SCHED */
9445 for_each_possible_cpu(i) {
9449 spin_lock_init(&rq->lock);
9451 rq->calc_load_active = 0;
9452 rq->calc_load_update = jiffies + LOAD_FREQ;
9453 init_cfs_rq(&rq->cfs, rq);
9454 init_rt_rq(&rq->rt, rq);
9455 #ifdef CONFIG_FAIR_GROUP_SCHED
9456 init_task_group.shares = init_task_group_load;
9457 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9458 #ifdef CONFIG_CGROUP_SCHED
9460 * How much cpu bandwidth does init_task_group get?
9462 * In case of task-groups formed thr' the cgroup filesystem, it
9463 * gets 100% of the cpu resources in the system. This overall
9464 * system cpu resource is divided among the tasks of
9465 * init_task_group and its child task-groups in a fair manner,
9466 * based on each entity's (task or task-group's) weight
9467 * (se->load.weight).
9469 * In other words, if init_task_group has 10 tasks of weight
9470 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9471 * then A0's share of the cpu resource is:
9473 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9475 * We achieve this by letting init_task_group's tasks sit
9476 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9478 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
9479 #elif defined CONFIG_USER_SCHED
9480 root_task_group.shares = NICE_0_LOAD;
9481 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
9483 * In case of task-groups formed thr' the user id of tasks,
9484 * init_task_group represents tasks belonging to root user.
9485 * Hence it forms a sibling of all subsequent groups formed.
9486 * In this case, init_task_group gets only a fraction of overall
9487 * system cpu resource, based on the weight assigned to root
9488 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9489 * by letting tasks of init_task_group sit in a separate cfs_rq
9490 * (init_tg_cfs_rq) and having one entity represent this group of
9491 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9493 init_tg_cfs_entry(&init_task_group,
9494 &per_cpu(init_tg_cfs_rq, i),
9495 &per_cpu(init_sched_entity, i), i, 1,
9496 root_task_group.se[i]);
9499 #endif /* CONFIG_FAIR_GROUP_SCHED */
9501 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9502 #ifdef CONFIG_RT_GROUP_SCHED
9503 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
9504 #ifdef CONFIG_CGROUP_SCHED
9505 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
9506 #elif defined CONFIG_USER_SCHED
9507 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
9508 init_tg_rt_entry(&init_task_group,
9509 &per_cpu(init_rt_rq, i),
9510 &per_cpu(init_sched_rt_entity, i), i, 1,
9511 root_task_group.rt_se[i]);
9515 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
9516 rq->cpu_load[j] = 0;
9520 rq->post_schedule = 0;
9521 rq->active_balance = 0;
9522 rq->next_balance = jiffies;
9526 rq->migration_thread = NULL;
9527 INIT_LIST_HEAD(&rq->migration_queue);
9528 rq_attach_root(rq, &def_root_domain);
9531 atomic_set(&rq->nr_iowait, 0);
9534 set_load_weight(&init_task);
9536 #ifdef CONFIG_PREEMPT_NOTIFIERS
9537 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
9541 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9544 #ifdef CONFIG_RT_MUTEXES
9545 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
9549 * The boot idle thread does lazy MMU switching as well:
9551 atomic_inc(&init_mm.mm_count);
9552 enter_lazy_tlb(&init_mm, current);
9555 * Make us the idle thread. Technically, schedule() should not be
9556 * called from this thread, however somewhere below it might be,
9557 * but because we are the idle thread, we just pick up running again
9558 * when this runqueue becomes "idle".
9560 init_idle(current, smp_processor_id());
9562 calc_load_update = jiffies + LOAD_FREQ;
9565 * During early bootup we pretend to be a normal task:
9567 current->sched_class = &fair_sched_class;
9569 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9570 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
9573 zalloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
9574 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
9576 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
9581 scheduler_running = 1;
9584 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9585 static inline int preempt_count_equals(int preempt_offset)
9587 int nested = preempt_count() & ~PREEMPT_ACTIVE;
9589 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
9592 void __might_sleep(char *file, int line, int preempt_offset)
9595 static unsigned long prev_jiffy; /* ratelimiting */
9597 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
9598 system_state != SYSTEM_RUNNING || oops_in_progress)
9600 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9602 prev_jiffy = jiffies;
9605 "BUG: sleeping function called from invalid context at %s:%d\n",
9608 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9609 in_atomic(), irqs_disabled(),
9610 current->pid, current->comm);
9612 debug_show_held_locks(current);
9613 if (irqs_disabled())
9614 print_irqtrace_events(current);
9618 EXPORT_SYMBOL(__might_sleep);
9621 #ifdef CONFIG_MAGIC_SYSRQ
9622 static void normalize_task(struct rq *rq, struct task_struct *p)
9626 update_rq_clock(rq);
9627 on_rq = p->se.on_rq;
9629 deactivate_task(rq, p, 0);
9630 __setscheduler(rq, p, SCHED_NORMAL, 0);
9632 activate_task(rq, p, 0);
9633 resched_task(rq->curr);
9637 void normalize_rt_tasks(void)
9639 struct task_struct *g, *p;
9640 unsigned long flags;
9643 read_lock_irqsave(&tasklist_lock, flags);
9644 do_each_thread(g, p) {
9646 * Only normalize user tasks:
9651 p->se.exec_start = 0;
9652 #ifdef CONFIG_SCHEDSTATS
9653 p->se.wait_start = 0;
9654 p->se.sleep_start = 0;
9655 p->se.block_start = 0;
9660 * Renice negative nice level userspace
9663 if (TASK_NICE(p) < 0 && p->mm)
9664 set_user_nice(p, 0);
9668 spin_lock(&p->pi_lock);
9669 rq = __task_rq_lock(p);
9671 normalize_task(rq, p);
9673 __task_rq_unlock(rq);
9674 spin_unlock(&p->pi_lock);
9675 } while_each_thread(g, p);
9677 read_unlock_irqrestore(&tasklist_lock, flags);
9680 #endif /* CONFIG_MAGIC_SYSRQ */
9684 * These functions are only useful for the IA64 MCA handling.
9686 * They can only be called when the whole system has been
9687 * stopped - every CPU needs to be quiescent, and no scheduling
9688 * activity can take place. Using them for anything else would
9689 * be a serious bug, and as a result, they aren't even visible
9690 * under any other configuration.
9694 * curr_task - return the current task for a given cpu.
9695 * @cpu: the processor in question.
9697 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9699 struct task_struct *curr_task(int cpu)
9701 return cpu_curr(cpu);
9705 * set_curr_task - set the current task for a given cpu.
9706 * @cpu: the processor in question.
9707 * @p: the task pointer to set.
9709 * Description: This function must only be used when non-maskable interrupts
9710 * are serviced on a separate stack. It allows the architecture to switch the
9711 * notion of the current task on a cpu in a non-blocking manner. This function
9712 * must be called with all CPU's synchronized, and interrupts disabled, the
9713 * and caller must save the original value of the current task (see
9714 * curr_task() above) and restore that value before reenabling interrupts and
9715 * re-starting the system.
9717 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9719 void set_curr_task(int cpu, struct task_struct *p)
9726 #ifdef CONFIG_FAIR_GROUP_SCHED
9727 static void free_fair_sched_group(struct task_group *tg)
9731 for_each_possible_cpu(i) {
9733 kfree(tg->cfs_rq[i]);
9743 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9745 struct cfs_rq *cfs_rq;
9746 struct sched_entity *se;
9750 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9753 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9757 tg->shares = NICE_0_LOAD;
9759 for_each_possible_cpu(i) {
9762 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9763 GFP_KERNEL, cpu_to_node(i));
9767 se = kzalloc_node(sizeof(struct sched_entity),
9768 GFP_KERNEL, cpu_to_node(i));
9772 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
9781 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9783 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
9784 &cpu_rq(cpu)->leaf_cfs_rq_list);
9787 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9789 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
9791 #else /* !CONFG_FAIR_GROUP_SCHED */
9792 static inline void free_fair_sched_group(struct task_group *tg)
9797 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9802 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9806 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9809 #endif /* CONFIG_FAIR_GROUP_SCHED */
9811 #ifdef CONFIG_RT_GROUP_SCHED
9812 static void free_rt_sched_group(struct task_group *tg)
9816 destroy_rt_bandwidth(&tg->rt_bandwidth);
9818 for_each_possible_cpu(i) {
9820 kfree(tg->rt_rq[i]);
9822 kfree(tg->rt_se[i]);
9830 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9832 struct rt_rq *rt_rq;
9833 struct sched_rt_entity *rt_se;
9837 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
9840 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
9844 init_rt_bandwidth(&tg->rt_bandwidth,
9845 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
9847 for_each_possible_cpu(i) {
9850 rt_rq = kzalloc_node(sizeof(struct rt_rq),
9851 GFP_KERNEL, cpu_to_node(i));
9855 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
9856 GFP_KERNEL, cpu_to_node(i));
9860 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
9869 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9871 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
9872 &cpu_rq(cpu)->leaf_rt_rq_list);
9875 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9877 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
9879 #else /* !CONFIG_RT_GROUP_SCHED */
9880 static inline void free_rt_sched_group(struct task_group *tg)
9885 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9890 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9894 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9897 #endif /* CONFIG_RT_GROUP_SCHED */
9899 #ifdef CONFIG_GROUP_SCHED
9900 static void free_sched_group(struct task_group *tg)
9902 free_fair_sched_group(tg);
9903 free_rt_sched_group(tg);
9907 /* allocate runqueue etc for a new task group */
9908 struct task_group *sched_create_group(struct task_group *parent)
9910 struct task_group *tg;
9911 unsigned long flags;
9914 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
9916 return ERR_PTR(-ENOMEM);
9918 if (!alloc_fair_sched_group(tg, parent))
9921 if (!alloc_rt_sched_group(tg, parent))
9924 spin_lock_irqsave(&task_group_lock, flags);
9925 for_each_possible_cpu(i) {
9926 register_fair_sched_group(tg, i);
9927 register_rt_sched_group(tg, i);
9929 list_add_rcu(&tg->list, &task_groups);
9931 WARN_ON(!parent); /* root should already exist */
9933 tg->parent = parent;
9934 INIT_LIST_HEAD(&tg->children);
9935 list_add_rcu(&tg->siblings, &parent->children);
9936 spin_unlock_irqrestore(&task_group_lock, flags);
9941 free_sched_group(tg);
9942 return ERR_PTR(-ENOMEM);
9945 /* rcu callback to free various structures associated with a task group */
9946 static void free_sched_group_rcu(struct rcu_head *rhp)
9948 /* now it should be safe to free those cfs_rqs */
9949 free_sched_group(container_of(rhp, struct task_group, rcu));
9952 /* Destroy runqueue etc associated with a task group */
9953 void sched_destroy_group(struct task_group *tg)
9955 unsigned long flags;
9958 spin_lock_irqsave(&task_group_lock, flags);
9959 for_each_possible_cpu(i) {
9960 unregister_fair_sched_group(tg, i);
9961 unregister_rt_sched_group(tg, i);
9963 list_del_rcu(&tg->list);
9964 list_del_rcu(&tg->siblings);
9965 spin_unlock_irqrestore(&task_group_lock, flags);
9967 /* wait for possible concurrent references to cfs_rqs complete */
9968 call_rcu(&tg->rcu, free_sched_group_rcu);
9971 /* change task's runqueue when it moves between groups.
9972 * The caller of this function should have put the task in its new group
9973 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
9974 * reflect its new group.
9976 void sched_move_task(struct task_struct *tsk)
9979 unsigned long flags;
9982 rq = task_rq_lock(tsk, &flags);
9984 update_rq_clock(rq);
9986 running = task_current(rq, tsk);
9987 on_rq = tsk->se.on_rq;
9990 dequeue_task(rq, tsk, 0);
9991 if (unlikely(running))
9992 tsk->sched_class->put_prev_task(rq, tsk);
9994 set_task_rq(tsk, task_cpu(tsk));
9996 #ifdef CONFIG_FAIR_GROUP_SCHED
9997 if (tsk->sched_class->moved_group)
9998 tsk->sched_class->moved_group(tsk);
10001 if (unlikely(running))
10002 tsk->sched_class->set_curr_task(rq);
10004 enqueue_task(rq, tsk, 0);
10006 task_rq_unlock(rq, &flags);
10008 #endif /* CONFIG_GROUP_SCHED */
10010 #ifdef CONFIG_FAIR_GROUP_SCHED
10011 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
10013 struct cfs_rq *cfs_rq = se->cfs_rq;
10018 dequeue_entity(cfs_rq, se, 0);
10020 se->load.weight = shares;
10021 se->load.inv_weight = 0;
10024 enqueue_entity(cfs_rq, se, 0);
10027 static void set_se_shares(struct sched_entity *se, unsigned long shares)
10029 struct cfs_rq *cfs_rq = se->cfs_rq;
10030 struct rq *rq = cfs_rq->rq;
10031 unsigned long flags;
10033 spin_lock_irqsave(&rq->lock, flags);
10034 __set_se_shares(se, shares);
10035 spin_unlock_irqrestore(&rq->lock, flags);
10038 static DEFINE_MUTEX(shares_mutex);
10040 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10043 unsigned long flags;
10046 * We can't change the weight of the root cgroup.
10051 if (shares < MIN_SHARES)
10052 shares = MIN_SHARES;
10053 else if (shares > MAX_SHARES)
10054 shares = MAX_SHARES;
10056 mutex_lock(&shares_mutex);
10057 if (tg->shares == shares)
10060 spin_lock_irqsave(&task_group_lock, flags);
10061 for_each_possible_cpu(i)
10062 unregister_fair_sched_group(tg, i);
10063 list_del_rcu(&tg->siblings);
10064 spin_unlock_irqrestore(&task_group_lock, flags);
10066 /* wait for any ongoing reference to this group to finish */
10067 synchronize_sched();
10070 * Now we are free to modify the group's share on each cpu
10071 * w/o tripping rebalance_share or load_balance_fair.
10073 tg->shares = shares;
10074 for_each_possible_cpu(i) {
10076 * force a rebalance
10078 cfs_rq_set_shares(tg->cfs_rq[i], 0);
10079 set_se_shares(tg->se[i], shares);
10083 * Enable load balance activity on this group, by inserting it back on
10084 * each cpu's rq->leaf_cfs_rq_list.
10086 spin_lock_irqsave(&task_group_lock, flags);
10087 for_each_possible_cpu(i)
10088 register_fair_sched_group(tg, i);
10089 list_add_rcu(&tg->siblings, &tg->parent->children);
10090 spin_unlock_irqrestore(&task_group_lock, flags);
10092 mutex_unlock(&shares_mutex);
10096 unsigned long sched_group_shares(struct task_group *tg)
10102 #ifdef CONFIG_RT_GROUP_SCHED
10104 * Ensure that the real time constraints are schedulable.
10106 static DEFINE_MUTEX(rt_constraints_mutex);
10108 static unsigned long to_ratio(u64 period, u64 runtime)
10110 if (runtime == RUNTIME_INF)
10113 return div64_u64(runtime << 20, period);
10116 /* Must be called with tasklist_lock held */
10117 static inline int tg_has_rt_tasks(struct task_group *tg)
10119 struct task_struct *g, *p;
10121 do_each_thread(g, p) {
10122 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
10124 } while_each_thread(g, p);
10129 struct rt_schedulable_data {
10130 struct task_group *tg;
10135 static int tg_schedulable(struct task_group *tg, void *data)
10137 struct rt_schedulable_data *d = data;
10138 struct task_group *child;
10139 unsigned long total, sum = 0;
10140 u64 period, runtime;
10142 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10143 runtime = tg->rt_bandwidth.rt_runtime;
10146 period = d->rt_period;
10147 runtime = d->rt_runtime;
10150 #ifdef CONFIG_USER_SCHED
10151 if (tg == &root_task_group) {
10152 period = global_rt_period();
10153 runtime = global_rt_runtime();
10158 * Cannot have more runtime than the period.
10160 if (runtime > period && runtime != RUNTIME_INF)
10164 * Ensure we don't starve existing RT tasks.
10166 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
10169 total = to_ratio(period, runtime);
10172 * Nobody can have more than the global setting allows.
10174 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
10178 * The sum of our children's runtime should not exceed our own.
10180 list_for_each_entry_rcu(child, &tg->children, siblings) {
10181 period = ktime_to_ns(child->rt_bandwidth.rt_period);
10182 runtime = child->rt_bandwidth.rt_runtime;
10184 if (child == d->tg) {
10185 period = d->rt_period;
10186 runtime = d->rt_runtime;
10189 sum += to_ratio(period, runtime);
10198 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
10200 struct rt_schedulable_data data = {
10202 .rt_period = period,
10203 .rt_runtime = runtime,
10206 return walk_tg_tree(tg_schedulable, tg_nop, &data);
10209 static int tg_set_bandwidth(struct task_group *tg,
10210 u64 rt_period, u64 rt_runtime)
10214 mutex_lock(&rt_constraints_mutex);
10215 read_lock(&tasklist_lock);
10216 err = __rt_schedulable(tg, rt_period, rt_runtime);
10220 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10221 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
10222 tg->rt_bandwidth.rt_runtime = rt_runtime;
10224 for_each_possible_cpu(i) {
10225 struct rt_rq *rt_rq = tg->rt_rq[i];
10227 spin_lock(&rt_rq->rt_runtime_lock);
10228 rt_rq->rt_runtime = rt_runtime;
10229 spin_unlock(&rt_rq->rt_runtime_lock);
10231 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10233 read_unlock(&tasklist_lock);
10234 mutex_unlock(&rt_constraints_mutex);
10239 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
10241 u64 rt_runtime, rt_period;
10243 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10244 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
10245 if (rt_runtime_us < 0)
10246 rt_runtime = RUNTIME_INF;
10248 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10251 long sched_group_rt_runtime(struct task_group *tg)
10255 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
10258 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
10259 do_div(rt_runtime_us, NSEC_PER_USEC);
10260 return rt_runtime_us;
10263 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
10265 u64 rt_runtime, rt_period;
10267 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
10268 rt_runtime = tg->rt_bandwidth.rt_runtime;
10270 if (rt_period == 0)
10273 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10276 long sched_group_rt_period(struct task_group *tg)
10280 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
10281 do_div(rt_period_us, NSEC_PER_USEC);
10282 return rt_period_us;
10285 static int sched_rt_global_constraints(void)
10287 u64 runtime, period;
10290 if (sysctl_sched_rt_period <= 0)
10293 runtime = global_rt_runtime();
10294 period = global_rt_period();
10297 * Sanity check on the sysctl variables.
10299 if (runtime > period && runtime != RUNTIME_INF)
10302 mutex_lock(&rt_constraints_mutex);
10303 read_lock(&tasklist_lock);
10304 ret = __rt_schedulable(NULL, 0, 0);
10305 read_unlock(&tasklist_lock);
10306 mutex_unlock(&rt_constraints_mutex);
10311 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
10313 /* Don't accept realtime tasks when there is no way for them to run */
10314 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
10320 #else /* !CONFIG_RT_GROUP_SCHED */
10321 static int sched_rt_global_constraints(void)
10323 unsigned long flags;
10326 if (sysctl_sched_rt_period <= 0)
10330 * There's always some RT tasks in the root group
10331 * -- migration, kstopmachine etc..
10333 if (sysctl_sched_rt_runtime == 0)
10336 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
10337 for_each_possible_cpu(i) {
10338 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
10340 spin_lock(&rt_rq->rt_runtime_lock);
10341 rt_rq->rt_runtime = global_rt_runtime();
10342 spin_unlock(&rt_rq->rt_runtime_lock);
10344 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
10348 #endif /* CONFIG_RT_GROUP_SCHED */
10350 int sched_rt_handler(struct ctl_table *table, int write,
10351 void __user *buffer, size_t *lenp,
10355 int old_period, old_runtime;
10356 static DEFINE_MUTEX(mutex);
10358 mutex_lock(&mutex);
10359 old_period = sysctl_sched_rt_period;
10360 old_runtime = sysctl_sched_rt_runtime;
10362 ret = proc_dointvec(table, write, buffer, lenp, ppos);
10364 if (!ret && write) {
10365 ret = sched_rt_global_constraints();
10367 sysctl_sched_rt_period = old_period;
10368 sysctl_sched_rt_runtime = old_runtime;
10370 def_rt_bandwidth.rt_runtime = global_rt_runtime();
10371 def_rt_bandwidth.rt_period =
10372 ns_to_ktime(global_rt_period());
10375 mutex_unlock(&mutex);
10380 #ifdef CONFIG_CGROUP_SCHED
10382 /* return corresponding task_group object of a cgroup */
10383 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
10385 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
10386 struct task_group, css);
10389 static struct cgroup_subsys_state *
10390 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
10392 struct task_group *tg, *parent;
10394 if (!cgrp->parent) {
10395 /* This is early initialization for the top cgroup */
10396 return &init_task_group.css;
10399 parent = cgroup_tg(cgrp->parent);
10400 tg = sched_create_group(parent);
10402 return ERR_PTR(-ENOMEM);
10408 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10410 struct task_group *tg = cgroup_tg(cgrp);
10412 sched_destroy_group(tg);
10416 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
10418 #ifdef CONFIG_RT_GROUP_SCHED
10419 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
10422 /* We don't support RT-tasks being in separate groups */
10423 if (tsk->sched_class != &fair_sched_class)
10430 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10431 struct task_struct *tsk, bool threadgroup)
10433 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
10437 struct task_struct *c;
10439 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
10440 retval = cpu_cgroup_can_attach_task(cgrp, c);
10452 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10453 struct cgroup *old_cont, struct task_struct *tsk,
10456 sched_move_task(tsk);
10458 struct task_struct *c;
10460 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
10461 sched_move_task(c);
10467 #ifdef CONFIG_FAIR_GROUP_SCHED
10468 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
10471 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
10474 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
10476 struct task_group *tg = cgroup_tg(cgrp);
10478 return (u64) tg->shares;
10480 #endif /* CONFIG_FAIR_GROUP_SCHED */
10482 #ifdef CONFIG_RT_GROUP_SCHED
10483 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
10486 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
10489 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
10491 return sched_group_rt_runtime(cgroup_tg(cgrp));
10494 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
10497 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
10500 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
10502 return sched_group_rt_period(cgroup_tg(cgrp));
10504 #endif /* CONFIG_RT_GROUP_SCHED */
10506 static struct cftype cpu_files[] = {
10507 #ifdef CONFIG_FAIR_GROUP_SCHED
10510 .read_u64 = cpu_shares_read_u64,
10511 .write_u64 = cpu_shares_write_u64,
10514 #ifdef CONFIG_RT_GROUP_SCHED
10516 .name = "rt_runtime_us",
10517 .read_s64 = cpu_rt_runtime_read,
10518 .write_s64 = cpu_rt_runtime_write,
10521 .name = "rt_period_us",
10522 .read_u64 = cpu_rt_period_read_uint,
10523 .write_u64 = cpu_rt_period_write_uint,
10528 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
10530 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
10533 struct cgroup_subsys cpu_cgroup_subsys = {
10535 .create = cpu_cgroup_create,
10536 .destroy = cpu_cgroup_destroy,
10537 .can_attach = cpu_cgroup_can_attach,
10538 .attach = cpu_cgroup_attach,
10539 .populate = cpu_cgroup_populate,
10540 .subsys_id = cpu_cgroup_subsys_id,
10544 #endif /* CONFIG_CGROUP_SCHED */
10546 #ifdef CONFIG_CGROUP_CPUACCT
10549 * CPU accounting code for task groups.
10551 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10552 * (balbir@in.ibm.com).
10555 /* track cpu usage of a group of tasks and its child groups */
10557 struct cgroup_subsys_state css;
10558 /* cpuusage holds pointer to a u64-type object on every cpu */
10560 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
10561 struct cpuacct *parent;
10564 struct cgroup_subsys cpuacct_subsys;
10566 /* return cpu accounting group corresponding to this container */
10567 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
10569 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
10570 struct cpuacct, css);
10573 /* return cpu accounting group to which this task belongs */
10574 static inline struct cpuacct *task_ca(struct task_struct *tsk)
10576 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
10577 struct cpuacct, css);
10580 /* create a new cpu accounting group */
10581 static struct cgroup_subsys_state *cpuacct_create(
10582 struct cgroup_subsys *ss, struct cgroup *cgrp)
10584 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
10590 ca->cpuusage = alloc_percpu(u64);
10594 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10595 if (percpu_counter_init(&ca->cpustat[i], 0))
10596 goto out_free_counters;
10599 ca->parent = cgroup_ca(cgrp->parent);
10605 percpu_counter_destroy(&ca->cpustat[i]);
10606 free_percpu(ca->cpuusage);
10610 return ERR_PTR(-ENOMEM);
10613 /* destroy an existing cpu accounting group */
10615 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10617 struct cpuacct *ca = cgroup_ca(cgrp);
10620 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10621 percpu_counter_destroy(&ca->cpustat[i]);
10622 free_percpu(ca->cpuusage);
10626 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
10628 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10631 #ifndef CONFIG_64BIT
10633 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10635 spin_lock_irq(&cpu_rq(cpu)->lock);
10637 spin_unlock_irq(&cpu_rq(cpu)->lock);
10645 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
10647 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10649 #ifndef CONFIG_64BIT
10651 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10653 spin_lock_irq(&cpu_rq(cpu)->lock);
10655 spin_unlock_irq(&cpu_rq(cpu)->lock);
10661 /* return total cpu usage (in nanoseconds) of a group */
10662 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
10664 struct cpuacct *ca = cgroup_ca(cgrp);
10665 u64 totalcpuusage = 0;
10668 for_each_present_cpu(i)
10669 totalcpuusage += cpuacct_cpuusage_read(ca, i);
10671 return totalcpuusage;
10674 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
10677 struct cpuacct *ca = cgroup_ca(cgrp);
10686 for_each_present_cpu(i)
10687 cpuacct_cpuusage_write(ca, i, 0);
10693 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
10694 struct seq_file *m)
10696 struct cpuacct *ca = cgroup_ca(cgroup);
10700 for_each_present_cpu(i) {
10701 percpu = cpuacct_cpuusage_read(ca, i);
10702 seq_printf(m, "%llu ", (unsigned long long) percpu);
10704 seq_printf(m, "\n");
10708 static const char *cpuacct_stat_desc[] = {
10709 [CPUACCT_STAT_USER] = "user",
10710 [CPUACCT_STAT_SYSTEM] = "system",
10713 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
10714 struct cgroup_map_cb *cb)
10716 struct cpuacct *ca = cgroup_ca(cgrp);
10719 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
10720 s64 val = percpu_counter_read(&ca->cpustat[i]);
10721 val = cputime64_to_clock_t(val);
10722 cb->fill(cb, cpuacct_stat_desc[i], val);
10727 static struct cftype files[] = {
10730 .read_u64 = cpuusage_read,
10731 .write_u64 = cpuusage_write,
10734 .name = "usage_percpu",
10735 .read_seq_string = cpuacct_percpu_seq_read,
10739 .read_map = cpuacct_stats_show,
10743 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
10745 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
10749 * charge this task's execution time to its accounting group.
10751 * called with rq->lock held.
10753 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
10755 struct cpuacct *ca;
10758 if (unlikely(!cpuacct_subsys.active))
10761 cpu = task_cpu(tsk);
10767 for (; ca; ca = ca->parent) {
10768 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10769 *cpuusage += cputime;
10776 * Charge the system/user time to the task's accounting group.
10778 static void cpuacct_update_stats(struct task_struct *tsk,
10779 enum cpuacct_stat_index idx, cputime_t val)
10781 struct cpuacct *ca;
10783 if (unlikely(!cpuacct_subsys.active))
10790 percpu_counter_add(&ca->cpustat[idx], val);
10796 struct cgroup_subsys cpuacct_subsys = {
10798 .create = cpuacct_create,
10799 .destroy = cpuacct_destroy,
10800 .populate = cpuacct_populate,
10801 .subsys_id = cpuacct_subsys_id,
10803 #endif /* CONFIG_CGROUP_CPUACCT */
10807 int rcu_expedited_torture_stats(char *page)
10811 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
10813 void synchronize_sched_expedited(void)
10816 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
10818 #else /* #ifndef CONFIG_SMP */
10820 static DEFINE_PER_CPU(struct migration_req, rcu_migration_req);
10821 static DEFINE_MUTEX(rcu_sched_expedited_mutex);
10823 #define RCU_EXPEDITED_STATE_POST -2
10824 #define RCU_EXPEDITED_STATE_IDLE -1
10826 static int rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
10828 int rcu_expedited_torture_stats(char *page)
10833 cnt += sprintf(&page[cnt], "state: %d /", rcu_expedited_state);
10834 for_each_online_cpu(cpu) {
10835 cnt += sprintf(&page[cnt], " %d:%d",
10836 cpu, per_cpu(rcu_migration_req, cpu).dest_cpu);
10838 cnt += sprintf(&page[cnt], "\n");
10841 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
10843 static long synchronize_sched_expedited_count;
10846 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
10847 * approach to force grace period to end quickly. This consumes
10848 * significant time on all CPUs, and is thus not recommended for
10849 * any sort of common-case code.
10851 * Note that it is illegal to call this function while holding any
10852 * lock that is acquired by a CPU-hotplug notifier. Failing to
10853 * observe this restriction will result in deadlock.
10855 void synchronize_sched_expedited(void)
10858 unsigned long flags;
10859 bool need_full_sync = 0;
10861 struct migration_req *req;
10865 smp_mb(); /* ensure prior mod happens before capturing snap. */
10866 snap = ACCESS_ONCE(synchronize_sched_expedited_count) + 1;
10868 while (!mutex_trylock(&rcu_sched_expedited_mutex)) {
10870 if (trycount++ < 10)
10871 udelay(trycount * num_online_cpus());
10873 synchronize_sched();
10876 if (ACCESS_ONCE(synchronize_sched_expedited_count) - snap > 0) {
10877 smp_mb(); /* ensure test happens before caller kfree */
10882 rcu_expedited_state = RCU_EXPEDITED_STATE_POST;
10883 for_each_online_cpu(cpu) {
10885 req = &per_cpu(rcu_migration_req, cpu);
10886 init_completion(&req->done);
10888 req->dest_cpu = RCU_MIGRATION_NEED_QS;
10889 spin_lock_irqsave(&rq->lock, flags);
10890 list_add(&req->list, &rq->migration_queue);
10891 spin_unlock_irqrestore(&rq->lock, flags);
10892 wake_up_process(rq->migration_thread);
10894 for_each_online_cpu(cpu) {
10895 rcu_expedited_state = cpu;
10896 req = &per_cpu(rcu_migration_req, cpu);
10898 wait_for_completion(&req->done);
10899 spin_lock_irqsave(&rq->lock, flags);
10900 if (unlikely(req->dest_cpu == RCU_MIGRATION_MUST_SYNC))
10901 need_full_sync = 1;
10902 req->dest_cpu = RCU_MIGRATION_IDLE;
10903 spin_unlock_irqrestore(&rq->lock, flags);
10905 rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
10906 mutex_unlock(&rcu_sched_expedited_mutex);
10908 if (need_full_sync)
10909 synchronize_sched();
10911 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
10913 #endif /* #else #ifndef CONFIG_SMP */