4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/perf_event.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/kthread.h>
59 #include <linux/proc_fs.h>
60 #include <linux/seq_file.h>
61 #include <linux/sysctl.h>
62 #include <linux/syscalls.h>
63 #include <linux/times.h>
64 #include <linux/tsacct_kern.h>
65 #include <linux/kprobes.h>
66 #include <linux/delayacct.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
73 #include <linux/ftrace.h>
76 #include <asm/irq_regs.h>
78 #include "sched_cpupri.h"
80 #define CREATE_TRACE_POINTS
81 #include <trace/events/sched.h>
84 * Convert user-nice values [ -20 ... 0 ... 19 ]
85 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
88 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
89 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
90 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
93 * 'User priority' is the nice value converted to something we
94 * can work with better when scaling various scheduler parameters,
95 * it's a [ 0 ... 39 ] range.
97 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
98 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
99 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
102 * Helpers for converting nanosecond timing to jiffy resolution
104 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
106 #define NICE_0_LOAD SCHED_LOAD_SCALE
107 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
110 * These are the 'tuning knobs' of the scheduler:
112 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
113 * Timeslices get refilled after they expire.
115 #define DEF_TIMESLICE (100 * HZ / 1000)
118 * single value that denotes runtime == period, ie unlimited time.
120 #define RUNTIME_INF ((u64)~0ULL)
122 static inline int rt_policy(int policy)
124 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
129 static inline int task_has_rt_policy(struct task_struct *p)
131 return rt_policy(p->policy);
135 * This is the priority-queue data structure of the RT scheduling class:
137 struct rt_prio_array {
138 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
139 struct list_head queue[MAX_RT_PRIO];
142 struct rt_bandwidth {
143 /* nests inside the rq lock: */
144 spinlock_t rt_runtime_lock;
147 struct hrtimer rt_period_timer;
150 static struct rt_bandwidth def_rt_bandwidth;
152 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
154 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
156 struct rt_bandwidth *rt_b =
157 container_of(timer, struct rt_bandwidth, rt_period_timer);
163 now = hrtimer_cb_get_time(timer);
164 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
169 idle = do_sched_rt_period_timer(rt_b, overrun);
172 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
176 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
178 rt_b->rt_period = ns_to_ktime(period);
179 rt_b->rt_runtime = runtime;
181 spin_lock_init(&rt_b->rt_runtime_lock);
183 hrtimer_init(&rt_b->rt_period_timer,
184 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
185 rt_b->rt_period_timer.function = sched_rt_period_timer;
188 static inline int rt_bandwidth_enabled(void)
190 return sysctl_sched_rt_runtime >= 0;
193 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
197 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
200 if (hrtimer_active(&rt_b->rt_period_timer))
203 spin_lock(&rt_b->rt_runtime_lock);
208 if (hrtimer_active(&rt_b->rt_period_timer))
211 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
212 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
214 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
215 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
216 delta = ktime_to_ns(ktime_sub(hard, soft));
217 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
218 HRTIMER_MODE_ABS_PINNED, 0);
220 spin_unlock(&rt_b->rt_runtime_lock);
223 #ifdef CONFIG_RT_GROUP_SCHED
224 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
226 hrtimer_cancel(&rt_b->rt_period_timer);
231 * sched_domains_mutex serializes calls to arch_init_sched_domains,
232 * detach_destroy_domains and partition_sched_domains.
234 static DEFINE_MUTEX(sched_domains_mutex);
236 #ifdef CONFIG_GROUP_SCHED
238 #include <linux/cgroup.h>
242 static LIST_HEAD(task_groups);
244 /* task group related information */
246 #ifdef CONFIG_CGROUP_SCHED
247 struct cgroup_subsys_state css;
250 #ifdef CONFIG_USER_SCHED
254 #ifdef CONFIG_FAIR_GROUP_SCHED
255 /* schedulable entities of this group on each cpu */
256 struct sched_entity **se;
257 /* runqueue "owned" by this group on each cpu */
258 struct cfs_rq **cfs_rq;
259 unsigned long shares;
262 #ifdef CONFIG_RT_GROUP_SCHED
263 struct sched_rt_entity **rt_se;
264 struct rt_rq **rt_rq;
266 struct rt_bandwidth rt_bandwidth;
270 struct list_head list;
272 struct task_group *parent;
273 struct list_head siblings;
274 struct list_head children;
277 #ifdef CONFIG_USER_SCHED
279 /* Helper function to pass uid information to create_sched_user() */
280 void set_tg_uid(struct user_struct *user)
282 user->tg->uid = user->uid;
287 * Every UID task group (including init_task_group aka UID-0) will
288 * be a child to this group.
290 struct task_group root_task_group;
292 #ifdef CONFIG_FAIR_GROUP_SCHED
293 /* Default task group's sched entity on each cpu */
294 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
295 /* Default task group's cfs_rq on each cpu */
296 static DEFINE_PER_CPU_SHARED_ALIGNED(struct cfs_rq, init_tg_cfs_rq);
297 #endif /* CONFIG_FAIR_GROUP_SCHED */
299 #ifdef CONFIG_RT_GROUP_SCHED
300 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
301 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rt_rq, init_rt_rq);
302 #endif /* CONFIG_RT_GROUP_SCHED */
303 #else /* !CONFIG_USER_SCHED */
304 #define root_task_group init_task_group
305 #endif /* CONFIG_USER_SCHED */
307 /* task_group_lock serializes add/remove of task groups and also changes to
308 * a task group's cpu shares.
310 static DEFINE_SPINLOCK(task_group_lock);
312 #ifdef CONFIG_FAIR_GROUP_SCHED
315 static int root_task_group_empty(void)
317 return list_empty(&root_task_group.children);
321 #ifdef CONFIG_USER_SCHED
322 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
323 #else /* !CONFIG_USER_SCHED */
324 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
325 #endif /* CONFIG_USER_SCHED */
328 * A weight of 0 or 1 can cause arithmetics problems.
329 * A weight of a cfs_rq is the sum of weights of which entities
330 * are queued on this cfs_rq, so a weight of a entity should not be
331 * too large, so as the shares value of a task group.
332 * (The default weight is 1024 - so there's no practical
333 * limitation from this.)
336 #define MAX_SHARES (1UL << 18)
338 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
341 /* Default task group.
342 * Every task in system belong to this group at bootup.
344 struct task_group init_task_group;
346 /* return group to which a task belongs */
347 static inline struct task_group *task_group(struct task_struct *p)
349 struct task_group *tg;
351 #ifdef CONFIG_USER_SCHED
353 tg = __task_cred(p)->user->tg;
355 #elif defined(CONFIG_CGROUP_SCHED)
356 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
357 struct task_group, css);
359 tg = &init_task_group;
364 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
365 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
367 #ifdef CONFIG_FAIR_GROUP_SCHED
368 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
369 p->se.parent = task_group(p)->se[cpu];
372 #ifdef CONFIG_RT_GROUP_SCHED
373 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
374 p->rt.parent = task_group(p)->rt_se[cpu];
380 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
381 static inline struct task_group *task_group(struct task_struct *p)
386 #endif /* CONFIG_GROUP_SCHED */
388 /* CFS-related fields in a runqueue */
390 struct load_weight load;
391 unsigned long nr_running;
396 struct rb_root tasks_timeline;
397 struct rb_node *rb_leftmost;
399 struct list_head tasks;
400 struct list_head *balance_iterator;
403 * 'curr' points to currently running entity on this cfs_rq.
404 * It is set to NULL otherwise (i.e when none are currently running).
406 struct sched_entity *curr, *next, *last;
408 unsigned int nr_spread_over;
410 #ifdef CONFIG_FAIR_GROUP_SCHED
411 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
414 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
415 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
416 * (like users, containers etc.)
418 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
419 * list is used during load balance.
421 struct list_head leaf_cfs_rq_list;
422 struct task_group *tg; /* group that "owns" this runqueue */
426 * the part of load.weight contributed by tasks
428 unsigned long task_weight;
431 * h_load = weight * f(tg)
433 * Where f(tg) is the recursive weight fraction assigned to
436 unsigned long h_load;
439 * this cpu's part of tg->shares
441 unsigned long shares;
444 * load.weight at the time we set shares
446 unsigned long rq_weight;
451 /* Real-Time classes' related field in a runqueue: */
453 struct rt_prio_array active;
454 unsigned long rt_nr_running;
455 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
457 int curr; /* highest queued rt task prio */
459 int next; /* next highest */
464 unsigned long rt_nr_migratory;
465 unsigned long rt_nr_total;
467 struct plist_head pushable_tasks;
472 /* Nests inside the rq lock: */
473 spinlock_t rt_runtime_lock;
475 #ifdef CONFIG_RT_GROUP_SCHED
476 unsigned long rt_nr_boosted;
479 struct list_head leaf_rt_rq_list;
480 struct task_group *tg;
481 struct sched_rt_entity *rt_se;
488 * We add the notion of a root-domain which will be used to define per-domain
489 * variables. Each exclusive cpuset essentially defines an island domain by
490 * fully partitioning the member cpus from any other cpuset. Whenever a new
491 * exclusive cpuset is created, we also create and attach a new root-domain
498 cpumask_var_t online;
501 * The "RT overload" flag: it gets set if a CPU has more than
502 * one runnable RT task.
504 cpumask_var_t rto_mask;
507 struct cpupri cpupri;
512 * By default the system creates a single root-domain with all cpus as
513 * members (mimicking the global state we have today).
515 static struct root_domain def_root_domain;
520 * This is the main, per-CPU runqueue data structure.
522 * Locking rule: those places that want to lock multiple runqueues
523 * (such as the load balancing or the thread migration code), lock
524 * acquire operations must be ordered by ascending &runqueue.
531 * nr_running and cpu_load should be in the same cacheline because
532 * remote CPUs use both these fields when doing load calculation.
534 unsigned long nr_running;
535 #define CPU_LOAD_IDX_MAX 5
536 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
538 unsigned char in_nohz_recently;
540 /* capture load from *all* tasks on this cpu: */
541 struct load_weight load;
542 unsigned long nr_load_updates;
548 #ifdef CONFIG_FAIR_GROUP_SCHED
549 /* list of leaf cfs_rq on this cpu: */
550 struct list_head leaf_cfs_rq_list;
552 #ifdef CONFIG_RT_GROUP_SCHED
553 struct list_head leaf_rt_rq_list;
557 * This is part of a global counter where only the total sum
558 * over all CPUs matters. A task can increase this counter on
559 * one CPU and if it got migrated afterwards it may decrease
560 * it on another CPU. Always updated under the runqueue lock:
562 unsigned long nr_uninterruptible;
564 struct task_struct *curr, *idle;
565 unsigned long next_balance;
566 struct mm_struct *prev_mm;
573 struct root_domain *rd;
574 struct sched_domain *sd;
576 unsigned char idle_at_tick;
577 /* For active balancing */
581 /* cpu of this runqueue: */
585 unsigned long avg_load_per_task;
587 struct task_struct *migration_thread;
588 struct list_head migration_queue;
596 /* calc_load related fields */
597 unsigned long calc_load_update;
598 long calc_load_active;
600 #ifdef CONFIG_SCHED_HRTICK
602 int hrtick_csd_pending;
603 struct call_single_data hrtick_csd;
605 struct hrtimer hrtick_timer;
608 #ifdef CONFIG_SCHEDSTATS
610 struct sched_info rq_sched_info;
611 unsigned long long rq_cpu_time;
612 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
614 /* sys_sched_yield() stats */
615 unsigned int yld_count;
617 /* schedule() stats */
618 unsigned int sched_switch;
619 unsigned int sched_count;
620 unsigned int sched_goidle;
622 /* try_to_wake_up() stats */
623 unsigned int ttwu_count;
624 unsigned int ttwu_local;
627 unsigned int bkl_count;
631 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
634 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
636 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
639 static inline int cpu_of(struct rq *rq)
649 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
650 * See detach_destroy_domains: synchronize_sched for details.
652 * The domain tree of any CPU may only be accessed from within
653 * preempt-disabled sections.
655 #define for_each_domain(cpu, __sd) \
656 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
658 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
659 #define this_rq() (&__get_cpu_var(runqueues))
660 #define task_rq(p) cpu_rq(task_cpu(p))
661 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
662 #define raw_rq() (&__raw_get_cpu_var(runqueues))
664 inline void update_rq_clock(struct rq *rq)
666 rq->clock = sched_clock_cpu(cpu_of(rq));
670 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
672 #ifdef CONFIG_SCHED_DEBUG
673 # define const_debug __read_mostly
675 # define const_debug static const
680 * @cpu: the processor in question.
682 * Returns true if the current cpu runqueue is locked.
683 * This interface allows printk to be called with the runqueue lock
684 * held and know whether or not it is OK to wake up the klogd.
686 int runqueue_is_locked(int cpu)
688 return spin_is_locked(&cpu_rq(cpu)->lock);
692 * Debugging: various feature bits
695 #define SCHED_FEAT(name, enabled) \
696 __SCHED_FEAT_##name ,
699 #include "sched_features.h"
704 #define SCHED_FEAT(name, enabled) \
705 (1UL << __SCHED_FEAT_##name) * enabled |
707 const_debug unsigned int sysctl_sched_features =
708 #include "sched_features.h"
713 #ifdef CONFIG_SCHED_DEBUG
714 #define SCHED_FEAT(name, enabled) \
717 static __read_mostly char *sched_feat_names[] = {
718 #include "sched_features.h"
724 static int sched_feat_show(struct seq_file *m, void *v)
728 for (i = 0; sched_feat_names[i]; i++) {
729 if (!(sysctl_sched_features & (1UL << i)))
731 seq_printf(m, "%s ", sched_feat_names[i]);
739 sched_feat_write(struct file *filp, const char __user *ubuf,
740 size_t cnt, loff_t *ppos)
750 if (copy_from_user(&buf, ubuf, cnt))
755 if (strncmp(buf, "NO_", 3) == 0) {
760 for (i = 0; sched_feat_names[i]; i++) {
761 int len = strlen(sched_feat_names[i]);
763 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
765 sysctl_sched_features &= ~(1UL << i);
767 sysctl_sched_features |= (1UL << i);
772 if (!sched_feat_names[i])
780 static int sched_feat_open(struct inode *inode, struct file *filp)
782 return single_open(filp, sched_feat_show, NULL);
785 static const struct file_operations sched_feat_fops = {
786 .open = sched_feat_open,
787 .write = sched_feat_write,
790 .release = single_release,
793 static __init int sched_init_debug(void)
795 debugfs_create_file("sched_features", 0644, NULL, NULL,
800 late_initcall(sched_init_debug);
804 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
807 * Number of tasks to iterate in a single balance run.
808 * Limited because this is done with IRQs disabled.
810 const_debug unsigned int sysctl_sched_nr_migrate = 32;
813 * ratelimit for updating the group shares.
816 unsigned int sysctl_sched_shares_ratelimit = 250000;
819 * Inject some fuzzyness into changing the per-cpu group shares
820 * this avoids remote rq-locks at the expense of fairness.
823 unsigned int sysctl_sched_shares_thresh = 4;
826 * period over which we average the RT time consumption, measured
831 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
834 * period over which we measure -rt task cpu usage in us.
837 unsigned int sysctl_sched_rt_period = 1000000;
839 static __read_mostly int scheduler_running;
842 * part of the period that we allow rt tasks to run in us.
845 int sysctl_sched_rt_runtime = 950000;
847 static inline u64 global_rt_period(void)
849 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
852 static inline u64 global_rt_runtime(void)
854 if (sysctl_sched_rt_runtime < 0)
857 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
860 #ifndef prepare_arch_switch
861 # define prepare_arch_switch(next) do { } while (0)
863 #ifndef finish_arch_switch
864 # define finish_arch_switch(prev) do { } while (0)
867 static inline int task_current(struct rq *rq, struct task_struct *p)
869 return rq->curr == p;
872 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
873 static inline int task_running(struct rq *rq, struct task_struct *p)
875 return task_current(rq, p);
878 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
882 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
884 #ifdef CONFIG_DEBUG_SPINLOCK
885 /* this is a valid case when another task releases the spinlock */
886 rq->lock.owner = current;
889 * If we are tracking spinlock dependencies then we have to
890 * fix up the runqueue lock - which gets 'carried over' from
893 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
895 spin_unlock_irq(&rq->lock);
898 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
899 static inline int task_running(struct rq *rq, struct task_struct *p)
904 return task_current(rq, p);
908 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
912 * We can optimise this out completely for !SMP, because the
913 * SMP rebalancing from interrupt is the only thing that cares
918 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
919 spin_unlock_irq(&rq->lock);
921 spin_unlock(&rq->lock);
925 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
929 * After ->oncpu is cleared, the task can be moved to a different CPU.
930 * We must ensure this doesn't happen until the switch is completely
936 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
940 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
943 * __task_rq_lock - lock the runqueue a given task resides on.
944 * Must be called interrupts disabled.
946 static inline struct rq *__task_rq_lock(struct task_struct *p)
950 struct rq *rq = task_rq(p);
951 spin_lock(&rq->lock);
952 if (likely(rq == task_rq(p)))
954 spin_unlock(&rq->lock);
959 * task_rq_lock - lock the runqueue a given task resides on and disable
960 * interrupts. Note the ordering: we can safely lookup the task_rq without
961 * explicitly disabling preemption.
963 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
969 local_irq_save(*flags);
971 spin_lock(&rq->lock);
972 if (likely(rq == task_rq(p)))
974 spin_unlock_irqrestore(&rq->lock, *flags);
978 void task_rq_unlock_wait(struct task_struct *p)
980 struct rq *rq = task_rq(p);
982 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
983 spin_unlock_wait(&rq->lock);
986 static void __task_rq_unlock(struct rq *rq)
989 spin_unlock(&rq->lock);
992 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
995 spin_unlock_irqrestore(&rq->lock, *flags);
999 * this_rq_lock - lock this runqueue and disable interrupts.
1001 static struct rq *this_rq_lock(void)
1002 __acquires(rq->lock)
1006 local_irq_disable();
1008 spin_lock(&rq->lock);
1013 #ifdef CONFIG_SCHED_HRTICK
1015 * Use HR-timers to deliver accurate preemption points.
1017 * Its all a bit involved since we cannot program an hrt while holding the
1018 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1021 * When we get rescheduled we reprogram the hrtick_timer outside of the
1027 * - enabled by features
1028 * - hrtimer is actually high res
1030 static inline int hrtick_enabled(struct rq *rq)
1032 if (!sched_feat(HRTICK))
1034 if (!cpu_active(cpu_of(rq)))
1036 return hrtimer_is_hres_active(&rq->hrtick_timer);
1039 static void hrtick_clear(struct rq *rq)
1041 if (hrtimer_active(&rq->hrtick_timer))
1042 hrtimer_cancel(&rq->hrtick_timer);
1046 * High-resolution timer tick.
1047 * Runs from hardirq context with interrupts disabled.
1049 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1051 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1053 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1055 spin_lock(&rq->lock);
1056 update_rq_clock(rq);
1057 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1058 spin_unlock(&rq->lock);
1060 return HRTIMER_NORESTART;
1065 * called from hardirq (IPI) context
1067 static void __hrtick_start(void *arg)
1069 struct rq *rq = arg;
1071 spin_lock(&rq->lock);
1072 hrtimer_restart(&rq->hrtick_timer);
1073 rq->hrtick_csd_pending = 0;
1074 spin_unlock(&rq->lock);
1078 * Called to set the hrtick timer state.
1080 * called with rq->lock held and irqs disabled
1082 static void hrtick_start(struct rq *rq, u64 delay)
1084 struct hrtimer *timer = &rq->hrtick_timer;
1085 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1087 hrtimer_set_expires(timer, time);
1089 if (rq == this_rq()) {
1090 hrtimer_restart(timer);
1091 } else if (!rq->hrtick_csd_pending) {
1092 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1093 rq->hrtick_csd_pending = 1;
1098 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1100 int cpu = (int)(long)hcpu;
1103 case CPU_UP_CANCELED:
1104 case CPU_UP_CANCELED_FROZEN:
1105 case CPU_DOWN_PREPARE:
1106 case CPU_DOWN_PREPARE_FROZEN:
1108 case CPU_DEAD_FROZEN:
1109 hrtick_clear(cpu_rq(cpu));
1116 static __init void init_hrtick(void)
1118 hotcpu_notifier(hotplug_hrtick, 0);
1122 * Called to set the hrtick timer state.
1124 * called with rq->lock held and irqs disabled
1126 static void hrtick_start(struct rq *rq, u64 delay)
1128 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1129 HRTIMER_MODE_REL_PINNED, 0);
1132 static inline void init_hrtick(void)
1135 #endif /* CONFIG_SMP */
1137 static void init_rq_hrtick(struct rq *rq)
1140 rq->hrtick_csd_pending = 0;
1142 rq->hrtick_csd.flags = 0;
1143 rq->hrtick_csd.func = __hrtick_start;
1144 rq->hrtick_csd.info = rq;
1147 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1148 rq->hrtick_timer.function = hrtick;
1150 #else /* CONFIG_SCHED_HRTICK */
1151 static inline void hrtick_clear(struct rq *rq)
1155 static inline void init_rq_hrtick(struct rq *rq)
1159 static inline void init_hrtick(void)
1162 #endif /* CONFIG_SCHED_HRTICK */
1165 * resched_task - mark a task 'to be rescheduled now'.
1167 * On UP this means the setting of the need_resched flag, on SMP it
1168 * might also involve a cross-CPU call to trigger the scheduler on
1173 #ifndef tsk_is_polling
1174 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1177 static void resched_task(struct task_struct *p)
1181 assert_spin_locked(&task_rq(p)->lock);
1183 if (test_tsk_need_resched(p))
1186 set_tsk_need_resched(p);
1189 if (cpu == smp_processor_id())
1192 /* NEED_RESCHED must be visible before we test polling */
1194 if (!tsk_is_polling(p))
1195 smp_send_reschedule(cpu);
1198 static void resched_cpu(int cpu)
1200 struct rq *rq = cpu_rq(cpu);
1201 unsigned long flags;
1203 if (!spin_trylock_irqsave(&rq->lock, flags))
1205 resched_task(cpu_curr(cpu));
1206 spin_unlock_irqrestore(&rq->lock, flags);
1211 * When add_timer_on() enqueues a timer into the timer wheel of an
1212 * idle CPU then this timer might expire before the next timer event
1213 * which is scheduled to wake up that CPU. In case of a completely
1214 * idle system the next event might even be infinite time into the
1215 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1216 * leaves the inner idle loop so the newly added timer is taken into
1217 * account when the CPU goes back to idle and evaluates the timer
1218 * wheel for the next timer event.
1220 void wake_up_idle_cpu(int cpu)
1222 struct rq *rq = cpu_rq(cpu);
1224 if (cpu == smp_processor_id())
1228 * This is safe, as this function is called with the timer
1229 * wheel base lock of (cpu) held. When the CPU is on the way
1230 * to idle and has not yet set rq->curr to idle then it will
1231 * be serialized on the timer wheel base lock and take the new
1232 * timer into account automatically.
1234 if (rq->curr != rq->idle)
1238 * We can set TIF_RESCHED on the idle task of the other CPU
1239 * lockless. The worst case is that the other CPU runs the
1240 * idle task through an additional NOOP schedule()
1242 set_tsk_need_resched(rq->idle);
1244 /* NEED_RESCHED must be visible before we test polling */
1246 if (!tsk_is_polling(rq->idle))
1247 smp_send_reschedule(cpu);
1249 #endif /* CONFIG_NO_HZ */
1251 static u64 sched_avg_period(void)
1253 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1256 static void sched_avg_update(struct rq *rq)
1258 s64 period = sched_avg_period();
1260 while ((s64)(rq->clock - rq->age_stamp) > period) {
1261 rq->age_stamp += period;
1266 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1268 rq->rt_avg += rt_delta;
1269 sched_avg_update(rq);
1272 #else /* !CONFIG_SMP */
1273 static void resched_task(struct task_struct *p)
1275 assert_spin_locked(&task_rq(p)->lock);
1276 set_tsk_need_resched(p);
1279 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1282 #endif /* CONFIG_SMP */
1284 #if BITS_PER_LONG == 32
1285 # define WMULT_CONST (~0UL)
1287 # define WMULT_CONST (1UL << 32)
1290 #define WMULT_SHIFT 32
1293 * Shift right and round:
1295 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1298 * delta *= weight / lw
1300 static unsigned long
1301 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1302 struct load_weight *lw)
1306 if (!lw->inv_weight) {
1307 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1310 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1314 tmp = (u64)delta_exec * weight;
1316 * Check whether we'd overflow the 64-bit multiplication:
1318 if (unlikely(tmp > WMULT_CONST))
1319 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1322 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1324 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1327 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1333 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1340 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1341 * of tasks with abnormal "nice" values across CPUs the contribution that
1342 * each task makes to its run queue's load is weighted according to its
1343 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1344 * scaled version of the new time slice allocation that they receive on time
1348 #define WEIGHT_IDLEPRIO 3
1349 #define WMULT_IDLEPRIO 1431655765
1352 * Nice levels are multiplicative, with a gentle 10% change for every
1353 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1354 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1355 * that remained on nice 0.
1357 * The "10% effect" is relative and cumulative: from _any_ nice level,
1358 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1359 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1360 * If a task goes up by ~10% and another task goes down by ~10% then
1361 * the relative distance between them is ~25%.)
1363 static const int prio_to_weight[40] = {
1364 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1365 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1366 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1367 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1368 /* 0 */ 1024, 820, 655, 526, 423,
1369 /* 5 */ 335, 272, 215, 172, 137,
1370 /* 10 */ 110, 87, 70, 56, 45,
1371 /* 15 */ 36, 29, 23, 18, 15,
1375 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1377 * In cases where the weight does not change often, we can use the
1378 * precalculated inverse to speed up arithmetics by turning divisions
1379 * into multiplications:
1381 static const u32 prio_to_wmult[40] = {
1382 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1383 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1384 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1385 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1386 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1387 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1388 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1389 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1392 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1395 * runqueue iterator, to support SMP load-balancing between different
1396 * scheduling classes, without having to expose their internal data
1397 * structures to the load-balancing proper:
1399 struct rq_iterator {
1401 struct task_struct *(*start)(void *);
1402 struct task_struct *(*next)(void *);
1406 static unsigned long
1407 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1408 unsigned long max_load_move, struct sched_domain *sd,
1409 enum cpu_idle_type idle, int *all_pinned,
1410 int *this_best_prio, struct rq_iterator *iterator);
1413 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1414 struct sched_domain *sd, enum cpu_idle_type idle,
1415 struct rq_iterator *iterator);
1418 /* Time spent by the tasks of the cpu accounting group executing in ... */
1419 enum cpuacct_stat_index {
1420 CPUACCT_STAT_USER, /* ... user mode */
1421 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1423 CPUACCT_STAT_NSTATS,
1426 #ifdef CONFIG_CGROUP_CPUACCT
1427 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1428 static void cpuacct_update_stats(struct task_struct *tsk,
1429 enum cpuacct_stat_index idx, cputime_t val);
1431 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1432 static inline void cpuacct_update_stats(struct task_struct *tsk,
1433 enum cpuacct_stat_index idx, cputime_t val) {}
1436 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1438 update_load_add(&rq->load, load);
1441 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1443 update_load_sub(&rq->load, load);
1446 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1447 typedef int (*tg_visitor)(struct task_group *, void *);
1450 * Iterate the full tree, calling @down when first entering a node and @up when
1451 * leaving it for the final time.
1453 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1455 struct task_group *parent, *child;
1459 parent = &root_task_group;
1461 ret = (*down)(parent, data);
1464 list_for_each_entry_rcu(child, &parent->children, siblings) {
1471 ret = (*up)(parent, data);
1476 parent = parent->parent;
1485 static int tg_nop(struct task_group *tg, void *data)
1492 /* Used instead of source_load when we know the type == 0 */
1493 static unsigned long weighted_cpuload(const int cpu)
1495 return cpu_rq(cpu)->load.weight;
1499 * Return a low guess at the load of a migration-source cpu weighted
1500 * according to the scheduling class and "nice" value.
1502 * We want to under-estimate the load of migration sources, to
1503 * balance conservatively.
1505 static unsigned long source_load(int cpu, int type)
1507 struct rq *rq = cpu_rq(cpu);
1508 unsigned long total = weighted_cpuload(cpu);
1510 if (type == 0 || !sched_feat(LB_BIAS))
1513 return min(rq->cpu_load[type-1], total);
1517 * Return a high guess at the load of a migration-target cpu weighted
1518 * according to the scheduling class and "nice" value.
1520 static unsigned long target_load(int cpu, int type)
1522 struct rq *rq = cpu_rq(cpu);
1523 unsigned long total = weighted_cpuload(cpu);
1525 if (type == 0 || !sched_feat(LB_BIAS))
1528 return max(rq->cpu_load[type-1], total);
1531 static struct sched_group *group_of(int cpu)
1533 struct sched_domain *sd = rcu_dereference(cpu_rq(cpu)->sd);
1541 static unsigned long power_of(int cpu)
1543 struct sched_group *group = group_of(cpu);
1546 return SCHED_LOAD_SCALE;
1548 return group->cpu_power;
1551 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1553 static unsigned long cpu_avg_load_per_task(int cpu)
1555 struct rq *rq = cpu_rq(cpu);
1556 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1559 rq->avg_load_per_task = rq->load.weight / nr_running;
1561 rq->avg_load_per_task = 0;
1563 return rq->avg_load_per_task;
1566 #ifdef CONFIG_FAIR_GROUP_SCHED
1568 static __read_mostly unsigned long *update_shares_data;
1570 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1573 * Calculate and set the cpu's group shares.
1575 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1576 unsigned long sd_shares,
1577 unsigned long sd_rq_weight,
1578 unsigned long *usd_rq_weight)
1580 unsigned long shares, rq_weight;
1583 rq_weight = usd_rq_weight[cpu];
1586 rq_weight = NICE_0_LOAD;
1590 * \Sum_j shares_j * rq_weight_i
1591 * shares_i = -----------------------------
1592 * \Sum_j rq_weight_j
1594 shares = (sd_shares * rq_weight) / sd_rq_weight;
1595 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1597 if (abs(shares - tg->se[cpu]->load.weight) >
1598 sysctl_sched_shares_thresh) {
1599 struct rq *rq = cpu_rq(cpu);
1600 unsigned long flags;
1602 spin_lock_irqsave(&rq->lock, flags);
1603 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1604 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1605 __set_se_shares(tg->se[cpu], shares);
1606 spin_unlock_irqrestore(&rq->lock, flags);
1611 * Re-compute the task group their per cpu shares over the given domain.
1612 * This needs to be done in a bottom-up fashion because the rq weight of a
1613 * parent group depends on the shares of its child groups.
1615 static int tg_shares_up(struct task_group *tg, void *data)
1617 unsigned long weight, rq_weight = 0, sum_weight = 0, shares = 0;
1618 unsigned long *usd_rq_weight;
1619 struct sched_domain *sd = data;
1620 unsigned long flags;
1626 local_irq_save(flags);
1627 usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id());
1629 for_each_cpu(i, sched_domain_span(sd)) {
1630 weight = tg->cfs_rq[i]->load.weight;
1631 usd_rq_weight[i] = weight;
1633 rq_weight += weight;
1635 * If there are currently no tasks on the cpu pretend there
1636 * is one of average load so that when a new task gets to
1637 * run here it will not get delayed by group starvation.
1640 weight = NICE_0_LOAD;
1642 sum_weight += weight;
1643 shares += tg->cfs_rq[i]->shares;
1647 rq_weight = sum_weight;
1649 if ((!shares && rq_weight) || shares > tg->shares)
1650 shares = tg->shares;
1652 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1653 shares = tg->shares;
1655 for_each_cpu(i, sched_domain_span(sd))
1656 update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight);
1658 local_irq_restore(flags);
1664 * Compute the cpu's hierarchical load factor for each task group.
1665 * This needs to be done in a top-down fashion because the load of a child
1666 * group is a fraction of its parents load.
1668 static int tg_load_down(struct task_group *tg, void *data)
1671 long cpu = (long)data;
1674 load = cpu_rq(cpu)->load.weight;
1676 load = tg->parent->cfs_rq[cpu]->h_load;
1677 load *= tg->cfs_rq[cpu]->shares;
1678 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1681 tg->cfs_rq[cpu]->h_load = load;
1686 static void update_shares(struct sched_domain *sd)
1691 if (root_task_group_empty())
1694 now = cpu_clock(raw_smp_processor_id());
1695 elapsed = now - sd->last_update;
1697 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1698 sd->last_update = now;
1699 walk_tg_tree(tg_nop, tg_shares_up, sd);
1703 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1705 if (root_task_group_empty())
1708 spin_unlock(&rq->lock);
1710 spin_lock(&rq->lock);
1713 static void update_h_load(long cpu)
1715 if (root_task_group_empty())
1718 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1723 static inline void update_shares(struct sched_domain *sd)
1727 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1733 #ifdef CONFIG_PREEMPT
1735 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1738 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1739 * way at the expense of forcing extra atomic operations in all
1740 * invocations. This assures that the double_lock is acquired using the
1741 * same underlying policy as the spinlock_t on this architecture, which
1742 * reduces latency compared to the unfair variant below. However, it
1743 * also adds more overhead and therefore may reduce throughput.
1745 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1746 __releases(this_rq->lock)
1747 __acquires(busiest->lock)
1748 __acquires(this_rq->lock)
1750 spin_unlock(&this_rq->lock);
1751 double_rq_lock(this_rq, busiest);
1758 * Unfair double_lock_balance: Optimizes throughput at the expense of
1759 * latency by eliminating extra atomic operations when the locks are
1760 * already in proper order on entry. This favors lower cpu-ids and will
1761 * grant the double lock to lower cpus over higher ids under contention,
1762 * regardless of entry order into the function.
1764 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1765 __releases(this_rq->lock)
1766 __acquires(busiest->lock)
1767 __acquires(this_rq->lock)
1771 if (unlikely(!spin_trylock(&busiest->lock))) {
1772 if (busiest < this_rq) {
1773 spin_unlock(&this_rq->lock);
1774 spin_lock(&busiest->lock);
1775 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
1778 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
1783 #endif /* CONFIG_PREEMPT */
1786 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1788 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1790 if (unlikely(!irqs_disabled())) {
1791 /* printk() doesn't work good under rq->lock */
1792 spin_unlock(&this_rq->lock);
1796 return _double_lock_balance(this_rq, busiest);
1799 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1800 __releases(busiest->lock)
1802 spin_unlock(&busiest->lock);
1803 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1807 #ifdef CONFIG_FAIR_GROUP_SCHED
1808 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1811 cfs_rq->shares = shares;
1816 static void calc_load_account_active(struct rq *this_rq);
1818 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1820 set_task_rq(p, cpu);
1823 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1824 * successfuly executed on another CPU. We must ensure that updates of
1825 * per-task data have been completed by this moment.
1828 task_thread_info(p)->cpu = cpu;
1832 #include "sched_stats.h"
1833 #include "sched_idletask.c"
1834 #include "sched_fair.c"
1835 #include "sched_rt.c"
1836 #ifdef CONFIG_SCHED_DEBUG
1837 # include "sched_debug.c"
1840 #define sched_class_highest (&rt_sched_class)
1841 #define for_each_class(class) \
1842 for (class = sched_class_highest; class; class = class->next)
1844 static void inc_nr_running(struct rq *rq)
1849 static void dec_nr_running(struct rq *rq)
1854 static void set_load_weight(struct task_struct *p)
1856 if (task_has_rt_policy(p)) {
1857 p->se.load.weight = prio_to_weight[0] * 2;
1858 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1863 * SCHED_IDLE tasks get minimal weight:
1865 if (p->policy == SCHED_IDLE) {
1866 p->se.load.weight = WEIGHT_IDLEPRIO;
1867 p->se.load.inv_weight = WMULT_IDLEPRIO;
1871 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1872 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1875 static void update_avg(u64 *avg, u64 sample)
1877 s64 diff = sample - *avg;
1881 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1884 p->se.start_runtime = p->se.sum_exec_runtime;
1886 sched_info_queued(p);
1887 p->sched_class->enqueue_task(rq, p, wakeup);
1891 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1894 if (p->se.last_wakeup) {
1895 update_avg(&p->se.avg_overlap,
1896 p->se.sum_exec_runtime - p->se.last_wakeup);
1897 p->se.last_wakeup = 0;
1899 update_avg(&p->se.avg_wakeup,
1900 sysctl_sched_wakeup_granularity);
1904 sched_info_dequeued(p);
1905 p->sched_class->dequeue_task(rq, p, sleep);
1910 * __normal_prio - return the priority that is based on the static prio
1912 static inline int __normal_prio(struct task_struct *p)
1914 return p->static_prio;
1918 * Calculate the expected normal priority: i.e. priority
1919 * without taking RT-inheritance into account. Might be
1920 * boosted by interactivity modifiers. Changes upon fork,
1921 * setprio syscalls, and whenever the interactivity
1922 * estimator recalculates.
1924 static inline int normal_prio(struct task_struct *p)
1928 if (task_has_rt_policy(p))
1929 prio = MAX_RT_PRIO-1 - p->rt_priority;
1931 prio = __normal_prio(p);
1936 * Calculate the current priority, i.e. the priority
1937 * taken into account by the scheduler. This value might
1938 * be boosted by RT tasks, or might be boosted by
1939 * interactivity modifiers. Will be RT if the task got
1940 * RT-boosted. If not then it returns p->normal_prio.
1942 static int effective_prio(struct task_struct *p)
1944 p->normal_prio = normal_prio(p);
1946 * If we are RT tasks or we were boosted to RT priority,
1947 * keep the priority unchanged. Otherwise, update priority
1948 * to the normal priority:
1950 if (!rt_prio(p->prio))
1951 return p->normal_prio;
1956 * activate_task - move a task to the runqueue.
1958 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1960 if (task_contributes_to_load(p))
1961 rq->nr_uninterruptible--;
1963 enqueue_task(rq, p, wakeup);
1968 * deactivate_task - remove a task from the runqueue.
1970 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1972 if (task_contributes_to_load(p))
1973 rq->nr_uninterruptible++;
1975 dequeue_task(rq, p, sleep);
1980 * task_curr - is this task currently executing on a CPU?
1981 * @p: the task in question.
1983 inline int task_curr(const struct task_struct *p)
1985 return cpu_curr(task_cpu(p)) == p;
1988 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1989 const struct sched_class *prev_class,
1990 int oldprio, int running)
1992 if (prev_class != p->sched_class) {
1993 if (prev_class->switched_from)
1994 prev_class->switched_from(rq, p, running);
1995 p->sched_class->switched_to(rq, p, running);
1997 p->sched_class->prio_changed(rq, p, oldprio, running);
2001 * kthread_bind - bind a just-created kthread to a cpu.
2002 * @p: thread created by kthread_create().
2003 * @cpu: cpu (might not be online, must be possible) for @k to run on.
2005 * Description: This function is equivalent to set_cpus_allowed(),
2006 * except that @cpu doesn't need to be online, and the thread must be
2007 * stopped (i.e., just returned from kthread_create()).
2009 * Function lives here instead of kthread.c because it messes with
2010 * scheduler internals which require locking.
2012 void kthread_bind(struct task_struct *p, unsigned int cpu)
2014 struct rq *rq = cpu_rq(cpu);
2015 unsigned long flags;
2017 /* Must have done schedule() in kthread() before we set_task_cpu */
2018 if (!wait_task_inactive(p, TASK_UNINTERRUPTIBLE)) {
2023 spin_lock_irqsave(&rq->lock, flags);
2024 update_rq_clock(rq);
2025 set_task_cpu(p, cpu);
2026 p->cpus_allowed = cpumask_of_cpu(cpu);
2027 p->rt.nr_cpus_allowed = 1;
2028 p->flags |= PF_THREAD_BOUND;
2029 spin_unlock_irqrestore(&rq->lock, flags);
2031 EXPORT_SYMBOL(kthread_bind);
2035 * Is this task likely cache-hot:
2038 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2043 * Buddy candidates are cache hot:
2045 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2046 (&p->se == cfs_rq_of(&p->se)->next ||
2047 &p->se == cfs_rq_of(&p->se)->last))
2050 if (p->sched_class != &fair_sched_class)
2053 if (sysctl_sched_migration_cost == -1)
2055 if (sysctl_sched_migration_cost == 0)
2058 delta = now - p->se.exec_start;
2060 return delta < (s64)sysctl_sched_migration_cost;
2064 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2066 int old_cpu = task_cpu(p);
2067 struct rq *old_rq = cpu_rq(old_cpu);
2068 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
2069 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
2071 trace_sched_migrate_task(p, new_cpu);
2073 if (old_cpu != new_cpu) {
2074 p->se.nr_migrations++;
2075 #ifdef CONFIG_SCHEDSTATS
2076 if (task_hot(p, old_rq->clock, NULL))
2077 schedstat_inc(p, se.nr_forced2_migrations);
2079 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS,
2082 p->se.vruntime -= old_cfsrq->min_vruntime -
2083 new_cfsrq->min_vruntime;
2085 __set_task_cpu(p, new_cpu);
2088 struct migration_req {
2089 struct list_head list;
2091 struct task_struct *task;
2094 struct completion done;
2098 * The task's runqueue lock must be held.
2099 * Returns true if you have to wait for migration thread.
2102 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2104 struct rq *rq = task_rq(p);
2107 * If the task is not on a runqueue (and not running), then
2108 * it is sufficient to simply update the task's cpu field.
2110 if (!p->se.on_rq && !task_running(rq, p)) {
2111 update_rq_clock(rq);
2112 set_task_cpu(p, dest_cpu);
2116 init_completion(&req->done);
2118 req->dest_cpu = dest_cpu;
2119 list_add(&req->list, &rq->migration_queue);
2125 * wait_task_context_switch - wait for a thread to complete at least one
2128 * @p must not be current.
2130 void wait_task_context_switch(struct task_struct *p)
2132 unsigned long nvcsw, nivcsw, flags;
2140 * The runqueue is assigned before the actual context
2141 * switch. We need to take the runqueue lock.
2143 * We could check initially without the lock but it is
2144 * very likely that we need to take the lock in every
2147 rq = task_rq_lock(p, &flags);
2148 running = task_running(rq, p);
2149 task_rq_unlock(rq, &flags);
2151 if (likely(!running))
2154 * The switch count is incremented before the actual
2155 * context switch. We thus wait for two switches to be
2156 * sure at least one completed.
2158 if ((p->nvcsw - nvcsw) > 1)
2160 if ((p->nivcsw - nivcsw) > 1)
2168 * wait_task_inactive - wait for a thread to unschedule.
2170 * If @match_state is nonzero, it's the @p->state value just checked and
2171 * not expected to change. If it changes, i.e. @p might have woken up,
2172 * then return zero. When we succeed in waiting for @p to be off its CPU,
2173 * we return a positive number (its total switch count). If a second call
2174 * a short while later returns the same number, the caller can be sure that
2175 * @p has remained unscheduled the whole time.
2177 * The caller must ensure that the task *will* unschedule sometime soon,
2178 * else this function might spin for a *long* time. This function can't
2179 * be called with interrupts off, or it may introduce deadlock with
2180 * smp_call_function() if an IPI is sent by the same process we are
2181 * waiting to become inactive.
2183 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2185 unsigned long flags;
2192 * We do the initial early heuristics without holding
2193 * any task-queue locks at all. We'll only try to get
2194 * the runqueue lock when things look like they will
2200 * If the task is actively running on another CPU
2201 * still, just relax and busy-wait without holding
2204 * NOTE! Since we don't hold any locks, it's not
2205 * even sure that "rq" stays as the right runqueue!
2206 * But we don't care, since "task_running()" will
2207 * return false if the runqueue has changed and p
2208 * is actually now running somewhere else!
2210 while (task_running(rq, p)) {
2211 if (match_state && unlikely(p->state != match_state))
2217 * Ok, time to look more closely! We need the rq
2218 * lock now, to be *sure*. If we're wrong, we'll
2219 * just go back and repeat.
2221 rq = task_rq_lock(p, &flags);
2222 trace_sched_wait_task(rq, p);
2223 running = task_running(rq, p);
2224 on_rq = p->se.on_rq;
2226 if (!match_state || p->state == match_state)
2227 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2228 task_rq_unlock(rq, &flags);
2231 * If it changed from the expected state, bail out now.
2233 if (unlikely(!ncsw))
2237 * Was it really running after all now that we
2238 * checked with the proper locks actually held?
2240 * Oops. Go back and try again..
2242 if (unlikely(running)) {
2248 * It's not enough that it's not actively running,
2249 * it must be off the runqueue _entirely_, and not
2252 * So if it was still runnable (but just not actively
2253 * running right now), it's preempted, and we should
2254 * yield - it could be a while.
2256 if (unlikely(on_rq)) {
2257 schedule_timeout_uninterruptible(1);
2262 * Ahh, all good. It wasn't running, and it wasn't
2263 * runnable, which means that it will never become
2264 * running in the future either. We're all done!
2273 * kick_process - kick a running thread to enter/exit the kernel
2274 * @p: the to-be-kicked thread
2276 * Cause a process which is running on another CPU to enter
2277 * kernel-mode, without any delay. (to get signals handled.)
2279 * NOTE: this function doesnt have to take the runqueue lock,
2280 * because all it wants to ensure is that the remote task enters
2281 * the kernel. If the IPI races and the task has been migrated
2282 * to another CPU then no harm is done and the purpose has been
2285 void kick_process(struct task_struct *p)
2291 if ((cpu != smp_processor_id()) && task_curr(p))
2292 smp_send_reschedule(cpu);
2295 EXPORT_SYMBOL_GPL(kick_process);
2296 #endif /* CONFIG_SMP */
2299 * task_oncpu_function_call - call a function on the cpu on which a task runs
2300 * @p: the task to evaluate
2301 * @func: the function to be called
2302 * @info: the function call argument
2304 * Calls the function @func when the task is currently running. This might
2305 * be on the current CPU, which just calls the function directly
2307 void task_oncpu_function_call(struct task_struct *p,
2308 void (*func) (void *info), void *info)
2315 smp_call_function_single(cpu, func, info, 1);
2321 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2323 return p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2328 * try_to_wake_up - wake up a thread
2329 * @p: the to-be-woken-up thread
2330 * @state: the mask of task states that can be woken
2331 * @sync: do a synchronous wakeup?
2333 * Put it on the run-queue if it's not already there. The "current"
2334 * thread is always on the run-queue (except when the actual
2335 * re-schedule is in progress), and as such you're allowed to do
2336 * the simpler "current->state = TASK_RUNNING" to mark yourself
2337 * runnable without the overhead of this.
2339 * returns failure only if the task is already active.
2341 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2344 int cpu, orig_cpu, this_cpu, success = 0;
2345 unsigned long flags;
2346 struct rq *rq, *orig_rq;
2348 if (!sched_feat(SYNC_WAKEUPS))
2349 wake_flags &= ~WF_SYNC;
2351 this_cpu = get_cpu();
2354 rq = orig_rq = task_rq_lock(p, &flags);
2355 update_rq_clock(rq);
2356 if (!(p->state & state))
2366 if (unlikely(task_running(rq, p)))
2370 * In order to handle concurrent wakeups and release the rq->lock
2371 * we put the task in TASK_WAKING state.
2373 * First fix up the nr_uninterruptible count:
2375 if (task_contributes_to_load(p))
2376 rq->nr_uninterruptible--;
2377 p->state = TASK_WAKING;
2378 __task_rq_unlock(rq);
2380 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2381 if (cpu != orig_cpu)
2382 set_task_cpu(p, cpu);
2384 rq = __task_rq_lock(p);
2385 update_rq_clock(rq);
2387 WARN_ON(p->state != TASK_WAKING);
2390 #ifdef CONFIG_SCHEDSTATS
2391 schedstat_inc(rq, ttwu_count);
2392 if (cpu == this_cpu)
2393 schedstat_inc(rq, ttwu_local);
2395 struct sched_domain *sd;
2396 for_each_domain(this_cpu, sd) {
2397 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2398 schedstat_inc(sd, ttwu_wake_remote);
2403 #endif /* CONFIG_SCHEDSTATS */
2406 #endif /* CONFIG_SMP */
2407 schedstat_inc(p, se.nr_wakeups);
2408 if (wake_flags & WF_SYNC)
2409 schedstat_inc(p, se.nr_wakeups_sync);
2410 if (orig_cpu != cpu)
2411 schedstat_inc(p, se.nr_wakeups_migrate);
2412 if (cpu == this_cpu)
2413 schedstat_inc(p, se.nr_wakeups_local);
2415 schedstat_inc(p, se.nr_wakeups_remote);
2416 activate_task(rq, p, 1);
2420 * Only attribute actual wakeups done by this task.
2422 if (!in_interrupt()) {
2423 struct sched_entity *se = ¤t->se;
2424 u64 sample = se->sum_exec_runtime;
2426 if (se->last_wakeup)
2427 sample -= se->last_wakeup;
2429 sample -= se->start_runtime;
2430 update_avg(&se->avg_wakeup, sample);
2432 se->last_wakeup = se->sum_exec_runtime;
2436 trace_sched_wakeup(rq, p, success);
2437 check_preempt_curr(rq, p, wake_flags);
2439 p->state = TASK_RUNNING;
2441 if (p->sched_class->task_wake_up)
2442 p->sched_class->task_wake_up(rq, p);
2444 if (unlikely(rq->idle_stamp)) {
2445 u64 delta = rq->clock - rq->idle_stamp;
2446 u64 max = 2*sysctl_sched_migration_cost;
2451 update_avg(&rq->avg_idle, delta);
2456 task_rq_unlock(rq, &flags);
2463 * wake_up_process - Wake up a specific process
2464 * @p: The process to be woken up.
2466 * Attempt to wake up the nominated process and move it to the set of runnable
2467 * processes. Returns 1 if the process was woken up, 0 if it was already
2470 * It may be assumed that this function implies a write memory barrier before
2471 * changing the task state if and only if any tasks are woken up.
2473 int wake_up_process(struct task_struct *p)
2475 return try_to_wake_up(p, TASK_ALL, 0);
2477 EXPORT_SYMBOL(wake_up_process);
2479 int wake_up_state(struct task_struct *p, unsigned int state)
2481 return try_to_wake_up(p, state, 0);
2485 * Perform scheduler related setup for a newly forked process p.
2486 * p is forked by current.
2488 * __sched_fork() is basic setup used by init_idle() too:
2490 static void __sched_fork(struct task_struct *p)
2492 p->se.exec_start = 0;
2493 p->se.sum_exec_runtime = 0;
2494 p->se.prev_sum_exec_runtime = 0;
2495 p->se.nr_migrations = 0;
2496 p->se.last_wakeup = 0;
2497 p->se.avg_overlap = 0;
2498 p->se.start_runtime = 0;
2499 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2501 #ifdef CONFIG_SCHEDSTATS
2502 p->se.wait_start = 0;
2504 p->se.wait_count = 0;
2507 p->se.sleep_start = 0;
2508 p->se.sleep_max = 0;
2509 p->se.sum_sleep_runtime = 0;
2511 p->se.block_start = 0;
2512 p->se.block_max = 0;
2514 p->se.slice_max = 0;
2516 p->se.nr_migrations_cold = 0;
2517 p->se.nr_failed_migrations_affine = 0;
2518 p->se.nr_failed_migrations_running = 0;
2519 p->se.nr_failed_migrations_hot = 0;
2520 p->se.nr_forced_migrations = 0;
2521 p->se.nr_forced2_migrations = 0;
2523 p->se.nr_wakeups = 0;
2524 p->se.nr_wakeups_sync = 0;
2525 p->se.nr_wakeups_migrate = 0;
2526 p->se.nr_wakeups_local = 0;
2527 p->se.nr_wakeups_remote = 0;
2528 p->se.nr_wakeups_affine = 0;
2529 p->se.nr_wakeups_affine_attempts = 0;
2530 p->se.nr_wakeups_passive = 0;
2531 p->se.nr_wakeups_idle = 0;
2535 INIT_LIST_HEAD(&p->rt.run_list);
2537 INIT_LIST_HEAD(&p->se.group_node);
2539 #ifdef CONFIG_PREEMPT_NOTIFIERS
2540 INIT_HLIST_HEAD(&p->preempt_notifiers);
2544 * We mark the process as running here, but have not actually
2545 * inserted it onto the runqueue yet. This guarantees that
2546 * nobody will actually run it, and a signal or other external
2547 * event cannot wake it up and insert it on the runqueue either.
2549 p->state = TASK_RUNNING;
2553 * fork()/clone()-time setup:
2555 void sched_fork(struct task_struct *p, int clone_flags)
2557 int cpu = get_cpu();
2562 * Revert to default priority/policy on fork if requested.
2564 if (unlikely(p->sched_reset_on_fork)) {
2565 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2566 p->policy = SCHED_NORMAL;
2567 p->normal_prio = p->static_prio;
2570 if (PRIO_TO_NICE(p->static_prio) < 0) {
2571 p->static_prio = NICE_TO_PRIO(0);
2572 p->normal_prio = p->static_prio;
2577 * We don't need the reset flag anymore after the fork. It has
2578 * fulfilled its duty:
2580 p->sched_reset_on_fork = 0;
2584 * Make sure we do not leak PI boosting priority to the child.
2586 p->prio = current->normal_prio;
2588 if (!rt_prio(p->prio))
2589 p->sched_class = &fair_sched_class;
2591 if (p->sched_class->task_fork)
2592 p->sched_class->task_fork(p);
2595 cpu = select_task_rq(p, SD_BALANCE_FORK, 0);
2597 set_task_cpu(p, cpu);
2599 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2600 if (likely(sched_info_on()))
2601 memset(&p->sched_info, 0, sizeof(p->sched_info));
2603 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2606 #ifdef CONFIG_PREEMPT
2607 /* Want to start with kernel preemption disabled. */
2608 task_thread_info(p)->preempt_count = 1;
2610 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2616 * wake_up_new_task - wake up a newly created task for the first time.
2618 * This function will do some initial scheduler statistics housekeeping
2619 * that must be done for every newly created context, then puts the task
2620 * on the runqueue and wakes it.
2622 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2624 unsigned long flags;
2627 rq = task_rq_lock(p, &flags);
2628 BUG_ON(p->state != TASK_RUNNING);
2629 update_rq_clock(rq);
2630 activate_task(rq, p, 0);
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);
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 (likely(!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 * Update rq->cpu_load[] statistics. This function is usually called every
3025 * scheduler tick (TICK_NSEC).
3027 static void update_cpu_load(struct rq *this_rq)
3029 unsigned long this_load = this_rq->load.weight;
3032 this_rq->nr_load_updates++;
3034 /* Update our load: */
3035 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3036 unsigned long old_load, new_load;
3038 /* scale is effectively 1 << i now, and >> i divides by scale */
3040 old_load = this_rq->cpu_load[i];
3041 new_load = this_load;
3043 * Round up the averaging division if load is increasing. This
3044 * prevents us from getting stuck on 9 if the load is 10, for
3047 if (new_load > old_load)
3048 new_load += scale-1;
3049 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3052 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3053 this_rq->calc_load_update += LOAD_FREQ;
3054 calc_load_account_active(this_rq);
3061 * double_rq_lock - safely lock two runqueues
3063 * Note this does not disable interrupts like task_rq_lock,
3064 * you need to do so manually before calling.
3066 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
3067 __acquires(rq1->lock)
3068 __acquires(rq2->lock)
3070 BUG_ON(!irqs_disabled());
3072 spin_lock(&rq1->lock);
3073 __acquire(rq2->lock); /* Fake it out ;) */
3076 spin_lock(&rq1->lock);
3077 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
3079 spin_lock(&rq2->lock);
3080 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
3083 update_rq_clock(rq1);
3084 update_rq_clock(rq2);
3088 * double_rq_unlock - safely unlock two runqueues
3090 * Note this does not restore interrupts like task_rq_unlock,
3091 * you need to do so manually after calling.
3093 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
3094 __releases(rq1->lock)
3095 __releases(rq2->lock)
3097 spin_unlock(&rq1->lock);
3099 spin_unlock(&rq2->lock);
3101 __release(rq2->lock);
3105 * If dest_cpu is allowed for this process, migrate the task to it.
3106 * This is accomplished by forcing the cpu_allowed mask to only
3107 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3108 * the cpu_allowed mask is restored.
3110 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
3112 struct migration_req req;
3113 unsigned long flags;
3116 rq = task_rq_lock(p, &flags);
3117 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
3118 || unlikely(!cpu_active(dest_cpu)))
3121 /* force the process onto the specified CPU */
3122 if (migrate_task(p, dest_cpu, &req)) {
3123 /* Need to wait for migration thread (might exit: take ref). */
3124 struct task_struct *mt = rq->migration_thread;
3126 get_task_struct(mt);
3127 task_rq_unlock(rq, &flags);
3128 wake_up_process(mt);
3129 put_task_struct(mt);
3130 wait_for_completion(&req.done);
3135 task_rq_unlock(rq, &flags);
3139 * sched_exec - execve() is a valuable balancing opportunity, because at
3140 * this point the task has the smallest effective memory and cache footprint.
3142 void sched_exec(void)
3144 int new_cpu, this_cpu = get_cpu();
3145 new_cpu = select_task_rq(current, SD_BALANCE_EXEC, 0);
3147 if (new_cpu != this_cpu)
3148 sched_migrate_task(current, new_cpu);
3152 * pull_task - move a task from a remote runqueue to the local runqueue.
3153 * Both runqueues must be locked.
3155 static void pull_task(struct rq *src_rq, struct task_struct *p,
3156 struct rq *this_rq, int this_cpu)
3158 deactivate_task(src_rq, p, 0);
3159 set_task_cpu(p, this_cpu);
3160 activate_task(this_rq, p, 0);
3162 * Note that idle threads have a prio of MAX_PRIO, for this test
3163 * to be always true for them.
3165 check_preempt_curr(this_rq, p, 0);
3169 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3172 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3173 struct sched_domain *sd, enum cpu_idle_type idle,
3176 int tsk_cache_hot = 0;
3178 * We do not migrate tasks that are:
3179 * 1) running (obviously), or
3180 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3181 * 3) are cache-hot on their current CPU.
3183 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
3184 schedstat_inc(p, se.nr_failed_migrations_affine);
3189 if (task_running(rq, p)) {
3190 schedstat_inc(p, se.nr_failed_migrations_running);
3195 * Aggressive migration if:
3196 * 1) task is cache cold, or
3197 * 2) too many balance attempts have failed.
3200 tsk_cache_hot = task_hot(p, rq->clock, sd);
3201 if (!tsk_cache_hot ||
3202 sd->nr_balance_failed > sd->cache_nice_tries) {
3203 #ifdef CONFIG_SCHEDSTATS
3204 if (tsk_cache_hot) {
3205 schedstat_inc(sd, lb_hot_gained[idle]);
3206 schedstat_inc(p, se.nr_forced_migrations);
3212 if (tsk_cache_hot) {
3213 schedstat_inc(p, se.nr_failed_migrations_hot);
3219 static unsigned long
3220 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3221 unsigned long max_load_move, struct sched_domain *sd,
3222 enum cpu_idle_type idle, int *all_pinned,
3223 int *this_best_prio, struct rq_iterator *iterator)
3225 int loops = 0, pulled = 0, pinned = 0;
3226 struct task_struct *p;
3227 long rem_load_move = max_load_move;
3229 if (max_load_move == 0)
3235 * Start the load-balancing iterator:
3237 p = iterator->start(iterator->arg);
3239 if (!p || loops++ > sysctl_sched_nr_migrate)
3242 if ((p->se.load.weight >> 1) > rem_load_move ||
3243 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3244 p = iterator->next(iterator->arg);
3248 pull_task(busiest, p, this_rq, this_cpu);
3250 rem_load_move -= p->se.load.weight;
3252 #ifdef CONFIG_PREEMPT
3254 * NEWIDLE balancing is a source of latency, so preemptible kernels
3255 * will stop after the first task is pulled to minimize the critical
3258 if (idle == CPU_NEWLY_IDLE)
3263 * We only want to steal up to the prescribed amount of weighted load.
3265 if (rem_load_move > 0) {
3266 if (p->prio < *this_best_prio)
3267 *this_best_prio = p->prio;
3268 p = iterator->next(iterator->arg);
3273 * Right now, this is one of only two places pull_task() is called,
3274 * so we can safely collect pull_task() stats here rather than
3275 * inside pull_task().
3277 schedstat_add(sd, lb_gained[idle], pulled);
3280 *all_pinned = pinned;
3282 return max_load_move - rem_load_move;
3286 * move_tasks tries to move up to max_load_move weighted load from busiest to
3287 * this_rq, as part of a balancing operation within domain "sd".
3288 * Returns 1 if successful and 0 otherwise.
3290 * Called with both runqueues locked.
3292 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3293 unsigned long max_load_move,
3294 struct sched_domain *sd, enum cpu_idle_type idle,
3297 const struct sched_class *class = sched_class_highest;
3298 unsigned long total_load_moved = 0;
3299 int this_best_prio = this_rq->curr->prio;
3303 class->load_balance(this_rq, this_cpu, busiest,
3304 max_load_move - total_load_moved,
3305 sd, idle, all_pinned, &this_best_prio);
3306 class = class->next;
3308 #ifdef CONFIG_PREEMPT
3310 * NEWIDLE balancing is a source of latency, so preemptible
3311 * kernels will stop after the first task is pulled to minimize
3312 * the critical section.
3314 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3317 } while (class && max_load_move > total_load_moved);
3319 return total_load_moved > 0;
3323 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3324 struct sched_domain *sd, enum cpu_idle_type idle,
3325 struct rq_iterator *iterator)
3327 struct task_struct *p = iterator->start(iterator->arg);
3331 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3332 pull_task(busiest, p, this_rq, this_cpu);
3334 * Right now, this is only the second place pull_task()
3335 * is called, so we can safely collect pull_task()
3336 * stats here rather than inside pull_task().
3338 schedstat_inc(sd, lb_gained[idle]);
3342 p = iterator->next(iterator->arg);
3349 * move_one_task tries to move exactly one task from busiest to this_rq, as
3350 * part of active balancing operations within "domain".
3351 * Returns 1 if successful and 0 otherwise.
3353 * Called with both runqueues locked.
3355 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3356 struct sched_domain *sd, enum cpu_idle_type idle)
3358 const struct sched_class *class;
3360 for_each_class(class) {
3361 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3367 /********** Helpers for find_busiest_group ************************/
3369 * sd_lb_stats - Structure to store the statistics of a sched_domain
3370 * during load balancing.
3372 struct sd_lb_stats {
3373 struct sched_group *busiest; /* Busiest group in this sd */
3374 struct sched_group *this; /* Local group in this sd */
3375 unsigned long total_load; /* Total load of all groups in sd */
3376 unsigned long total_pwr; /* Total power of all groups in sd */
3377 unsigned long avg_load; /* Average load across all groups in sd */
3379 /** Statistics of this group */
3380 unsigned long this_load;
3381 unsigned long this_load_per_task;
3382 unsigned long this_nr_running;
3384 /* Statistics of the busiest group */
3385 unsigned long max_load;
3386 unsigned long busiest_load_per_task;
3387 unsigned long busiest_nr_running;
3389 int group_imb; /* Is there imbalance in this sd */
3390 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3391 int power_savings_balance; /* Is powersave balance needed for this sd */
3392 struct sched_group *group_min; /* Least loaded group in sd */
3393 struct sched_group *group_leader; /* Group which relieves group_min */
3394 unsigned long min_load_per_task; /* load_per_task in group_min */
3395 unsigned long leader_nr_running; /* Nr running of group_leader */
3396 unsigned long min_nr_running; /* Nr running of group_min */
3401 * sg_lb_stats - stats of a sched_group required for load_balancing
3403 struct sg_lb_stats {
3404 unsigned long avg_load; /*Avg load across the CPUs of the group */
3405 unsigned long group_load; /* Total load over the CPUs of the group */
3406 unsigned long sum_nr_running; /* Nr tasks running in the group */
3407 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3408 unsigned long group_capacity;
3409 int group_imb; /* Is there an imbalance in the group ? */
3413 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3414 * @group: The group whose first cpu is to be returned.
3416 static inline unsigned int group_first_cpu(struct sched_group *group)
3418 return cpumask_first(sched_group_cpus(group));
3422 * get_sd_load_idx - Obtain the load index for a given sched domain.
3423 * @sd: The sched_domain whose load_idx is to be obtained.
3424 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3426 static inline int get_sd_load_idx(struct sched_domain *sd,
3427 enum cpu_idle_type idle)
3433 load_idx = sd->busy_idx;
3436 case CPU_NEWLY_IDLE:
3437 load_idx = sd->newidle_idx;
3440 load_idx = sd->idle_idx;
3448 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3450 * init_sd_power_savings_stats - Initialize power savings statistics for
3451 * the given sched_domain, during load balancing.
3453 * @sd: Sched domain whose power-savings statistics are to be initialized.
3454 * @sds: Variable containing the statistics for sd.
3455 * @idle: Idle status of the CPU at which we're performing load-balancing.
3457 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3458 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3461 * Busy processors will not participate in power savings
3464 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3465 sds->power_savings_balance = 0;
3467 sds->power_savings_balance = 1;
3468 sds->min_nr_running = ULONG_MAX;
3469 sds->leader_nr_running = 0;
3474 * update_sd_power_savings_stats - Update the power saving stats for a
3475 * sched_domain while performing load balancing.
3477 * @group: sched_group belonging to the sched_domain under consideration.
3478 * @sds: Variable containing the statistics of the sched_domain
3479 * @local_group: Does group contain the CPU for which we're performing
3481 * @sgs: Variable containing the statistics of the group.
3483 static inline void update_sd_power_savings_stats(struct sched_group *group,
3484 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3487 if (!sds->power_savings_balance)
3491 * If the local group is idle or completely loaded
3492 * no need to do power savings balance at this domain
3494 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3495 !sds->this_nr_running))
3496 sds->power_savings_balance = 0;
3499 * If a group is already running at full capacity or idle,
3500 * don't include that group in power savings calculations
3502 if (!sds->power_savings_balance ||
3503 sgs->sum_nr_running >= sgs->group_capacity ||
3504 !sgs->sum_nr_running)
3508 * Calculate the group which has the least non-idle load.
3509 * This is the group from where we need to pick up the load
3512 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3513 (sgs->sum_nr_running == sds->min_nr_running &&
3514 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3515 sds->group_min = group;
3516 sds->min_nr_running = sgs->sum_nr_running;
3517 sds->min_load_per_task = sgs->sum_weighted_load /
3518 sgs->sum_nr_running;
3522 * Calculate the group which is almost near its
3523 * capacity but still has some space to pick up some load
3524 * from other group and save more power
3526 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
3529 if (sgs->sum_nr_running > sds->leader_nr_running ||
3530 (sgs->sum_nr_running == sds->leader_nr_running &&
3531 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3532 sds->group_leader = group;
3533 sds->leader_nr_running = sgs->sum_nr_running;
3538 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3539 * @sds: Variable containing the statistics of the sched_domain
3540 * under consideration.
3541 * @this_cpu: Cpu at which we're currently performing load-balancing.
3542 * @imbalance: Variable to store the imbalance.
3545 * Check if we have potential to perform some power-savings balance.
3546 * If yes, set the busiest group to be the least loaded group in the
3547 * sched_domain, so that it's CPUs can be put to idle.
3549 * Returns 1 if there is potential to perform power-savings balance.
3552 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3553 int this_cpu, unsigned long *imbalance)
3555 if (!sds->power_savings_balance)
3558 if (sds->this != sds->group_leader ||
3559 sds->group_leader == sds->group_min)
3562 *imbalance = sds->min_load_per_task;
3563 sds->busiest = sds->group_min;
3568 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3569 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3570 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3575 static inline void update_sd_power_savings_stats(struct sched_group *group,
3576 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3581 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3582 int this_cpu, unsigned long *imbalance)
3586 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3589 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
3591 return SCHED_LOAD_SCALE;
3594 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
3596 return default_scale_freq_power(sd, cpu);
3599 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
3601 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3602 unsigned long smt_gain = sd->smt_gain;
3609 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3611 return default_scale_smt_power(sd, cpu);
3614 unsigned long scale_rt_power(int cpu)
3616 struct rq *rq = cpu_rq(cpu);
3617 u64 total, available;
3619 sched_avg_update(rq);
3621 total = sched_avg_period() + (rq->clock - rq->age_stamp);
3622 available = total - rq->rt_avg;
3624 if (unlikely((s64)total < SCHED_LOAD_SCALE))
3625 total = SCHED_LOAD_SCALE;
3627 total >>= SCHED_LOAD_SHIFT;
3629 return div_u64(available, total);
3632 static void update_cpu_power(struct sched_domain *sd, int cpu)
3634 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3635 unsigned long power = SCHED_LOAD_SCALE;
3636 struct sched_group *sdg = sd->groups;
3638 if (sched_feat(ARCH_POWER))
3639 power *= arch_scale_freq_power(sd, cpu);
3641 power *= default_scale_freq_power(sd, cpu);
3643 power >>= SCHED_LOAD_SHIFT;
3645 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
3646 if (sched_feat(ARCH_POWER))
3647 power *= arch_scale_smt_power(sd, cpu);
3649 power *= default_scale_smt_power(sd, cpu);
3651 power >>= SCHED_LOAD_SHIFT;
3654 power *= scale_rt_power(cpu);
3655 power >>= SCHED_LOAD_SHIFT;
3660 sdg->cpu_power = power;
3663 static void update_group_power(struct sched_domain *sd, int cpu)
3665 struct sched_domain *child = sd->child;
3666 struct sched_group *group, *sdg = sd->groups;
3667 unsigned long power;
3670 update_cpu_power(sd, cpu);
3676 group = child->groups;
3678 power += group->cpu_power;
3679 group = group->next;
3680 } while (group != child->groups);
3682 sdg->cpu_power = power;
3686 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3687 * @sd: The sched_domain whose statistics are to be updated.
3688 * @group: sched_group whose statistics are to be updated.
3689 * @this_cpu: Cpu for which load balance is currently performed.
3690 * @idle: Idle status of this_cpu
3691 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3692 * @sd_idle: Idle status of the sched_domain containing group.
3693 * @local_group: Does group contain this_cpu.
3694 * @cpus: Set of cpus considered for load balancing.
3695 * @balance: Should we balance.
3696 * @sgs: variable to hold the statistics for this group.
3698 static inline void update_sg_lb_stats(struct sched_domain *sd,
3699 struct sched_group *group, int this_cpu,
3700 enum cpu_idle_type idle, int load_idx, int *sd_idle,
3701 int local_group, const struct cpumask *cpus,
3702 int *balance, struct sg_lb_stats *sgs)
3704 unsigned long load, max_cpu_load, min_cpu_load;
3706 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3707 unsigned long sum_avg_load_per_task;
3708 unsigned long avg_load_per_task;
3711 balance_cpu = group_first_cpu(group);
3712 if (balance_cpu == this_cpu)
3713 update_group_power(sd, this_cpu);
3716 /* Tally up the load of all CPUs in the group */
3717 sum_avg_load_per_task = avg_load_per_task = 0;
3719 min_cpu_load = ~0UL;
3721 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3722 struct rq *rq = cpu_rq(i);
3724 if (*sd_idle && rq->nr_running)
3727 /* Bias balancing toward cpus of our domain */
3729 if (idle_cpu(i) && !first_idle_cpu) {
3734 load = target_load(i, load_idx);
3736 load = source_load(i, load_idx);
3737 if (load > max_cpu_load)
3738 max_cpu_load = load;
3739 if (min_cpu_load > load)
3740 min_cpu_load = load;
3743 sgs->group_load += load;
3744 sgs->sum_nr_running += rq->nr_running;
3745 sgs->sum_weighted_load += weighted_cpuload(i);
3747 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3751 * First idle cpu or the first cpu(busiest) in this sched group
3752 * is eligible for doing load balancing at this and above
3753 * domains. In the newly idle case, we will allow all the cpu's
3754 * to do the newly idle load balance.
3756 if (idle != CPU_NEWLY_IDLE && local_group &&
3757 balance_cpu != this_cpu && balance) {
3762 /* Adjust by relative CPU power of the group */
3763 sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power;
3767 * Consider the group unbalanced when the imbalance is larger
3768 * than the average weight of two tasks.
3770 * APZ: with cgroup the avg task weight can vary wildly and
3771 * might not be a suitable number - should we keep a
3772 * normalized nr_running number somewhere that negates
3775 avg_load_per_task = (sum_avg_load_per_task * SCHED_LOAD_SCALE) /
3778 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3781 sgs->group_capacity =
3782 DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
3786 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3787 * @sd: sched_domain whose statistics are to be updated.
3788 * @this_cpu: Cpu for which load balance is currently performed.
3789 * @idle: Idle status of this_cpu
3790 * @sd_idle: Idle status of the sched_domain containing group.
3791 * @cpus: Set of cpus considered for load balancing.
3792 * @balance: Should we balance.
3793 * @sds: variable to hold the statistics for this sched_domain.
3795 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3796 enum cpu_idle_type idle, int *sd_idle,
3797 const struct cpumask *cpus, int *balance,
3798 struct sd_lb_stats *sds)
3800 struct sched_domain *child = sd->child;
3801 struct sched_group *group = sd->groups;
3802 struct sg_lb_stats sgs;
3803 int load_idx, prefer_sibling = 0;
3805 if (child && child->flags & SD_PREFER_SIBLING)
3808 init_sd_power_savings_stats(sd, sds, idle);
3809 load_idx = get_sd_load_idx(sd, idle);
3814 local_group = cpumask_test_cpu(this_cpu,
3815 sched_group_cpus(group));
3816 memset(&sgs, 0, sizeof(sgs));
3817 update_sg_lb_stats(sd, group, this_cpu, idle, load_idx, sd_idle,
3818 local_group, cpus, balance, &sgs);
3820 if (local_group && balance && !(*balance))
3823 sds->total_load += sgs.group_load;
3824 sds->total_pwr += group->cpu_power;
3827 * In case the child domain prefers tasks go to siblings
3828 * first, lower the group capacity to one so that we'll try
3829 * and move all the excess tasks away.
3832 sgs.group_capacity = min(sgs.group_capacity, 1UL);
3835 sds->this_load = sgs.avg_load;
3837 sds->this_nr_running = sgs.sum_nr_running;
3838 sds->this_load_per_task = sgs.sum_weighted_load;
3839 } else if (sgs.avg_load > sds->max_load &&
3840 (sgs.sum_nr_running > sgs.group_capacity ||
3842 sds->max_load = sgs.avg_load;
3843 sds->busiest = group;
3844 sds->busiest_nr_running = sgs.sum_nr_running;
3845 sds->busiest_load_per_task = sgs.sum_weighted_load;
3846 sds->group_imb = sgs.group_imb;
3849 update_sd_power_savings_stats(group, sds, local_group, &sgs);
3850 group = group->next;
3851 } while (group != sd->groups);
3855 * fix_small_imbalance - Calculate the minor imbalance that exists
3856 * amongst the groups of a sched_domain, during
3858 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3859 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3860 * @imbalance: Variable to store the imbalance.
3862 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3863 int this_cpu, unsigned long *imbalance)
3865 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3866 unsigned int imbn = 2;
3868 if (sds->this_nr_running) {
3869 sds->this_load_per_task /= sds->this_nr_running;
3870 if (sds->busiest_load_per_task >
3871 sds->this_load_per_task)
3874 sds->this_load_per_task =
3875 cpu_avg_load_per_task(this_cpu);
3877 if (sds->max_load - sds->this_load + sds->busiest_load_per_task >=
3878 sds->busiest_load_per_task * imbn) {
3879 *imbalance = sds->busiest_load_per_task;
3884 * OK, we don't have enough imbalance to justify moving tasks,
3885 * however we may be able to increase total CPU power used by
3889 pwr_now += sds->busiest->cpu_power *
3890 min(sds->busiest_load_per_task, sds->max_load);
3891 pwr_now += sds->this->cpu_power *
3892 min(sds->this_load_per_task, sds->this_load);
3893 pwr_now /= SCHED_LOAD_SCALE;
3895 /* Amount of load we'd subtract */
3896 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3897 sds->busiest->cpu_power;
3898 if (sds->max_load > tmp)
3899 pwr_move += sds->busiest->cpu_power *
3900 min(sds->busiest_load_per_task, sds->max_load - tmp);
3902 /* Amount of load we'd add */
3903 if (sds->max_load * sds->busiest->cpu_power <
3904 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
3905 tmp = (sds->max_load * sds->busiest->cpu_power) /
3906 sds->this->cpu_power;
3908 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3909 sds->this->cpu_power;
3910 pwr_move += sds->this->cpu_power *
3911 min(sds->this_load_per_task, sds->this_load + tmp);
3912 pwr_move /= SCHED_LOAD_SCALE;
3914 /* Move if we gain throughput */
3915 if (pwr_move > pwr_now)
3916 *imbalance = sds->busiest_load_per_task;
3920 * calculate_imbalance - Calculate the amount of imbalance present within the
3921 * groups of a given sched_domain during load balance.
3922 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3923 * @this_cpu: Cpu for which currently load balance is being performed.
3924 * @imbalance: The variable to store the imbalance.
3926 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3927 unsigned long *imbalance)
3929 unsigned long max_pull;
3931 * In the presence of smp nice balancing, certain scenarios can have
3932 * max load less than avg load(as we skip the groups at or below
3933 * its cpu_power, while calculating max_load..)
3935 if (sds->max_load < sds->avg_load) {
3937 return fix_small_imbalance(sds, this_cpu, imbalance);
3940 /* Don't want to pull so many tasks that a group would go idle */
3941 max_pull = min(sds->max_load - sds->avg_load,
3942 sds->max_load - sds->busiest_load_per_task);
3944 /* How much load to actually move to equalise the imbalance */
3945 *imbalance = min(max_pull * sds->busiest->cpu_power,
3946 (sds->avg_load - sds->this_load) * sds->this->cpu_power)
3950 * if *imbalance is less than the average load per runnable task
3951 * there is no gaurantee that any tasks will be moved so we'll have
3952 * a think about bumping its value to force at least one task to be
3955 if (*imbalance < sds->busiest_load_per_task)
3956 return fix_small_imbalance(sds, this_cpu, imbalance);
3959 /******* find_busiest_group() helpers end here *********************/
3962 * find_busiest_group - Returns the busiest group within the sched_domain
3963 * if there is an imbalance. If there isn't an imbalance, and
3964 * the user has opted for power-savings, it returns a group whose
3965 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3966 * such a group exists.
3968 * Also calculates the amount of weighted load which should be moved
3969 * to restore balance.
3971 * @sd: The sched_domain whose busiest group is to be returned.
3972 * @this_cpu: The cpu for which load balancing is currently being performed.
3973 * @imbalance: Variable which stores amount of weighted load which should
3974 * be moved to restore balance/put a group to idle.
3975 * @idle: The idle status of this_cpu.
3976 * @sd_idle: The idleness of sd
3977 * @cpus: The set of CPUs under consideration for load-balancing.
3978 * @balance: Pointer to a variable indicating if this_cpu
3979 * is the appropriate cpu to perform load balancing at this_level.
3981 * Returns: - the busiest group if imbalance exists.
3982 * - If no imbalance and user has opted for power-savings balance,
3983 * return the least loaded group whose CPUs can be
3984 * put to idle by rebalancing its tasks onto our group.
3986 static struct sched_group *
3987 find_busiest_group(struct sched_domain *sd, int this_cpu,
3988 unsigned long *imbalance, enum cpu_idle_type idle,
3989 int *sd_idle, const struct cpumask *cpus, int *balance)
3991 struct sd_lb_stats sds;
3993 memset(&sds, 0, sizeof(sds));
3996 * Compute the various statistics relavent for load balancing at
3999 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
4002 /* Cases where imbalance does not exist from POV of this_cpu */
4003 /* 1) this_cpu is not the appropriate cpu to perform load balancing
4005 * 2) There is no busy sibling group to pull from.
4006 * 3) This group is the busiest group.
4007 * 4) This group is more busy than the avg busieness at this
4009 * 5) The imbalance is within the specified limit.
4010 * 6) Any rebalance would lead to ping-pong
4012 if (balance && !(*balance))
4015 if (!sds.busiest || sds.busiest_nr_running == 0)
4018 if (sds.this_load >= sds.max_load)
4021 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
4023 if (sds.this_load >= sds.avg_load)
4026 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
4029 sds.busiest_load_per_task /= sds.busiest_nr_running;
4031 sds.busiest_load_per_task =
4032 min(sds.busiest_load_per_task, sds.avg_load);
4035 * We're trying to get all the cpus to the average_load, so we don't
4036 * want to push ourselves above the average load, nor do we wish to
4037 * reduce the max loaded cpu below the average load, as either of these
4038 * actions would just result in more rebalancing later, and ping-pong
4039 * tasks around. Thus we look for the minimum possible imbalance.
4040 * Negative imbalances (*we* are more loaded than anyone else) will
4041 * be counted as no imbalance for these purposes -- we can't fix that
4042 * by pulling tasks to us. Be careful of negative numbers as they'll
4043 * appear as very large values with unsigned longs.
4045 if (sds.max_load <= sds.busiest_load_per_task)
4048 /* Looks like there is an imbalance. Compute it */
4049 calculate_imbalance(&sds, this_cpu, imbalance);
4054 * There is no obvious imbalance. But check if we can do some balancing
4057 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
4065 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4068 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
4069 unsigned long imbalance, const struct cpumask *cpus)
4071 struct rq *busiest = NULL, *rq;
4072 unsigned long max_load = 0;
4075 for_each_cpu(i, sched_group_cpus(group)) {
4076 unsigned long power = power_of(i);
4077 unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
4080 if (!cpumask_test_cpu(i, cpus))
4084 wl = weighted_cpuload(i) * SCHED_LOAD_SCALE;
4087 if (capacity && rq->nr_running == 1 && wl > imbalance)
4090 if (wl > max_load) {
4100 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4101 * so long as it is large enough.
4103 #define MAX_PINNED_INTERVAL 512
4105 /* Working cpumask for load_balance and load_balance_newidle. */
4106 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4109 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4110 * tasks if there is an imbalance.
4112 static int load_balance(int this_cpu, struct rq *this_rq,
4113 struct sched_domain *sd, enum cpu_idle_type idle,
4116 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
4117 struct sched_group *group;
4118 unsigned long imbalance;
4120 unsigned long flags;
4121 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4123 cpumask_copy(cpus, cpu_active_mask);
4126 * When power savings policy is enabled for the parent domain, idle
4127 * sibling can pick up load irrespective of busy siblings. In this case,
4128 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4129 * portraying it as CPU_NOT_IDLE.
4131 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
4132 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4135 schedstat_inc(sd, lb_count[idle]);
4139 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
4146 schedstat_inc(sd, lb_nobusyg[idle]);
4150 busiest = find_busiest_queue(group, idle, imbalance, cpus);
4152 schedstat_inc(sd, lb_nobusyq[idle]);
4156 BUG_ON(busiest == this_rq);
4158 schedstat_add(sd, lb_imbalance[idle], imbalance);
4161 if (busiest->nr_running > 1) {
4163 * Attempt to move tasks. If find_busiest_group has found
4164 * an imbalance but busiest->nr_running <= 1, the group is
4165 * still unbalanced. ld_moved simply stays zero, so it is
4166 * correctly treated as an imbalance.
4168 local_irq_save(flags);
4169 double_rq_lock(this_rq, busiest);
4170 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4171 imbalance, sd, idle, &all_pinned);
4172 double_rq_unlock(this_rq, busiest);
4173 local_irq_restore(flags);
4176 * some other cpu did the load balance for us.
4178 if (ld_moved && this_cpu != smp_processor_id())
4179 resched_cpu(this_cpu);
4181 /* All tasks on this runqueue were pinned by CPU affinity */
4182 if (unlikely(all_pinned)) {
4183 cpumask_clear_cpu(cpu_of(busiest), cpus);
4184 if (!cpumask_empty(cpus))
4191 schedstat_inc(sd, lb_failed[idle]);
4192 sd->nr_balance_failed++;
4194 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
4196 spin_lock_irqsave(&busiest->lock, flags);
4198 /* don't kick the migration_thread, if the curr
4199 * task on busiest cpu can't be moved to this_cpu
4201 if (!cpumask_test_cpu(this_cpu,
4202 &busiest->curr->cpus_allowed)) {
4203 spin_unlock_irqrestore(&busiest->lock, flags);
4205 goto out_one_pinned;
4208 if (!busiest->active_balance) {
4209 busiest->active_balance = 1;
4210 busiest->push_cpu = this_cpu;
4213 spin_unlock_irqrestore(&busiest->lock, flags);
4215 wake_up_process(busiest->migration_thread);
4218 * We've kicked active balancing, reset the failure
4221 sd->nr_balance_failed = sd->cache_nice_tries+1;
4224 sd->nr_balance_failed = 0;
4226 if (likely(!active_balance)) {
4227 /* We were unbalanced, so reset the balancing interval */
4228 sd->balance_interval = sd->min_interval;
4231 * If we've begun active balancing, start to back off. This
4232 * case may not be covered by the all_pinned logic if there
4233 * is only 1 task on the busy runqueue (because we don't call
4236 if (sd->balance_interval < sd->max_interval)
4237 sd->balance_interval *= 2;
4240 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4241 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4247 schedstat_inc(sd, lb_balanced[idle]);
4249 sd->nr_balance_failed = 0;
4252 /* tune up the balancing interval */
4253 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4254 (sd->balance_interval < sd->max_interval))
4255 sd->balance_interval *= 2;
4257 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4258 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4269 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4270 * tasks if there is an imbalance.
4272 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4273 * this_rq is locked.
4276 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
4278 struct sched_group *group;
4279 struct rq *busiest = NULL;
4280 unsigned long imbalance;
4284 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4286 cpumask_copy(cpus, cpu_active_mask);
4289 * When power savings policy is enabled for the parent domain, idle
4290 * sibling can pick up load irrespective of busy siblings. In this case,
4291 * let the state of idle sibling percolate up as IDLE, instead of
4292 * portraying it as CPU_NOT_IDLE.
4294 if (sd->flags & SD_SHARE_CPUPOWER &&
4295 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4298 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
4300 update_shares_locked(this_rq, sd);
4301 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
4302 &sd_idle, cpus, NULL);
4304 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
4308 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
4310 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
4314 BUG_ON(busiest == this_rq);
4316 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
4319 if (busiest->nr_running > 1) {
4320 /* Attempt to move tasks */
4321 double_lock_balance(this_rq, busiest);
4322 /* this_rq->clock is already updated */
4323 update_rq_clock(busiest);
4324 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4325 imbalance, sd, CPU_NEWLY_IDLE,
4327 double_unlock_balance(this_rq, busiest);
4329 if (unlikely(all_pinned)) {
4330 cpumask_clear_cpu(cpu_of(busiest), cpus);
4331 if (!cpumask_empty(cpus))
4337 int active_balance = 0;
4339 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
4340 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4341 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4344 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4347 if (sd->nr_balance_failed++ < 2)
4351 * The only task running in a non-idle cpu can be moved to this
4352 * cpu in an attempt to completely freeup the other CPU
4353 * package. The same method used to move task in load_balance()
4354 * have been extended for load_balance_newidle() to speedup
4355 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4357 * The package power saving logic comes from
4358 * find_busiest_group(). If there are no imbalance, then
4359 * f_b_g() will return NULL. However when sched_mc={1,2} then
4360 * f_b_g() will select a group from which a running task may be
4361 * pulled to this cpu in order to make the other package idle.
4362 * If there is no opportunity to make a package idle and if
4363 * there are no imbalance, then f_b_g() will return NULL and no
4364 * action will be taken in load_balance_newidle().
4366 * Under normal task pull operation due to imbalance, there
4367 * will be more than one task in the source run queue and
4368 * move_tasks() will succeed. ld_moved will be true and this
4369 * active balance code will not be triggered.
4372 /* Lock busiest in correct order while this_rq is held */
4373 double_lock_balance(this_rq, busiest);
4376 * don't kick the migration_thread, if the curr
4377 * task on busiest cpu can't be moved to this_cpu
4379 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
4380 double_unlock_balance(this_rq, busiest);
4385 if (!busiest->active_balance) {
4386 busiest->active_balance = 1;
4387 busiest->push_cpu = this_cpu;
4391 double_unlock_balance(this_rq, busiest);
4393 * Should not call ttwu while holding a rq->lock
4395 spin_unlock(&this_rq->lock);
4397 wake_up_process(busiest->migration_thread);
4398 spin_lock(&this_rq->lock);
4401 sd->nr_balance_failed = 0;
4403 update_shares_locked(this_rq, sd);
4407 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
4408 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4409 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4411 sd->nr_balance_failed = 0;
4417 * idle_balance is called by schedule() if this_cpu is about to become
4418 * idle. Attempts to pull tasks from other CPUs.
4420 static void idle_balance(int this_cpu, struct rq *this_rq)
4422 struct sched_domain *sd;
4423 int pulled_task = 0;
4424 unsigned long next_balance = jiffies + HZ;
4426 this_rq->idle_stamp = this_rq->clock;
4428 if (this_rq->avg_idle < sysctl_sched_migration_cost)
4431 for_each_domain(this_cpu, sd) {
4432 unsigned long interval;
4434 if (!(sd->flags & SD_LOAD_BALANCE))
4437 if (sd->flags & SD_BALANCE_NEWIDLE)
4438 /* If we've pulled tasks over stop searching: */
4439 pulled_task = load_balance_newidle(this_cpu, this_rq,
4442 interval = msecs_to_jiffies(sd->balance_interval);
4443 if (time_after(next_balance, sd->last_balance + interval))
4444 next_balance = sd->last_balance + interval;
4446 this_rq->idle_stamp = 0;
4450 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4452 * We are going idle. next_balance may be set based on
4453 * a busy processor. So reset next_balance.
4455 this_rq->next_balance = next_balance;
4460 * active_load_balance is run by migration threads. It pushes running tasks
4461 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4462 * running on each physical CPU where possible, and avoids physical /
4463 * logical imbalances.
4465 * Called with busiest_rq locked.
4467 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
4469 int target_cpu = busiest_rq->push_cpu;
4470 struct sched_domain *sd;
4471 struct rq *target_rq;
4473 /* Is there any task to move? */
4474 if (busiest_rq->nr_running <= 1)
4477 target_rq = cpu_rq(target_cpu);
4480 * This condition is "impossible", if it occurs
4481 * we need to fix it. Originally reported by
4482 * Bjorn Helgaas on a 128-cpu setup.
4484 BUG_ON(busiest_rq == target_rq);
4486 /* move a task from busiest_rq to target_rq */
4487 double_lock_balance(busiest_rq, target_rq);
4488 update_rq_clock(busiest_rq);
4489 update_rq_clock(target_rq);
4491 /* Search for an sd spanning us and the target CPU. */
4492 for_each_domain(target_cpu, sd) {
4493 if ((sd->flags & SD_LOAD_BALANCE) &&
4494 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4499 schedstat_inc(sd, alb_count);
4501 if (move_one_task(target_rq, target_cpu, busiest_rq,
4503 schedstat_inc(sd, alb_pushed);
4505 schedstat_inc(sd, alb_failed);
4507 double_unlock_balance(busiest_rq, target_rq);
4512 atomic_t load_balancer;
4513 cpumask_var_t cpu_mask;
4514 cpumask_var_t ilb_grp_nohz_mask;
4515 } nohz ____cacheline_aligned = {
4516 .load_balancer = ATOMIC_INIT(-1),
4519 int get_nohz_load_balancer(void)
4521 return atomic_read(&nohz.load_balancer);
4524 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4526 * lowest_flag_domain - Return lowest sched_domain containing flag.
4527 * @cpu: The cpu whose lowest level of sched domain is to
4529 * @flag: The flag to check for the lowest sched_domain
4530 * for the given cpu.
4532 * Returns the lowest sched_domain of a cpu which contains the given flag.
4534 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4536 struct sched_domain *sd;
4538 for_each_domain(cpu, sd)
4539 if (sd && (sd->flags & flag))
4546 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4547 * @cpu: The cpu whose domains we're iterating over.
4548 * @sd: variable holding the value of the power_savings_sd
4550 * @flag: The flag to filter the sched_domains to be iterated.
4552 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4553 * set, starting from the lowest sched_domain to the highest.
4555 #define for_each_flag_domain(cpu, sd, flag) \
4556 for (sd = lowest_flag_domain(cpu, flag); \
4557 (sd && (sd->flags & flag)); sd = sd->parent)
4560 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4561 * @ilb_group: group to be checked for semi-idleness
4563 * Returns: 1 if the group is semi-idle. 0 otherwise.
4565 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4566 * and atleast one non-idle CPU. This helper function checks if the given
4567 * sched_group is semi-idle or not.
4569 static inline int is_semi_idle_group(struct sched_group *ilb_group)
4571 cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask,
4572 sched_group_cpus(ilb_group));
4575 * A sched_group is semi-idle when it has atleast one busy cpu
4576 * and atleast one idle cpu.
4578 if (cpumask_empty(nohz.ilb_grp_nohz_mask))
4581 if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group)))
4587 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4588 * @cpu: The cpu which is nominating a new idle_load_balancer.
4590 * Returns: Returns the id of the idle load balancer if it exists,
4591 * Else, returns >= nr_cpu_ids.
4593 * This algorithm picks the idle load balancer such that it belongs to a
4594 * semi-idle powersavings sched_domain. The idea is to try and avoid
4595 * completely idle packages/cores just for the purpose of idle load balancing
4596 * when there are other idle cpu's which are better suited for that job.
4598 static int find_new_ilb(int cpu)
4600 struct sched_domain *sd;
4601 struct sched_group *ilb_group;
4604 * Have idle load balancer selection from semi-idle packages only
4605 * when power-aware load balancing is enabled
4607 if (!(sched_smt_power_savings || sched_mc_power_savings))
4611 * Optimize for the case when we have no idle CPUs or only one
4612 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4614 if (cpumask_weight(nohz.cpu_mask) < 2)
4617 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4618 ilb_group = sd->groups;
4621 if (is_semi_idle_group(ilb_group))
4622 return cpumask_first(nohz.ilb_grp_nohz_mask);
4624 ilb_group = ilb_group->next;
4626 } while (ilb_group != sd->groups);
4630 return cpumask_first(nohz.cpu_mask);
4632 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4633 static inline int find_new_ilb(int call_cpu)
4635 return cpumask_first(nohz.cpu_mask);
4640 * This routine will try to nominate the ilb (idle load balancing)
4641 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4642 * load balancing on behalf of all those cpus. If all the cpus in the system
4643 * go into this tickless mode, then there will be no ilb owner (as there is
4644 * no need for one) and all the cpus will sleep till the next wakeup event
4647 * For the ilb owner, tick is not stopped. And this tick will be used
4648 * for idle load balancing. ilb owner will still be part of
4651 * While stopping the tick, this cpu will become the ilb owner if there
4652 * is no other owner. And will be the owner till that cpu becomes busy
4653 * or if all cpus in the system stop their ticks at which point
4654 * there is no need for ilb owner.
4656 * When the ilb owner becomes busy, it nominates another owner, during the
4657 * next busy scheduler_tick()
4659 int select_nohz_load_balancer(int stop_tick)
4661 int cpu = smp_processor_id();
4664 cpu_rq(cpu)->in_nohz_recently = 1;
4666 if (!cpu_active(cpu)) {
4667 if (atomic_read(&nohz.load_balancer) != cpu)
4671 * If we are going offline and still the leader,
4674 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4680 cpumask_set_cpu(cpu, nohz.cpu_mask);
4682 /* time for ilb owner also to sleep */
4683 if (cpumask_weight(nohz.cpu_mask) == num_active_cpus()) {
4684 if (atomic_read(&nohz.load_balancer) == cpu)
4685 atomic_set(&nohz.load_balancer, -1);
4689 if (atomic_read(&nohz.load_balancer) == -1) {
4690 /* make me the ilb owner */
4691 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4693 } else if (atomic_read(&nohz.load_balancer) == cpu) {
4696 if (!(sched_smt_power_savings ||
4697 sched_mc_power_savings))
4700 * Check to see if there is a more power-efficient
4703 new_ilb = find_new_ilb(cpu);
4704 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4705 atomic_set(&nohz.load_balancer, -1);
4706 resched_cpu(new_ilb);
4712 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4715 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4717 if (atomic_read(&nohz.load_balancer) == cpu)
4718 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4725 static DEFINE_SPINLOCK(balancing);
4728 * It checks each scheduling domain to see if it is due to be balanced,
4729 * and initiates a balancing operation if so.
4731 * Balancing parameters are set up in arch_init_sched_domains.
4733 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4736 struct rq *rq = cpu_rq(cpu);
4737 unsigned long interval;
4738 struct sched_domain *sd;
4739 /* Earliest time when we have to do rebalance again */
4740 unsigned long next_balance = jiffies + 60*HZ;
4741 int update_next_balance = 0;
4744 for_each_domain(cpu, sd) {
4745 if (!(sd->flags & SD_LOAD_BALANCE))
4748 interval = sd->balance_interval;
4749 if (idle != CPU_IDLE)
4750 interval *= sd->busy_factor;
4752 /* scale ms to jiffies */
4753 interval = msecs_to_jiffies(interval);
4754 if (unlikely(!interval))
4756 if (interval > HZ*NR_CPUS/10)
4757 interval = HZ*NR_CPUS/10;
4759 need_serialize = sd->flags & SD_SERIALIZE;
4761 if (need_serialize) {
4762 if (!spin_trylock(&balancing))
4766 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4767 if (load_balance(cpu, rq, sd, idle, &balance)) {
4769 * We've pulled tasks over so either we're no
4770 * longer idle, or one of our SMT siblings is
4773 idle = CPU_NOT_IDLE;
4775 sd->last_balance = jiffies;
4778 spin_unlock(&balancing);
4780 if (time_after(next_balance, sd->last_balance + interval)) {
4781 next_balance = sd->last_balance + interval;
4782 update_next_balance = 1;
4786 * Stop the load balance at this level. There is another
4787 * CPU in our sched group which is doing load balancing more
4795 * next_balance will be updated only when there is a need.
4796 * When the cpu is attached to null domain for ex, it will not be
4799 if (likely(update_next_balance))
4800 rq->next_balance = next_balance;
4804 * run_rebalance_domains is triggered when needed from the scheduler tick.
4805 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4806 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4808 static void run_rebalance_domains(struct softirq_action *h)
4810 int this_cpu = smp_processor_id();
4811 struct rq *this_rq = cpu_rq(this_cpu);
4812 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4813 CPU_IDLE : CPU_NOT_IDLE;
4815 rebalance_domains(this_cpu, idle);
4819 * If this cpu is the owner for idle load balancing, then do the
4820 * balancing on behalf of the other idle cpus whose ticks are
4823 if (this_rq->idle_at_tick &&
4824 atomic_read(&nohz.load_balancer) == this_cpu) {
4828 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4829 if (balance_cpu == this_cpu)
4833 * If this cpu gets work to do, stop the load balancing
4834 * work being done for other cpus. Next load
4835 * balancing owner will pick it up.
4840 rebalance_domains(balance_cpu, CPU_IDLE);
4842 rq = cpu_rq(balance_cpu);
4843 if (time_after(this_rq->next_balance, rq->next_balance))
4844 this_rq->next_balance = rq->next_balance;
4850 static inline int on_null_domain(int cpu)
4852 return !rcu_dereference(cpu_rq(cpu)->sd);
4856 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4858 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4859 * idle load balancing owner or decide to stop the periodic load balancing,
4860 * if the whole system is idle.
4862 static inline void trigger_load_balance(struct rq *rq, int cpu)
4866 * If we were in the nohz mode recently and busy at the current
4867 * scheduler tick, then check if we need to nominate new idle
4870 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4871 rq->in_nohz_recently = 0;
4873 if (atomic_read(&nohz.load_balancer) == cpu) {
4874 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4875 atomic_set(&nohz.load_balancer, -1);
4878 if (atomic_read(&nohz.load_balancer) == -1) {
4879 int ilb = find_new_ilb(cpu);
4881 if (ilb < nr_cpu_ids)
4887 * If this cpu is idle and doing idle load balancing for all the
4888 * cpus with ticks stopped, is it time for that to stop?
4890 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4891 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4897 * If this cpu is idle and the idle load balancing is done by
4898 * someone else, then no need raise the SCHED_SOFTIRQ
4900 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4901 cpumask_test_cpu(cpu, nohz.cpu_mask))
4904 /* Don't need to rebalance while attached to NULL domain */
4905 if (time_after_eq(jiffies, rq->next_balance) &&
4906 likely(!on_null_domain(cpu)))
4907 raise_softirq(SCHED_SOFTIRQ);
4910 #else /* CONFIG_SMP */
4913 * on UP we do not need to balance between CPUs:
4915 static inline void idle_balance(int cpu, struct rq *rq)
4921 DEFINE_PER_CPU(struct kernel_stat, kstat);
4923 EXPORT_PER_CPU_SYMBOL(kstat);
4926 * Return any ns on the sched_clock that have not yet been accounted in
4927 * @p in case that task is currently running.
4929 * Called with task_rq_lock() held on @rq.
4931 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
4935 if (task_current(rq, p)) {
4936 update_rq_clock(rq);
4937 ns = rq->clock - p->se.exec_start;
4945 unsigned long long task_delta_exec(struct task_struct *p)
4947 unsigned long flags;
4951 rq = task_rq_lock(p, &flags);
4952 ns = do_task_delta_exec(p, rq);
4953 task_rq_unlock(rq, &flags);
4959 * Return accounted runtime for the task.
4960 * In case the task is currently running, return the runtime plus current's
4961 * pending runtime that have not been accounted yet.
4963 unsigned long long task_sched_runtime(struct task_struct *p)
4965 unsigned long flags;
4969 rq = task_rq_lock(p, &flags);
4970 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
4971 task_rq_unlock(rq, &flags);
4977 * Return sum_exec_runtime for the thread group.
4978 * In case the task is currently running, return the sum plus current's
4979 * pending runtime that have not been accounted yet.
4981 * Note that the thread group might have other running tasks as well,
4982 * so the return value not includes other pending runtime that other
4983 * running tasks might have.
4985 unsigned long long thread_group_sched_runtime(struct task_struct *p)
4987 struct task_cputime totals;
4988 unsigned long flags;
4992 rq = task_rq_lock(p, &flags);
4993 thread_group_cputime(p, &totals);
4994 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
4995 task_rq_unlock(rq, &flags);
5001 * Account user cpu time to a process.
5002 * @p: the process that the cpu time gets accounted to
5003 * @cputime: the cpu time spent in user space since the last update
5004 * @cputime_scaled: cputime scaled by cpu frequency
5006 void account_user_time(struct task_struct *p, cputime_t cputime,
5007 cputime_t cputime_scaled)
5009 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5012 /* Add user time to process. */
5013 p->utime = cputime_add(p->utime, cputime);
5014 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5015 account_group_user_time(p, cputime);
5017 /* Add user time to cpustat. */
5018 tmp = cputime_to_cputime64(cputime);
5019 if (TASK_NICE(p) > 0)
5020 cpustat->nice = cputime64_add(cpustat->nice, tmp);
5022 cpustat->user = cputime64_add(cpustat->user, tmp);
5024 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
5025 /* Account for user time used */
5026 acct_update_integrals(p);
5030 * Account guest cpu time to a process.
5031 * @p: the process that the cpu time gets accounted to
5032 * @cputime: the cpu time spent in virtual machine since the last update
5033 * @cputime_scaled: cputime scaled by cpu frequency
5035 static void account_guest_time(struct task_struct *p, cputime_t cputime,
5036 cputime_t cputime_scaled)
5039 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5041 tmp = cputime_to_cputime64(cputime);
5043 /* Add guest time to process. */
5044 p->utime = cputime_add(p->utime, cputime);
5045 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5046 account_group_user_time(p, cputime);
5047 p->gtime = cputime_add(p->gtime, cputime);
5049 /* Add guest time to cpustat. */
5050 if (TASK_NICE(p) > 0) {
5051 cpustat->nice = cputime64_add(cpustat->nice, tmp);
5052 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
5054 cpustat->user = cputime64_add(cpustat->user, tmp);
5055 cpustat->guest = cputime64_add(cpustat->guest, tmp);
5060 * Account system cpu time to a process.
5061 * @p: the process that the cpu time gets accounted to
5062 * @hardirq_offset: the offset to subtract from hardirq_count()
5063 * @cputime: the cpu time spent in kernel space since the last update
5064 * @cputime_scaled: cputime scaled by cpu frequency
5066 void account_system_time(struct task_struct *p, int hardirq_offset,
5067 cputime_t cputime, cputime_t cputime_scaled)
5069 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5072 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
5073 account_guest_time(p, cputime, cputime_scaled);
5077 /* Add system time to process. */
5078 p->stime = cputime_add(p->stime, cputime);
5079 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
5080 account_group_system_time(p, cputime);
5082 /* Add system time to cpustat. */
5083 tmp = cputime_to_cputime64(cputime);
5084 if (hardirq_count() - hardirq_offset)
5085 cpustat->irq = cputime64_add(cpustat->irq, tmp);
5086 else if (softirq_count())
5087 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
5089 cpustat->system = cputime64_add(cpustat->system, tmp);
5091 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
5093 /* Account for system time used */
5094 acct_update_integrals(p);
5098 * Account for involuntary wait time.
5099 * @steal: the cpu time spent in involuntary wait
5101 void account_steal_time(cputime_t cputime)
5103 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5104 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5106 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
5110 * Account for idle time.
5111 * @cputime: the cpu time spent in idle wait
5113 void account_idle_time(cputime_t cputime)
5115 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5116 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5117 struct rq *rq = this_rq();
5119 if (atomic_read(&rq->nr_iowait) > 0)
5120 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
5122 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
5125 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
5128 * Account a single tick of cpu time.
5129 * @p: the process that the cpu time gets accounted to
5130 * @user_tick: indicates if the tick is a user or a system tick
5132 void account_process_tick(struct task_struct *p, int user_tick)
5134 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
5135 struct rq *rq = this_rq();
5138 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
5139 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
5140 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
5143 account_idle_time(cputime_one_jiffy);
5147 * Account multiple ticks of steal time.
5148 * @p: the process from which the cpu time has been stolen
5149 * @ticks: number of stolen ticks
5151 void account_steal_ticks(unsigned long ticks)
5153 account_steal_time(jiffies_to_cputime(ticks));
5157 * Account multiple ticks of idle time.
5158 * @ticks: number of stolen ticks
5160 void account_idle_ticks(unsigned long ticks)
5162 account_idle_time(jiffies_to_cputime(ticks));
5168 * Use precise platform statistics if available:
5170 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
5171 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5177 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5179 struct task_cputime cputime;
5181 thread_group_cputime(p, &cputime);
5183 *ut = cputime.utime;
5184 *st = cputime.stime;
5188 #ifndef nsecs_to_cputime
5189 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
5192 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5194 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
5197 * Use CFS's precise accounting:
5199 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
5204 temp = (u64)(rtime * utime);
5205 do_div(temp, total);
5206 utime = (cputime_t)temp;
5211 * Compare with previous values, to keep monotonicity:
5213 p->prev_utime = max(p->prev_utime, utime);
5214 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
5216 *ut = p->prev_utime;
5217 *st = p->prev_stime;
5221 * Must be called with siglock held.
5223 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5225 struct signal_struct *sig = p->signal;
5226 struct task_cputime cputime;
5227 cputime_t rtime, utime, total;
5229 thread_group_cputime(p, &cputime);
5231 total = cputime_add(cputime.utime, cputime.stime);
5232 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
5237 temp = (u64)(rtime * cputime.utime);
5238 do_div(temp, total);
5239 utime = (cputime_t)temp;
5243 sig->prev_utime = max(sig->prev_utime, utime);
5244 sig->prev_stime = max(sig->prev_stime,
5245 cputime_sub(rtime, sig->prev_utime));
5247 *ut = sig->prev_utime;
5248 *st = sig->prev_stime;
5253 * This function gets called by the timer code, with HZ frequency.
5254 * We call it with interrupts disabled.
5256 * It also gets called by the fork code, when changing the parent's
5259 void scheduler_tick(void)
5261 int cpu = smp_processor_id();
5262 struct rq *rq = cpu_rq(cpu);
5263 struct task_struct *curr = rq->curr;
5267 spin_lock(&rq->lock);
5268 update_rq_clock(rq);
5269 update_cpu_load(rq);
5270 curr->sched_class->task_tick(rq, curr, 0);
5271 spin_unlock(&rq->lock);
5273 perf_event_task_tick(curr, cpu);
5276 rq->idle_at_tick = idle_cpu(cpu);
5277 trigger_load_balance(rq, cpu);
5281 notrace unsigned long get_parent_ip(unsigned long addr)
5283 if (in_lock_functions(addr)) {
5284 addr = CALLER_ADDR2;
5285 if (in_lock_functions(addr))
5286 addr = CALLER_ADDR3;
5291 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5292 defined(CONFIG_PREEMPT_TRACER))
5294 void __kprobes add_preempt_count(int val)
5296 #ifdef CONFIG_DEBUG_PREEMPT
5300 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5303 preempt_count() += val;
5304 #ifdef CONFIG_DEBUG_PREEMPT
5306 * Spinlock count overflowing soon?
5308 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5311 if (preempt_count() == val)
5312 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5314 EXPORT_SYMBOL(add_preempt_count);
5316 void __kprobes sub_preempt_count(int val)
5318 #ifdef CONFIG_DEBUG_PREEMPT
5322 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5325 * Is the spinlock portion underflowing?
5327 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5328 !(preempt_count() & PREEMPT_MASK)))
5332 if (preempt_count() == val)
5333 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5334 preempt_count() -= val;
5336 EXPORT_SYMBOL(sub_preempt_count);
5341 * Print scheduling while atomic bug:
5343 static noinline void __schedule_bug(struct task_struct *prev)
5345 struct pt_regs *regs = get_irq_regs();
5347 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5348 prev->comm, prev->pid, preempt_count());
5350 debug_show_held_locks(prev);
5352 if (irqs_disabled())
5353 print_irqtrace_events(prev);
5362 * Various schedule()-time debugging checks and statistics:
5364 static inline void schedule_debug(struct task_struct *prev)
5367 * Test if we are atomic. Since do_exit() needs to call into
5368 * schedule() atomically, we ignore that path for now.
5369 * Otherwise, whine if we are scheduling when we should not be.
5371 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
5372 __schedule_bug(prev);
5374 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5376 schedstat_inc(this_rq(), sched_count);
5377 #ifdef CONFIG_SCHEDSTATS
5378 if (unlikely(prev->lock_depth >= 0)) {
5379 schedstat_inc(this_rq(), bkl_count);
5380 schedstat_inc(prev, sched_info.bkl_count);
5385 static void put_prev_task(struct rq *rq, struct task_struct *prev)
5387 if (prev->state == TASK_RUNNING) {
5388 u64 runtime = prev->se.sum_exec_runtime;
5390 runtime -= prev->se.prev_sum_exec_runtime;
5391 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
5394 * In order to avoid avg_overlap growing stale when we are
5395 * indeed overlapping and hence not getting put to sleep, grow
5396 * the avg_overlap on preemption.
5398 * We use the average preemption runtime because that
5399 * correlates to the amount of cache footprint a task can
5402 update_avg(&prev->se.avg_overlap, runtime);
5404 prev->sched_class->put_prev_task(rq, prev);
5408 * Pick up the highest-prio task:
5410 static inline struct task_struct *
5411 pick_next_task(struct rq *rq)
5413 const struct sched_class *class;
5414 struct task_struct *p;
5417 * Optimization: we know that if all tasks are in
5418 * the fair class we can call that function directly:
5420 if (likely(rq->nr_running == rq->cfs.nr_running)) {
5421 p = fair_sched_class.pick_next_task(rq);
5426 class = sched_class_highest;
5428 p = class->pick_next_task(rq);
5432 * Will never be NULL as the idle class always
5433 * returns a non-NULL p:
5435 class = class->next;
5440 * schedule() is the main scheduler function.
5442 asmlinkage void __sched schedule(void)
5444 struct task_struct *prev, *next;
5445 unsigned long *switch_count;
5451 cpu = smp_processor_id();
5455 switch_count = &prev->nivcsw;
5457 release_kernel_lock(prev);
5458 need_resched_nonpreemptible:
5460 schedule_debug(prev);
5462 if (sched_feat(HRTICK))
5465 spin_lock_irq(&rq->lock);
5466 update_rq_clock(rq);
5467 clear_tsk_need_resched(prev);
5469 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
5470 if (unlikely(signal_pending_state(prev->state, prev)))
5471 prev->state = TASK_RUNNING;
5473 deactivate_task(rq, prev, 1);
5474 switch_count = &prev->nvcsw;
5477 pre_schedule(rq, prev);
5479 if (unlikely(!rq->nr_running))
5480 idle_balance(cpu, rq);
5482 put_prev_task(rq, prev);
5483 next = pick_next_task(rq);
5485 if (likely(prev != next)) {
5486 sched_info_switch(prev, next);
5487 perf_event_task_sched_out(prev, next, cpu);
5493 context_switch(rq, prev, next); /* unlocks the rq */
5495 * the context switch might have flipped the stack from under
5496 * us, hence refresh the local variables.
5498 cpu = smp_processor_id();
5501 spin_unlock_irq(&rq->lock);
5505 if (unlikely(reacquire_kernel_lock(current) < 0))
5506 goto need_resched_nonpreemptible;
5508 preempt_enable_no_resched();
5512 EXPORT_SYMBOL(schedule);
5514 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
5516 * Look out! "owner" is an entirely speculative pointer
5517 * access and not reliable.
5519 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5524 if (!sched_feat(OWNER_SPIN))
5527 #ifdef CONFIG_DEBUG_PAGEALLOC
5529 * Need to access the cpu field knowing that
5530 * DEBUG_PAGEALLOC could have unmapped it if
5531 * the mutex owner just released it and exited.
5533 if (probe_kernel_address(&owner->cpu, cpu))
5540 * Even if the access succeeded (likely case),
5541 * the cpu field may no longer be valid.
5543 if (cpu >= nr_cpumask_bits)
5547 * We need to validate that we can do a
5548 * get_cpu() and that we have the percpu area.
5550 if (!cpu_online(cpu))
5557 * Owner changed, break to re-assess state.
5559 if (lock->owner != owner)
5563 * Is that owner really running on that cpu?
5565 if (task_thread_info(rq->curr) != owner || need_resched())
5575 #ifdef CONFIG_PREEMPT
5577 * this is the entry point to schedule() from in-kernel preemption
5578 * off of preempt_enable. Kernel preemptions off return from interrupt
5579 * occur there and call schedule directly.
5581 asmlinkage void __sched preempt_schedule(void)
5583 struct thread_info *ti = current_thread_info();
5586 * If there is a non-zero preempt_count or interrupts are disabled,
5587 * we do not want to preempt the current task. Just return..
5589 if (likely(ti->preempt_count || irqs_disabled()))
5593 add_preempt_count(PREEMPT_ACTIVE);
5595 sub_preempt_count(PREEMPT_ACTIVE);
5598 * Check again in case we missed a preemption opportunity
5599 * between schedule and now.
5602 } while (need_resched());
5604 EXPORT_SYMBOL(preempt_schedule);
5607 * this is the entry point to schedule() from kernel preemption
5608 * off of irq context.
5609 * Note, that this is called and return with irqs disabled. This will
5610 * protect us against recursive calling from irq.
5612 asmlinkage void __sched preempt_schedule_irq(void)
5614 struct thread_info *ti = current_thread_info();
5616 /* Catch callers which need to be fixed */
5617 BUG_ON(ti->preempt_count || !irqs_disabled());
5620 add_preempt_count(PREEMPT_ACTIVE);
5623 local_irq_disable();
5624 sub_preempt_count(PREEMPT_ACTIVE);
5627 * Check again in case we missed a preemption opportunity
5628 * between schedule and now.
5631 } while (need_resched());
5634 #endif /* CONFIG_PREEMPT */
5636 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
5639 return try_to_wake_up(curr->private, mode, wake_flags);
5641 EXPORT_SYMBOL(default_wake_function);
5644 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5645 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5646 * number) then we wake all the non-exclusive tasks and one exclusive task.
5648 * There are circumstances in which we can try to wake a task which has already
5649 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5650 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5652 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
5653 int nr_exclusive, int wake_flags, void *key)
5655 wait_queue_t *curr, *next;
5657 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
5658 unsigned flags = curr->flags;
5660 if (curr->func(curr, mode, wake_flags, key) &&
5661 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
5667 * __wake_up - wake up threads blocked on a waitqueue.
5669 * @mode: which threads
5670 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5671 * @key: is directly passed to the wakeup function
5673 * It may be assumed that this function implies a write memory barrier before
5674 * changing the task state if and only if any tasks are woken up.
5676 void __wake_up(wait_queue_head_t *q, unsigned int mode,
5677 int nr_exclusive, void *key)
5679 unsigned long flags;
5681 spin_lock_irqsave(&q->lock, flags);
5682 __wake_up_common(q, mode, nr_exclusive, 0, key);
5683 spin_unlock_irqrestore(&q->lock, flags);
5685 EXPORT_SYMBOL(__wake_up);
5688 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5690 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
5692 __wake_up_common(q, mode, 1, 0, NULL);
5695 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
5697 __wake_up_common(q, mode, 1, 0, key);
5701 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5703 * @mode: which threads
5704 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5705 * @key: opaque value to be passed to wakeup targets
5707 * The sync wakeup differs that the waker knows that it will schedule
5708 * away soon, so while the target thread will be woken up, it will not
5709 * be migrated to another CPU - ie. the two threads are 'synchronized'
5710 * with each other. This can prevent needless bouncing between CPUs.
5712 * On UP it can prevent extra preemption.
5714 * It may be assumed that this function implies a write memory barrier before
5715 * changing the task state if and only if any tasks are woken up.
5717 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
5718 int nr_exclusive, void *key)
5720 unsigned long flags;
5721 int wake_flags = WF_SYNC;
5726 if (unlikely(!nr_exclusive))
5729 spin_lock_irqsave(&q->lock, flags);
5730 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
5731 spin_unlock_irqrestore(&q->lock, flags);
5733 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
5736 * __wake_up_sync - see __wake_up_sync_key()
5738 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
5740 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
5742 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
5745 * complete: - signals a single thread waiting on this completion
5746 * @x: holds the state of this particular completion
5748 * This will wake up a single thread waiting on this completion. Threads will be
5749 * awakened in the same order in which they were queued.
5751 * See also complete_all(), wait_for_completion() and related routines.
5753 * It may be assumed that this function implies a write memory barrier before
5754 * changing the task state if and only if any tasks are woken up.
5756 void complete(struct completion *x)
5758 unsigned long flags;
5760 spin_lock_irqsave(&x->wait.lock, flags);
5762 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
5763 spin_unlock_irqrestore(&x->wait.lock, flags);
5765 EXPORT_SYMBOL(complete);
5768 * complete_all: - signals all threads waiting on this completion
5769 * @x: holds the state of this particular completion
5771 * This will wake up all threads waiting on this particular completion event.
5773 * It may be assumed that this function implies a write memory barrier before
5774 * changing the task state if and only if any tasks are woken up.
5776 void complete_all(struct completion *x)
5778 unsigned long flags;
5780 spin_lock_irqsave(&x->wait.lock, flags);
5781 x->done += UINT_MAX/2;
5782 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
5783 spin_unlock_irqrestore(&x->wait.lock, flags);
5785 EXPORT_SYMBOL(complete_all);
5787 static inline long __sched
5788 do_wait_for_common(struct completion *x, long timeout, int state)
5791 DECLARE_WAITQUEUE(wait, current);
5793 wait.flags |= WQ_FLAG_EXCLUSIVE;
5794 __add_wait_queue_tail(&x->wait, &wait);
5796 if (signal_pending_state(state, current)) {
5797 timeout = -ERESTARTSYS;
5800 __set_current_state(state);
5801 spin_unlock_irq(&x->wait.lock);
5802 timeout = schedule_timeout(timeout);
5803 spin_lock_irq(&x->wait.lock);
5804 } while (!x->done && timeout);
5805 __remove_wait_queue(&x->wait, &wait);
5810 return timeout ?: 1;
5814 wait_for_common(struct completion *x, long timeout, int state)
5818 spin_lock_irq(&x->wait.lock);
5819 timeout = do_wait_for_common(x, timeout, state);
5820 spin_unlock_irq(&x->wait.lock);
5825 * wait_for_completion: - waits for completion of a task
5826 * @x: holds the state of this particular completion
5828 * This waits to be signaled for completion of a specific task. It is NOT
5829 * interruptible and there is no timeout.
5831 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5832 * and interrupt capability. Also see complete().
5834 void __sched wait_for_completion(struct completion *x)
5836 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
5838 EXPORT_SYMBOL(wait_for_completion);
5841 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5842 * @x: holds the state of this particular completion
5843 * @timeout: timeout value in jiffies
5845 * This waits for either a completion of a specific task to be signaled or for a
5846 * specified timeout to expire. The timeout is in jiffies. It is not
5849 unsigned long __sched
5850 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5852 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5854 EXPORT_SYMBOL(wait_for_completion_timeout);
5857 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5858 * @x: holds the state of this particular completion
5860 * This waits for completion of a specific task to be signaled. It is
5863 int __sched wait_for_completion_interruptible(struct completion *x)
5865 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5866 if (t == -ERESTARTSYS)
5870 EXPORT_SYMBOL(wait_for_completion_interruptible);
5873 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5874 * @x: holds the state of this particular completion
5875 * @timeout: timeout value in jiffies
5877 * This waits for either a completion of a specific task to be signaled or for a
5878 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5880 unsigned long __sched
5881 wait_for_completion_interruptible_timeout(struct completion *x,
5882 unsigned long timeout)
5884 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
5886 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5889 * wait_for_completion_killable: - waits for completion of a task (killable)
5890 * @x: holds the state of this particular completion
5892 * This waits to be signaled for completion of a specific task. It can be
5893 * interrupted by a kill signal.
5895 int __sched wait_for_completion_killable(struct completion *x)
5897 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5898 if (t == -ERESTARTSYS)
5902 EXPORT_SYMBOL(wait_for_completion_killable);
5905 * try_wait_for_completion - try to decrement a completion without blocking
5906 * @x: completion structure
5908 * Returns: 0 if a decrement cannot be done without blocking
5909 * 1 if a decrement succeeded.
5911 * If a completion is being used as a counting completion,
5912 * attempt to decrement the counter without blocking. This
5913 * enables us to avoid waiting if the resource the completion
5914 * is protecting is not available.
5916 bool try_wait_for_completion(struct completion *x)
5920 spin_lock_irq(&x->wait.lock);
5925 spin_unlock_irq(&x->wait.lock);
5928 EXPORT_SYMBOL(try_wait_for_completion);
5931 * completion_done - Test to see if a completion has any waiters
5932 * @x: completion structure
5934 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5935 * 1 if there are no waiters.
5938 bool completion_done(struct completion *x)
5942 spin_lock_irq(&x->wait.lock);
5945 spin_unlock_irq(&x->wait.lock);
5948 EXPORT_SYMBOL(completion_done);
5951 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
5953 unsigned long flags;
5956 init_waitqueue_entry(&wait, current);
5958 __set_current_state(state);
5960 spin_lock_irqsave(&q->lock, flags);
5961 __add_wait_queue(q, &wait);
5962 spin_unlock(&q->lock);
5963 timeout = schedule_timeout(timeout);
5964 spin_lock_irq(&q->lock);
5965 __remove_wait_queue(q, &wait);
5966 spin_unlock_irqrestore(&q->lock, flags);
5971 void __sched interruptible_sleep_on(wait_queue_head_t *q)
5973 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5975 EXPORT_SYMBOL(interruptible_sleep_on);
5978 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
5980 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
5982 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
5984 void __sched sleep_on(wait_queue_head_t *q)
5986 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5988 EXPORT_SYMBOL(sleep_on);
5990 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
5992 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
5994 EXPORT_SYMBOL(sleep_on_timeout);
5996 #ifdef CONFIG_RT_MUTEXES
5999 * rt_mutex_setprio - set the current priority of a task
6001 * @prio: prio value (kernel-internal form)
6003 * This function changes the 'effective' priority of a task. It does
6004 * not touch ->normal_prio like __setscheduler().
6006 * Used by the rt_mutex code to implement priority inheritance logic.
6008 void rt_mutex_setprio(struct task_struct *p, int prio)
6010 unsigned long flags;
6011 int oldprio, on_rq, running;
6013 const struct sched_class *prev_class = p->sched_class;
6015 BUG_ON(prio < 0 || prio > MAX_PRIO);
6017 rq = task_rq_lock(p, &flags);
6018 update_rq_clock(rq);
6021 on_rq = p->se.on_rq;
6022 running = task_current(rq, p);
6024 dequeue_task(rq, p, 0);
6026 p->sched_class->put_prev_task(rq, p);
6029 p->sched_class = &rt_sched_class;
6031 p->sched_class = &fair_sched_class;
6036 p->sched_class->set_curr_task(rq);
6038 enqueue_task(rq, p, 0);
6040 check_class_changed(rq, p, prev_class, oldprio, running);
6042 task_rq_unlock(rq, &flags);
6047 void set_user_nice(struct task_struct *p, long nice)
6049 int old_prio, delta, on_rq;
6050 unsigned long flags;
6053 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
6056 * We have to be careful, if called from sys_setpriority(),
6057 * the task might be in the middle of scheduling on another CPU.
6059 rq = task_rq_lock(p, &flags);
6060 update_rq_clock(rq);
6062 * The RT priorities are set via sched_setscheduler(), but we still
6063 * allow the 'normal' nice value to be set - but as expected
6064 * it wont have any effect on scheduling until the task is
6065 * SCHED_FIFO/SCHED_RR:
6067 if (task_has_rt_policy(p)) {
6068 p->static_prio = NICE_TO_PRIO(nice);
6071 on_rq = p->se.on_rq;
6073 dequeue_task(rq, p, 0);
6075 p->static_prio = NICE_TO_PRIO(nice);
6078 p->prio = effective_prio(p);
6079 delta = p->prio - old_prio;
6082 enqueue_task(rq, p, 0);
6084 * If the task increased its priority or is running and
6085 * lowered its priority, then reschedule its CPU:
6087 if (delta < 0 || (delta > 0 && task_running(rq, p)))
6088 resched_task(rq->curr);
6091 task_rq_unlock(rq, &flags);
6093 EXPORT_SYMBOL(set_user_nice);
6096 * can_nice - check if a task can reduce its nice value
6100 int can_nice(const struct task_struct *p, const int nice)
6102 /* convert nice value [19,-20] to rlimit style value [1,40] */
6103 int nice_rlim = 20 - nice;
6105 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
6106 capable(CAP_SYS_NICE));
6109 #ifdef __ARCH_WANT_SYS_NICE
6112 * sys_nice - change the priority of the current process.
6113 * @increment: priority increment
6115 * sys_setpriority is a more generic, but much slower function that
6116 * does similar things.
6118 SYSCALL_DEFINE1(nice, int, increment)
6123 * Setpriority might change our priority at the same moment.
6124 * We don't have to worry. Conceptually one call occurs first
6125 * and we have a single winner.
6127 if (increment < -40)
6132 nice = TASK_NICE(current) + increment;
6138 if (increment < 0 && !can_nice(current, nice))
6141 retval = security_task_setnice(current, nice);
6145 set_user_nice(current, nice);
6152 * task_prio - return the priority value of a given task.
6153 * @p: the task in question.
6155 * This is the priority value as seen by users in /proc.
6156 * RT tasks are offset by -200. Normal tasks are centered
6157 * around 0, value goes from -16 to +15.
6159 int task_prio(const struct task_struct *p)
6161 return p->prio - MAX_RT_PRIO;
6165 * task_nice - return the nice value of a given task.
6166 * @p: the task in question.
6168 int task_nice(const struct task_struct *p)
6170 return TASK_NICE(p);
6172 EXPORT_SYMBOL(task_nice);
6175 * idle_cpu - is a given cpu idle currently?
6176 * @cpu: the processor in question.
6178 int idle_cpu(int cpu)
6180 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
6184 * idle_task - return the idle task for a given cpu.
6185 * @cpu: the processor in question.
6187 struct task_struct *idle_task(int cpu)
6189 return cpu_rq(cpu)->idle;
6193 * find_process_by_pid - find a process with a matching PID value.
6194 * @pid: the pid in question.
6196 static struct task_struct *find_process_by_pid(pid_t pid)
6198 return pid ? find_task_by_vpid(pid) : current;
6201 /* Actually do priority change: must hold rq lock. */
6203 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
6205 BUG_ON(p->se.on_rq);
6208 p->rt_priority = prio;
6209 p->normal_prio = normal_prio(p);
6210 /* we are holding p->pi_lock already */
6211 p->prio = rt_mutex_getprio(p);
6212 if (rt_prio(p->prio))
6213 p->sched_class = &rt_sched_class;
6215 p->sched_class = &fair_sched_class;
6220 * check the target process has a UID that matches the current process's
6222 static bool check_same_owner(struct task_struct *p)
6224 const struct cred *cred = current_cred(), *pcred;
6228 pcred = __task_cred(p);
6229 match = (cred->euid == pcred->euid ||
6230 cred->euid == pcred->uid);
6235 static int __sched_setscheduler(struct task_struct *p, int policy,
6236 struct sched_param *param, bool user)
6238 int retval, oldprio, oldpolicy = -1, on_rq, running;
6239 unsigned long flags;
6240 const struct sched_class *prev_class = p->sched_class;
6244 /* may grab non-irq protected spin_locks */
6245 BUG_ON(in_interrupt());
6247 /* double check policy once rq lock held */
6249 reset_on_fork = p->sched_reset_on_fork;
6250 policy = oldpolicy = p->policy;
6252 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
6253 policy &= ~SCHED_RESET_ON_FORK;
6255 if (policy != SCHED_FIFO && policy != SCHED_RR &&
6256 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
6257 policy != SCHED_IDLE)
6262 * Valid priorities for SCHED_FIFO and SCHED_RR are
6263 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6264 * SCHED_BATCH and SCHED_IDLE is 0.
6266 if (param->sched_priority < 0 ||
6267 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
6268 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
6270 if (rt_policy(policy) != (param->sched_priority != 0))
6274 * Allow unprivileged RT tasks to decrease priority:
6276 if (user && !capable(CAP_SYS_NICE)) {
6277 if (rt_policy(policy)) {
6278 unsigned long rlim_rtprio;
6280 if (!lock_task_sighand(p, &flags))
6282 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
6283 unlock_task_sighand(p, &flags);
6285 /* can't set/change the rt policy */
6286 if (policy != p->policy && !rlim_rtprio)
6289 /* can't increase priority */
6290 if (param->sched_priority > p->rt_priority &&
6291 param->sched_priority > rlim_rtprio)
6295 * Like positive nice levels, dont allow tasks to
6296 * move out of SCHED_IDLE either:
6298 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
6301 /* can't change other user's priorities */
6302 if (!check_same_owner(p))
6305 /* Normal users shall not reset the sched_reset_on_fork flag */
6306 if (p->sched_reset_on_fork && !reset_on_fork)
6311 #ifdef CONFIG_RT_GROUP_SCHED
6313 * Do not allow realtime tasks into groups that have no runtime
6316 if (rt_bandwidth_enabled() && rt_policy(policy) &&
6317 task_group(p)->rt_bandwidth.rt_runtime == 0)
6321 retval = security_task_setscheduler(p, policy, param);
6327 * make sure no PI-waiters arrive (or leave) while we are
6328 * changing the priority of the task:
6330 spin_lock_irqsave(&p->pi_lock, flags);
6332 * To be able to change p->policy safely, the apropriate
6333 * runqueue lock must be held.
6335 rq = __task_rq_lock(p);
6336 /* recheck policy now with rq lock held */
6337 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
6338 policy = oldpolicy = -1;
6339 __task_rq_unlock(rq);
6340 spin_unlock_irqrestore(&p->pi_lock, flags);
6343 update_rq_clock(rq);
6344 on_rq = p->se.on_rq;
6345 running = task_current(rq, p);
6347 deactivate_task(rq, p, 0);
6349 p->sched_class->put_prev_task(rq, p);
6351 p->sched_reset_on_fork = reset_on_fork;
6354 __setscheduler(rq, p, policy, param->sched_priority);
6357 p->sched_class->set_curr_task(rq);
6359 activate_task(rq, p, 0);
6361 check_class_changed(rq, p, prev_class, oldprio, running);
6363 __task_rq_unlock(rq);
6364 spin_unlock_irqrestore(&p->pi_lock, flags);
6366 rt_mutex_adjust_pi(p);
6372 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6373 * @p: the task in question.
6374 * @policy: new policy.
6375 * @param: structure containing the new RT priority.
6377 * NOTE that the task may be already dead.
6379 int sched_setscheduler(struct task_struct *p, int policy,
6380 struct sched_param *param)
6382 return __sched_setscheduler(p, policy, param, true);
6384 EXPORT_SYMBOL_GPL(sched_setscheduler);
6387 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6388 * @p: the task in question.
6389 * @policy: new policy.
6390 * @param: structure containing the new RT priority.
6392 * Just like sched_setscheduler, only don't bother checking if the
6393 * current context has permission. For example, this is needed in
6394 * stop_machine(): we create temporary high priority worker threads,
6395 * but our caller might not have that capability.
6397 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
6398 struct sched_param *param)
6400 return __sched_setscheduler(p, policy, param, false);
6404 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
6406 struct sched_param lparam;
6407 struct task_struct *p;
6410 if (!param || pid < 0)
6412 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
6417 p = find_process_by_pid(pid);
6419 retval = sched_setscheduler(p, policy, &lparam);
6426 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6427 * @pid: the pid in question.
6428 * @policy: new policy.
6429 * @param: structure containing the new RT priority.
6431 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
6432 struct sched_param __user *, param)
6434 /* negative values for policy are not valid */
6438 return do_sched_setscheduler(pid, policy, param);
6442 * sys_sched_setparam - set/change the RT priority of a thread
6443 * @pid: the pid in question.
6444 * @param: structure containing the new RT priority.
6446 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
6448 return do_sched_setscheduler(pid, -1, param);
6452 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6453 * @pid: the pid in question.
6455 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
6457 struct task_struct *p;
6464 read_lock(&tasklist_lock);
6465 p = find_process_by_pid(pid);
6467 retval = security_task_getscheduler(p);
6470 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
6472 read_unlock(&tasklist_lock);
6477 * sys_sched_getparam - get the RT priority of a thread
6478 * @pid: the pid in question.
6479 * @param: structure containing the RT priority.
6481 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
6483 struct sched_param lp;
6484 struct task_struct *p;
6487 if (!param || pid < 0)
6490 read_lock(&tasklist_lock);
6491 p = find_process_by_pid(pid);
6496 retval = security_task_getscheduler(p);
6500 lp.sched_priority = p->rt_priority;
6501 read_unlock(&tasklist_lock);
6504 * This one might sleep, we cannot do it with a spinlock held ...
6506 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
6511 read_unlock(&tasklist_lock);
6515 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
6517 cpumask_var_t cpus_allowed, new_mask;
6518 struct task_struct *p;
6522 read_lock(&tasklist_lock);
6524 p = find_process_by_pid(pid);
6526 read_unlock(&tasklist_lock);
6532 * It is not safe to call set_cpus_allowed with the
6533 * tasklist_lock held. We will bump the task_struct's
6534 * usage count and then drop tasklist_lock.
6537 read_unlock(&tasklist_lock);
6539 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6543 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
6545 goto out_free_cpus_allowed;
6548 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
6551 retval = security_task_setscheduler(p, 0, NULL);
6555 cpuset_cpus_allowed(p, cpus_allowed);
6556 cpumask_and(new_mask, in_mask, cpus_allowed);
6558 retval = set_cpus_allowed_ptr(p, new_mask);
6561 cpuset_cpus_allowed(p, cpus_allowed);
6562 if (!cpumask_subset(new_mask, cpus_allowed)) {
6564 * We must have raced with a concurrent cpuset
6565 * update. Just reset the cpus_allowed to the
6566 * cpuset's cpus_allowed
6568 cpumask_copy(new_mask, cpus_allowed);
6573 free_cpumask_var(new_mask);
6574 out_free_cpus_allowed:
6575 free_cpumask_var(cpus_allowed);
6582 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
6583 struct cpumask *new_mask)
6585 if (len < cpumask_size())
6586 cpumask_clear(new_mask);
6587 else if (len > cpumask_size())
6588 len = cpumask_size();
6590 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6594 * sys_sched_setaffinity - set the cpu affinity of a process
6595 * @pid: pid of the process
6596 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6597 * @user_mask_ptr: user-space pointer to the new cpu mask
6599 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6600 unsigned long __user *, user_mask_ptr)
6602 cpumask_var_t new_mask;
6605 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6608 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6610 retval = sched_setaffinity(pid, new_mask);
6611 free_cpumask_var(new_mask);
6615 long sched_getaffinity(pid_t pid, struct cpumask *mask)
6617 struct task_struct *p;
6618 unsigned long flags;
6623 read_lock(&tasklist_lock);
6626 p = find_process_by_pid(pid);
6630 retval = security_task_getscheduler(p);
6634 rq = task_rq_lock(p, &flags);
6635 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
6636 task_rq_unlock(rq, &flags);
6639 read_unlock(&tasklist_lock);
6646 * sys_sched_getaffinity - get the cpu affinity of a process
6647 * @pid: pid of the process
6648 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6649 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6651 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6652 unsigned long __user *, user_mask_ptr)
6657 if (len < cpumask_size())
6660 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6663 ret = sched_getaffinity(pid, mask);
6665 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
6668 ret = cpumask_size();
6670 free_cpumask_var(mask);
6676 * sys_sched_yield - yield the current processor to other threads.
6678 * This function yields the current CPU to other tasks. If there are no
6679 * other threads running on this CPU then this function will return.
6681 SYSCALL_DEFINE0(sched_yield)
6683 struct rq *rq = this_rq_lock();
6685 schedstat_inc(rq, yld_count);
6686 current->sched_class->yield_task(rq);
6689 * Since we are going to call schedule() anyway, there's
6690 * no need to preempt or enable interrupts:
6692 __release(rq->lock);
6693 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
6694 _raw_spin_unlock(&rq->lock);
6695 preempt_enable_no_resched();
6702 static inline int should_resched(void)
6704 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
6707 static void __cond_resched(void)
6709 add_preempt_count(PREEMPT_ACTIVE);
6711 sub_preempt_count(PREEMPT_ACTIVE);
6714 int __sched _cond_resched(void)
6716 if (should_resched()) {
6722 EXPORT_SYMBOL(_cond_resched);
6725 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
6726 * call schedule, and on return reacquire the lock.
6728 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6729 * operations here to prevent schedule() from being called twice (once via
6730 * spin_unlock(), once by hand).
6732 int __cond_resched_lock(spinlock_t *lock)
6734 int resched = should_resched();
6737 lockdep_assert_held(lock);
6739 if (spin_needbreak(lock) || resched) {
6750 EXPORT_SYMBOL(__cond_resched_lock);
6752 int __sched __cond_resched_softirq(void)
6754 BUG_ON(!in_softirq());
6756 if (should_resched()) {
6764 EXPORT_SYMBOL(__cond_resched_softirq);
6767 * yield - yield the current processor to other threads.
6769 * This is a shortcut for kernel-space yielding - it marks the
6770 * thread runnable and calls sys_sched_yield().
6772 void __sched yield(void)
6774 set_current_state(TASK_RUNNING);
6777 EXPORT_SYMBOL(yield);
6780 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6781 * that process accounting knows that this is a task in IO wait state.
6783 void __sched io_schedule(void)
6785 struct rq *rq = raw_rq();
6787 delayacct_blkio_start();
6788 atomic_inc(&rq->nr_iowait);
6789 current->in_iowait = 1;
6791 current->in_iowait = 0;
6792 atomic_dec(&rq->nr_iowait);
6793 delayacct_blkio_end();
6795 EXPORT_SYMBOL(io_schedule);
6797 long __sched io_schedule_timeout(long timeout)
6799 struct rq *rq = raw_rq();
6802 delayacct_blkio_start();
6803 atomic_inc(&rq->nr_iowait);
6804 current->in_iowait = 1;
6805 ret = schedule_timeout(timeout);
6806 current->in_iowait = 0;
6807 atomic_dec(&rq->nr_iowait);
6808 delayacct_blkio_end();
6813 * sys_sched_get_priority_max - return maximum RT priority.
6814 * @policy: scheduling class.
6816 * this syscall returns the maximum rt_priority that can be used
6817 * by a given scheduling class.
6819 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6826 ret = MAX_USER_RT_PRIO-1;
6838 * sys_sched_get_priority_min - return minimum RT priority.
6839 * @policy: scheduling class.
6841 * this syscall returns the minimum rt_priority that can be used
6842 * by a given scheduling class.
6844 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6862 * sys_sched_rr_get_interval - return the default timeslice of a process.
6863 * @pid: pid of the process.
6864 * @interval: userspace pointer to the timeslice value.
6866 * this syscall writes the default timeslice value of a given process
6867 * into the user-space timespec buffer. A value of '0' means infinity.
6869 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6870 struct timespec __user *, interval)
6872 struct task_struct *p;
6873 unsigned int time_slice;
6874 unsigned long flags;
6883 read_lock(&tasklist_lock);
6884 p = find_process_by_pid(pid);
6888 retval = security_task_getscheduler(p);
6892 rq = task_rq_lock(p, &flags);
6893 time_slice = p->sched_class->get_rr_interval(rq, p);
6894 task_rq_unlock(rq, &flags);
6896 read_unlock(&tasklist_lock);
6897 jiffies_to_timespec(time_slice, &t);
6898 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
6902 read_unlock(&tasklist_lock);
6906 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
6908 void sched_show_task(struct task_struct *p)
6910 unsigned long free = 0;
6913 state = p->state ? __ffs(p->state) + 1 : 0;
6914 printk(KERN_INFO "%-13.13s %c", p->comm,
6915 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
6916 #if BITS_PER_LONG == 32
6917 if (state == TASK_RUNNING)
6918 printk(KERN_CONT " running ");
6920 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
6922 if (state == TASK_RUNNING)
6923 printk(KERN_CONT " running task ");
6925 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
6927 #ifdef CONFIG_DEBUG_STACK_USAGE
6928 free = stack_not_used(p);
6930 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
6931 task_pid_nr(p), task_pid_nr(p->real_parent),
6932 (unsigned long)task_thread_info(p)->flags);
6934 show_stack(p, NULL);
6937 void show_state_filter(unsigned long state_filter)
6939 struct task_struct *g, *p;
6941 #if BITS_PER_LONG == 32
6943 " task PC stack pid father\n");
6946 " task PC stack pid father\n");
6948 read_lock(&tasklist_lock);
6949 do_each_thread(g, p) {
6951 * reset the NMI-timeout, listing all files on a slow
6952 * console might take alot of time:
6954 touch_nmi_watchdog();
6955 if (!state_filter || (p->state & state_filter))
6957 } while_each_thread(g, p);
6959 touch_all_softlockup_watchdogs();
6961 #ifdef CONFIG_SCHED_DEBUG
6962 sysrq_sched_debug_show();
6964 read_unlock(&tasklist_lock);
6966 * Only show locks if all tasks are dumped:
6969 debug_show_all_locks();
6972 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6974 idle->sched_class = &idle_sched_class;
6978 * init_idle - set up an idle thread for a given CPU
6979 * @idle: task in question
6980 * @cpu: cpu the idle task belongs to
6982 * NOTE: this function does not set the idle thread's NEED_RESCHED
6983 * flag, to make booting more robust.
6985 void __cpuinit init_idle(struct task_struct *idle, int cpu)
6987 struct rq *rq = cpu_rq(cpu);
6988 unsigned long flags;
6990 spin_lock_irqsave(&rq->lock, flags);
6993 idle->se.exec_start = sched_clock();
6995 idle->prio = idle->normal_prio = MAX_PRIO;
6996 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
6997 __set_task_cpu(idle, cpu);
6999 rq->curr = rq->idle = idle;
7000 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
7003 spin_unlock_irqrestore(&rq->lock, flags);
7005 /* Set the preempt count _outside_ the spinlocks! */
7006 #if defined(CONFIG_PREEMPT)
7007 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
7009 task_thread_info(idle)->preempt_count = 0;
7012 * The idle tasks have their own, simple scheduling class:
7014 idle->sched_class = &idle_sched_class;
7015 ftrace_graph_init_task(idle);
7019 * In a system that switches off the HZ timer nohz_cpu_mask
7020 * indicates which cpus entered this state. This is used
7021 * in the rcu update to wait only for active cpus. For system
7022 * which do not switch off the HZ timer nohz_cpu_mask should
7023 * always be CPU_BITS_NONE.
7025 cpumask_var_t nohz_cpu_mask;
7028 * Increase the granularity value when there are more CPUs,
7029 * because with more CPUs the 'effective latency' as visible
7030 * to users decreases. But the relationship is not linear,
7031 * so pick a second-best guess by going with the log2 of the
7034 * This idea comes from the SD scheduler of Con Kolivas:
7036 static inline void sched_init_granularity(void)
7038 unsigned int factor = 1 + ilog2(num_online_cpus());
7039 const unsigned long limit = 200000000;
7041 sysctl_sched_min_granularity *= factor;
7042 if (sysctl_sched_min_granularity > limit)
7043 sysctl_sched_min_granularity = limit;
7045 sysctl_sched_latency *= factor;
7046 if (sysctl_sched_latency > limit)
7047 sysctl_sched_latency = limit;
7049 sysctl_sched_wakeup_granularity *= factor;
7051 sysctl_sched_shares_ratelimit *= factor;
7056 * This is how migration works:
7058 * 1) we queue a struct migration_req structure in the source CPU's
7059 * runqueue and wake up that CPU's migration thread.
7060 * 2) we down() the locked semaphore => thread blocks.
7061 * 3) migration thread wakes up (implicitly it forces the migrated
7062 * thread off the CPU)
7063 * 4) it gets the migration request and checks whether the migrated
7064 * task is still in the wrong runqueue.
7065 * 5) if it's in the wrong runqueue then the migration thread removes
7066 * it and puts it into the right queue.
7067 * 6) migration thread up()s the semaphore.
7068 * 7) we wake up and the migration is done.
7072 * Change a given task's CPU affinity. Migrate the thread to a
7073 * proper CPU and schedule it away if the CPU it's executing on
7074 * is removed from the allowed bitmask.
7076 * NOTE: the caller must have a valid reference to the task, the
7077 * task must not exit() & deallocate itself prematurely. The
7078 * call is not atomic; no spinlocks may be held.
7080 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
7082 struct migration_req req;
7083 unsigned long flags;
7087 rq = task_rq_lock(p, &flags);
7088 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
7093 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
7094 !cpumask_equal(&p->cpus_allowed, new_mask))) {
7099 if (p->sched_class->set_cpus_allowed)
7100 p->sched_class->set_cpus_allowed(p, new_mask);
7102 cpumask_copy(&p->cpus_allowed, new_mask);
7103 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
7106 /* Can the task run on the task's current CPU? If so, we're done */
7107 if (cpumask_test_cpu(task_cpu(p), new_mask))
7110 if (migrate_task(p, cpumask_any_and(cpu_active_mask, new_mask), &req)) {
7111 /* Need help from migration thread: drop lock and wait. */
7112 struct task_struct *mt = rq->migration_thread;
7114 get_task_struct(mt);
7115 task_rq_unlock(rq, &flags);
7116 wake_up_process(rq->migration_thread);
7117 put_task_struct(mt);
7118 wait_for_completion(&req.done);
7119 tlb_migrate_finish(p->mm);
7123 task_rq_unlock(rq, &flags);
7127 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
7130 * Move (not current) task off this cpu, onto dest cpu. We're doing
7131 * this because either it can't run here any more (set_cpus_allowed()
7132 * away from this CPU, or CPU going down), or because we're
7133 * attempting to rebalance this task on exec (sched_exec).
7135 * So we race with normal scheduler movements, but that's OK, as long
7136 * as the task is no longer on this CPU.
7138 * Returns non-zero if task was successfully migrated.
7140 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
7142 struct rq *rq_dest, *rq_src;
7145 if (unlikely(!cpu_active(dest_cpu)))
7148 rq_src = cpu_rq(src_cpu);
7149 rq_dest = cpu_rq(dest_cpu);
7151 double_rq_lock(rq_src, rq_dest);
7152 /* Already moved. */
7153 if (task_cpu(p) != src_cpu)
7155 /* Affinity changed (again). */
7156 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7159 on_rq = p->se.on_rq;
7161 deactivate_task(rq_src, p, 0);
7163 set_task_cpu(p, dest_cpu);
7165 activate_task(rq_dest, p, 0);
7166 check_preempt_curr(rq_dest, p, 0);
7171 double_rq_unlock(rq_src, rq_dest);
7175 #define RCU_MIGRATION_IDLE 0
7176 #define RCU_MIGRATION_NEED_QS 1
7177 #define RCU_MIGRATION_GOT_QS 2
7178 #define RCU_MIGRATION_MUST_SYNC 3
7181 * migration_thread - this is a highprio system thread that performs
7182 * thread migration by bumping thread off CPU then 'pushing' onto
7185 static int migration_thread(void *data)
7188 int cpu = (long)data;
7192 BUG_ON(rq->migration_thread != current);
7194 set_current_state(TASK_INTERRUPTIBLE);
7195 while (!kthread_should_stop()) {
7196 struct migration_req *req;
7197 struct list_head *head;
7199 spin_lock_irq(&rq->lock);
7201 if (cpu_is_offline(cpu)) {
7202 spin_unlock_irq(&rq->lock);
7206 if (rq->active_balance) {
7207 active_load_balance(rq, cpu);
7208 rq->active_balance = 0;
7211 head = &rq->migration_queue;
7213 if (list_empty(head)) {
7214 spin_unlock_irq(&rq->lock);
7216 set_current_state(TASK_INTERRUPTIBLE);
7219 req = list_entry(head->next, struct migration_req, list);
7220 list_del_init(head->next);
7222 if (req->task != NULL) {
7223 spin_unlock(&rq->lock);
7224 __migrate_task(req->task, cpu, req->dest_cpu);
7225 } else if (likely(cpu == (badcpu = smp_processor_id()))) {
7226 req->dest_cpu = RCU_MIGRATION_GOT_QS;
7227 spin_unlock(&rq->lock);
7229 req->dest_cpu = RCU_MIGRATION_MUST_SYNC;
7230 spin_unlock(&rq->lock);
7231 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu, cpu);
7235 complete(&req->done);
7237 __set_current_state(TASK_RUNNING);
7242 #ifdef CONFIG_HOTPLUG_CPU
7244 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
7248 local_irq_disable();
7249 ret = __migrate_task(p, src_cpu, dest_cpu);
7255 * Figure out where task on dead CPU should go, use force if necessary.
7257 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
7260 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(dead_cpu));
7263 /* Look for allowed, online CPU in same node. */
7264 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
7265 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7268 /* Any allowed, online CPU? */
7269 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
7270 if (dest_cpu < nr_cpu_ids)
7273 /* No more Mr. Nice Guy. */
7274 if (dest_cpu >= nr_cpu_ids) {
7275 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
7276 dest_cpu = cpumask_any_and(cpu_active_mask, &p->cpus_allowed);
7279 * Don't tell them about moving exiting tasks or
7280 * kernel threads (both mm NULL), since they never
7283 if (p->mm && printk_ratelimit()) {
7284 printk(KERN_INFO "process %d (%s) no "
7285 "longer affine to cpu%d\n",
7286 task_pid_nr(p), p->comm, dead_cpu);
7291 /* It can have affinity changed while we were choosing. */
7292 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
7297 * While a dead CPU has no uninterruptible tasks queued at this point,
7298 * it might still have a nonzero ->nr_uninterruptible counter, because
7299 * for performance reasons the counter is not stricly tracking tasks to
7300 * their home CPUs. So we just add the counter to another CPU's counter,
7301 * to keep the global sum constant after CPU-down:
7303 static void migrate_nr_uninterruptible(struct rq *rq_src)
7305 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
7306 unsigned long flags;
7308 local_irq_save(flags);
7309 double_rq_lock(rq_src, rq_dest);
7310 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
7311 rq_src->nr_uninterruptible = 0;
7312 double_rq_unlock(rq_src, rq_dest);
7313 local_irq_restore(flags);
7316 /* Run through task list and migrate tasks from the dead cpu. */
7317 static void migrate_live_tasks(int src_cpu)
7319 struct task_struct *p, *t;
7321 read_lock(&tasklist_lock);
7323 do_each_thread(t, p) {
7327 if (task_cpu(p) == src_cpu)
7328 move_task_off_dead_cpu(src_cpu, p);
7329 } while_each_thread(t, p);
7331 read_unlock(&tasklist_lock);
7335 * Schedules idle task to be the next runnable task on current CPU.
7336 * It does so by boosting its priority to highest possible.
7337 * Used by CPU offline code.
7339 void sched_idle_next(void)
7341 int this_cpu = smp_processor_id();
7342 struct rq *rq = cpu_rq(this_cpu);
7343 struct task_struct *p = rq->idle;
7344 unsigned long flags;
7346 /* cpu has to be offline */
7347 BUG_ON(cpu_online(this_cpu));
7350 * Strictly not necessary since rest of the CPUs are stopped by now
7351 * and interrupts disabled on the current cpu.
7353 spin_lock_irqsave(&rq->lock, flags);
7355 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7357 update_rq_clock(rq);
7358 activate_task(rq, p, 0);
7360 spin_unlock_irqrestore(&rq->lock, flags);
7364 * Ensures that the idle task is using init_mm right before its cpu goes
7367 void idle_task_exit(void)
7369 struct mm_struct *mm = current->active_mm;
7371 BUG_ON(cpu_online(smp_processor_id()));
7374 switch_mm(mm, &init_mm, current);
7378 /* called under rq->lock with disabled interrupts */
7379 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
7381 struct rq *rq = cpu_rq(dead_cpu);
7383 /* Must be exiting, otherwise would be on tasklist. */
7384 BUG_ON(!p->exit_state);
7386 /* Cannot have done final schedule yet: would have vanished. */
7387 BUG_ON(p->state == TASK_DEAD);
7392 * Drop lock around migration; if someone else moves it,
7393 * that's OK. No task can be added to this CPU, so iteration is
7396 spin_unlock_irq(&rq->lock);
7397 move_task_off_dead_cpu(dead_cpu, p);
7398 spin_lock_irq(&rq->lock);
7403 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7404 static void migrate_dead_tasks(unsigned int dead_cpu)
7406 struct rq *rq = cpu_rq(dead_cpu);
7407 struct task_struct *next;
7410 if (!rq->nr_running)
7412 update_rq_clock(rq);
7413 next = pick_next_task(rq);
7416 next->sched_class->put_prev_task(rq, next);
7417 migrate_dead(dead_cpu, next);
7423 * remove the tasks which were accounted by rq from calc_load_tasks.
7425 static void calc_global_load_remove(struct rq *rq)
7427 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
7428 rq->calc_load_active = 0;
7430 #endif /* CONFIG_HOTPLUG_CPU */
7432 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7434 static struct ctl_table sd_ctl_dir[] = {
7436 .procname = "sched_domain",
7442 static struct ctl_table sd_ctl_root[] = {
7444 .ctl_name = CTL_KERN,
7445 .procname = "kernel",
7447 .child = sd_ctl_dir,
7452 static struct ctl_table *sd_alloc_ctl_entry(int n)
7454 struct ctl_table *entry =
7455 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
7460 static void sd_free_ctl_entry(struct ctl_table **tablep)
7462 struct ctl_table *entry;
7465 * In the intermediate directories, both the child directory and
7466 * procname are dynamically allocated and could fail but the mode
7467 * will always be set. In the lowest directory the names are
7468 * static strings and all have proc handlers.
7470 for (entry = *tablep; entry->mode; entry++) {
7472 sd_free_ctl_entry(&entry->child);
7473 if (entry->proc_handler == NULL)
7474 kfree(entry->procname);
7482 set_table_entry(struct ctl_table *entry,
7483 const char *procname, void *data, int maxlen,
7484 mode_t mode, proc_handler *proc_handler)
7486 entry->procname = procname;
7488 entry->maxlen = maxlen;
7490 entry->proc_handler = proc_handler;
7493 static struct ctl_table *
7494 sd_alloc_ctl_domain_table(struct sched_domain *sd)
7496 struct ctl_table *table = sd_alloc_ctl_entry(13);
7501 set_table_entry(&table[0], "min_interval", &sd->min_interval,
7502 sizeof(long), 0644, proc_doulongvec_minmax);
7503 set_table_entry(&table[1], "max_interval", &sd->max_interval,
7504 sizeof(long), 0644, proc_doulongvec_minmax);
7505 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
7506 sizeof(int), 0644, proc_dointvec_minmax);
7507 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
7508 sizeof(int), 0644, proc_dointvec_minmax);
7509 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
7510 sizeof(int), 0644, proc_dointvec_minmax);
7511 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
7512 sizeof(int), 0644, proc_dointvec_minmax);
7513 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
7514 sizeof(int), 0644, proc_dointvec_minmax);
7515 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
7516 sizeof(int), 0644, proc_dointvec_minmax);
7517 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
7518 sizeof(int), 0644, proc_dointvec_minmax);
7519 set_table_entry(&table[9], "cache_nice_tries",
7520 &sd->cache_nice_tries,
7521 sizeof(int), 0644, proc_dointvec_minmax);
7522 set_table_entry(&table[10], "flags", &sd->flags,
7523 sizeof(int), 0644, proc_dointvec_minmax);
7524 set_table_entry(&table[11], "name", sd->name,
7525 CORENAME_MAX_SIZE, 0444, proc_dostring);
7526 /* &table[12] is terminator */
7531 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
7533 struct ctl_table *entry, *table;
7534 struct sched_domain *sd;
7535 int domain_num = 0, i;
7538 for_each_domain(cpu, sd)
7540 entry = table = sd_alloc_ctl_entry(domain_num + 1);
7545 for_each_domain(cpu, sd) {
7546 snprintf(buf, 32, "domain%d", i);
7547 entry->procname = kstrdup(buf, GFP_KERNEL);
7549 entry->child = sd_alloc_ctl_domain_table(sd);
7556 static struct ctl_table_header *sd_sysctl_header;
7557 static void register_sched_domain_sysctl(void)
7559 int i, cpu_num = num_possible_cpus();
7560 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
7563 WARN_ON(sd_ctl_dir[0].child);
7564 sd_ctl_dir[0].child = entry;
7569 for_each_possible_cpu(i) {
7570 snprintf(buf, 32, "cpu%d", i);
7571 entry->procname = kstrdup(buf, GFP_KERNEL);
7573 entry->child = sd_alloc_ctl_cpu_table(i);
7577 WARN_ON(sd_sysctl_header);
7578 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
7581 /* may be called multiple times per register */
7582 static void unregister_sched_domain_sysctl(void)
7584 if (sd_sysctl_header)
7585 unregister_sysctl_table(sd_sysctl_header);
7586 sd_sysctl_header = NULL;
7587 if (sd_ctl_dir[0].child)
7588 sd_free_ctl_entry(&sd_ctl_dir[0].child);
7591 static void register_sched_domain_sysctl(void)
7594 static void unregister_sched_domain_sysctl(void)
7599 static void set_rq_online(struct rq *rq)
7602 const struct sched_class *class;
7604 cpumask_set_cpu(rq->cpu, rq->rd->online);
7607 for_each_class(class) {
7608 if (class->rq_online)
7609 class->rq_online(rq);
7614 static void set_rq_offline(struct rq *rq)
7617 const struct sched_class *class;
7619 for_each_class(class) {
7620 if (class->rq_offline)
7621 class->rq_offline(rq);
7624 cpumask_clear_cpu(rq->cpu, rq->rd->online);
7630 * migration_call - callback that gets triggered when a CPU is added.
7631 * Here we can start up the necessary migration thread for the new CPU.
7633 static int __cpuinit
7634 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
7636 struct task_struct *p;
7637 int cpu = (long)hcpu;
7638 unsigned long flags;
7643 case CPU_UP_PREPARE:
7644 case CPU_UP_PREPARE_FROZEN:
7645 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
7648 kthread_bind(p, cpu);
7649 /* Must be high prio: stop_machine expects to yield to it. */
7650 rq = task_rq_lock(p, &flags);
7651 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7652 task_rq_unlock(rq, &flags);
7654 cpu_rq(cpu)->migration_thread = p;
7655 rq->calc_load_update = calc_load_update;
7659 case CPU_ONLINE_FROZEN:
7660 /* Strictly unnecessary, as first user will wake it. */
7661 wake_up_process(cpu_rq(cpu)->migration_thread);
7663 /* Update our root-domain */
7665 spin_lock_irqsave(&rq->lock, flags);
7667 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7671 spin_unlock_irqrestore(&rq->lock, flags);
7674 #ifdef CONFIG_HOTPLUG_CPU
7675 case CPU_UP_CANCELED:
7676 case CPU_UP_CANCELED_FROZEN:
7677 if (!cpu_rq(cpu)->migration_thread)
7679 /* Unbind it from offline cpu so it can run. Fall thru. */
7680 kthread_bind(cpu_rq(cpu)->migration_thread,
7681 cpumask_any(cpu_online_mask));
7682 kthread_stop(cpu_rq(cpu)->migration_thread);
7683 put_task_struct(cpu_rq(cpu)->migration_thread);
7684 cpu_rq(cpu)->migration_thread = NULL;
7688 case CPU_DEAD_FROZEN:
7689 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7690 migrate_live_tasks(cpu);
7692 kthread_stop(rq->migration_thread);
7693 put_task_struct(rq->migration_thread);
7694 rq->migration_thread = NULL;
7695 /* Idle task back to normal (off runqueue, low prio) */
7696 spin_lock_irq(&rq->lock);
7697 update_rq_clock(rq);
7698 deactivate_task(rq, rq->idle, 0);
7699 rq->idle->static_prio = MAX_PRIO;
7700 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
7701 rq->idle->sched_class = &idle_sched_class;
7702 migrate_dead_tasks(cpu);
7703 spin_unlock_irq(&rq->lock);
7705 migrate_nr_uninterruptible(rq);
7706 BUG_ON(rq->nr_running != 0);
7707 calc_global_load_remove(rq);
7709 * No need to migrate the tasks: it was best-effort if
7710 * they didn't take sched_hotcpu_mutex. Just wake up
7713 spin_lock_irq(&rq->lock);
7714 while (!list_empty(&rq->migration_queue)) {
7715 struct migration_req *req;
7717 req = list_entry(rq->migration_queue.next,
7718 struct migration_req, list);
7719 list_del_init(&req->list);
7720 spin_unlock_irq(&rq->lock);
7721 complete(&req->done);
7722 spin_lock_irq(&rq->lock);
7724 spin_unlock_irq(&rq->lock);
7728 case CPU_DYING_FROZEN:
7729 /* Update our root-domain */
7731 spin_lock_irqsave(&rq->lock, flags);
7733 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7736 spin_unlock_irqrestore(&rq->lock, flags);
7744 * Register at high priority so that task migration (migrate_all_tasks)
7745 * happens before everything else. This has to be lower priority than
7746 * the notifier in the perf_event subsystem, though.
7748 static struct notifier_block __cpuinitdata migration_notifier = {
7749 .notifier_call = migration_call,
7753 static int __init migration_init(void)
7755 void *cpu = (void *)(long)smp_processor_id();
7758 /* Start one for the boot CPU: */
7759 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
7760 BUG_ON(err == NOTIFY_BAD);
7761 migration_call(&migration_notifier, CPU_ONLINE, cpu);
7762 register_cpu_notifier(&migration_notifier);
7766 early_initcall(migration_init);
7771 #ifdef CONFIG_SCHED_DEBUG
7773 static __read_mostly int sched_domain_debug_enabled;
7775 static int __init sched_domain_debug_setup(char *str)
7777 sched_domain_debug_enabled = 1;
7781 early_param("sched_debug", sched_domain_debug_setup);
7783 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
7784 struct cpumask *groupmask)
7786 struct sched_group *group = sd->groups;
7789 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
7790 cpumask_clear(groupmask);
7792 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
7794 if (!(sd->flags & SD_LOAD_BALANCE)) {
7795 printk("does not load-balance\n");
7797 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
7802 printk(KERN_CONT "span %s level %s\n", str, sd->name);
7804 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
7805 printk(KERN_ERR "ERROR: domain->span does not contain "
7808 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
7809 printk(KERN_ERR "ERROR: domain->groups does not contain"
7813 printk(KERN_DEBUG "%*s groups:", level + 1, "");
7817 printk(KERN_ERR "ERROR: group is NULL\n");
7821 if (!group->cpu_power) {
7822 printk(KERN_CONT "\n");
7823 printk(KERN_ERR "ERROR: domain->cpu_power not "
7828 if (!cpumask_weight(sched_group_cpus(group))) {
7829 printk(KERN_CONT "\n");
7830 printk(KERN_ERR "ERROR: empty group\n");
7834 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
7835 printk(KERN_CONT "\n");
7836 printk(KERN_ERR "ERROR: repeated CPUs\n");
7840 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
7842 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
7844 printk(KERN_CONT " %s", str);
7845 if (group->cpu_power != SCHED_LOAD_SCALE) {
7846 printk(KERN_CONT " (cpu_power = %d)",
7850 group = group->next;
7851 } while (group != sd->groups);
7852 printk(KERN_CONT "\n");
7854 if (!cpumask_equal(sched_domain_span(sd), groupmask))
7855 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
7858 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
7859 printk(KERN_ERR "ERROR: parent span is not a superset "
7860 "of domain->span\n");
7864 static void sched_domain_debug(struct sched_domain *sd, int cpu)
7866 cpumask_var_t groupmask;
7869 if (!sched_domain_debug_enabled)
7873 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7877 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7879 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
7880 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
7885 if (sched_domain_debug_one(sd, cpu, level, groupmask))
7892 free_cpumask_var(groupmask);
7894 #else /* !CONFIG_SCHED_DEBUG */
7895 # define sched_domain_debug(sd, cpu) do { } while (0)
7896 #endif /* CONFIG_SCHED_DEBUG */
7898 static int sd_degenerate(struct sched_domain *sd)
7900 if (cpumask_weight(sched_domain_span(sd)) == 1)
7903 /* Following flags need at least 2 groups */
7904 if (sd->flags & (SD_LOAD_BALANCE |
7905 SD_BALANCE_NEWIDLE |
7909 SD_SHARE_PKG_RESOURCES)) {
7910 if (sd->groups != sd->groups->next)
7914 /* Following flags don't use groups */
7915 if (sd->flags & (SD_WAKE_AFFINE))
7922 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
7924 unsigned long cflags = sd->flags, pflags = parent->flags;
7926 if (sd_degenerate(parent))
7929 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
7932 /* Flags needing groups don't count if only 1 group in parent */
7933 if (parent->groups == parent->groups->next) {
7934 pflags &= ~(SD_LOAD_BALANCE |
7935 SD_BALANCE_NEWIDLE |
7939 SD_SHARE_PKG_RESOURCES);
7940 if (nr_node_ids == 1)
7941 pflags &= ~SD_SERIALIZE;
7943 if (~cflags & pflags)
7949 static void free_rootdomain(struct root_domain *rd)
7951 synchronize_sched();
7953 cpupri_cleanup(&rd->cpupri);
7955 free_cpumask_var(rd->rto_mask);
7956 free_cpumask_var(rd->online);
7957 free_cpumask_var(rd->span);
7961 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7963 struct root_domain *old_rd = NULL;
7964 unsigned long flags;
7966 spin_lock_irqsave(&rq->lock, flags);
7971 if (cpumask_test_cpu(rq->cpu, old_rd->online))
7974 cpumask_clear_cpu(rq->cpu, old_rd->span);
7977 * If we dont want to free the old_rt yet then
7978 * set old_rd to NULL to skip the freeing later
7981 if (!atomic_dec_and_test(&old_rd->refcount))
7985 atomic_inc(&rd->refcount);
7988 cpumask_set_cpu(rq->cpu, rd->span);
7989 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
7992 spin_unlock_irqrestore(&rq->lock, flags);
7995 free_rootdomain(old_rd);
7998 static int init_rootdomain(struct root_domain *rd, bool bootmem)
8000 gfp_t gfp = GFP_KERNEL;
8002 memset(rd, 0, sizeof(*rd));
8007 if (!alloc_cpumask_var(&rd->span, gfp))
8009 if (!alloc_cpumask_var(&rd->online, gfp))
8011 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
8014 if (cpupri_init(&rd->cpupri, bootmem) != 0)
8019 free_cpumask_var(rd->rto_mask);
8021 free_cpumask_var(rd->online);
8023 free_cpumask_var(rd->span);
8028 static void init_defrootdomain(void)
8030 init_rootdomain(&def_root_domain, true);
8032 atomic_set(&def_root_domain.refcount, 1);
8035 static struct root_domain *alloc_rootdomain(void)
8037 struct root_domain *rd;
8039 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
8043 if (init_rootdomain(rd, false) != 0) {
8052 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
8053 * hold the hotplug lock.
8056 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
8058 struct rq *rq = cpu_rq(cpu);
8059 struct sched_domain *tmp;
8061 /* Remove the sched domains which do not contribute to scheduling. */
8062 for (tmp = sd; tmp; ) {
8063 struct sched_domain *parent = tmp->parent;
8067 if (sd_parent_degenerate(tmp, parent)) {
8068 tmp->parent = parent->parent;
8070 parent->parent->child = tmp;
8075 if (sd && sd_degenerate(sd)) {
8081 sched_domain_debug(sd, cpu);
8083 rq_attach_root(rq, rd);
8084 rcu_assign_pointer(rq->sd, sd);
8087 /* cpus with isolated domains */
8088 static cpumask_var_t cpu_isolated_map;
8090 /* Setup the mask of cpus configured for isolated domains */
8091 static int __init isolated_cpu_setup(char *str)
8093 alloc_bootmem_cpumask_var(&cpu_isolated_map);
8094 cpulist_parse(str, cpu_isolated_map);
8098 __setup("isolcpus=", isolated_cpu_setup);
8101 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
8102 * to a function which identifies what group(along with sched group) a CPU
8103 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
8104 * (due to the fact that we keep track of groups covered with a struct cpumask).
8106 * init_sched_build_groups will build a circular linked list of the groups
8107 * covered by the given span, and will set each group's ->cpumask correctly,
8108 * and ->cpu_power to 0.
8111 init_sched_build_groups(const struct cpumask *span,
8112 const struct cpumask *cpu_map,
8113 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
8114 struct sched_group **sg,
8115 struct cpumask *tmpmask),
8116 struct cpumask *covered, struct cpumask *tmpmask)
8118 struct sched_group *first = NULL, *last = NULL;
8121 cpumask_clear(covered);
8123 for_each_cpu(i, span) {
8124 struct sched_group *sg;
8125 int group = group_fn(i, cpu_map, &sg, tmpmask);
8128 if (cpumask_test_cpu(i, covered))
8131 cpumask_clear(sched_group_cpus(sg));
8134 for_each_cpu(j, span) {
8135 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
8138 cpumask_set_cpu(j, covered);
8139 cpumask_set_cpu(j, sched_group_cpus(sg));
8150 #define SD_NODES_PER_DOMAIN 16
8155 * find_next_best_node - find the next node to include in a sched_domain
8156 * @node: node whose sched_domain we're building
8157 * @used_nodes: nodes already in the sched_domain
8159 * Find the next node to include in a given scheduling domain. Simply
8160 * finds the closest node not already in the @used_nodes map.
8162 * Should use nodemask_t.
8164 static int find_next_best_node(int node, nodemask_t *used_nodes)
8166 int i, n, val, min_val, best_node = 0;
8170 for (i = 0; i < nr_node_ids; i++) {
8171 /* Start at @node */
8172 n = (node + i) % nr_node_ids;
8174 if (!nr_cpus_node(n))
8177 /* Skip already used nodes */
8178 if (node_isset(n, *used_nodes))
8181 /* Simple min distance search */
8182 val = node_distance(node, n);
8184 if (val < min_val) {
8190 node_set(best_node, *used_nodes);
8195 * sched_domain_node_span - get a cpumask for a node's sched_domain
8196 * @node: node whose cpumask we're constructing
8197 * @span: resulting cpumask
8199 * Given a node, construct a good cpumask for its sched_domain to span. It
8200 * should be one that prevents unnecessary balancing, but also spreads tasks
8203 static void sched_domain_node_span(int node, struct cpumask *span)
8205 nodemask_t used_nodes;
8208 cpumask_clear(span);
8209 nodes_clear(used_nodes);
8211 cpumask_or(span, span, cpumask_of_node(node));
8212 node_set(node, used_nodes);
8214 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
8215 int next_node = find_next_best_node(node, &used_nodes);
8217 cpumask_or(span, span, cpumask_of_node(next_node));
8220 #endif /* CONFIG_NUMA */
8222 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
8225 * The cpus mask in sched_group and sched_domain hangs off the end.
8227 * ( See the the comments in include/linux/sched.h:struct sched_group
8228 * and struct sched_domain. )
8230 struct static_sched_group {
8231 struct sched_group sg;
8232 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
8235 struct static_sched_domain {
8236 struct sched_domain sd;
8237 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
8243 cpumask_var_t domainspan;
8244 cpumask_var_t covered;
8245 cpumask_var_t notcovered;
8247 cpumask_var_t nodemask;
8248 cpumask_var_t this_sibling_map;
8249 cpumask_var_t this_core_map;
8250 cpumask_var_t send_covered;
8251 cpumask_var_t tmpmask;
8252 struct sched_group **sched_group_nodes;
8253 struct root_domain *rd;
8257 sa_sched_groups = 0,
8262 sa_this_sibling_map,
8264 sa_sched_group_nodes,
8274 * SMT sched-domains:
8276 #ifdef CONFIG_SCHED_SMT
8277 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
8278 static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
8281 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
8282 struct sched_group **sg, struct cpumask *unused)
8285 *sg = &per_cpu(sched_group_cpus, cpu).sg;
8288 #endif /* CONFIG_SCHED_SMT */
8291 * multi-core sched-domains:
8293 #ifdef CONFIG_SCHED_MC
8294 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
8295 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
8296 #endif /* CONFIG_SCHED_MC */
8298 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8300 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8301 struct sched_group **sg, struct cpumask *mask)
8305 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8306 group = cpumask_first(mask);
8308 *sg = &per_cpu(sched_group_core, group).sg;
8311 #elif defined(CONFIG_SCHED_MC)
8313 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8314 struct sched_group **sg, struct cpumask *unused)
8317 *sg = &per_cpu(sched_group_core, cpu).sg;
8322 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
8323 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
8326 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
8327 struct sched_group **sg, struct cpumask *mask)
8330 #ifdef CONFIG_SCHED_MC
8331 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
8332 group = cpumask_first(mask);
8333 #elif defined(CONFIG_SCHED_SMT)
8334 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8335 group = cpumask_first(mask);
8340 *sg = &per_cpu(sched_group_phys, group).sg;
8346 * The init_sched_build_groups can't handle what we want to do with node
8347 * groups, so roll our own. Now each node has its own list of groups which
8348 * gets dynamically allocated.
8350 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
8351 static struct sched_group ***sched_group_nodes_bycpu;
8353 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
8354 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
8356 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
8357 struct sched_group **sg,
8358 struct cpumask *nodemask)
8362 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
8363 group = cpumask_first(nodemask);
8366 *sg = &per_cpu(sched_group_allnodes, group).sg;
8370 static void init_numa_sched_groups_power(struct sched_group *group_head)
8372 struct sched_group *sg = group_head;
8378 for_each_cpu(j, sched_group_cpus(sg)) {
8379 struct sched_domain *sd;
8381 sd = &per_cpu(phys_domains, j).sd;
8382 if (j != group_first_cpu(sd->groups)) {
8384 * Only add "power" once for each
8390 sg->cpu_power += sd->groups->cpu_power;
8393 } while (sg != group_head);
8396 static int build_numa_sched_groups(struct s_data *d,
8397 const struct cpumask *cpu_map, int num)
8399 struct sched_domain *sd;
8400 struct sched_group *sg, *prev;
8403 cpumask_clear(d->covered);
8404 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
8405 if (cpumask_empty(d->nodemask)) {
8406 d->sched_group_nodes[num] = NULL;
8410 sched_domain_node_span(num, d->domainspan);
8411 cpumask_and(d->domainspan, d->domainspan, cpu_map);
8413 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8416 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
8420 d->sched_group_nodes[num] = sg;
8422 for_each_cpu(j, d->nodemask) {
8423 sd = &per_cpu(node_domains, j).sd;
8428 cpumask_copy(sched_group_cpus(sg), d->nodemask);
8430 cpumask_or(d->covered, d->covered, d->nodemask);
8433 for (j = 0; j < nr_node_ids; j++) {
8434 n = (num + j) % nr_node_ids;
8435 cpumask_complement(d->notcovered, d->covered);
8436 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
8437 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
8438 if (cpumask_empty(d->tmpmask))
8440 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
8441 if (cpumask_empty(d->tmpmask))
8443 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8447 "Can not alloc domain group for node %d\n", j);
8451 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
8452 sg->next = prev->next;
8453 cpumask_or(d->covered, d->covered, d->tmpmask);
8460 #endif /* CONFIG_NUMA */
8463 /* Free memory allocated for various sched_group structures */
8464 static void free_sched_groups(const struct cpumask *cpu_map,
8465 struct cpumask *nodemask)
8469 for_each_cpu(cpu, cpu_map) {
8470 struct sched_group **sched_group_nodes
8471 = sched_group_nodes_bycpu[cpu];
8473 if (!sched_group_nodes)
8476 for (i = 0; i < nr_node_ids; i++) {
8477 struct sched_group *oldsg, *sg = sched_group_nodes[i];
8479 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8480 if (cpumask_empty(nodemask))
8490 if (oldsg != sched_group_nodes[i])
8493 kfree(sched_group_nodes);
8494 sched_group_nodes_bycpu[cpu] = NULL;
8497 #else /* !CONFIG_NUMA */
8498 static void free_sched_groups(const struct cpumask *cpu_map,
8499 struct cpumask *nodemask)
8502 #endif /* CONFIG_NUMA */
8505 * Initialize sched groups cpu_power.
8507 * cpu_power indicates the capacity of sched group, which is used while
8508 * distributing the load between different sched groups in a sched domain.
8509 * Typically cpu_power for all the groups in a sched domain will be same unless
8510 * there are asymmetries in the topology. If there are asymmetries, group
8511 * having more cpu_power will pickup more load compared to the group having
8514 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
8516 struct sched_domain *child;
8517 struct sched_group *group;
8521 WARN_ON(!sd || !sd->groups);
8523 if (cpu != group_first_cpu(sd->groups))
8528 sd->groups->cpu_power = 0;
8531 power = SCHED_LOAD_SCALE;
8532 weight = cpumask_weight(sched_domain_span(sd));
8534 * SMT siblings share the power of a single core.
8535 * Usually multiple threads get a better yield out of
8536 * that one core than a single thread would have,
8537 * reflect that in sd->smt_gain.
8539 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
8540 power *= sd->smt_gain;
8542 power >>= SCHED_LOAD_SHIFT;
8544 sd->groups->cpu_power += power;
8549 * Add cpu_power of each child group to this groups cpu_power.
8551 group = child->groups;
8553 sd->groups->cpu_power += group->cpu_power;
8554 group = group->next;
8555 } while (group != child->groups);
8559 * Initializers for schedule domains
8560 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8563 #ifdef CONFIG_SCHED_DEBUG
8564 # define SD_INIT_NAME(sd, type) sd->name = #type
8566 # define SD_INIT_NAME(sd, type) do { } while (0)
8569 #define SD_INIT(sd, type) sd_init_##type(sd)
8571 #define SD_INIT_FUNC(type) \
8572 static noinline void sd_init_##type(struct sched_domain *sd) \
8574 memset(sd, 0, sizeof(*sd)); \
8575 *sd = SD_##type##_INIT; \
8576 sd->level = SD_LV_##type; \
8577 SD_INIT_NAME(sd, type); \
8582 SD_INIT_FUNC(ALLNODES)
8585 #ifdef CONFIG_SCHED_SMT
8586 SD_INIT_FUNC(SIBLING)
8588 #ifdef CONFIG_SCHED_MC
8592 static int default_relax_domain_level = -1;
8594 static int __init setup_relax_domain_level(char *str)
8598 val = simple_strtoul(str, NULL, 0);
8599 if (val < SD_LV_MAX)
8600 default_relax_domain_level = val;
8604 __setup("relax_domain_level=", setup_relax_domain_level);
8606 static void set_domain_attribute(struct sched_domain *sd,
8607 struct sched_domain_attr *attr)
8611 if (!attr || attr->relax_domain_level < 0) {
8612 if (default_relax_domain_level < 0)
8615 request = default_relax_domain_level;
8617 request = attr->relax_domain_level;
8618 if (request < sd->level) {
8619 /* turn off idle balance on this domain */
8620 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
8622 /* turn on idle balance on this domain */
8623 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
8627 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
8628 const struct cpumask *cpu_map)
8631 case sa_sched_groups:
8632 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
8633 d->sched_group_nodes = NULL;
8635 free_rootdomain(d->rd); /* fall through */
8637 free_cpumask_var(d->tmpmask); /* fall through */
8638 case sa_send_covered:
8639 free_cpumask_var(d->send_covered); /* fall through */
8640 case sa_this_core_map:
8641 free_cpumask_var(d->this_core_map); /* fall through */
8642 case sa_this_sibling_map:
8643 free_cpumask_var(d->this_sibling_map); /* fall through */
8645 free_cpumask_var(d->nodemask); /* fall through */
8646 case sa_sched_group_nodes:
8648 kfree(d->sched_group_nodes); /* fall through */
8650 free_cpumask_var(d->notcovered); /* fall through */
8652 free_cpumask_var(d->covered); /* fall through */
8654 free_cpumask_var(d->domainspan); /* fall through */
8661 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
8662 const struct cpumask *cpu_map)
8665 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
8667 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
8668 return sa_domainspan;
8669 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
8671 /* Allocate the per-node list of sched groups */
8672 d->sched_group_nodes = kcalloc(nr_node_ids,
8673 sizeof(struct sched_group *), GFP_KERNEL);
8674 if (!d->sched_group_nodes) {
8675 printk(KERN_WARNING "Can not alloc sched group node list\n");
8676 return sa_notcovered;
8678 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
8680 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
8681 return sa_sched_group_nodes;
8682 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
8684 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
8685 return sa_this_sibling_map;
8686 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
8687 return sa_this_core_map;
8688 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
8689 return sa_send_covered;
8690 d->rd = alloc_rootdomain();
8692 printk(KERN_WARNING "Cannot alloc root domain\n");
8695 return sa_rootdomain;
8698 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
8699 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
8701 struct sched_domain *sd = NULL;
8703 struct sched_domain *parent;
8706 if (cpumask_weight(cpu_map) >
8707 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
8708 sd = &per_cpu(allnodes_domains, i).sd;
8709 SD_INIT(sd, ALLNODES);
8710 set_domain_attribute(sd, attr);
8711 cpumask_copy(sched_domain_span(sd), cpu_map);
8712 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
8717 sd = &per_cpu(node_domains, i).sd;
8719 set_domain_attribute(sd, attr);
8720 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
8721 sd->parent = parent;
8724 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
8729 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
8730 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8731 struct sched_domain *parent, int i)
8733 struct sched_domain *sd;
8734 sd = &per_cpu(phys_domains, i).sd;
8736 set_domain_attribute(sd, attr);
8737 cpumask_copy(sched_domain_span(sd), d->nodemask);
8738 sd->parent = parent;
8741 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
8745 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
8746 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8747 struct sched_domain *parent, int i)
8749 struct sched_domain *sd = parent;
8750 #ifdef CONFIG_SCHED_MC
8751 sd = &per_cpu(core_domains, i).sd;
8753 set_domain_attribute(sd, attr);
8754 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
8755 sd->parent = parent;
8757 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
8762 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
8763 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8764 struct sched_domain *parent, int i)
8766 struct sched_domain *sd = parent;
8767 #ifdef CONFIG_SCHED_SMT
8768 sd = &per_cpu(cpu_domains, i).sd;
8769 SD_INIT(sd, SIBLING);
8770 set_domain_attribute(sd, attr);
8771 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
8772 sd->parent = parent;
8774 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
8779 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
8780 const struct cpumask *cpu_map, int cpu)
8783 #ifdef CONFIG_SCHED_SMT
8784 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
8785 cpumask_and(d->this_sibling_map, cpu_map,
8786 topology_thread_cpumask(cpu));
8787 if (cpu == cpumask_first(d->this_sibling_map))
8788 init_sched_build_groups(d->this_sibling_map, cpu_map,
8790 d->send_covered, d->tmpmask);
8793 #ifdef CONFIG_SCHED_MC
8794 case SD_LV_MC: /* set up multi-core groups */
8795 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
8796 if (cpu == cpumask_first(d->this_core_map))
8797 init_sched_build_groups(d->this_core_map, cpu_map,
8799 d->send_covered, d->tmpmask);
8802 case SD_LV_CPU: /* set up physical groups */
8803 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
8804 if (!cpumask_empty(d->nodemask))
8805 init_sched_build_groups(d->nodemask, cpu_map,
8807 d->send_covered, d->tmpmask);
8810 case SD_LV_ALLNODES:
8811 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
8812 d->send_covered, d->tmpmask);
8821 * Build sched domains for a given set of cpus and attach the sched domains
8822 * to the individual cpus
8824 static int __build_sched_domains(const struct cpumask *cpu_map,
8825 struct sched_domain_attr *attr)
8827 enum s_alloc alloc_state = sa_none;
8829 struct sched_domain *sd;
8835 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
8836 if (alloc_state != sa_rootdomain)
8838 alloc_state = sa_sched_groups;
8841 * Set up domains for cpus specified by the cpu_map.
8843 for_each_cpu(i, cpu_map) {
8844 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
8847 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
8848 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
8849 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
8850 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
8853 for_each_cpu(i, cpu_map) {
8854 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
8855 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
8858 /* Set up physical groups */
8859 for (i = 0; i < nr_node_ids; i++)
8860 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
8863 /* Set up node groups */
8865 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
8867 for (i = 0; i < nr_node_ids; i++)
8868 if (build_numa_sched_groups(&d, cpu_map, i))
8872 /* Calculate CPU power for physical packages and nodes */
8873 #ifdef CONFIG_SCHED_SMT
8874 for_each_cpu(i, cpu_map) {
8875 sd = &per_cpu(cpu_domains, i).sd;
8876 init_sched_groups_power(i, sd);
8879 #ifdef CONFIG_SCHED_MC
8880 for_each_cpu(i, cpu_map) {
8881 sd = &per_cpu(core_domains, i).sd;
8882 init_sched_groups_power(i, sd);
8886 for_each_cpu(i, cpu_map) {
8887 sd = &per_cpu(phys_domains, i).sd;
8888 init_sched_groups_power(i, sd);
8892 for (i = 0; i < nr_node_ids; i++)
8893 init_numa_sched_groups_power(d.sched_group_nodes[i]);
8895 if (d.sd_allnodes) {
8896 struct sched_group *sg;
8898 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
8900 init_numa_sched_groups_power(sg);
8904 /* Attach the domains */
8905 for_each_cpu(i, cpu_map) {
8906 #ifdef CONFIG_SCHED_SMT
8907 sd = &per_cpu(cpu_domains, i).sd;
8908 #elif defined(CONFIG_SCHED_MC)
8909 sd = &per_cpu(core_domains, i).sd;
8911 sd = &per_cpu(phys_domains, i).sd;
8913 cpu_attach_domain(sd, d.rd, i);
8916 d.sched_group_nodes = NULL; /* don't free this we still need it */
8917 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
8921 __free_domain_allocs(&d, alloc_state, cpu_map);
8925 static int build_sched_domains(const struct cpumask *cpu_map)
8927 return __build_sched_domains(cpu_map, NULL);
8930 static cpumask_var_t *doms_cur; /* current sched domains */
8931 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
8932 static struct sched_domain_attr *dattr_cur;
8933 /* attribues of custom domains in 'doms_cur' */
8936 * Special case: If a kmalloc of a doms_cur partition (array of
8937 * cpumask) fails, then fallback to a single sched domain,
8938 * as determined by the single cpumask fallback_doms.
8940 static cpumask_var_t fallback_doms;
8943 * arch_update_cpu_topology lets virtualized architectures update the
8944 * cpu core maps. It is supposed to return 1 if the topology changed
8945 * or 0 if it stayed the same.
8947 int __attribute__((weak)) arch_update_cpu_topology(void)
8952 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
8955 cpumask_var_t *doms;
8957 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
8960 for (i = 0; i < ndoms; i++) {
8961 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
8962 free_sched_domains(doms, i);
8969 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
8972 for (i = 0; i < ndoms; i++)
8973 free_cpumask_var(doms[i]);
8978 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
8979 * For now this just excludes isolated cpus, but could be used to
8980 * exclude other special cases in the future.
8982 static int arch_init_sched_domains(const struct cpumask *cpu_map)
8986 arch_update_cpu_topology();
8988 doms_cur = alloc_sched_domains(ndoms_cur);
8990 doms_cur = &fallback_doms;
8991 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
8993 err = build_sched_domains(doms_cur[0]);
8994 register_sched_domain_sysctl();
8999 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
9000 struct cpumask *tmpmask)
9002 free_sched_groups(cpu_map, tmpmask);
9006 * Detach sched domains from a group of cpus specified in cpu_map
9007 * These cpus will now be attached to the NULL domain
9009 static void detach_destroy_domains(const struct cpumask *cpu_map)
9011 /* Save because hotplug lock held. */
9012 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
9015 for_each_cpu(i, cpu_map)
9016 cpu_attach_domain(NULL, &def_root_domain, i);
9017 synchronize_sched();
9018 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
9021 /* handle null as "default" */
9022 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
9023 struct sched_domain_attr *new, int idx_new)
9025 struct sched_domain_attr tmp;
9032 return !memcmp(cur ? (cur + idx_cur) : &tmp,
9033 new ? (new + idx_new) : &tmp,
9034 sizeof(struct sched_domain_attr));
9038 * Partition sched domains as specified by the 'ndoms_new'
9039 * cpumasks in the array doms_new[] of cpumasks. This compares
9040 * doms_new[] to the current sched domain partitioning, doms_cur[].
9041 * It destroys each deleted domain and builds each new domain.
9043 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
9044 * The masks don't intersect (don't overlap.) We should setup one
9045 * sched domain for each mask. CPUs not in any of the cpumasks will
9046 * not be load balanced. If the same cpumask appears both in the
9047 * current 'doms_cur' domains and in the new 'doms_new', we can leave
9050 * The passed in 'doms_new' should be allocated using
9051 * alloc_sched_domains. This routine takes ownership of it and will
9052 * free_sched_domains it when done with it. If the caller failed the
9053 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
9054 * and partition_sched_domains() will fallback to the single partition
9055 * 'fallback_doms', it also forces the domains to be rebuilt.
9057 * If doms_new == NULL it will be replaced with cpu_online_mask.
9058 * ndoms_new == 0 is a special case for destroying existing domains,
9059 * and it will not create the default domain.
9061 * Call with hotplug lock held
9063 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
9064 struct sched_domain_attr *dattr_new)
9069 mutex_lock(&sched_domains_mutex);
9071 /* always unregister in case we don't destroy any domains */
9072 unregister_sched_domain_sysctl();
9074 /* Let architecture update cpu core mappings. */
9075 new_topology = arch_update_cpu_topology();
9077 n = doms_new ? ndoms_new : 0;
9079 /* Destroy deleted domains */
9080 for (i = 0; i < ndoms_cur; i++) {
9081 for (j = 0; j < n && !new_topology; j++) {
9082 if (cpumask_equal(doms_cur[i], doms_new[j])
9083 && dattrs_equal(dattr_cur, i, dattr_new, j))
9086 /* no match - a current sched domain not in new doms_new[] */
9087 detach_destroy_domains(doms_cur[i]);
9092 if (doms_new == NULL) {
9094 doms_new = &fallback_doms;
9095 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
9096 WARN_ON_ONCE(dattr_new);
9099 /* Build new domains */
9100 for (i = 0; i < ndoms_new; i++) {
9101 for (j = 0; j < ndoms_cur && !new_topology; j++) {
9102 if (cpumask_equal(doms_new[i], doms_cur[j])
9103 && dattrs_equal(dattr_new, i, dattr_cur, j))
9106 /* no match - add a new doms_new */
9107 __build_sched_domains(doms_new[i],
9108 dattr_new ? dattr_new + i : NULL);
9113 /* Remember the new sched domains */
9114 if (doms_cur != &fallback_doms)
9115 free_sched_domains(doms_cur, ndoms_cur);
9116 kfree(dattr_cur); /* kfree(NULL) is safe */
9117 doms_cur = doms_new;
9118 dattr_cur = dattr_new;
9119 ndoms_cur = ndoms_new;
9121 register_sched_domain_sysctl();
9123 mutex_unlock(&sched_domains_mutex);
9126 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
9127 static void arch_reinit_sched_domains(void)
9131 /* Destroy domains first to force the rebuild */
9132 partition_sched_domains(0, NULL, NULL);
9134 rebuild_sched_domains();
9138 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
9140 unsigned int level = 0;
9142 if (sscanf(buf, "%u", &level) != 1)
9146 * level is always be positive so don't check for
9147 * level < POWERSAVINGS_BALANCE_NONE which is 0
9148 * What happens on 0 or 1 byte write,
9149 * need to check for count as well?
9152 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
9156 sched_smt_power_savings = level;
9158 sched_mc_power_savings = level;
9160 arch_reinit_sched_domains();
9165 #ifdef CONFIG_SCHED_MC
9166 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
9169 return sprintf(page, "%u\n", sched_mc_power_savings);
9171 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
9172 const char *buf, size_t count)
9174 return sched_power_savings_store(buf, count, 0);
9176 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
9177 sched_mc_power_savings_show,
9178 sched_mc_power_savings_store);
9181 #ifdef CONFIG_SCHED_SMT
9182 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
9185 return sprintf(page, "%u\n", sched_smt_power_savings);
9187 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
9188 const char *buf, size_t count)
9190 return sched_power_savings_store(buf, count, 1);
9192 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
9193 sched_smt_power_savings_show,
9194 sched_smt_power_savings_store);
9197 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
9201 #ifdef CONFIG_SCHED_SMT
9203 err = sysfs_create_file(&cls->kset.kobj,
9204 &attr_sched_smt_power_savings.attr);
9206 #ifdef CONFIG_SCHED_MC
9207 if (!err && mc_capable())
9208 err = sysfs_create_file(&cls->kset.kobj,
9209 &attr_sched_mc_power_savings.attr);
9213 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
9215 #ifndef CONFIG_CPUSETS
9217 * Add online and remove offline CPUs from the scheduler domains.
9218 * When cpusets are enabled they take over this function.
9220 static int update_sched_domains(struct notifier_block *nfb,
9221 unsigned long action, void *hcpu)
9225 case CPU_ONLINE_FROZEN:
9226 case CPU_DOWN_PREPARE:
9227 case CPU_DOWN_PREPARE_FROZEN:
9228 case CPU_DOWN_FAILED:
9229 case CPU_DOWN_FAILED_FROZEN:
9230 partition_sched_domains(1, NULL, NULL);
9239 static int update_runtime(struct notifier_block *nfb,
9240 unsigned long action, void *hcpu)
9242 int cpu = (int)(long)hcpu;
9245 case CPU_DOWN_PREPARE:
9246 case CPU_DOWN_PREPARE_FROZEN:
9247 disable_runtime(cpu_rq(cpu));
9250 case CPU_DOWN_FAILED:
9251 case CPU_DOWN_FAILED_FROZEN:
9253 case CPU_ONLINE_FROZEN:
9254 enable_runtime(cpu_rq(cpu));
9262 void __init sched_init_smp(void)
9264 cpumask_var_t non_isolated_cpus;
9266 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
9267 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
9269 #if defined(CONFIG_NUMA)
9270 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
9272 BUG_ON(sched_group_nodes_bycpu == NULL);
9275 mutex_lock(&sched_domains_mutex);
9276 arch_init_sched_domains(cpu_active_mask);
9277 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
9278 if (cpumask_empty(non_isolated_cpus))
9279 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
9280 mutex_unlock(&sched_domains_mutex);
9283 #ifndef CONFIG_CPUSETS
9284 /* XXX: Theoretical race here - CPU may be hotplugged now */
9285 hotcpu_notifier(update_sched_domains, 0);
9288 /* RT runtime code needs to handle some hotplug events */
9289 hotcpu_notifier(update_runtime, 0);
9293 /* Move init over to a non-isolated CPU */
9294 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
9296 sched_init_granularity();
9297 free_cpumask_var(non_isolated_cpus);
9299 init_sched_rt_class();
9302 void __init sched_init_smp(void)
9304 sched_init_granularity();
9306 #endif /* CONFIG_SMP */
9308 const_debug unsigned int sysctl_timer_migration = 1;
9310 int in_sched_functions(unsigned long addr)
9312 return in_lock_functions(addr) ||
9313 (addr >= (unsigned long)__sched_text_start
9314 && addr < (unsigned long)__sched_text_end);
9317 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
9319 cfs_rq->tasks_timeline = RB_ROOT;
9320 INIT_LIST_HEAD(&cfs_rq->tasks);
9321 #ifdef CONFIG_FAIR_GROUP_SCHED
9324 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9327 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
9329 struct rt_prio_array *array;
9332 array = &rt_rq->active;
9333 for (i = 0; i < MAX_RT_PRIO; i++) {
9334 INIT_LIST_HEAD(array->queue + i);
9335 __clear_bit(i, array->bitmap);
9337 /* delimiter for bitsearch: */
9338 __set_bit(MAX_RT_PRIO, array->bitmap);
9340 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9341 rt_rq->highest_prio.curr = MAX_RT_PRIO;
9343 rt_rq->highest_prio.next = MAX_RT_PRIO;
9347 rt_rq->rt_nr_migratory = 0;
9348 rt_rq->overloaded = 0;
9349 plist_head_init(&rt_rq->pushable_tasks, &rq->lock);
9353 rt_rq->rt_throttled = 0;
9354 rt_rq->rt_runtime = 0;
9355 spin_lock_init(&rt_rq->rt_runtime_lock);
9357 #ifdef CONFIG_RT_GROUP_SCHED
9358 rt_rq->rt_nr_boosted = 0;
9363 #ifdef CONFIG_FAIR_GROUP_SCHED
9364 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9365 struct sched_entity *se, int cpu, int add,
9366 struct sched_entity *parent)
9368 struct rq *rq = cpu_rq(cpu);
9369 tg->cfs_rq[cpu] = cfs_rq;
9370 init_cfs_rq(cfs_rq, rq);
9373 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
9376 /* se could be NULL for init_task_group */
9381 se->cfs_rq = &rq->cfs;
9383 se->cfs_rq = parent->my_q;
9386 se->load.weight = tg->shares;
9387 se->load.inv_weight = 0;
9388 se->parent = parent;
9392 #ifdef CONFIG_RT_GROUP_SCHED
9393 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
9394 struct sched_rt_entity *rt_se, int cpu, int add,
9395 struct sched_rt_entity *parent)
9397 struct rq *rq = cpu_rq(cpu);
9399 tg->rt_rq[cpu] = rt_rq;
9400 init_rt_rq(rt_rq, rq);
9402 rt_rq->rt_se = rt_se;
9403 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
9405 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
9407 tg->rt_se[cpu] = rt_se;
9412 rt_se->rt_rq = &rq->rt;
9414 rt_se->rt_rq = parent->my_q;
9416 rt_se->my_q = rt_rq;
9417 rt_se->parent = parent;
9418 INIT_LIST_HEAD(&rt_se->run_list);
9422 void __init sched_init(void)
9425 unsigned long alloc_size = 0, ptr;
9427 #ifdef CONFIG_FAIR_GROUP_SCHED
9428 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9430 #ifdef CONFIG_RT_GROUP_SCHED
9431 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9433 #ifdef CONFIG_USER_SCHED
9436 #ifdef CONFIG_CPUMASK_OFFSTACK
9437 alloc_size += num_possible_cpus() * cpumask_size();
9440 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
9442 #ifdef CONFIG_FAIR_GROUP_SCHED
9443 init_task_group.se = (struct sched_entity **)ptr;
9444 ptr += nr_cpu_ids * sizeof(void **);
9446 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
9447 ptr += nr_cpu_ids * sizeof(void **);
9449 #ifdef CONFIG_USER_SCHED
9450 root_task_group.se = (struct sched_entity **)ptr;
9451 ptr += nr_cpu_ids * sizeof(void **);
9453 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9454 ptr += nr_cpu_ids * sizeof(void **);
9455 #endif /* CONFIG_USER_SCHED */
9456 #endif /* CONFIG_FAIR_GROUP_SCHED */
9457 #ifdef CONFIG_RT_GROUP_SCHED
9458 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
9459 ptr += nr_cpu_ids * sizeof(void **);
9461 init_task_group.rt_rq = (struct rt_rq **)ptr;
9462 ptr += nr_cpu_ids * sizeof(void **);
9464 #ifdef CONFIG_USER_SCHED
9465 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9466 ptr += nr_cpu_ids * sizeof(void **);
9468 root_task_group.rt_rq = (struct rt_rq **)ptr;
9469 ptr += nr_cpu_ids * sizeof(void **);
9470 #endif /* CONFIG_USER_SCHED */
9471 #endif /* CONFIG_RT_GROUP_SCHED */
9472 #ifdef CONFIG_CPUMASK_OFFSTACK
9473 for_each_possible_cpu(i) {
9474 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
9475 ptr += cpumask_size();
9477 #endif /* CONFIG_CPUMASK_OFFSTACK */
9481 init_defrootdomain();
9484 init_rt_bandwidth(&def_rt_bandwidth,
9485 global_rt_period(), global_rt_runtime());
9487 #ifdef CONFIG_RT_GROUP_SCHED
9488 init_rt_bandwidth(&init_task_group.rt_bandwidth,
9489 global_rt_period(), global_rt_runtime());
9490 #ifdef CONFIG_USER_SCHED
9491 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9492 global_rt_period(), RUNTIME_INF);
9493 #endif /* CONFIG_USER_SCHED */
9494 #endif /* CONFIG_RT_GROUP_SCHED */
9496 #ifdef CONFIG_GROUP_SCHED
9497 list_add(&init_task_group.list, &task_groups);
9498 INIT_LIST_HEAD(&init_task_group.children);
9500 #ifdef CONFIG_USER_SCHED
9501 INIT_LIST_HEAD(&root_task_group.children);
9502 init_task_group.parent = &root_task_group;
9503 list_add(&init_task_group.siblings, &root_task_group.children);
9504 #endif /* CONFIG_USER_SCHED */
9505 #endif /* CONFIG_GROUP_SCHED */
9507 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
9508 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
9509 __alignof__(unsigned long));
9511 for_each_possible_cpu(i) {
9515 spin_lock_init(&rq->lock);
9517 rq->calc_load_active = 0;
9518 rq->calc_load_update = jiffies + LOAD_FREQ;
9519 init_cfs_rq(&rq->cfs, rq);
9520 init_rt_rq(&rq->rt, rq);
9521 #ifdef CONFIG_FAIR_GROUP_SCHED
9522 init_task_group.shares = init_task_group_load;
9523 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9524 #ifdef CONFIG_CGROUP_SCHED
9526 * How much cpu bandwidth does init_task_group get?
9528 * In case of task-groups formed thr' the cgroup filesystem, it
9529 * gets 100% of the cpu resources in the system. This overall
9530 * system cpu resource is divided among the tasks of
9531 * init_task_group and its child task-groups in a fair manner,
9532 * based on each entity's (task or task-group's) weight
9533 * (se->load.weight).
9535 * In other words, if init_task_group has 10 tasks of weight
9536 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9537 * then A0's share of the cpu resource is:
9539 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9541 * We achieve this by letting init_task_group's tasks sit
9542 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9544 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
9545 #elif defined CONFIG_USER_SCHED
9546 root_task_group.shares = NICE_0_LOAD;
9547 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
9549 * In case of task-groups formed thr' the user id of tasks,
9550 * init_task_group represents tasks belonging to root user.
9551 * Hence it forms a sibling of all subsequent groups formed.
9552 * In this case, init_task_group gets only a fraction of overall
9553 * system cpu resource, based on the weight assigned to root
9554 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9555 * by letting tasks of init_task_group sit in a separate cfs_rq
9556 * (init_tg_cfs_rq) and having one entity represent this group of
9557 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9559 init_tg_cfs_entry(&init_task_group,
9560 &per_cpu(init_tg_cfs_rq, i),
9561 &per_cpu(init_sched_entity, i), i, 1,
9562 root_task_group.se[i]);
9565 #endif /* CONFIG_FAIR_GROUP_SCHED */
9567 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9568 #ifdef CONFIG_RT_GROUP_SCHED
9569 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
9570 #ifdef CONFIG_CGROUP_SCHED
9571 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
9572 #elif defined CONFIG_USER_SCHED
9573 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
9574 init_tg_rt_entry(&init_task_group,
9575 &per_cpu(init_rt_rq, i),
9576 &per_cpu(init_sched_rt_entity, i), i, 1,
9577 root_task_group.rt_se[i]);
9581 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
9582 rq->cpu_load[j] = 0;
9586 rq->post_schedule = 0;
9587 rq->active_balance = 0;
9588 rq->next_balance = jiffies;
9592 rq->migration_thread = NULL;
9594 rq->avg_idle = 2*sysctl_sched_migration_cost;
9595 INIT_LIST_HEAD(&rq->migration_queue);
9596 rq_attach_root(rq, &def_root_domain);
9599 atomic_set(&rq->nr_iowait, 0);
9602 set_load_weight(&init_task);
9604 #ifdef CONFIG_PREEMPT_NOTIFIERS
9605 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
9609 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9612 #ifdef CONFIG_RT_MUTEXES
9613 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
9617 * The boot idle thread does lazy MMU switching as well:
9619 atomic_inc(&init_mm.mm_count);
9620 enter_lazy_tlb(&init_mm, current);
9623 * Make us the idle thread. Technically, schedule() should not be
9624 * called from this thread, however somewhere below it might be,
9625 * but because we are the idle thread, we just pick up running again
9626 * when this runqueue becomes "idle".
9628 init_idle(current, smp_processor_id());
9630 calc_load_update = jiffies + LOAD_FREQ;
9633 * During early bootup we pretend to be a normal task:
9635 current->sched_class = &fair_sched_class;
9637 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9638 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
9641 zalloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
9642 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
9644 /* May be allocated at isolcpus cmdline parse time */
9645 if (cpu_isolated_map == NULL)
9646 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
9651 scheduler_running = 1;
9654 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9655 static inline int preempt_count_equals(int preempt_offset)
9657 int nested = preempt_count() & ~PREEMPT_ACTIVE;
9659 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
9662 void __might_sleep(char *file, int line, int preempt_offset)
9665 static unsigned long prev_jiffy; /* ratelimiting */
9667 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
9668 system_state != SYSTEM_RUNNING || oops_in_progress)
9670 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9672 prev_jiffy = jiffies;
9675 "BUG: sleeping function called from invalid context at %s:%d\n",
9678 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9679 in_atomic(), irqs_disabled(),
9680 current->pid, current->comm);
9682 debug_show_held_locks(current);
9683 if (irqs_disabled())
9684 print_irqtrace_events(current);
9688 EXPORT_SYMBOL(__might_sleep);
9691 #ifdef CONFIG_MAGIC_SYSRQ
9692 static void normalize_task(struct rq *rq, struct task_struct *p)
9696 update_rq_clock(rq);
9697 on_rq = p->se.on_rq;
9699 deactivate_task(rq, p, 0);
9700 __setscheduler(rq, p, SCHED_NORMAL, 0);
9702 activate_task(rq, p, 0);
9703 resched_task(rq->curr);
9707 void normalize_rt_tasks(void)
9709 struct task_struct *g, *p;
9710 unsigned long flags;
9713 read_lock_irqsave(&tasklist_lock, flags);
9714 do_each_thread(g, p) {
9716 * Only normalize user tasks:
9721 p->se.exec_start = 0;
9722 #ifdef CONFIG_SCHEDSTATS
9723 p->se.wait_start = 0;
9724 p->se.sleep_start = 0;
9725 p->se.block_start = 0;
9730 * Renice negative nice level userspace
9733 if (TASK_NICE(p) < 0 && p->mm)
9734 set_user_nice(p, 0);
9738 spin_lock(&p->pi_lock);
9739 rq = __task_rq_lock(p);
9741 normalize_task(rq, p);
9743 __task_rq_unlock(rq);
9744 spin_unlock(&p->pi_lock);
9745 } while_each_thread(g, p);
9747 read_unlock_irqrestore(&tasklist_lock, flags);
9750 #endif /* CONFIG_MAGIC_SYSRQ */
9754 * These functions are only useful for the IA64 MCA handling.
9756 * They can only be called when the whole system has been
9757 * stopped - every CPU needs to be quiescent, and no scheduling
9758 * activity can take place. Using them for anything else would
9759 * be a serious bug, and as a result, they aren't even visible
9760 * under any other configuration.
9764 * curr_task - return the current task for a given cpu.
9765 * @cpu: the processor in question.
9767 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9769 struct task_struct *curr_task(int cpu)
9771 return cpu_curr(cpu);
9775 * set_curr_task - set the current task for a given cpu.
9776 * @cpu: the processor in question.
9777 * @p: the task pointer to set.
9779 * Description: This function must only be used when non-maskable interrupts
9780 * are serviced on a separate stack. It allows the architecture to switch the
9781 * notion of the current task on a cpu in a non-blocking manner. This function
9782 * must be called with all CPU's synchronized, and interrupts disabled, the
9783 * and caller must save the original value of the current task (see
9784 * curr_task() above) and restore that value before reenabling interrupts and
9785 * re-starting the system.
9787 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9789 void set_curr_task(int cpu, struct task_struct *p)
9796 #ifdef CONFIG_FAIR_GROUP_SCHED
9797 static void free_fair_sched_group(struct task_group *tg)
9801 for_each_possible_cpu(i) {
9803 kfree(tg->cfs_rq[i]);
9813 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9815 struct cfs_rq *cfs_rq;
9816 struct sched_entity *se;
9820 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9823 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9827 tg->shares = NICE_0_LOAD;
9829 for_each_possible_cpu(i) {
9832 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9833 GFP_KERNEL, cpu_to_node(i));
9837 se = kzalloc_node(sizeof(struct sched_entity),
9838 GFP_KERNEL, cpu_to_node(i));
9842 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
9851 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9853 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
9854 &cpu_rq(cpu)->leaf_cfs_rq_list);
9857 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9859 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
9861 #else /* !CONFG_FAIR_GROUP_SCHED */
9862 static inline void free_fair_sched_group(struct task_group *tg)
9867 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9872 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9876 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9879 #endif /* CONFIG_FAIR_GROUP_SCHED */
9881 #ifdef CONFIG_RT_GROUP_SCHED
9882 static void free_rt_sched_group(struct task_group *tg)
9886 destroy_rt_bandwidth(&tg->rt_bandwidth);
9888 for_each_possible_cpu(i) {
9890 kfree(tg->rt_rq[i]);
9892 kfree(tg->rt_se[i]);
9900 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9902 struct rt_rq *rt_rq;
9903 struct sched_rt_entity *rt_se;
9907 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
9910 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
9914 init_rt_bandwidth(&tg->rt_bandwidth,
9915 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
9917 for_each_possible_cpu(i) {
9920 rt_rq = kzalloc_node(sizeof(struct rt_rq),
9921 GFP_KERNEL, cpu_to_node(i));
9925 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
9926 GFP_KERNEL, cpu_to_node(i));
9930 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
9939 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9941 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
9942 &cpu_rq(cpu)->leaf_rt_rq_list);
9945 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9947 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
9949 #else /* !CONFIG_RT_GROUP_SCHED */
9950 static inline void free_rt_sched_group(struct task_group *tg)
9955 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9960 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9964 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9967 #endif /* CONFIG_RT_GROUP_SCHED */
9969 #ifdef CONFIG_GROUP_SCHED
9970 static void free_sched_group(struct task_group *tg)
9972 free_fair_sched_group(tg);
9973 free_rt_sched_group(tg);
9977 /* allocate runqueue etc for a new task group */
9978 struct task_group *sched_create_group(struct task_group *parent)
9980 struct task_group *tg;
9981 unsigned long flags;
9984 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
9986 return ERR_PTR(-ENOMEM);
9988 if (!alloc_fair_sched_group(tg, parent))
9991 if (!alloc_rt_sched_group(tg, parent))
9994 spin_lock_irqsave(&task_group_lock, flags);
9995 for_each_possible_cpu(i) {
9996 register_fair_sched_group(tg, i);
9997 register_rt_sched_group(tg, i);
9999 list_add_rcu(&tg->list, &task_groups);
10001 WARN_ON(!parent); /* root should already exist */
10003 tg->parent = parent;
10004 INIT_LIST_HEAD(&tg->children);
10005 list_add_rcu(&tg->siblings, &parent->children);
10006 spin_unlock_irqrestore(&task_group_lock, flags);
10011 free_sched_group(tg);
10012 return ERR_PTR(-ENOMEM);
10015 /* rcu callback to free various structures associated with a task group */
10016 static void free_sched_group_rcu(struct rcu_head *rhp)
10018 /* now it should be safe to free those cfs_rqs */
10019 free_sched_group(container_of(rhp, struct task_group, rcu));
10022 /* Destroy runqueue etc associated with a task group */
10023 void sched_destroy_group(struct task_group *tg)
10025 unsigned long flags;
10028 spin_lock_irqsave(&task_group_lock, flags);
10029 for_each_possible_cpu(i) {
10030 unregister_fair_sched_group(tg, i);
10031 unregister_rt_sched_group(tg, i);
10033 list_del_rcu(&tg->list);
10034 list_del_rcu(&tg->siblings);
10035 spin_unlock_irqrestore(&task_group_lock, flags);
10037 /* wait for possible concurrent references to cfs_rqs complete */
10038 call_rcu(&tg->rcu, free_sched_group_rcu);
10041 /* change task's runqueue when it moves between groups.
10042 * The caller of this function should have put the task in its new group
10043 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
10044 * reflect its new group.
10046 void sched_move_task(struct task_struct *tsk)
10048 int on_rq, running;
10049 unsigned long flags;
10052 rq = task_rq_lock(tsk, &flags);
10054 update_rq_clock(rq);
10056 running = task_current(rq, tsk);
10057 on_rq = tsk->se.on_rq;
10060 dequeue_task(rq, tsk, 0);
10061 if (unlikely(running))
10062 tsk->sched_class->put_prev_task(rq, tsk);
10064 set_task_rq(tsk, task_cpu(tsk));
10066 #ifdef CONFIG_FAIR_GROUP_SCHED
10067 if (tsk->sched_class->moved_group)
10068 tsk->sched_class->moved_group(tsk);
10071 if (unlikely(running))
10072 tsk->sched_class->set_curr_task(rq);
10074 enqueue_task(rq, tsk, 0);
10076 task_rq_unlock(rq, &flags);
10078 #endif /* CONFIG_GROUP_SCHED */
10080 #ifdef CONFIG_FAIR_GROUP_SCHED
10081 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
10083 struct cfs_rq *cfs_rq = se->cfs_rq;
10088 dequeue_entity(cfs_rq, se, 0);
10090 se->load.weight = shares;
10091 se->load.inv_weight = 0;
10094 enqueue_entity(cfs_rq, se, 0);
10097 static void set_se_shares(struct sched_entity *se, unsigned long shares)
10099 struct cfs_rq *cfs_rq = se->cfs_rq;
10100 struct rq *rq = cfs_rq->rq;
10101 unsigned long flags;
10103 spin_lock_irqsave(&rq->lock, flags);
10104 __set_se_shares(se, shares);
10105 spin_unlock_irqrestore(&rq->lock, flags);
10108 static DEFINE_MUTEX(shares_mutex);
10110 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10113 unsigned long flags;
10116 * We can't change the weight of the root cgroup.
10121 if (shares < MIN_SHARES)
10122 shares = MIN_SHARES;
10123 else if (shares > MAX_SHARES)
10124 shares = MAX_SHARES;
10126 mutex_lock(&shares_mutex);
10127 if (tg->shares == shares)
10130 spin_lock_irqsave(&task_group_lock, flags);
10131 for_each_possible_cpu(i)
10132 unregister_fair_sched_group(tg, i);
10133 list_del_rcu(&tg->siblings);
10134 spin_unlock_irqrestore(&task_group_lock, flags);
10136 /* wait for any ongoing reference to this group to finish */
10137 synchronize_sched();
10140 * Now we are free to modify the group's share on each cpu
10141 * w/o tripping rebalance_share or load_balance_fair.
10143 tg->shares = shares;
10144 for_each_possible_cpu(i) {
10146 * force a rebalance
10148 cfs_rq_set_shares(tg->cfs_rq[i], 0);
10149 set_se_shares(tg->se[i], shares);
10153 * Enable load balance activity on this group, by inserting it back on
10154 * each cpu's rq->leaf_cfs_rq_list.
10156 spin_lock_irqsave(&task_group_lock, flags);
10157 for_each_possible_cpu(i)
10158 register_fair_sched_group(tg, i);
10159 list_add_rcu(&tg->siblings, &tg->parent->children);
10160 spin_unlock_irqrestore(&task_group_lock, flags);
10162 mutex_unlock(&shares_mutex);
10166 unsigned long sched_group_shares(struct task_group *tg)
10172 #ifdef CONFIG_RT_GROUP_SCHED
10174 * Ensure that the real time constraints are schedulable.
10176 static DEFINE_MUTEX(rt_constraints_mutex);
10178 static unsigned long to_ratio(u64 period, u64 runtime)
10180 if (runtime == RUNTIME_INF)
10183 return div64_u64(runtime << 20, period);
10186 /* Must be called with tasklist_lock held */
10187 static inline int tg_has_rt_tasks(struct task_group *tg)
10189 struct task_struct *g, *p;
10191 do_each_thread(g, p) {
10192 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
10194 } while_each_thread(g, p);
10199 struct rt_schedulable_data {
10200 struct task_group *tg;
10205 static int tg_schedulable(struct task_group *tg, void *data)
10207 struct rt_schedulable_data *d = data;
10208 struct task_group *child;
10209 unsigned long total, sum = 0;
10210 u64 period, runtime;
10212 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10213 runtime = tg->rt_bandwidth.rt_runtime;
10216 period = d->rt_period;
10217 runtime = d->rt_runtime;
10220 #ifdef CONFIG_USER_SCHED
10221 if (tg == &root_task_group) {
10222 period = global_rt_period();
10223 runtime = global_rt_runtime();
10228 * Cannot have more runtime than the period.
10230 if (runtime > period && runtime != RUNTIME_INF)
10234 * Ensure we don't starve existing RT tasks.
10236 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
10239 total = to_ratio(period, runtime);
10242 * Nobody can have more than the global setting allows.
10244 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
10248 * The sum of our children's runtime should not exceed our own.
10250 list_for_each_entry_rcu(child, &tg->children, siblings) {
10251 period = ktime_to_ns(child->rt_bandwidth.rt_period);
10252 runtime = child->rt_bandwidth.rt_runtime;
10254 if (child == d->tg) {
10255 period = d->rt_period;
10256 runtime = d->rt_runtime;
10259 sum += to_ratio(period, runtime);
10268 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
10270 struct rt_schedulable_data data = {
10272 .rt_period = period,
10273 .rt_runtime = runtime,
10276 return walk_tg_tree(tg_schedulable, tg_nop, &data);
10279 static int tg_set_bandwidth(struct task_group *tg,
10280 u64 rt_period, u64 rt_runtime)
10284 mutex_lock(&rt_constraints_mutex);
10285 read_lock(&tasklist_lock);
10286 err = __rt_schedulable(tg, rt_period, rt_runtime);
10290 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10291 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
10292 tg->rt_bandwidth.rt_runtime = rt_runtime;
10294 for_each_possible_cpu(i) {
10295 struct rt_rq *rt_rq = tg->rt_rq[i];
10297 spin_lock(&rt_rq->rt_runtime_lock);
10298 rt_rq->rt_runtime = rt_runtime;
10299 spin_unlock(&rt_rq->rt_runtime_lock);
10301 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10303 read_unlock(&tasklist_lock);
10304 mutex_unlock(&rt_constraints_mutex);
10309 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
10311 u64 rt_runtime, rt_period;
10313 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10314 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
10315 if (rt_runtime_us < 0)
10316 rt_runtime = RUNTIME_INF;
10318 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10321 long sched_group_rt_runtime(struct task_group *tg)
10325 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
10328 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
10329 do_div(rt_runtime_us, NSEC_PER_USEC);
10330 return rt_runtime_us;
10333 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
10335 u64 rt_runtime, rt_period;
10337 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
10338 rt_runtime = tg->rt_bandwidth.rt_runtime;
10340 if (rt_period == 0)
10343 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10346 long sched_group_rt_period(struct task_group *tg)
10350 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
10351 do_div(rt_period_us, NSEC_PER_USEC);
10352 return rt_period_us;
10355 static int sched_rt_global_constraints(void)
10357 u64 runtime, period;
10360 if (sysctl_sched_rt_period <= 0)
10363 runtime = global_rt_runtime();
10364 period = global_rt_period();
10367 * Sanity check on the sysctl variables.
10369 if (runtime > period && runtime != RUNTIME_INF)
10372 mutex_lock(&rt_constraints_mutex);
10373 read_lock(&tasklist_lock);
10374 ret = __rt_schedulable(NULL, 0, 0);
10375 read_unlock(&tasklist_lock);
10376 mutex_unlock(&rt_constraints_mutex);
10381 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
10383 /* Don't accept realtime tasks when there is no way for them to run */
10384 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
10390 #else /* !CONFIG_RT_GROUP_SCHED */
10391 static int sched_rt_global_constraints(void)
10393 unsigned long flags;
10396 if (sysctl_sched_rt_period <= 0)
10400 * There's always some RT tasks in the root group
10401 * -- migration, kstopmachine etc..
10403 if (sysctl_sched_rt_runtime == 0)
10406 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
10407 for_each_possible_cpu(i) {
10408 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
10410 spin_lock(&rt_rq->rt_runtime_lock);
10411 rt_rq->rt_runtime = global_rt_runtime();
10412 spin_unlock(&rt_rq->rt_runtime_lock);
10414 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
10418 #endif /* CONFIG_RT_GROUP_SCHED */
10420 int sched_rt_handler(struct ctl_table *table, int write,
10421 void __user *buffer, size_t *lenp,
10425 int old_period, old_runtime;
10426 static DEFINE_MUTEX(mutex);
10428 mutex_lock(&mutex);
10429 old_period = sysctl_sched_rt_period;
10430 old_runtime = sysctl_sched_rt_runtime;
10432 ret = proc_dointvec(table, write, buffer, lenp, ppos);
10434 if (!ret && write) {
10435 ret = sched_rt_global_constraints();
10437 sysctl_sched_rt_period = old_period;
10438 sysctl_sched_rt_runtime = old_runtime;
10440 def_rt_bandwidth.rt_runtime = global_rt_runtime();
10441 def_rt_bandwidth.rt_period =
10442 ns_to_ktime(global_rt_period());
10445 mutex_unlock(&mutex);
10450 #ifdef CONFIG_CGROUP_SCHED
10452 /* return corresponding task_group object of a cgroup */
10453 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
10455 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
10456 struct task_group, css);
10459 static struct cgroup_subsys_state *
10460 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
10462 struct task_group *tg, *parent;
10464 if (!cgrp->parent) {
10465 /* This is early initialization for the top cgroup */
10466 return &init_task_group.css;
10469 parent = cgroup_tg(cgrp->parent);
10470 tg = sched_create_group(parent);
10472 return ERR_PTR(-ENOMEM);
10478 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10480 struct task_group *tg = cgroup_tg(cgrp);
10482 sched_destroy_group(tg);
10486 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
10488 #ifdef CONFIG_RT_GROUP_SCHED
10489 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
10492 /* We don't support RT-tasks being in separate groups */
10493 if (tsk->sched_class != &fair_sched_class)
10500 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10501 struct task_struct *tsk, bool threadgroup)
10503 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
10507 struct task_struct *c;
10509 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
10510 retval = cpu_cgroup_can_attach_task(cgrp, c);
10522 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10523 struct cgroup *old_cont, struct task_struct *tsk,
10526 sched_move_task(tsk);
10528 struct task_struct *c;
10530 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
10531 sched_move_task(c);
10537 #ifdef CONFIG_FAIR_GROUP_SCHED
10538 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
10541 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
10544 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
10546 struct task_group *tg = cgroup_tg(cgrp);
10548 return (u64) tg->shares;
10550 #endif /* CONFIG_FAIR_GROUP_SCHED */
10552 #ifdef CONFIG_RT_GROUP_SCHED
10553 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
10556 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
10559 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
10561 return sched_group_rt_runtime(cgroup_tg(cgrp));
10564 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
10567 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
10570 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
10572 return sched_group_rt_period(cgroup_tg(cgrp));
10574 #endif /* CONFIG_RT_GROUP_SCHED */
10576 static struct cftype cpu_files[] = {
10577 #ifdef CONFIG_FAIR_GROUP_SCHED
10580 .read_u64 = cpu_shares_read_u64,
10581 .write_u64 = cpu_shares_write_u64,
10584 #ifdef CONFIG_RT_GROUP_SCHED
10586 .name = "rt_runtime_us",
10587 .read_s64 = cpu_rt_runtime_read,
10588 .write_s64 = cpu_rt_runtime_write,
10591 .name = "rt_period_us",
10592 .read_u64 = cpu_rt_period_read_uint,
10593 .write_u64 = cpu_rt_period_write_uint,
10598 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
10600 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
10603 struct cgroup_subsys cpu_cgroup_subsys = {
10605 .create = cpu_cgroup_create,
10606 .destroy = cpu_cgroup_destroy,
10607 .can_attach = cpu_cgroup_can_attach,
10608 .attach = cpu_cgroup_attach,
10609 .populate = cpu_cgroup_populate,
10610 .subsys_id = cpu_cgroup_subsys_id,
10614 #endif /* CONFIG_CGROUP_SCHED */
10616 #ifdef CONFIG_CGROUP_CPUACCT
10619 * CPU accounting code for task groups.
10621 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10622 * (balbir@in.ibm.com).
10625 /* track cpu usage of a group of tasks and its child groups */
10627 struct cgroup_subsys_state css;
10628 /* cpuusage holds pointer to a u64-type object on every cpu */
10630 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
10631 struct cpuacct *parent;
10634 struct cgroup_subsys cpuacct_subsys;
10636 /* return cpu accounting group corresponding to this container */
10637 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
10639 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
10640 struct cpuacct, css);
10643 /* return cpu accounting group to which this task belongs */
10644 static inline struct cpuacct *task_ca(struct task_struct *tsk)
10646 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
10647 struct cpuacct, css);
10650 /* create a new cpu accounting group */
10651 static struct cgroup_subsys_state *cpuacct_create(
10652 struct cgroup_subsys *ss, struct cgroup *cgrp)
10654 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
10660 ca->cpuusage = alloc_percpu(u64);
10664 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10665 if (percpu_counter_init(&ca->cpustat[i], 0))
10666 goto out_free_counters;
10669 ca->parent = cgroup_ca(cgrp->parent);
10675 percpu_counter_destroy(&ca->cpustat[i]);
10676 free_percpu(ca->cpuusage);
10680 return ERR_PTR(-ENOMEM);
10683 /* destroy an existing cpu accounting group */
10685 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10687 struct cpuacct *ca = cgroup_ca(cgrp);
10690 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10691 percpu_counter_destroy(&ca->cpustat[i]);
10692 free_percpu(ca->cpuusage);
10696 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
10698 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10701 #ifndef CONFIG_64BIT
10703 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10705 spin_lock_irq(&cpu_rq(cpu)->lock);
10707 spin_unlock_irq(&cpu_rq(cpu)->lock);
10715 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
10717 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10719 #ifndef CONFIG_64BIT
10721 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10723 spin_lock_irq(&cpu_rq(cpu)->lock);
10725 spin_unlock_irq(&cpu_rq(cpu)->lock);
10731 /* return total cpu usage (in nanoseconds) of a group */
10732 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
10734 struct cpuacct *ca = cgroup_ca(cgrp);
10735 u64 totalcpuusage = 0;
10738 for_each_present_cpu(i)
10739 totalcpuusage += cpuacct_cpuusage_read(ca, i);
10741 return totalcpuusage;
10744 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
10747 struct cpuacct *ca = cgroup_ca(cgrp);
10756 for_each_present_cpu(i)
10757 cpuacct_cpuusage_write(ca, i, 0);
10763 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
10764 struct seq_file *m)
10766 struct cpuacct *ca = cgroup_ca(cgroup);
10770 for_each_present_cpu(i) {
10771 percpu = cpuacct_cpuusage_read(ca, i);
10772 seq_printf(m, "%llu ", (unsigned long long) percpu);
10774 seq_printf(m, "\n");
10778 static const char *cpuacct_stat_desc[] = {
10779 [CPUACCT_STAT_USER] = "user",
10780 [CPUACCT_STAT_SYSTEM] = "system",
10783 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
10784 struct cgroup_map_cb *cb)
10786 struct cpuacct *ca = cgroup_ca(cgrp);
10789 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
10790 s64 val = percpu_counter_read(&ca->cpustat[i]);
10791 val = cputime64_to_clock_t(val);
10792 cb->fill(cb, cpuacct_stat_desc[i], val);
10797 static struct cftype files[] = {
10800 .read_u64 = cpuusage_read,
10801 .write_u64 = cpuusage_write,
10804 .name = "usage_percpu",
10805 .read_seq_string = cpuacct_percpu_seq_read,
10809 .read_map = cpuacct_stats_show,
10813 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
10815 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
10819 * charge this task's execution time to its accounting group.
10821 * called with rq->lock held.
10823 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
10825 struct cpuacct *ca;
10828 if (unlikely(!cpuacct_subsys.active))
10831 cpu = task_cpu(tsk);
10837 for (; ca; ca = ca->parent) {
10838 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10839 *cpuusage += cputime;
10846 * Charge the system/user time to the task's accounting group.
10848 static void cpuacct_update_stats(struct task_struct *tsk,
10849 enum cpuacct_stat_index idx, cputime_t val)
10851 struct cpuacct *ca;
10853 if (unlikely(!cpuacct_subsys.active))
10860 percpu_counter_add(&ca->cpustat[idx], val);
10866 struct cgroup_subsys cpuacct_subsys = {
10868 .create = cpuacct_create,
10869 .destroy = cpuacct_destroy,
10870 .populate = cpuacct_populate,
10871 .subsys_id = cpuacct_subsys_id,
10873 #endif /* CONFIG_CGROUP_CPUACCT */
10877 int rcu_expedited_torture_stats(char *page)
10881 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
10883 void synchronize_sched_expedited(void)
10886 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
10888 #else /* #ifndef CONFIG_SMP */
10890 static DEFINE_PER_CPU(struct migration_req, rcu_migration_req);
10891 static DEFINE_MUTEX(rcu_sched_expedited_mutex);
10893 #define RCU_EXPEDITED_STATE_POST -2
10894 #define RCU_EXPEDITED_STATE_IDLE -1
10896 static int rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
10898 int rcu_expedited_torture_stats(char *page)
10903 cnt += sprintf(&page[cnt], "state: %d /", rcu_expedited_state);
10904 for_each_online_cpu(cpu) {
10905 cnt += sprintf(&page[cnt], " %d:%d",
10906 cpu, per_cpu(rcu_migration_req, cpu).dest_cpu);
10908 cnt += sprintf(&page[cnt], "\n");
10911 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
10913 static long synchronize_sched_expedited_count;
10916 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
10917 * approach to force grace period to end quickly. This consumes
10918 * significant time on all CPUs, and is thus not recommended for
10919 * any sort of common-case code.
10921 * Note that it is illegal to call this function while holding any
10922 * lock that is acquired by a CPU-hotplug notifier. Failing to
10923 * observe this restriction will result in deadlock.
10925 void synchronize_sched_expedited(void)
10928 unsigned long flags;
10929 bool need_full_sync = 0;
10931 struct migration_req *req;
10935 smp_mb(); /* ensure prior mod happens before capturing snap. */
10936 snap = ACCESS_ONCE(synchronize_sched_expedited_count) + 1;
10938 while (!mutex_trylock(&rcu_sched_expedited_mutex)) {
10940 if (trycount++ < 10)
10941 udelay(trycount * num_online_cpus());
10943 synchronize_sched();
10946 if (ACCESS_ONCE(synchronize_sched_expedited_count) - snap > 0) {
10947 smp_mb(); /* ensure test happens before caller kfree */
10952 rcu_expedited_state = RCU_EXPEDITED_STATE_POST;
10953 for_each_online_cpu(cpu) {
10955 req = &per_cpu(rcu_migration_req, cpu);
10956 init_completion(&req->done);
10958 req->dest_cpu = RCU_MIGRATION_NEED_QS;
10959 spin_lock_irqsave(&rq->lock, flags);
10960 list_add(&req->list, &rq->migration_queue);
10961 spin_unlock_irqrestore(&rq->lock, flags);
10962 wake_up_process(rq->migration_thread);
10964 for_each_online_cpu(cpu) {
10965 rcu_expedited_state = cpu;
10966 req = &per_cpu(rcu_migration_req, cpu);
10968 wait_for_completion(&req->done);
10969 spin_lock_irqsave(&rq->lock, flags);
10970 if (unlikely(req->dest_cpu == RCU_MIGRATION_MUST_SYNC))
10971 need_full_sync = 1;
10972 req->dest_cpu = RCU_MIGRATION_IDLE;
10973 spin_unlock_irqrestore(&rq->lock, flags);
10975 rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
10976 synchronize_sched_expedited_count++;
10977 mutex_unlock(&rcu_sched_expedited_mutex);
10979 if (need_full_sync)
10980 synchronize_sched();
10982 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
10984 #endif /* #else #ifndef CONFIG_SMP */