4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/perf_event.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/kthread.h>
59 #include <linux/proc_fs.h>
60 #include <linux/seq_file.h>
61 #include <linux/sysctl.h>
62 #include <linux/syscalls.h>
63 #include <linux/times.h>
64 #include <linux/tsacct_kern.h>
65 #include <linux/kprobes.h>
66 #include <linux/delayacct.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
73 #include <linux/ftrace.h>
76 #include <asm/irq_regs.h>
78 #include "sched_cpupri.h"
80 #define CREATE_TRACE_POINTS
81 #include <trace/events/sched.h>
84 * Convert user-nice values [ -20 ... 0 ... 19 ]
85 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
88 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
89 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
90 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
93 * 'User priority' is the nice value converted to something we
94 * can work with better when scaling various scheduler parameters,
95 * it's a [ 0 ... 39 ] range.
97 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
98 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
99 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
102 * Helpers for converting nanosecond timing to jiffy resolution
104 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
106 #define NICE_0_LOAD SCHED_LOAD_SCALE
107 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
110 * These are the 'tuning knobs' of the scheduler:
112 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
113 * Timeslices get refilled after they expire.
115 #define DEF_TIMESLICE (100 * HZ / 1000)
118 * single value that denotes runtime == period, ie unlimited time.
120 #define RUNTIME_INF ((u64)~0ULL)
122 static inline int rt_policy(int policy)
124 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
129 static inline int task_has_rt_policy(struct task_struct *p)
131 return rt_policy(p->policy);
135 * This is the priority-queue data structure of the RT scheduling class:
137 struct rt_prio_array {
138 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
139 struct list_head queue[MAX_RT_PRIO];
142 struct rt_bandwidth {
143 /* nests inside the rq lock: */
144 spinlock_t rt_runtime_lock;
147 struct hrtimer rt_period_timer;
150 static struct rt_bandwidth def_rt_bandwidth;
152 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
154 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
156 struct rt_bandwidth *rt_b =
157 container_of(timer, struct rt_bandwidth, rt_period_timer);
163 now = hrtimer_cb_get_time(timer);
164 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
169 idle = do_sched_rt_period_timer(rt_b, overrun);
172 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
176 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
178 rt_b->rt_period = ns_to_ktime(period);
179 rt_b->rt_runtime = runtime;
181 spin_lock_init(&rt_b->rt_runtime_lock);
183 hrtimer_init(&rt_b->rt_period_timer,
184 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
185 rt_b->rt_period_timer.function = sched_rt_period_timer;
188 static inline int rt_bandwidth_enabled(void)
190 return sysctl_sched_rt_runtime >= 0;
193 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
197 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
200 if (hrtimer_active(&rt_b->rt_period_timer))
203 spin_lock(&rt_b->rt_runtime_lock);
208 if (hrtimer_active(&rt_b->rt_period_timer))
211 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
212 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
214 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
215 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
216 delta = ktime_to_ns(ktime_sub(hard, soft));
217 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
218 HRTIMER_MODE_ABS_PINNED, 0);
220 spin_unlock(&rt_b->rt_runtime_lock);
223 #ifdef CONFIG_RT_GROUP_SCHED
224 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
226 hrtimer_cancel(&rt_b->rt_period_timer);
231 * sched_domains_mutex serializes calls to arch_init_sched_domains,
232 * detach_destroy_domains and partition_sched_domains.
234 static DEFINE_MUTEX(sched_domains_mutex);
236 #ifdef CONFIG_GROUP_SCHED
238 #include <linux/cgroup.h>
242 static LIST_HEAD(task_groups);
244 /* task group related information */
246 #ifdef CONFIG_CGROUP_SCHED
247 struct cgroup_subsys_state css;
250 #ifdef CONFIG_USER_SCHED
254 #ifdef CONFIG_FAIR_GROUP_SCHED
255 /* schedulable entities of this group on each cpu */
256 struct sched_entity **se;
257 /* runqueue "owned" by this group on each cpu */
258 struct cfs_rq **cfs_rq;
259 unsigned long shares;
262 #ifdef CONFIG_RT_GROUP_SCHED
263 struct sched_rt_entity **rt_se;
264 struct rt_rq **rt_rq;
266 struct rt_bandwidth rt_bandwidth;
270 struct list_head list;
272 struct task_group *parent;
273 struct list_head siblings;
274 struct list_head children;
277 #ifdef CONFIG_USER_SCHED
279 /* Helper function to pass uid information to create_sched_user() */
280 void set_tg_uid(struct user_struct *user)
282 user->tg->uid = user->uid;
287 * Every UID task group (including init_task_group aka UID-0) will
288 * be a child to this group.
290 struct task_group root_task_group;
292 #ifdef CONFIG_FAIR_GROUP_SCHED
293 /* Default task group's sched entity on each cpu */
294 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
295 /* Default task group's cfs_rq on each cpu */
296 static DEFINE_PER_CPU_SHARED_ALIGNED(struct cfs_rq, init_tg_cfs_rq);
297 #endif /* CONFIG_FAIR_GROUP_SCHED */
299 #ifdef CONFIG_RT_GROUP_SCHED
300 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
301 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rt_rq, init_rt_rq);
302 #endif /* CONFIG_RT_GROUP_SCHED */
303 #else /* !CONFIG_USER_SCHED */
304 #define root_task_group init_task_group
305 #endif /* CONFIG_USER_SCHED */
307 /* task_group_lock serializes add/remove of task groups and also changes to
308 * a task group's cpu shares.
310 static DEFINE_SPINLOCK(task_group_lock);
312 #ifdef CONFIG_FAIR_GROUP_SCHED
315 static int root_task_group_empty(void)
317 return list_empty(&root_task_group.children);
321 #ifdef CONFIG_USER_SCHED
322 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
323 #else /* !CONFIG_USER_SCHED */
324 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
325 #endif /* CONFIG_USER_SCHED */
328 * A weight of 0 or 1 can cause arithmetics problems.
329 * A weight of a cfs_rq is the sum of weights of which entities
330 * are queued on this cfs_rq, so a weight of a entity should not be
331 * too large, so as the shares value of a task group.
332 * (The default weight is 1024 - so there's no practical
333 * limitation from this.)
336 #define MAX_SHARES (1UL << 18)
338 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
341 /* Default task group.
342 * Every task in system belong to this group at bootup.
344 struct task_group init_task_group;
346 /* return group to which a task belongs */
347 static inline struct task_group *task_group(struct task_struct *p)
349 struct task_group *tg;
351 #ifdef CONFIG_USER_SCHED
353 tg = __task_cred(p)->user->tg;
355 #elif defined(CONFIG_CGROUP_SCHED)
356 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
357 struct task_group, css);
359 tg = &init_task_group;
364 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
365 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
367 #ifdef CONFIG_FAIR_GROUP_SCHED
368 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
369 p->se.parent = task_group(p)->se[cpu];
372 #ifdef CONFIG_RT_GROUP_SCHED
373 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
374 p->rt.parent = task_group(p)->rt_se[cpu];
380 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
381 static inline struct task_group *task_group(struct task_struct *p)
386 #endif /* CONFIG_GROUP_SCHED */
388 /* CFS-related fields in a runqueue */
390 struct load_weight load;
391 unsigned long nr_running;
396 struct rb_root tasks_timeline;
397 struct rb_node *rb_leftmost;
399 struct list_head tasks;
400 struct list_head *balance_iterator;
403 * 'curr' points to currently running entity on this cfs_rq.
404 * It is set to NULL otherwise (i.e when none are currently running).
406 struct sched_entity *curr, *next, *last;
408 unsigned int nr_spread_over;
410 #ifdef CONFIG_FAIR_GROUP_SCHED
411 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
414 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
415 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
416 * (like users, containers etc.)
418 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
419 * list is used during load balance.
421 struct list_head leaf_cfs_rq_list;
422 struct task_group *tg; /* group that "owns" this runqueue */
426 * the part of load.weight contributed by tasks
428 unsigned long task_weight;
431 * h_load = weight * f(tg)
433 * Where f(tg) is the recursive weight fraction assigned to
436 unsigned long h_load;
439 * this cpu's part of tg->shares
441 unsigned long shares;
444 * load.weight at the time we set shares
446 unsigned long rq_weight;
451 /* Real-Time classes' related field in a runqueue: */
453 struct rt_prio_array active;
454 unsigned long rt_nr_running;
455 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
457 int curr; /* highest queued rt task prio */
459 int next; /* next highest */
464 unsigned long rt_nr_migratory;
465 unsigned long rt_nr_total;
467 struct plist_head pushable_tasks;
472 /* Nests inside the rq lock: */
473 spinlock_t rt_runtime_lock;
475 #ifdef CONFIG_RT_GROUP_SCHED
476 unsigned long rt_nr_boosted;
479 struct list_head leaf_rt_rq_list;
480 struct task_group *tg;
481 struct sched_rt_entity *rt_se;
488 * We add the notion of a root-domain which will be used to define per-domain
489 * variables. Each exclusive cpuset essentially defines an island domain by
490 * fully partitioning the member cpus from any other cpuset. Whenever a new
491 * exclusive cpuset is created, we also create and attach a new root-domain
498 cpumask_var_t online;
501 * The "RT overload" flag: it gets set if a CPU has more than
502 * one runnable RT task.
504 cpumask_var_t rto_mask;
507 struct cpupri cpupri;
512 * By default the system creates a single root-domain with all cpus as
513 * members (mimicking the global state we have today).
515 static struct root_domain def_root_domain;
520 * This is the main, per-CPU runqueue data structure.
522 * Locking rule: those places that want to lock multiple runqueues
523 * (such as the load balancing or the thread migration code), lock
524 * acquire operations must be ordered by ascending &runqueue.
531 * nr_running and cpu_load should be in the same cacheline because
532 * remote CPUs use both these fields when doing load calculation.
534 unsigned long nr_running;
535 #define CPU_LOAD_IDX_MAX 5
536 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
538 unsigned char in_nohz_recently;
540 /* capture load from *all* tasks on this cpu: */
541 struct load_weight load;
542 unsigned long nr_load_updates;
548 #ifdef CONFIG_FAIR_GROUP_SCHED
549 /* list of leaf cfs_rq on this cpu: */
550 struct list_head leaf_cfs_rq_list;
552 #ifdef CONFIG_RT_GROUP_SCHED
553 struct list_head leaf_rt_rq_list;
557 * This is part of a global counter where only the total sum
558 * over all CPUs matters. A task can increase this counter on
559 * one CPU and if it got migrated afterwards it may decrease
560 * it on another CPU. Always updated under the runqueue lock:
562 unsigned long nr_uninterruptible;
564 struct task_struct *curr, *idle;
565 unsigned long next_balance;
566 struct mm_struct *prev_mm;
573 struct root_domain *rd;
574 struct sched_domain *sd;
576 unsigned char idle_at_tick;
577 /* For active balancing */
581 /* cpu of this runqueue: */
585 unsigned long avg_load_per_task;
587 struct task_struct *migration_thread;
588 struct list_head migration_queue;
596 /* calc_load related fields */
597 unsigned long calc_load_update;
598 long calc_load_active;
600 #ifdef CONFIG_SCHED_HRTICK
602 int hrtick_csd_pending;
603 struct call_single_data hrtick_csd;
605 struct hrtimer hrtick_timer;
608 #ifdef CONFIG_SCHEDSTATS
610 struct sched_info rq_sched_info;
611 unsigned long long rq_cpu_time;
612 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
614 /* sys_sched_yield() stats */
615 unsigned int yld_count;
617 /* schedule() stats */
618 unsigned int sched_switch;
619 unsigned int sched_count;
620 unsigned int sched_goidle;
622 /* try_to_wake_up() stats */
623 unsigned int ttwu_count;
624 unsigned int ttwu_local;
627 unsigned int bkl_count;
631 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
634 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
636 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
639 static inline int cpu_of(struct rq *rq)
649 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
650 * See detach_destroy_domains: synchronize_sched for details.
652 * The domain tree of any CPU may only be accessed from within
653 * preempt-disabled sections.
655 #define for_each_domain(cpu, __sd) \
656 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
658 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
659 #define this_rq() (&__get_cpu_var(runqueues))
660 #define task_rq(p) cpu_rq(task_cpu(p))
661 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
662 #define raw_rq() (&__raw_get_cpu_var(runqueues))
664 inline void update_rq_clock(struct rq *rq)
666 rq->clock = sched_clock_cpu(cpu_of(rq));
670 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
672 #ifdef CONFIG_SCHED_DEBUG
673 # define const_debug __read_mostly
675 # define const_debug static const
680 * @cpu: the processor in question.
682 * Returns true if the current cpu runqueue is locked.
683 * This interface allows printk to be called with the runqueue lock
684 * held and know whether or not it is OK to wake up the klogd.
686 int runqueue_is_locked(int cpu)
688 return spin_is_locked(&cpu_rq(cpu)->lock);
692 * Debugging: various feature bits
695 #define SCHED_FEAT(name, enabled) \
696 __SCHED_FEAT_##name ,
699 #include "sched_features.h"
704 #define SCHED_FEAT(name, enabled) \
705 (1UL << __SCHED_FEAT_##name) * enabled |
707 const_debug unsigned int sysctl_sched_features =
708 #include "sched_features.h"
713 #ifdef CONFIG_SCHED_DEBUG
714 #define SCHED_FEAT(name, enabled) \
717 static __read_mostly char *sched_feat_names[] = {
718 #include "sched_features.h"
724 static int sched_feat_show(struct seq_file *m, void *v)
728 for (i = 0; sched_feat_names[i]; i++) {
729 if (!(sysctl_sched_features & (1UL << i)))
731 seq_printf(m, "%s ", sched_feat_names[i]);
739 sched_feat_write(struct file *filp, const char __user *ubuf,
740 size_t cnt, loff_t *ppos)
750 if (copy_from_user(&buf, ubuf, cnt))
755 if (strncmp(buf, "NO_", 3) == 0) {
760 for (i = 0; sched_feat_names[i]; i++) {
761 int len = strlen(sched_feat_names[i]);
763 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
765 sysctl_sched_features &= ~(1UL << i);
767 sysctl_sched_features |= (1UL << i);
772 if (!sched_feat_names[i])
780 static int sched_feat_open(struct inode *inode, struct file *filp)
782 return single_open(filp, sched_feat_show, NULL);
785 static const struct file_operations sched_feat_fops = {
786 .open = sched_feat_open,
787 .write = sched_feat_write,
790 .release = single_release,
793 static __init int sched_init_debug(void)
795 debugfs_create_file("sched_features", 0644, NULL, NULL,
800 late_initcall(sched_init_debug);
804 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
807 * Number of tasks to iterate in a single balance run.
808 * Limited because this is done with IRQs disabled.
810 const_debug unsigned int sysctl_sched_nr_migrate = 32;
813 * ratelimit for updating the group shares.
816 unsigned int sysctl_sched_shares_ratelimit = 250000;
819 * Inject some fuzzyness into changing the per-cpu group shares
820 * this avoids remote rq-locks at the expense of fairness.
823 unsigned int sysctl_sched_shares_thresh = 4;
826 * period over which we average the RT time consumption, measured
831 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
834 * period over which we measure -rt task cpu usage in us.
837 unsigned int sysctl_sched_rt_period = 1000000;
839 static __read_mostly int scheduler_running;
842 * part of the period that we allow rt tasks to run in us.
845 int sysctl_sched_rt_runtime = 950000;
847 static inline u64 global_rt_period(void)
849 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
852 static inline u64 global_rt_runtime(void)
854 if (sysctl_sched_rt_runtime < 0)
857 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
860 #ifndef prepare_arch_switch
861 # define prepare_arch_switch(next) do { } while (0)
863 #ifndef finish_arch_switch
864 # define finish_arch_switch(prev) do { } while (0)
867 static inline int task_current(struct rq *rq, struct task_struct *p)
869 return rq->curr == p;
872 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
873 static inline int task_running(struct rq *rq, struct task_struct *p)
875 return task_current(rq, p);
878 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
882 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
884 #ifdef CONFIG_DEBUG_SPINLOCK
885 /* this is a valid case when another task releases the spinlock */
886 rq->lock.owner = current;
889 * If we are tracking spinlock dependencies then we have to
890 * fix up the runqueue lock - which gets 'carried over' from
893 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
895 spin_unlock_irq(&rq->lock);
898 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
899 static inline int task_running(struct rq *rq, struct task_struct *p)
904 return task_current(rq, p);
908 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
912 * We can optimise this out completely for !SMP, because the
913 * SMP rebalancing from interrupt is the only thing that cares
918 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
919 spin_unlock_irq(&rq->lock);
921 spin_unlock(&rq->lock);
925 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
929 * After ->oncpu is cleared, the task can be moved to a different CPU.
930 * We must ensure this doesn't happen until the switch is completely
936 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
940 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
943 * __task_rq_lock - lock the runqueue a given task resides on.
944 * Must be called interrupts disabled.
946 static inline struct rq *__task_rq_lock(struct task_struct *p)
950 struct rq *rq = task_rq(p);
951 spin_lock(&rq->lock);
952 if (likely(rq == task_rq(p)))
954 spin_unlock(&rq->lock);
959 * task_rq_lock - lock the runqueue a given task resides on and disable
960 * interrupts. Note the ordering: we can safely lookup the task_rq without
961 * explicitly disabling preemption.
963 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
969 local_irq_save(*flags);
971 spin_lock(&rq->lock);
972 if (likely(rq == task_rq(p)))
974 spin_unlock_irqrestore(&rq->lock, *flags);
978 void task_rq_unlock_wait(struct task_struct *p)
980 struct rq *rq = task_rq(p);
982 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
983 spin_unlock_wait(&rq->lock);
986 static void __task_rq_unlock(struct rq *rq)
989 spin_unlock(&rq->lock);
992 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
995 spin_unlock_irqrestore(&rq->lock, *flags);
999 * this_rq_lock - lock this runqueue and disable interrupts.
1001 static struct rq *this_rq_lock(void)
1002 __acquires(rq->lock)
1006 local_irq_disable();
1008 spin_lock(&rq->lock);
1013 #ifdef CONFIG_SCHED_HRTICK
1015 * Use HR-timers to deliver accurate preemption points.
1017 * Its all a bit involved since we cannot program an hrt while holding the
1018 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1021 * When we get rescheduled we reprogram the hrtick_timer outside of the
1027 * - enabled by features
1028 * - hrtimer is actually high res
1030 static inline int hrtick_enabled(struct rq *rq)
1032 if (!sched_feat(HRTICK))
1034 if (!cpu_active(cpu_of(rq)))
1036 return hrtimer_is_hres_active(&rq->hrtick_timer);
1039 static void hrtick_clear(struct rq *rq)
1041 if (hrtimer_active(&rq->hrtick_timer))
1042 hrtimer_cancel(&rq->hrtick_timer);
1046 * High-resolution timer tick.
1047 * Runs from hardirq context with interrupts disabled.
1049 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1051 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1053 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1055 spin_lock(&rq->lock);
1056 update_rq_clock(rq);
1057 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1058 spin_unlock(&rq->lock);
1060 return HRTIMER_NORESTART;
1065 * called from hardirq (IPI) context
1067 static void __hrtick_start(void *arg)
1069 struct rq *rq = arg;
1071 spin_lock(&rq->lock);
1072 hrtimer_restart(&rq->hrtick_timer);
1073 rq->hrtick_csd_pending = 0;
1074 spin_unlock(&rq->lock);
1078 * Called to set the hrtick timer state.
1080 * called with rq->lock held and irqs disabled
1082 static void hrtick_start(struct rq *rq, u64 delay)
1084 struct hrtimer *timer = &rq->hrtick_timer;
1085 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1087 hrtimer_set_expires(timer, time);
1089 if (rq == this_rq()) {
1090 hrtimer_restart(timer);
1091 } else if (!rq->hrtick_csd_pending) {
1092 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1093 rq->hrtick_csd_pending = 1;
1098 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1100 int cpu = (int)(long)hcpu;
1103 case CPU_UP_CANCELED:
1104 case CPU_UP_CANCELED_FROZEN:
1105 case CPU_DOWN_PREPARE:
1106 case CPU_DOWN_PREPARE_FROZEN:
1108 case CPU_DEAD_FROZEN:
1109 hrtick_clear(cpu_rq(cpu));
1116 static __init void init_hrtick(void)
1118 hotcpu_notifier(hotplug_hrtick, 0);
1122 * Called to set the hrtick timer state.
1124 * called with rq->lock held and irqs disabled
1126 static void hrtick_start(struct rq *rq, u64 delay)
1128 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1129 HRTIMER_MODE_REL_PINNED, 0);
1132 static inline void init_hrtick(void)
1135 #endif /* CONFIG_SMP */
1137 static void init_rq_hrtick(struct rq *rq)
1140 rq->hrtick_csd_pending = 0;
1142 rq->hrtick_csd.flags = 0;
1143 rq->hrtick_csd.func = __hrtick_start;
1144 rq->hrtick_csd.info = rq;
1147 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1148 rq->hrtick_timer.function = hrtick;
1150 #else /* CONFIG_SCHED_HRTICK */
1151 static inline void hrtick_clear(struct rq *rq)
1155 static inline void init_rq_hrtick(struct rq *rq)
1159 static inline void init_hrtick(void)
1162 #endif /* CONFIG_SCHED_HRTICK */
1165 * resched_task - mark a task 'to be rescheduled now'.
1167 * On UP this means the setting of the need_resched flag, on SMP it
1168 * might also involve a cross-CPU call to trigger the scheduler on
1173 #ifndef tsk_is_polling
1174 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1177 static void resched_task(struct task_struct *p)
1181 assert_spin_locked(&task_rq(p)->lock);
1183 if (test_tsk_need_resched(p))
1186 set_tsk_need_resched(p);
1189 if (cpu == smp_processor_id())
1192 /* NEED_RESCHED must be visible before we test polling */
1194 if (!tsk_is_polling(p))
1195 smp_send_reschedule(cpu);
1198 static void resched_cpu(int cpu)
1200 struct rq *rq = cpu_rq(cpu);
1201 unsigned long flags;
1203 if (!spin_trylock_irqsave(&rq->lock, flags))
1205 resched_task(cpu_curr(cpu));
1206 spin_unlock_irqrestore(&rq->lock, flags);
1211 * When add_timer_on() enqueues a timer into the timer wheel of an
1212 * idle CPU then this timer might expire before the next timer event
1213 * which is scheduled to wake up that CPU. In case of a completely
1214 * idle system the next event might even be infinite time into the
1215 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1216 * leaves the inner idle loop so the newly added timer is taken into
1217 * account when the CPU goes back to idle and evaluates the timer
1218 * wheel for the next timer event.
1220 void wake_up_idle_cpu(int cpu)
1222 struct rq *rq = cpu_rq(cpu);
1224 if (cpu == smp_processor_id())
1228 * This is safe, as this function is called with the timer
1229 * wheel base lock of (cpu) held. When the CPU is on the way
1230 * to idle and has not yet set rq->curr to idle then it will
1231 * be serialized on the timer wheel base lock and take the new
1232 * timer into account automatically.
1234 if (rq->curr != rq->idle)
1238 * We can set TIF_RESCHED on the idle task of the other CPU
1239 * lockless. The worst case is that the other CPU runs the
1240 * idle task through an additional NOOP schedule()
1242 set_tsk_need_resched(rq->idle);
1244 /* NEED_RESCHED must be visible before we test polling */
1246 if (!tsk_is_polling(rq->idle))
1247 smp_send_reschedule(cpu);
1249 #endif /* CONFIG_NO_HZ */
1251 static u64 sched_avg_period(void)
1253 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1256 static void sched_avg_update(struct rq *rq)
1258 s64 period = sched_avg_period();
1260 while ((s64)(rq->clock - rq->age_stamp) > period) {
1261 rq->age_stamp += period;
1266 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1268 rq->rt_avg += rt_delta;
1269 sched_avg_update(rq);
1272 #else /* !CONFIG_SMP */
1273 static void resched_task(struct task_struct *p)
1275 assert_spin_locked(&task_rq(p)->lock);
1276 set_tsk_need_resched(p);
1279 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1282 #endif /* CONFIG_SMP */
1284 #if BITS_PER_LONG == 32
1285 # define WMULT_CONST (~0UL)
1287 # define WMULT_CONST (1UL << 32)
1290 #define WMULT_SHIFT 32
1293 * Shift right and round:
1295 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1298 * delta *= weight / lw
1300 static unsigned long
1301 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1302 struct load_weight *lw)
1306 if (!lw->inv_weight) {
1307 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1310 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1314 tmp = (u64)delta_exec * weight;
1316 * Check whether we'd overflow the 64-bit multiplication:
1318 if (unlikely(tmp > WMULT_CONST))
1319 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1322 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1324 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1327 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1333 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1340 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1341 * of tasks with abnormal "nice" values across CPUs the contribution that
1342 * each task makes to its run queue's load is weighted according to its
1343 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1344 * scaled version of the new time slice allocation that they receive on time
1348 #define WEIGHT_IDLEPRIO 3
1349 #define WMULT_IDLEPRIO 1431655765
1352 * Nice levels are multiplicative, with a gentle 10% change for every
1353 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1354 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1355 * that remained on nice 0.
1357 * The "10% effect" is relative and cumulative: from _any_ nice level,
1358 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1359 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1360 * If a task goes up by ~10% and another task goes down by ~10% then
1361 * the relative distance between them is ~25%.)
1363 static const int prio_to_weight[40] = {
1364 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1365 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1366 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1367 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1368 /* 0 */ 1024, 820, 655, 526, 423,
1369 /* 5 */ 335, 272, 215, 172, 137,
1370 /* 10 */ 110, 87, 70, 56, 45,
1371 /* 15 */ 36, 29, 23, 18, 15,
1375 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1377 * In cases where the weight does not change often, we can use the
1378 * precalculated inverse to speed up arithmetics by turning divisions
1379 * into multiplications:
1381 static const u32 prio_to_wmult[40] = {
1382 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1383 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1384 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1385 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1386 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1387 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1388 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1389 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1392 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1395 * runqueue iterator, to support SMP load-balancing between different
1396 * scheduling classes, without having to expose their internal data
1397 * structures to the load-balancing proper:
1399 struct rq_iterator {
1401 struct task_struct *(*start)(void *);
1402 struct task_struct *(*next)(void *);
1406 static unsigned long
1407 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1408 unsigned long max_load_move, struct sched_domain *sd,
1409 enum cpu_idle_type idle, int *all_pinned,
1410 int *this_best_prio, struct rq_iterator *iterator);
1413 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1414 struct sched_domain *sd, enum cpu_idle_type idle,
1415 struct rq_iterator *iterator);
1418 /* Time spent by the tasks of the cpu accounting group executing in ... */
1419 enum cpuacct_stat_index {
1420 CPUACCT_STAT_USER, /* ... user mode */
1421 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1423 CPUACCT_STAT_NSTATS,
1426 #ifdef CONFIG_CGROUP_CPUACCT
1427 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1428 static void cpuacct_update_stats(struct task_struct *tsk,
1429 enum cpuacct_stat_index idx, cputime_t val);
1431 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1432 static inline void cpuacct_update_stats(struct task_struct *tsk,
1433 enum cpuacct_stat_index idx, cputime_t val) {}
1436 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1438 update_load_add(&rq->load, load);
1441 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1443 update_load_sub(&rq->load, load);
1446 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1447 typedef int (*tg_visitor)(struct task_group *, void *);
1450 * Iterate the full tree, calling @down when first entering a node and @up when
1451 * leaving it for the final time.
1453 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1455 struct task_group *parent, *child;
1459 parent = &root_task_group;
1461 ret = (*down)(parent, data);
1464 list_for_each_entry_rcu(child, &parent->children, siblings) {
1471 ret = (*up)(parent, data);
1476 parent = parent->parent;
1485 static int tg_nop(struct task_group *tg, void *data)
1492 /* Used instead of source_load when we know the type == 0 */
1493 static unsigned long weighted_cpuload(const int cpu)
1495 return cpu_rq(cpu)->load.weight;
1499 * Return a low guess at the load of a migration-source cpu weighted
1500 * according to the scheduling class and "nice" value.
1502 * We want to under-estimate the load of migration sources, to
1503 * balance conservatively.
1505 static unsigned long source_load(int cpu, int type)
1507 struct rq *rq = cpu_rq(cpu);
1508 unsigned long total = weighted_cpuload(cpu);
1510 if (type == 0 || !sched_feat(LB_BIAS))
1513 return min(rq->cpu_load[type-1], total);
1517 * Return a high guess at the load of a migration-target cpu weighted
1518 * according to the scheduling class and "nice" value.
1520 static unsigned long target_load(int cpu, int type)
1522 struct rq *rq = cpu_rq(cpu);
1523 unsigned long total = weighted_cpuload(cpu);
1525 if (type == 0 || !sched_feat(LB_BIAS))
1528 return max(rq->cpu_load[type-1], total);
1531 static struct sched_group *group_of(int cpu)
1533 struct sched_domain *sd = rcu_dereference(cpu_rq(cpu)->sd);
1541 static unsigned long power_of(int cpu)
1543 struct sched_group *group = group_of(cpu);
1546 return SCHED_LOAD_SCALE;
1548 return group->cpu_power;
1551 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1553 static unsigned long cpu_avg_load_per_task(int cpu)
1555 struct rq *rq = cpu_rq(cpu);
1556 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1559 rq->avg_load_per_task = rq->load.weight / nr_running;
1561 rq->avg_load_per_task = 0;
1563 return rq->avg_load_per_task;
1566 #ifdef CONFIG_FAIR_GROUP_SCHED
1568 static __read_mostly unsigned long *update_shares_data;
1570 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1573 * Calculate and set the cpu's group shares.
1575 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1576 unsigned long sd_shares,
1577 unsigned long sd_rq_weight,
1578 unsigned long *usd_rq_weight)
1580 unsigned long shares, rq_weight;
1583 rq_weight = usd_rq_weight[cpu];
1586 rq_weight = NICE_0_LOAD;
1590 * \Sum_j shares_j * rq_weight_i
1591 * shares_i = -----------------------------
1592 * \Sum_j rq_weight_j
1594 shares = (sd_shares * rq_weight) / sd_rq_weight;
1595 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1597 if (abs(shares - tg->se[cpu]->load.weight) >
1598 sysctl_sched_shares_thresh) {
1599 struct rq *rq = cpu_rq(cpu);
1600 unsigned long flags;
1602 spin_lock_irqsave(&rq->lock, flags);
1603 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1604 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1605 __set_se_shares(tg->se[cpu], shares);
1606 spin_unlock_irqrestore(&rq->lock, flags);
1611 * Re-compute the task group their per cpu shares over the given domain.
1612 * This needs to be done in a bottom-up fashion because the rq weight of a
1613 * parent group depends on the shares of its child groups.
1615 static int tg_shares_up(struct task_group *tg, void *data)
1617 unsigned long weight, rq_weight = 0, shares = 0;
1618 unsigned long *usd_rq_weight;
1619 struct sched_domain *sd = data;
1620 unsigned long flags;
1626 local_irq_save(flags);
1627 usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id());
1629 for_each_cpu(i, sched_domain_span(sd)) {
1630 weight = tg->cfs_rq[i]->load.weight;
1631 usd_rq_weight[i] = weight;
1634 * If there are currently no tasks on the cpu pretend there
1635 * is one of average load so that when a new task gets to
1636 * run here it will not get delayed by group starvation.
1639 weight = NICE_0_LOAD;
1641 rq_weight += weight;
1642 shares += tg->cfs_rq[i]->shares;
1645 if ((!shares && rq_weight) || shares > tg->shares)
1646 shares = tg->shares;
1648 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1649 shares = tg->shares;
1651 for_each_cpu(i, sched_domain_span(sd))
1652 update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight);
1654 local_irq_restore(flags);
1660 * Compute the cpu's hierarchical load factor for each task group.
1661 * This needs to be done in a top-down fashion because the load of a child
1662 * group is a fraction of its parents load.
1664 static int tg_load_down(struct task_group *tg, void *data)
1667 long cpu = (long)data;
1670 load = cpu_rq(cpu)->load.weight;
1672 load = tg->parent->cfs_rq[cpu]->h_load;
1673 load *= tg->cfs_rq[cpu]->shares;
1674 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1677 tg->cfs_rq[cpu]->h_load = load;
1682 static void update_shares(struct sched_domain *sd)
1687 if (root_task_group_empty())
1690 now = cpu_clock(raw_smp_processor_id());
1691 elapsed = now - sd->last_update;
1693 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1694 sd->last_update = now;
1695 walk_tg_tree(tg_nop, tg_shares_up, sd);
1699 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1701 if (root_task_group_empty())
1704 spin_unlock(&rq->lock);
1706 spin_lock(&rq->lock);
1709 static void update_h_load(long cpu)
1711 if (root_task_group_empty())
1714 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1719 static inline void update_shares(struct sched_domain *sd)
1723 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1729 #ifdef CONFIG_PREEMPT
1731 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1734 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1735 * way at the expense of forcing extra atomic operations in all
1736 * invocations. This assures that the double_lock is acquired using the
1737 * same underlying policy as the spinlock_t on this architecture, which
1738 * reduces latency compared to the unfair variant below. However, it
1739 * also adds more overhead and therefore may reduce throughput.
1741 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1742 __releases(this_rq->lock)
1743 __acquires(busiest->lock)
1744 __acquires(this_rq->lock)
1746 spin_unlock(&this_rq->lock);
1747 double_rq_lock(this_rq, busiest);
1754 * Unfair double_lock_balance: Optimizes throughput at the expense of
1755 * latency by eliminating extra atomic operations when the locks are
1756 * already in proper order on entry. This favors lower cpu-ids and will
1757 * grant the double lock to lower cpus over higher ids under contention,
1758 * regardless of entry order into the function.
1760 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1761 __releases(this_rq->lock)
1762 __acquires(busiest->lock)
1763 __acquires(this_rq->lock)
1767 if (unlikely(!spin_trylock(&busiest->lock))) {
1768 if (busiest < this_rq) {
1769 spin_unlock(&this_rq->lock);
1770 spin_lock(&busiest->lock);
1771 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
1774 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
1779 #endif /* CONFIG_PREEMPT */
1782 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1784 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1786 if (unlikely(!irqs_disabled())) {
1787 /* printk() doesn't work good under rq->lock */
1788 spin_unlock(&this_rq->lock);
1792 return _double_lock_balance(this_rq, busiest);
1795 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1796 __releases(busiest->lock)
1798 spin_unlock(&busiest->lock);
1799 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1803 #ifdef CONFIG_FAIR_GROUP_SCHED
1804 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1807 cfs_rq->shares = shares;
1812 static void calc_load_account_active(struct rq *this_rq);
1814 #include "sched_stats.h"
1815 #include "sched_idletask.c"
1816 #include "sched_fair.c"
1817 #include "sched_rt.c"
1818 #ifdef CONFIG_SCHED_DEBUG
1819 # include "sched_debug.c"
1822 #define sched_class_highest (&rt_sched_class)
1823 #define for_each_class(class) \
1824 for (class = sched_class_highest; class; class = class->next)
1826 static void inc_nr_running(struct rq *rq)
1831 static void dec_nr_running(struct rq *rq)
1836 static void set_load_weight(struct task_struct *p)
1838 if (task_has_rt_policy(p)) {
1839 p->se.load.weight = prio_to_weight[0] * 2;
1840 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1845 * SCHED_IDLE tasks get minimal weight:
1847 if (p->policy == SCHED_IDLE) {
1848 p->se.load.weight = WEIGHT_IDLEPRIO;
1849 p->se.load.inv_weight = WMULT_IDLEPRIO;
1853 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1854 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1857 static void update_avg(u64 *avg, u64 sample)
1859 s64 diff = sample - *avg;
1863 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1866 p->se.start_runtime = p->se.sum_exec_runtime;
1868 sched_info_queued(p);
1869 p->sched_class->enqueue_task(rq, p, wakeup);
1873 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1876 if (p->se.last_wakeup) {
1877 update_avg(&p->se.avg_overlap,
1878 p->se.sum_exec_runtime - p->se.last_wakeup);
1879 p->se.last_wakeup = 0;
1881 update_avg(&p->se.avg_wakeup,
1882 sysctl_sched_wakeup_granularity);
1886 sched_info_dequeued(p);
1887 p->sched_class->dequeue_task(rq, p, sleep);
1892 * __normal_prio - return the priority that is based on the static prio
1894 static inline int __normal_prio(struct task_struct *p)
1896 return p->static_prio;
1900 * Calculate the expected normal priority: i.e. priority
1901 * without taking RT-inheritance into account. Might be
1902 * boosted by interactivity modifiers. Changes upon fork,
1903 * setprio syscalls, and whenever the interactivity
1904 * estimator recalculates.
1906 static inline int normal_prio(struct task_struct *p)
1910 if (task_has_rt_policy(p))
1911 prio = MAX_RT_PRIO-1 - p->rt_priority;
1913 prio = __normal_prio(p);
1918 * Calculate the current priority, i.e. the priority
1919 * taken into account by the scheduler. This value might
1920 * be boosted by RT tasks, or might be boosted by
1921 * interactivity modifiers. Will be RT if the task got
1922 * RT-boosted. If not then it returns p->normal_prio.
1924 static int effective_prio(struct task_struct *p)
1926 p->normal_prio = normal_prio(p);
1928 * If we are RT tasks or we were boosted to RT priority,
1929 * keep the priority unchanged. Otherwise, update priority
1930 * to the normal priority:
1932 if (!rt_prio(p->prio))
1933 return p->normal_prio;
1938 * activate_task - move a task to the runqueue.
1940 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1942 if (task_contributes_to_load(p))
1943 rq->nr_uninterruptible--;
1945 enqueue_task(rq, p, wakeup);
1950 * deactivate_task - remove a task from the runqueue.
1952 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1954 if (task_contributes_to_load(p))
1955 rq->nr_uninterruptible++;
1957 dequeue_task(rq, p, sleep);
1962 * task_curr - is this task currently executing on a CPU?
1963 * @p: the task in question.
1965 inline int task_curr(const struct task_struct *p)
1967 return cpu_curr(task_cpu(p)) == p;
1970 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1972 set_task_rq(p, cpu);
1975 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1976 * successfuly executed on another CPU. We must ensure that updates of
1977 * per-task data have been completed by this moment.
1980 task_thread_info(p)->cpu = cpu;
1984 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1985 const struct sched_class *prev_class,
1986 int oldprio, int running)
1988 if (prev_class != p->sched_class) {
1989 if (prev_class->switched_from)
1990 prev_class->switched_from(rq, p, running);
1991 p->sched_class->switched_to(rq, p, running);
1993 p->sched_class->prio_changed(rq, p, oldprio, running);
1997 * kthread_bind - bind a just-created kthread to a cpu.
1998 * @p: thread created by kthread_create().
1999 * @cpu: cpu (might not be online, must be possible) for @k to run on.
2001 * Description: This function is equivalent to set_cpus_allowed(),
2002 * except that @cpu doesn't need to be online, and the thread must be
2003 * stopped (i.e., just returned from kthread_create()).
2005 * Function lives here instead of kthread.c because it messes with
2006 * scheduler internals which require locking.
2008 void kthread_bind(struct task_struct *p, unsigned int cpu)
2010 struct rq *rq = cpu_rq(cpu);
2011 unsigned long flags;
2013 /* Must have done schedule() in kthread() before we set_task_cpu */
2014 if (!wait_task_inactive(p, TASK_UNINTERRUPTIBLE)) {
2019 spin_lock_irqsave(&rq->lock, flags);
2020 update_rq_clock(rq);
2021 set_task_cpu(p, cpu);
2022 p->cpus_allowed = cpumask_of_cpu(cpu);
2023 p->rt.nr_cpus_allowed = 1;
2024 p->flags |= PF_THREAD_BOUND;
2025 spin_unlock_irqrestore(&rq->lock, flags);
2027 EXPORT_SYMBOL(kthread_bind);
2031 * Is this task likely cache-hot:
2034 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2039 * Buddy candidates are cache hot:
2041 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2042 (&p->se == cfs_rq_of(&p->se)->next ||
2043 &p->se == cfs_rq_of(&p->se)->last))
2046 if (p->sched_class != &fair_sched_class)
2049 if (sysctl_sched_migration_cost == -1)
2051 if (sysctl_sched_migration_cost == 0)
2054 delta = now - p->se.exec_start;
2056 return delta < (s64)sysctl_sched_migration_cost;
2060 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2062 int old_cpu = task_cpu(p);
2063 struct rq *old_rq = cpu_rq(old_cpu);
2064 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
2065 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
2067 trace_sched_migrate_task(p, new_cpu);
2069 if (old_cpu != new_cpu) {
2070 p->se.nr_migrations++;
2071 #ifdef CONFIG_SCHEDSTATS
2072 if (task_hot(p, old_rq->clock, NULL))
2073 schedstat_inc(p, se.nr_forced2_migrations);
2075 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS,
2078 p->se.vruntime -= old_cfsrq->min_vruntime -
2079 new_cfsrq->min_vruntime;
2081 __set_task_cpu(p, new_cpu);
2084 struct migration_req {
2085 struct list_head list;
2087 struct task_struct *task;
2090 struct completion done;
2094 * The task's runqueue lock must be held.
2095 * Returns true if you have to wait for migration thread.
2098 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2100 struct rq *rq = task_rq(p);
2103 * If the task is not on a runqueue (and not running), then
2104 * it is sufficient to simply update the task's cpu field.
2106 if (!p->se.on_rq && !task_running(rq, p)) {
2107 update_rq_clock(rq);
2108 set_task_cpu(p, dest_cpu);
2112 init_completion(&req->done);
2114 req->dest_cpu = dest_cpu;
2115 list_add(&req->list, &rq->migration_queue);
2121 * wait_task_context_switch - wait for a thread to complete at least one
2124 * @p must not be current.
2126 void wait_task_context_switch(struct task_struct *p)
2128 unsigned long nvcsw, nivcsw, flags;
2136 * The runqueue is assigned before the actual context
2137 * switch. We need to take the runqueue lock.
2139 * We could check initially without the lock but it is
2140 * very likely that we need to take the lock in every
2143 rq = task_rq_lock(p, &flags);
2144 running = task_running(rq, p);
2145 task_rq_unlock(rq, &flags);
2147 if (likely(!running))
2150 * The switch count is incremented before the actual
2151 * context switch. We thus wait for two switches to be
2152 * sure at least one completed.
2154 if ((p->nvcsw - nvcsw) > 1)
2156 if ((p->nivcsw - nivcsw) > 1)
2164 * wait_task_inactive - wait for a thread to unschedule.
2166 * If @match_state is nonzero, it's the @p->state value just checked and
2167 * not expected to change. If it changes, i.e. @p might have woken up,
2168 * then return zero. When we succeed in waiting for @p to be off its CPU,
2169 * we return a positive number (its total switch count). If a second call
2170 * a short while later returns the same number, the caller can be sure that
2171 * @p has remained unscheduled the whole time.
2173 * The caller must ensure that the task *will* unschedule sometime soon,
2174 * else this function might spin for a *long* time. This function can't
2175 * be called with interrupts off, or it may introduce deadlock with
2176 * smp_call_function() if an IPI is sent by the same process we are
2177 * waiting to become inactive.
2179 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2181 unsigned long flags;
2188 * We do the initial early heuristics without holding
2189 * any task-queue locks at all. We'll only try to get
2190 * the runqueue lock when things look like they will
2196 * If the task is actively running on another CPU
2197 * still, just relax and busy-wait without holding
2200 * NOTE! Since we don't hold any locks, it's not
2201 * even sure that "rq" stays as the right runqueue!
2202 * But we don't care, since "task_running()" will
2203 * return false if the runqueue has changed and p
2204 * is actually now running somewhere else!
2206 while (task_running(rq, p)) {
2207 if (match_state && unlikely(p->state != match_state))
2213 * Ok, time to look more closely! We need the rq
2214 * lock now, to be *sure*. If we're wrong, we'll
2215 * just go back and repeat.
2217 rq = task_rq_lock(p, &flags);
2218 trace_sched_wait_task(rq, p);
2219 running = task_running(rq, p);
2220 on_rq = p->se.on_rq;
2222 if (!match_state || p->state == match_state)
2223 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2224 task_rq_unlock(rq, &flags);
2227 * If it changed from the expected state, bail out now.
2229 if (unlikely(!ncsw))
2233 * Was it really running after all now that we
2234 * checked with the proper locks actually held?
2236 * Oops. Go back and try again..
2238 if (unlikely(running)) {
2244 * It's not enough that it's not actively running,
2245 * it must be off the runqueue _entirely_, and not
2248 * So if it was still runnable (but just not actively
2249 * running right now), it's preempted, and we should
2250 * yield - it could be a while.
2252 if (unlikely(on_rq)) {
2253 schedule_timeout_uninterruptible(1);
2258 * Ahh, all good. It wasn't running, and it wasn't
2259 * runnable, which means that it will never become
2260 * running in the future either. We're all done!
2269 * kick_process - kick a running thread to enter/exit the kernel
2270 * @p: the to-be-kicked thread
2272 * Cause a process which is running on another CPU to enter
2273 * kernel-mode, without any delay. (to get signals handled.)
2275 * NOTE: this function doesnt have to take the runqueue lock,
2276 * because all it wants to ensure is that the remote task enters
2277 * the kernel. If the IPI races and the task has been migrated
2278 * to another CPU then no harm is done and the purpose has been
2281 void kick_process(struct task_struct *p)
2287 if ((cpu != smp_processor_id()) && task_curr(p))
2288 smp_send_reschedule(cpu);
2291 EXPORT_SYMBOL_GPL(kick_process);
2292 #endif /* CONFIG_SMP */
2295 * task_oncpu_function_call - call a function on the cpu on which a task runs
2296 * @p: the task to evaluate
2297 * @func: the function to be called
2298 * @info: the function call argument
2300 * Calls the function @func when the task is currently running. This might
2301 * be on the current CPU, which just calls the function directly
2303 void task_oncpu_function_call(struct task_struct *p,
2304 void (*func) (void *info), void *info)
2311 smp_call_function_single(cpu, func, info, 1);
2317 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2319 return p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2324 * try_to_wake_up - wake up a thread
2325 * @p: the to-be-woken-up thread
2326 * @state: the mask of task states that can be woken
2327 * @sync: do a synchronous wakeup?
2329 * Put it on the run-queue if it's not already there. The "current"
2330 * thread is always on the run-queue (except when the actual
2331 * re-schedule is in progress), and as such you're allowed to do
2332 * the simpler "current->state = TASK_RUNNING" to mark yourself
2333 * runnable without the overhead of this.
2335 * returns failure only if the task is already active.
2337 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2340 int cpu, orig_cpu, this_cpu, success = 0;
2341 unsigned long flags;
2342 struct rq *rq, *orig_rq;
2344 if (!sched_feat(SYNC_WAKEUPS))
2345 wake_flags &= ~WF_SYNC;
2347 this_cpu = get_cpu();
2350 rq = orig_rq = task_rq_lock(p, &flags);
2351 update_rq_clock(rq);
2352 if (!(p->state & state))
2362 if (unlikely(task_running(rq, p)))
2366 * In order to handle concurrent wakeups and release the rq->lock
2367 * we put the task in TASK_WAKING state.
2369 * First fix up the nr_uninterruptible count:
2371 if (task_contributes_to_load(p))
2372 rq->nr_uninterruptible--;
2373 p->state = TASK_WAKING;
2374 __task_rq_unlock(rq);
2376 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2377 if (cpu != orig_cpu)
2378 set_task_cpu(p, cpu);
2380 rq = __task_rq_lock(p);
2381 update_rq_clock(rq);
2383 WARN_ON(p->state != TASK_WAKING);
2386 #ifdef CONFIG_SCHEDSTATS
2387 schedstat_inc(rq, ttwu_count);
2388 if (cpu == this_cpu)
2389 schedstat_inc(rq, ttwu_local);
2391 struct sched_domain *sd;
2392 for_each_domain(this_cpu, sd) {
2393 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2394 schedstat_inc(sd, ttwu_wake_remote);
2399 #endif /* CONFIG_SCHEDSTATS */
2402 #endif /* CONFIG_SMP */
2403 schedstat_inc(p, se.nr_wakeups);
2404 if (wake_flags & WF_SYNC)
2405 schedstat_inc(p, se.nr_wakeups_sync);
2406 if (orig_cpu != cpu)
2407 schedstat_inc(p, se.nr_wakeups_migrate);
2408 if (cpu == this_cpu)
2409 schedstat_inc(p, se.nr_wakeups_local);
2411 schedstat_inc(p, se.nr_wakeups_remote);
2412 activate_task(rq, p, 1);
2416 * Only attribute actual wakeups done by this task.
2418 if (!in_interrupt()) {
2419 struct sched_entity *se = ¤t->se;
2420 u64 sample = se->sum_exec_runtime;
2422 if (se->last_wakeup)
2423 sample -= se->last_wakeup;
2425 sample -= se->start_runtime;
2426 update_avg(&se->avg_wakeup, sample);
2428 se->last_wakeup = se->sum_exec_runtime;
2432 trace_sched_wakeup(rq, p, success);
2433 check_preempt_curr(rq, p, wake_flags);
2435 p->state = TASK_RUNNING;
2437 if (p->sched_class->task_wake_up)
2438 p->sched_class->task_wake_up(rq, p);
2440 if (unlikely(rq->idle_stamp)) {
2441 u64 delta = rq->clock - rq->idle_stamp;
2442 u64 max = 2*sysctl_sched_migration_cost;
2447 update_avg(&rq->avg_idle, delta);
2452 task_rq_unlock(rq, &flags);
2459 * wake_up_process - Wake up a specific process
2460 * @p: The process to be woken up.
2462 * Attempt to wake up the nominated process and move it to the set of runnable
2463 * processes. Returns 1 if the process was woken up, 0 if it was already
2466 * It may be assumed that this function implies a write memory barrier before
2467 * changing the task state if and only if any tasks are woken up.
2469 int wake_up_process(struct task_struct *p)
2471 return try_to_wake_up(p, TASK_ALL, 0);
2473 EXPORT_SYMBOL(wake_up_process);
2475 int wake_up_state(struct task_struct *p, unsigned int state)
2477 return try_to_wake_up(p, state, 0);
2481 * Perform scheduler related setup for a newly forked process p.
2482 * p is forked by current.
2484 * __sched_fork() is basic setup used by init_idle() too:
2486 static void __sched_fork(struct task_struct *p)
2488 p->se.exec_start = 0;
2489 p->se.sum_exec_runtime = 0;
2490 p->se.prev_sum_exec_runtime = 0;
2491 p->se.nr_migrations = 0;
2492 p->se.last_wakeup = 0;
2493 p->se.avg_overlap = 0;
2494 p->se.start_runtime = 0;
2495 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2496 p->se.avg_running = 0;
2498 #ifdef CONFIG_SCHEDSTATS
2499 p->se.wait_start = 0;
2501 p->se.wait_count = 0;
2504 p->se.sleep_start = 0;
2505 p->se.sleep_max = 0;
2506 p->se.sum_sleep_runtime = 0;
2508 p->se.block_start = 0;
2509 p->se.block_max = 0;
2511 p->se.slice_max = 0;
2513 p->se.nr_migrations_cold = 0;
2514 p->se.nr_failed_migrations_affine = 0;
2515 p->se.nr_failed_migrations_running = 0;
2516 p->se.nr_failed_migrations_hot = 0;
2517 p->se.nr_forced_migrations = 0;
2518 p->se.nr_forced2_migrations = 0;
2520 p->se.nr_wakeups = 0;
2521 p->se.nr_wakeups_sync = 0;
2522 p->se.nr_wakeups_migrate = 0;
2523 p->se.nr_wakeups_local = 0;
2524 p->se.nr_wakeups_remote = 0;
2525 p->se.nr_wakeups_affine = 0;
2526 p->se.nr_wakeups_affine_attempts = 0;
2527 p->se.nr_wakeups_passive = 0;
2528 p->se.nr_wakeups_idle = 0;
2532 INIT_LIST_HEAD(&p->rt.run_list);
2534 INIT_LIST_HEAD(&p->se.group_node);
2536 #ifdef CONFIG_PREEMPT_NOTIFIERS
2537 INIT_HLIST_HEAD(&p->preempt_notifiers);
2541 * We mark the process as running here, but have not actually
2542 * inserted it onto the runqueue yet. This guarantees that
2543 * nobody will actually run it, and a signal or other external
2544 * event cannot wake it up and insert it on the runqueue either.
2546 p->state = TASK_RUNNING;
2550 * fork()/clone()-time setup:
2552 void sched_fork(struct task_struct *p, int clone_flags)
2554 int cpu = get_cpu();
2555 unsigned long flags;
2560 * Revert to default priority/policy on fork if requested.
2562 if (unlikely(p->sched_reset_on_fork)) {
2563 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2564 p->policy = SCHED_NORMAL;
2565 p->normal_prio = p->static_prio;
2568 if (PRIO_TO_NICE(p->static_prio) < 0) {
2569 p->static_prio = NICE_TO_PRIO(0);
2570 p->normal_prio = p->static_prio;
2575 * We don't need the reset flag anymore after the fork. It has
2576 * fulfilled its duty:
2578 p->sched_reset_on_fork = 0;
2582 * Make sure we do not leak PI boosting priority to the child.
2584 p->prio = current->normal_prio;
2586 if (!rt_prio(p->prio))
2587 p->sched_class = &fair_sched_class;
2590 cpu = select_task_rq(p, SD_BALANCE_FORK, 0);
2592 local_irq_save(flags);
2593 update_rq_clock(cpu_rq(cpu));
2594 set_task_cpu(p, cpu);
2595 local_irq_restore(flags);
2597 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2598 if (likely(sched_info_on()))
2599 memset(&p->sched_info, 0, sizeof(p->sched_info));
2601 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2604 #ifdef CONFIG_PREEMPT
2605 /* Want to start with kernel preemption disabled. */
2606 task_thread_info(p)->preempt_count = 1;
2608 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2614 * wake_up_new_task - wake up a newly created task for the first time.
2616 * This function will do some initial scheduler statistics housekeeping
2617 * that must be done for every newly created context, then puts the task
2618 * on the runqueue and wakes it.
2620 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2622 unsigned long flags;
2625 rq = task_rq_lock(p, &flags);
2626 BUG_ON(p->state != TASK_RUNNING);
2627 update_rq_clock(rq);
2629 if (!p->sched_class->task_new || !current->se.on_rq) {
2630 activate_task(rq, p, 0);
2633 * Let the scheduling class do new task startup
2634 * management (if any):
2636 p->sched_class->task_new(rq, p);
2639 trace_sched_wakeup_new(rq, p, 1);
2640 check_preempt_curr(rq, p, WF_FORK);
2642 if (p->sched_class->task_wake_up)
2643 p->sched_class->task_wake_up(rq, p);
2645 task_rq_unlock(rq, &flags);
2648 #ifdef CONFIG_PREEMPT_NOTIFIERS
2651 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2652 * @notifier: notifier struct to register
2654 void preempt_notifier_register(struct preempt_notifier *notifier)
2656 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2658 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2661 * preempt_notifier_unregister - no longer interested in preemption notifications
2662 * @notifier: notifier struct to unregister
2664 * This is safe to call from within a preemption notifier.
2666 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2668 hlist_del(¬ifier->link);
2670 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2672 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2674 struct preempt_notifier *notifier;
2675 struct hlist_node *node;
2677 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2678 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2682 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2683 struct task_struct *next)
2685 struct preempt_notifier *notifier;
2686 struct hlist_node *node;
2688 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2689 notifier->ops->sched_out(notifier, next);
2692 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2694 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2699 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2700 struct task_struct *next)
2704 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2707 * prepare_task_switch - prepare to switch tasks
2708 * @rq: the runqueue preparing to switch
2709 * @prev: the current task that is being switched out
2710 * @next: the task we are going to switch to.
2712 * This is called with the rq lock held and interrupts off. It must
2713 * be paired with a subsequent finish_task_switch after the context
2716 * prepare_task_switch sets up locking and calls architecture specific
2720 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2721 struct task_struct *next)
2723 fire_sched_out_preempt_notifiers(prev, next);
2724 prepare_lock_switch(rq, next);
2725 prepare_arch_switch(next);
2729 * finish_task_switch - clean up after a task-switch
2730 * @rq: runqueue associated with task-switch
2731 * @prev: the thread we just switched away from.
2733 * finish_task_switch must be called after the context switch, paired
2734 * with a prepare_task_switch call before the context switch.
2735 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2736 * and do any other architecture-specific cleanup actions.
2738 * Note that we may have delayed dropping an mm in context_switch(). If
2739 * so, we finish that here outside of the runqueue lock. (Doing it
2740 * with the lock held can cause deadlocks; see schedule() for
2743 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2744 __releases(rq->lock)
2746 struct mm_struct *mm = rq->prev_mm;
2752 * A task struct has one reference for the use as "current".
2753 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2754 * schedule one last time. The schedule call will never return, and
2755 * the scheduled task must drop that reference.
2756 * The test for TASK_DEAD must occur while the runqueue locks are
2757 * still held, otherwise prev could be scheduled on another cpu, die
2758 * there before we look at prev->state, and then the reference would
2760 * Manfred Spraul <manfred@colorfullife.com>
2762 prev_state = prev->state;
2763 finish_arch_switch(prev);
2764 perf_event_task_sched_in(current, cpu_of(rq));
2765 finish_lock_switch(rq, prev);
2767 fire_sched_in_preempt_notifiers(current);
2770 if (unlikely(prev_state == TASK_DEAD)) {
2772 * Remove function-return probe instances associated with this
2773 * task and put them back on the free list.
2775 kprobe_flush_task(prev);
2776 put_task_struct(prev);
2782 /* assumes rq->lock is held */
2783 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2785 if (prev->sched_class->pre_schedule)
2786 prev->sched_class->pre_schedule(rq, prev);
2789 /* rq->lock is NOT held, but preemption is disabled */
2790 static inline void post_schedule(struct rq *rq)
2792 if (rq->post_schedule) {
2793 unsigned long flags;
2795 spin_lock_irqsave(&rq->lock, flags);
2796 if (rq->curr->sched_class->post_schedule)
2797 rq->curr->sched_class->post_schedule(rq);
2798 spin_unlock_irqrestore(&rq->lock, flags);
2800 rq->post_schedule = 0;
2806 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2810 static inline void post_schedule(struct rq *rq)
2817 * schedule_tail - first thing a freshly forked thread must call.
2818 * @prev: the thread we just switched away from.
2820 asmlinkage void schedule_tail(struct task_struct *prev)
2821 __releases(rq->lock)
2823 struct rq *rq = this_rq();
2825 finish_task_switch(rq, prev);
2828 * FIXME: do we need to worry about rq being invalidated by the
2833 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2834 /* In this case, finish_task_switch does not reenable preemption */
2837 if (current->set_child_tid)
2838 put_user(task_pid_vnr(current), current->set_child_tid);
2842 * context_switch - switch to the new MM and the new
2843 * thread's register state.
2846 context_switch(struct rq *rq, struct task_struct *prev,
2847 struct task_struct *next)
2849 struct mm_struct *mm, *oldmm;
2851 prepare_task_switch(rq, prev, next);
2852 trace_sched_switch(rq, prev, next);
2854 oldmm = prev->active_mm;
2856 * For paravirt, this is coupled with an exit in switch_to to
2857 * combine the page table reload and the switch backend into
2860 arch_start_context_switch(prev);
2863 next->active_mm = oldmm;
2864 atomic_inc(&oldmm->mm_count);
2865 enter_lazy_tlb(oldmm, next);
2867 switch_mm(oldmm, mm, next);
2869 if (likely(!prev->mm)) {
2870 prev->active_mm = NULL;
2871 rq->prev_mm = oldmm;
2874 * Since the runqueue lock will be released by the next
2875 * task (which is an invalid locking op but in the case
2876 * of the scheduler it's an obvious special-case), so we
2877 * do an early lockdep release here:
2879 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2880 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2883 /* Here we just switch the register state and the stack. */
2884 switch_to(prev, next, prev);
2888 * this_rq must be evaluated again because prev may have moved
2889 * CPUs since it called schedule(), thus the 'rq' on its stack
2890 * frame will be invalid.
2892 finish_task_switch(this_rq(), prev);
2896 * nr_running, nr_uninterruptible and nr_context_switches:
2898 * externally visible scheduler statistics: current number of runnable
2899 * threads, current number of uninterruptible-sleeping threads, total
2900 * number of context switches performed since bootup.
2902 unsigned long nr_running(void)
2904 unsigned long i, sum = 0;
2906 for_each_online_cpu(i)
2907 sum += cpu_rq(i)->nr_running;
2912 unsigned long nr_uninterruptible(void)
2914 unsigned long i, sum = 0;
2916 for_each_possible_cpu(i)
2917 sum += cpu_rq(i)->nr_uninterruptible;
2920 * Since we read the counters lockless, it might be slightly
2921 * inaccurate. Do not allow it to go below zero though:
2923 if (unlikely((long)sum < 0))
2929 unsigned long long nr_context_switches(void)
2932 unsigned long long sum = 0;
2934 for_each_possible_cpu(i)
2935 sum += cpu_rq(i)->nr_switches;
2940 unsigned long nr_iowait(void)
2942 unsigned long i, sum = 0;
2944 for_each_possible_cpu(i)
2945 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2950 unsigned long nr_iowait_cpu(void)
2952 struct rq *this = this_rq();
2953 return atomic_read(&this->nr_iowait);
2956 unsigned long this_cpu_load(void)
2958 struct rq *this = this_rq();
2959 return this->cpu_load[0];
2963 /* Variables and functions for calc_load */
2964 static atomic_long_t calc_load_tasks;
2965 static unsigned long calc_load_update;
2966 unsigned long avenrun[3];
2967 EXPORT_SYMBOL(avenrun);
2970 * get_avenrun - get the load average array
2971 * @loads: pointer to dest load array
2972 * @offset: offset to add
2973 * @shift: shift count to shift the result left
2975 * These values are estimates at best, so no need for locking.
2977 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2979 loads[0] = (avenrun[0] + offset) << shift;
2980 loads[1] = (avenrun[1] + offset) << shift;
2981 loads[2] = (avenrun[2] + offset) << shift;
2984 static unsigned long
2985 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2988 load += active * (FIXED_1 - exp);
2989 return load >> FSHIFT;
2993 * calc_load - update the avenrun load estimates 10 ticks after the
2994 * CPUs have updated calc_load_tasks.
2996 void calc_global_load(void)
2998 unsigned long upd = calc_load_update + 10;
3001 if (time_before(jiffies, upd))
3004 active = atomic_long_read(&calc_load_tasks);
3005 active = active > 0 ? active * FIXED_1 : 0;
3007 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3008 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3009 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3011 calc_load_update += LOAD_FREQ;
3015 * Either called from update_cpu_load() or from a cpu going idle
3017 static void calc_load_account_active(struct rq *this_rq)
3019 long nr_active, delta;
3021 nr_active = this_rq->nr_running;
3022 nr_active += (long) this_rq->nr_uninterruptible;
3024 if (nr_active != this_rq->calc_load_active) {
3025 delta = nr_active - this_rq->calc_load_active;
3026 this_rq->calc_load_active = nr_active;
3027 atomic_long_add(delta, &calc_load_tasks);
3032 * Update rq->cpu_load[] statistics. This function is usually called every
3033 * scheduler tick (TICK_NSEC).
3035 static void update_cpu_load(struct rq *this_rq)
3037 unsigned long this_load = this_rq->load.weight;
3040 this_rq->nr_load_updates++;
3042 /* Update our load: */
3043 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3044 unsigned long old_load, new_load;
3046 /* scale is effectively 1 << i now, and >> i divides by scale */
3048 old_load = this_rq->cpu_load[i];
3049 new_load = this_load;
3051 * Round up the averaging division if load is increasing. This
3052 * prevents us from getting stuck on 9 if the load is 10, for
3055 if (new_load > old_load)
3056 new_load += scale-1;
3057 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3060 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3061 this_rq->calc_load_update += LOAD_FREQ;
3062 calc_load_account_active(this_rq);
3069 * double_rq_lock - safely lock two runqueues
3071 * Note this does not disable interrupts like task_rq_lock,
3072 * you need to do so manually before calling.
3074 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
3075 __acquires(rq1->lock)
3076 __acquires(rq2->lock)
3078 BUG_ON(!irqs_disabled());
3080 spin_lock(&rq1->lock);
3081 __acquire(rq2->lock); /* Fake it out ;) */
3084 spin_lock(&rq1->lock);
3085 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
3087 spin_lock(&rq2->lock);
3088 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
3091 update_rq_clock(rq1);
3092 update_rq_clock(rq2);
3096 * double_rq_unlock - safely unlock two runqueues
3098 * Note this does not restore interrupts like task_rq_unlock,
3099 * you need to do so manually after calling.
3101 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
3102 __releases(rq1->lock)
3103 __releases(rq2->lock)
3105 spin_unlock(&rq1->lock);
3107 spin_unlock(&rq2->lock);
3109 __release(rq2->lock);
3113 * If dest_cpu is allowed for this process, migrate the task to it.
3114 * This is accomplished by forcing the cpu_allowed mask to only
3115 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3116 * the cpu_allowed mask is restored.
3118 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
3120 struct migration_req req;
3121 unsigned long flags;
3124 rq = task_rq_lock(p, &flags);
3125 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
3126 || unlikely(!cpu_active(dest_cpu)))
3129 /* force the process onto the specified CPU */
3130 if (migrate_task(p, dest_cpu, &req)) {
3131 /* Need to wait for migration thread (might exit: take ref). */
3132 struct task_struct *mt = rq->migration_thread;
3134 get_task_struct(mt);
3135 task_rq_unlock(rq, &flags);
3136 wake_up_process(mt);
3137 put_task_struct(mt);
3138 wait_for_completion(&req.done);
3143 task_rq_unlock(rq, &flags);
3147 * sched_exec - execve() is a valuable balancing opportunity, because at
3148 * this point the task has the smallest effective memory and cache footprint.
3150 void sched_exec(void)
3152 int new_cpu, this_cpu = get_cpu();
3153 new_cpu = select_task_rq(current, SD_BALANCE_EXEC, 0);
3155 if (new_cpu != this_cpu)
3156 sched_migrate_task(current, new_cpu);
3160 * pull_task - move a task from a remote runqueue to the local runqueue.
3161 * Both runqueues must be locked.
3163 static void pull_task(struct rq *src_rq, struct task_struct *p,
3164 struct rq *this_rq, int this_cpu)
3166 deactivate_task(src_rq, p, 0);
3167 set_task_cpu(p, this_cpu);
3168 activate_task(this_rq, p, 0);
3170 * Note that idle threads have a prio of MAX_PRIO, for this test
3171 * to be always true for them.
3173 check_preempt_curr(this_rq, p, 0);
3177 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3180 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3181 struct sched_domain *sd, enum cpu_idle_type idle,
3184 int tsk_cache_hot = 0;
3186 * We do not migrate tasks that are:
3187 * 1) running (obviously), or
3188 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3189 * 3) are cache-hot on their current CPU.
3191 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
3192 schedstat_inc(p, se.nr_failed_migrations_affine);
3197 if (task_running(rq, p)) {
3198 schedstat_inc(p, se.nr_failed_migrations_running);
3203 * Aggressive migration if:
3204 * 1) task is cache cold, or
3205 * 2) too many balance attempts have failed.
3208 tsk_cache_hot = task_hot(p, rq->clock, sd);
3209 if (!tsk_cache_hot ||
3210 sd->nr_balance_failed > sd->cache_nice_tries) {
3211 #ifdef CONFIG_SCHEDSTATS
3212 if (tsk_cache_hot) {
3213 schedstat_inc(sd, lb_hot_gained[idle]);
3214 schedstat_inc(p, se.nr_forced_migrations);
3220 if (tsk_cache_hot) {
3221 schedstat_inc(p, se.nr_failed_migrations_hot);
3227 static unsigned long
3228 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3229 unsigned long max_load_move, struct sched_domain *sd,
3230 enum cpu_idle_type idle, int *all_pinned,
3231 int *this_best_prio, struct rq_iterator *iterator)
3233 int loops = 0, pulled = 0, pinned = 0;
3234 struct task_struct *p;
3235 long rem_load_move = max_load_move;
3237 if (max_load_move == 0)
3243 * Start the load-balancing iterator:
3245 p = iterator->start(iterator->arg);
3247 if (!p || loops++ > sysctl_sched_nr_migrate)
3250 if ((p->se.load.weight >> 1) > rem_load_move ||
3251 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3252 p = iterator->next(iterator->arg);
3256 pull_task(busiest, p, this_rq, this_cpu);
3258 rem_load_move -= p->se.load.weight;
3260 #ifdef CONFIG_PREEMPT
3262 * NEWIDLE balancing is a source of latency, so preemptible kernels
3263 * will stop after the first task is pulled to minimize the critical
3266 if (idle == CPU_NEWLY_IDLE)
3271 * We only want to steal up to the prescribed amount of weighted load.
3273 if (rem_load_move > 0) {
3274 if (p->prio < *this_best_prio)
3275 *this_best_prio = p->prio;
3276 p = iterator->next(iterator->arg);
3281 * Right now, this is one of only two places pull_task() is called,
3282 * so we can safely collect pull_task() stats here rather than
3283 * inside pull_task().
3285 schedstat_add(sd, lb_gained[idle], pulled);
3288 *all_pinned = pinned;
3290 return max_load_move - rem_load_move;
3294 * move_tasks tries to move up to max_load_move weighted load from busiest to
3295 * this_rq, as part of a balancing operation within domain "sd".
3296 * Returns 1 if successful and 0 otherwise.
3298 * Called with both runqueues locked.
3300 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3301 unsigned long max_load_move,
3302 struct sched_domain *sd, enum cpu_idle_type idle,
3305 const struct sched_class *class = sched_class_highest;
3306 unsigned long total_load_moved = 0;
3307 int this_best_prio = this_rq->curr->prio;
3311 class->load_balance(this_rq, this_cpu, busiest,
3312 max_load_move - total_load_moved,
3313 sd, idle, all_pinned, &this_best_prio);
3314 class = class->next;
3316 #ifdef CONFIG_PREEMPT
3318 * NEWIDLE balancing is a source of latency, so preemptible
3319 * kernels will stop after the first task is pulled to minimize
3320 * the critical section.
3322 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3325 } while (class && max_load_move > total_load_moved);
3327 return total_load_moved > 0;
3331 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3332 struct sched_domain *sd, enum cpu_idle_type idle,
3333 struct rq_iterator *iterator)
3335 struct task_struct *p = iterator->start(iterator->arg);
3339 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3340 pull_task(busiest, p, this_rq, this_cpu);
3342 * Right now, this is only the second place pull_task()
3343 * is called, so we can safely collect pull_task()
3344 * stats here rather than inside pull_task().
3346 schedstat_inc(sd, lb_gained[idle]);
3350 p = iterator->next(iterator->arg);
3357 * move_one_task tries to move exactly one task from busiest to this_rq, as
3358 * part of active balancing operations within "domain".
3359 * Returns 1 if successful and 0 otherwise.
3361 * Called with both runqueues locked.
3363 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3364 struct sched_domain *sd, enum cpu_idle_type idle)
3366 const struct sched_class *class;
3368 for_each_class(class) {
3369 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3375 /********** Helpers for find_busiest_group ************************/
3377 * sd_lb_stats - Structure to store the statistics of a sched_domain
3378 * during load balancing.
3380 struct sd_lb_stats {
3381 struct sched_group *busiest; /* Busiest group in this sd */
3382 struct sched_group *this; /* Local group in this sd */
3383 unsigned long total_load; /* Total load of all groups in sd */
3384 unsigned long total_pwr; /* Total power of all groups in sd */
3385 unsigned long avg_load; /* Average load across all groups in sd */
3387 /** Statistics of this group */
3388 unsigned long this_load;
3389 unsigned long this_load_per_task;
3390 unsigned long this_nr_running;
3392 /* Statistics of the busiest group */
3393 unsigned long max_load;
3394 unsigned long busiest_load_per_task;
3395 unsigned long busiest_nr_running;
3397 int group_imb; /* Is there imbalance in this sd */
3398 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3399 int power_savings_balance; /* Is powersave balance needed for this sd */
3400 struct sched_group *group_min; /* Least loaded group in sd */
3401 struct sched_group *group_leader; /* Group which relieves group_min */
3402 unsigned long min_load_per_task; /* load_per_task in group_min */
3403 unsigned long leader_nr_running; /* Nr running of group_leader */
3404 unsigned long min_nr_running; /* Nr running of group_min */
3409 * sg_lb_stats - stats of a sched_group required for load_balancing
3411 struct sg_lb_stats {
3412 unsigned long avg_load; /*Avg load across the CPUs of the group */
3413 unsigned long group_load; /* Total load over the CPUs of the group */
3414 unsigned long sum_nr_running; /* Nr tasks running in the group */
3415 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3416 unsigned long group_capacity;
3417 int group_imb; /* Is there an imbalance in the group ? */
3421 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3422 * @group: The group whose first cpu is to be returned.
3424 static inline unsigned int group_first_cpu(struct sched_group *group)
3426 return cpumask_first(sched_group_cpus(group));
3430 * get_sd_load_idx - Obtain the load index for a given sched domain.
3431 * @sd: The sched_domain whose load_idx is to be obtained.
3432 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3434 static inline int get_sd_load_idx(struct sched_domain *sd,
3435 enum cpu_idle_type idle)
3441 load_idx = sd->busy_idx;
3444 case CPU_NEWLY_IDLE:
3445 load_idx = sd->newidle_idx;
3448 load_idx = sd->idle_idx;
3456 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3458 * init_sd_power_savings_stats - Initialize power savings statistics for
3459 * the given sched_domain, during load balancing.
3461 * @sd: Sched domain whose power-savings statistics are to be initialized.
3462 * @sds: Variable containing the statistics for sd.
3463 * @idle: Idle status of the CPU at which we're performing load-balancing.
3465 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3466 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3469 * Busy processors will not participate in power savings
3472 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3473 sds->power_savings_balance = 0;
3475 sds->power_savings_balance = 1;
3476 sds->min_nr_running = ULONG_MAX;
3477 sds->leader_nr_running = 0;
3482 * update_sd_power_savings_stats - Update the power saving stats for a
3483 * sched_domain while performing load balancing.
3485 * @group: sched_group belonging to the sched_domain under consideration.
3486 * @sds: Variable containing the statistics of the sched_domain
3487 * @local_group: Does group contain the CPU for which we're performing
3489 * @sgs: Variable containing the statistics of the group.
3491 static inline void update_sd_power_savings_stats(struct sched_group *group,
3492 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3495 if (!sds->power_savings_balance)
3499 * If the local group is idle or completely loaded
3500 * no need to do power savings balance at this domain
3502 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3503 !sds->this_nr_running))
3504 sds->power_savings_balance = 0;
3507 * If a group is already running at full capacity or idle,
3508 * don't include that group in power savings calculations
3510 if (!sds->power_savings_balance ||
3511 sgs->sum_nr_running >= sgs->group_capacity ||
3512 !sgs->sum_nr_running)
3516 * Calculate the group which has the least non-idle load.
3517 * This is the group from where we need to pick up the load
3520 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3521 (sgs->sum_nr_running == sds->min_nr_running &&
3522 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3523 sds->group_min = group;
3524 sds->min_nr_running = sgs->sum_nr_running;
3525 sds->min_load_per_task = sgs->sum_weighted_load /
3526 sgs->sum_nr_running;
3530 * Calculate the group which is almost near its
3531 * capacity but still has some space to pick up some load
3532 * from other group and save more power
3534 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
3537 if (sgs->sum_nr_running > sds->leader_nr_running ||
3538 (sgs->sum_nr_running == sds->leader_nr_running &&
3539 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3540 sds->group_leader = group;
3541 sds->leader_nr_running = sgs->sum_nr_running;
3546 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3547 * @sds: Variable containing the statistics of the sched_domain
3548 * under consideration.
3549 * @this_cpu: Cpu at which we're currently performing load-balancing.
3550 * @imbalance: Variable to store the imbalance.
3553 * Check if we have potential to perform some power-savings balance.
3554 * If yes, set the busiest group to be the least loaded group in the
3555 * sched_domain, so that it's CPUs can be put to idle.
3557 * Returns 1 if there is potential to perform power-savings balance.
3560 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3561 int this_cpu, unsigned long *imbalance)
3563 if (!sds->power_savings_balance)
3566 if (sds->this != sds->group_leader ||
3567 sds->group_leader == sds->group_min)
3570 *imbalance = sds->min_load_per_task;
3571 sds->busiest = sds->group_min;
3576 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3577 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3578 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3583 static inline void update_sd_power_savings_stats(struct sched_group *group,
3584 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3589 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3590 int this_cpu, unsigned long *imbalance)
3594 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3597 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
3599 return SCHED_LOAD_SCALE;
3602 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
3604 return default_scale_freq_power(sd, cpu);
3607 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
3609 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3610 unsigned long smt_gain = sd->smt_gain;
3617 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3619 return default_scale_smt_power(sd, cpu);
3622 unsigned long scale_rt_power(int cpu)
3624 struct rq *rq = cpu_rq(cpu);
3625 u64 total, available;
3627 sched_avg_update(rq);
3629 total = sched_avg_period() + (rq->clock - rq->age_stamp);
3630 available = total - rq->rt_avg;
3632 if (unlikely((s64)total < SCHED_LOAD_SCALE))
3633 total = SCHED_LOAD_SCALE;
3635 total >>= SCHED_LOAD_SHIFT;
3637 return div_u64(available, total);
3640 static void update_cpu_power(struct sched_domain *sd, int cpu)
3642 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3643 unsigned long power = SCHED_LOAD_SCALE;
3644 struct sched_group *sdg = sd->groups;
3646 if (sched_feat(ARCH_POWER))
3647 power *= arch_scale_freq_power(sd, cpu);
3649 power *= default_scale_freq_power(sd, cpu);
3651 power >>= SCHED_LOAD_SHIFT;
3653 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
3654 if (sched_feat(ARCH_POWER))
3655 power *= arch_scale_smt_power(sd, cpu);
3657 power *= default_scale_smt_power(sd, cpu);
3659 power >>= SCHED_LOAD_SHIFT;
3662 power *= scale_rt_power(cpu);
3663 power >>= SCHED_LOAD_SHIFT;
3668 sdg->cpu_power = power;
3671 static void update_group_power(struct sched_domain *sd, int cpu)
3673 struct sched_domain *child = sd->child;
3674 struct sched_group *group, *sdg = sd->groups;
3675 unsigned long power;
3678 update_cpu_power(sd, cpu);
3684 group = child->groups;
3686 power += group->cpu_power;
3687 group = group->next;
3688 } while (group != child->groups);
3690 sdg->cpu_power = power;
3694 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3695 * @sd: The sched_domain whose statistics are to be updated.
3696 * @group: sched_group whose statistics are to be updated.
3697 * @this_cpu: Cpu for which load balance is currently performed.
3698 * @idle: Idle status of this_cpu
3699 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3700 * @sd_idle: Idle status of the sched_domain containing group.
3701 * @local_group: Does group contain this_cpu.
3702 * @cpus: Set of cpus considered for load balancing.
3703 * @balance: Should we balance.
3704 * @sgs: variable to hold the statistics for this group.
3706 static inline void update_sg_lb_stats(struct sched_domain *sd,
3707 struct sched_group *group, int this_cpu,
3708 enum cpu_idle_type idle, int load_idx, int *sd_idle,
3709 int local_group, const struct cpumask *cpus,
3710 int *balance, struct sg_lb_stats *sgs)
3712 unsigned long load, max_cpu_load, min_cpu_load;
3714 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3715 unsigned long sum_avg_load_per_task;
3716 unsigned long avg_load_per_task;
3719 balance_cpu = group_first_cpu(group);
3720 if (balance_cpu == this_cpu)
3721 update_group_power(sd, this_cpu);
3724 /* Tally up the load of all CPUs in the group */
3725 sum_avg_load_per_task = avg_load_per_task = 0;
3727 min_cpu_load = ~0UL;
3729 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3730 struct rq *rq = cpu_rq(i);
3732 if (*sd_idle && rq->nr_running)
3735 /* Bias balancing toward cpus of our domain */
3737 if (idle_cpu(i) && !first_idle_cpu) {
3742 load = target_load(i, load_idx);
3744 load = source_load(i, load_idx);
3745 if (load > max_cpu_load)
3746 max_cpu_load = load;
3747 if (min_cpu_load > load)
3748 min_cpu_load = load;
3751 sgs->group_load += load;
3752 sgs->sum_nr_running += rq->nr_running;
3753 sgs->sum_weighted_load += weighted_cpuload(i);
3755 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3759 * First idle cpu or the first cpu(busiest) in this sched group
3760 * is eligible for doing load balancing at this and above
3761 * domains. In the newly idle case, we will allow all the cpu's
3762 * to do the newly idle load balance.
3764 if (idle != CPU_NEWLY_IDLE && local_group &&
3765 balance_cpu != this_cpu && balance) {
3770 /* Adjust by relative CPU power of the group */
3771 sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power;
3775 * Consider the group unbalanced when the imbalance is larger
3776 * than the average weight of two tasks.
3778 * APZ: with cgroup the avg task weight can vary wildly and
3779 * might not be a suitable number - should we keep a
3780 * normalized nr_running number somewhere that negates
3783 avg_load_per_task = (sum_avg_load_per_task * SCHED_LOAD_SCALE) /
3786 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3789 sgs->group_capacity =
3790 DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
3794 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3795 * @sd: sched_domain whose statistics are to be updated.
3796 * @this_cpu: Cpu for which load balance is currently performed.
3797 * @idle: Idle status of this_cpu
3798 * @sd_idle: Idle status of the sched_domain containing group.
3799 * @cpus: Set of cpus considered for load balancing.
3800 * @balance: Should we balance.
3801 * @sds: variable to hold the statistics for this sched_domain.
3803 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3804 enum cpu_idle_type idle, int *sd_idle,
3805 const struct cpumask *cpus, int *balance,
3806 struct sd_lb_stats *sds)
3808 struct sched_domain *child = sd->child;
3809 struct sched_group *group = sd->groups;
3810 struct sg_lb_stats sgs;
3811 int load_idx, prefer_sibling = 0;
3813 if (child && child->flags & SD_PREFER_SIBLING)
3816 init_sd_power_savings_stats(sd, sds, idle);
3817 load_idx = get_sd_load_idx(sd, idle);
3822 local_group = cpumask_test_cpu(this_cpu,
3823 sched_group_cpus(group));
3824 memset(&sgs, 0, sizeof(sgs));
3825 update_sg_lb_stats(sd, group, this_cpu, idle, load_idx, sd_idle,
3826 local_group, cpus, balance, &sgs);
3828 if (local_group && balance && !(*balance))
3831 sds->total_load += sgs.group_load;
3832 sds->total_pwr += group->cpu_power;
3835 * In case the child domain prefers tasks go to siblings
3836 * first, lower the group capacity to one so that we'll try
3837 * and move all the excess tasks away.
3840 sgs.group_capacity = min(sgs.group_capacity, 1UL);
3843 sds->this_load = sgs.avg_load;
3845 sds->this_nr_running = sgs.sum_nr_running;
3846 sds->this_load_per_task = sgs.sum_weighted_load;
3847 } else if (sgs.avg_load > sds->max_load &&
3848 (sgs.sum_nr_running > sgs.group_capacity ||
3850 sds->max_load = sgs.avg_load;
3851 sds->busiest = group;
3852 sds->busiest_nr_running = sgs.sum_nr_running;
3853 sds->busiest_load_per_task = sgs.sum_weighted_load;
3854 sds->group_imb = sgs.group_imb;
3857 update_sd_power_savings_stats(group, sds, local_group, &sgs);
3858 group = group->next;
3859 } while (group != sd->groups);
3863 * fix_small_imbalance - Calculate the minor imbalance that exists
3864 * amongst the groups of a sched_domain, during
3866 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3867 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3868 * @imbalance: Variable to store the imbalance.
3870 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3871 int this_cpu, unsigned long *imbalance)
3873 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3874 unsigned int imbn = 2;
3876 if (sds->this_nr_running) {
3877 sds->this_load_per_task /= sds->this_nr_running;
3878 if (sds->busiest_load_per_task >
3879 sds->this_load_per_task)
3882 sds->this_load_per_task =
3883 cpu_avg_load_per_task(this_cpu);
3885 if (sds->max_load - sds->this_load + sds->busiest_load_per_task >=
3886 sds->busiest_load_per_task * imbn) {
3887 *imbalance = sds->busiest_load_per_task;
3892 * OK, we don't have enough imbalance to justify moving tasks,
3893 * however we may be able to increase total CPU power used by
3897 pwr_now += sds->busiest->cpu_power *
3898 min(sds->busiest_load_per_task, sds->max_load);
3899 pwr_now += sds->this->cpu_power *
3900 min(sds->this_load_per_task, sds->this_load);
3901 pwr_now /= SCHED_LOAD_SCALE;
3903 /* Amount of load we'd subtract */
3904 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3905 sds->busiest->cpu_power;
3906 if (sds->max_load > tmp)
3907 pwr_move += sds->busiest->cpu_power *
3908 min(sds->busiest_load_per_task, sds->max_load - tmp);
3910 /* Amount of load we'd add */
3911 if (sds->max_load * sds->busiest->cpu_power <
3912 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
3913 tmp = (sds->max_load * sds->busiest->cpu_power) /
3914 sds->this->cpu_power;
3916 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3917 sds->this->cpu_power;
3918 pwr_move += sds->this->cpu_power *
3919 min(sds->this_load_per_task, sds->this_load + tmp);
3920 pwr_move /= SCHED_LOAD_SCALE;
3922 /* Move if we gain throughput */
3923 if (pwr_move > pwr_now)
3924 *imbalance = sds->busiest_load_per_task;
3928 * calculate_imbalance - Calculate the amount of imbalance present within the
3929 * groups of a given sched_domain during load balance.
3930 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3931 * @this_cpu: Cpu for which currently load balance is being performed.
3932 * @imbalance: The variable to store the imbalance.
3934 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3935 unsigned long *imbalance)
3937 unsigned long max_pull;
3939 * In the presence of smp nice balancing, certain scenarios can have
3940 * max load less than avg load(as we skip the groups at or below
3941 * its cpu_power, while calculating max_load..)
3943 if (sds->max_load < sds->avg_load) {
3945 return fix_small_imbalance(sds, this_cpu, imbalance);
3948 /* Don't want to pull so many tasks that a group would go idle */
3949 max_pull = min(sds->max_load - sds->avg_load,
3950 sds->max_load - sds->busiest_load_per_task);
3952 /* How much load to actually move to equalise the imbalance */
3953 *imbalance = min(max_pull * sds->busiest->cpu_power,
3954 (sds->avg_load - sds->this_load) * sds->this->cpu_power)
3958 * if *imbalance is less than the average load per runnable task
3959 * there is no gaurantee that any tasks will be moved so we'll have
3960 * a think about bumping its value to force at least one task to be
3963 if (*imbalance < sds->busiest_load_per_task)
3964 return fix_small_imbalance(sds, this_cpu, imbalance);
3967 /******* find_busiest_group() helpers end here *********************/
3970 * find_busiest_group - Returns the busiest group within the sched_domain
3971 * if there is an imbalance. If there isn't an imbalance, and
3972 * the user has opted for power-savings, it returns a group whose
3973 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3974 * such a group exists.
3976 * Also calculates the amount of weighted load which should be moved
3977 * to restore balance.
3979 * @sd: The sched_domain whose busiest group is to be returned.
3980 * @this_cpu: The cpu for which load balancing is currently being performed.
3981 * @imbalance: Variable which stores amount of weighted load which should
3982 * be moved to restore balance/put a group to idle.
3983 * @idle: The idle status of this_cpu.
3984 * @sd_idle: The idleness of sd
3985 * @cpus: The set of CPUs under consideration for load-balancing.
3986 * @balance: Pointer to a variable indicating if this_cpu
3987 * is the appropriate cpu to perform load balancing at this_level.
3989 * Returns: - the busiest group if imbalance exists.
3990 * - If no imbalance and user has opted for power-savings balance,
3991 * return the least loaded group whose CPUs can be
3992 * put to idle by rebalancing its tasks onto our group.
3994 static struct sched_group *
3995 find_busiest_group(struct sched_domain *sd, int this_cpu,
3996 unsigned long *imbalance, enum cpu_idle_type idle,
3997 int *sd_idle, const struct cpumask *cpus, int *balance)
3999 struct sd_lb_stats sds;
4001 memset(&sds, 0, sizeof(sds));
4004 * Compute the various statistics relavent for load balancing at
4007 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
4010 /* Cases where imbalance does not exist from POV of this_cpu */
4011 /* 1) this_cpu is not the appropriate cpu to perform load balancing
4013 * 2) There is no busy sibling group to pull from.
4014 * 3) This group is the busiest group.
4015 * 4) This group is more busy than the avg busieness at this
4017 * 5) The imbalance is within the specified limit.
4018 * 6) Any rebalance would lead to ping-pong
4020 if (balance && !(*balance))
4023 if (!sds.busiest || sds.busiest_nr_running == 0)
4026 if (sds.this_load >= sds.max_load)
4029 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
4031 if (sds.this_load >= sds.avg_load)
4034 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
4037 sds.busiest_load_per_task /= sds.busiest_nr_running;
4039 sds.busiest_load_per_task =
4040 min(sds.busiest_load_per_task, sds.avg_load);
4043 * We're trying to get all the cpus to the average_load, so we don't
4044 * want to push ourselves above the average load, nor do we wish to
4045 * reduce the max loaded cpu below the average load, as either of these
4046 * actions would just result in more rebalancing later, and ping-pong
4047 * tasks around. Thus we look for the minimum possible imbalance.
4048 * Negative imbalances (*we* are more loaded than anyone else) will
4049 * be counted as no imbalance for these purposes -- we can't fix that
4050 * by pulling tasks to us. Be careful of negative numbers as they'll
4051 * appear as very large values with unsigned longs.
4053 if (sds.max_load <= sds.busiest_load_per_task)
4056 /* Looks like there is an imbalance. Compute it */
4057 calculate_imbalance(&sds, this_cpu, imbalance);
4062 * There is no obvious imbalance. But check if we can do some balancing
4065 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
4073 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4076 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
4077 unsigned long imbalance, const struct cpumask *cpus)
4079 struct rq *busiest = NULL, *rq;
4080 unsigned long max_load = 0;
4083 for_each_cpu(i, sched_group_cpus(group)) {
4084 unsigned long power = power_of(i);
4085 unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
4088 if (!cpumask_test_cpu(i, cpus))
4092 wl = weighted_cpuload(i) * SCHED_LOAD_SCALE;
4095 if (capacity && rq->nr_running == 1 && wl > imbalance)
4098 if (wl > max_load) {
4108 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4109 * so long as it is large enough.
4111 #define MAX_PINNED_INTERVAL 512
4113 /* Working cpumask for load_balance and load_balance_newidle. */
4114 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4117 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4118 * tasks if there is an imbalance.
4120 static int load_balance(int this_cpu, struct rq *this_rq,
4121 struct sched_domain *sd, enum cpu_idle_type idle,
4124 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
4125 struct sched_group *group;
4126 unsigned long imbalance;
4128 unsigned long flags;
4129 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4131 cpumask_copy(cpus, cpu_active_mask);
4134 * When power savings policy is enabled for the parent domain, idle
4135 * sibling can pick up load irrespective of busy siblings. In this case,
4136 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4137 * portraying it as CPU_NOT_IDLE.
4139 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
4140 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4143 schedstat_inc(sd, lb_count[idle]);
4147 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
4154 schedstat_inc(sd, lb_nobusyg[idle]);
4158 busiest = find_busiest_queue(group, idle, imbalance, cpus);
4160 schedstat_inc(sd, lb_nobusyq[idle]);
4164 BUG_ON(busiest == this_rq);
4166 schedstat_add(sd, lb_imbalance[idle], imbalance);
4169 if (busiest->nr_running > 1) {
4171 * Attempt to move tasks. If find_busiest_group has found
4172 * an imbalance but busiest->nr_running <= 1, the group is
4173 * still unbalanced. ld_moved simply stays zero, so it is
4174 * correctly treated as an imbalance.
4176 local_irq_save(flags);
4177 double_rq_lock(this_rq, busiest);
4178 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4179 imbalance, sd, idle, &all_pinned);
4180 double_rq_unlock(this_rq, busiest);
4181 local_irq_restore(flags);
4184 * some other cpu did the load balance for us.
4186 if (ld_moved && this_cpu != smp_processor_id())
4187 resched_cpu(this_cpu);
4189 /* All tasks on this runqueue were pinned by CPU affinity */
4190 if (unlikely(all_pinned)) {
4191 cpumask_clear_cpu(cpu_of(busiest), cpus);
4192 if (!cpumask_empty(cpus))
4199 schedstat_inc(sd, lb_failed[idle]);
4200 sd->nr_balance_failed++;
4202 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
4204 spin_lock_irqsave(&busiest->lock, flags);
4206 /* don't kick the migration_thread, if the curr
4207 * task on busiest cpu can't be moved to this_cpu
4209 if (!cpumask_test_cpu(this_cpu,
4210 &busiest->curr->cpus_allowed)) {
4211 spin_unlock_irqrestore(&busiest->lock, flags);
4213 goto out_one_pinned;
4216 if (!busiest->active_balance) {
4217 busiest->active_balance = 1;
4218 busiest->push_cpu = this_cpu;
4221 spin_unlock_irqrestore(&busiest->lock, flags);
4223 wake_up_process(busiest->migration_thread);
4226 * We've kicked active balancing, reset the failure
4229 sd->nr_balance_failed = sd->cache_nice_tries+1;
4232 sd->nr_balance_failed = 0;
4234 if (likely(!active_balance)) {
4235 /* We were unbalanced, so reset the balancing interval */
4236 sd->balance_interval = sd->min_interval;
4239 * If we've begun active balancing, start to back off. This
4240 * case may not be covered by the all_pinned logic if there
4241 * is only 1 task on the busy runqueue (because we don't call
4244 if (sd->balance_interval < sd->max_interval)
4245 sd->balance_interval *= 2;
4248 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4249 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4255 schedstat_inc(sd, lb_balanced[idle]);
4257 sd->nr_balance_failed = 0;
4260 /* tune up the balancing interval */
4261 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4262 (sd->balance_interval < sd->max_interval))
4263 sd->balance_interval *= 2;
4265 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4266 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4277 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4278 * tasks if there is an imbalance.
4280 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4281 * this_rq is locked.
4284 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
4286 struct sched_group *group;
4287 struct rq *busiest = NULL;
4288 unsigned long imbalance;
4292 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4294 cpumask_copy(cpus, cpu_active_mask);
4297 * When power savings policy is enabled for the parent domain, idle
4298 * sibling can pick up load irrespective of busy siblings. In this case,
4299 * let the state of idle sibling percolate up as IDLE, instead of
4300 * portraying it as CPU_NOT_IDLE.
4302 if (sd->flags & SD_SHARE_CPUPOWER &&
4303 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4306 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
4308 update_shares_locked(this_rq, sd);
4309 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
4310 &sd_idle, cpus, NULL);
4312 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
4316 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
4318 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
4322 BUG_ON(busiest == this_rq);
4324 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
4327 if (busiest->nr_running > 1) {
4328 /* Attempt to move tasks */
4329 double_lock_balance(this_rq, busiest);
4330 /* this_rq->clock is already updated */
4331 update_rq_clock(busiest);
4332 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4333 imbalance, sd, CPU_NEWLY_IDLE,
4335 double_unlock_balance(this_rq, busiest);
4337 if (unlikely(all_pinned)) {
4338 cpumask_clear_cpu(cpu_of(busiest), cpus);
4339 if (!cpumask_empty(cpus))
4345 int active_balance = 0;
4347 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
4348 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4349 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4352 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4355 if (sd->nr_balance_failed++ < 2)
4359 * The only task running in a non-idle cpu can be moved to this
4360 * cpu in an attempt to completely freeup the other CPU
4361 * package. The same method used to move task in load_balance()
4362 * have been extended for load_balance_newidle() to speedup
4363 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4365 * The package power saving logic comes from
4366 * find_busiest_group(). If there are no imbalance, then
4367 * f_b_g() will return NULL. However when sched_mc={1,2} then
4368 * f_b_g() will select a group from which a running task may be
4369 * pulled to this cpu in order to make the other package idle.
4370 * If there is no opportunity to make a package idle and if
4371 * there are no imbalance, then f_b_g() will return NULL and no
4372 * action will be taken in load_balance_newidle().
4374 * Under normal task pull operation due to imbalance, there
4375 * will be more than one task in the source run queue and
4376 * move_tasks() will succeed. ld_moved will be true and this
4377 * active balance code will not be triggered.
4380 /* Lock busiest in correct order while this_rq is held */
4381 double_lock_balance(this_rq, busiest);
4384 * don't kick the migration_thread, if the curr
4385 * task on busiest cpu can't be moved to this_cpu
4387 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
4388 double_unlock_balance(this_rq, busiest);
4393 if (!busiest->active_balance) {
4394 busiest->active_balance = 1;
4395 busiest->push_cpu = this_cpu;
4399 double_unlock_balance(this_rq, busiest);
4401 * Should not call ttwu while holding a rq->lock
4403 spin_unlock(&this_rq->lock);
4405 wake_up_process(busiest->migration_thread);
4406 spin_lock(&this_rq->lock);
4409 sd->nr_balance_failed = 0;
4411 update_shares_locked(this_rq, sd);
4415 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
4416 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4417 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4419 sd->nr_balance_failed = 0;
4425 * idle_balance is called by schedule() if this_cpu is about to become
4426 * idle. Attempts to pull tasks from other CPUs.
4428 static void idle_balance(int this_cpu, struct rq *this_rq)
4430 struct sched_domain *sd;
4431 int pulled_task = 0;
4432 unsigned long next_balance = jiffies + HZ;
4434 this_rq->idle_stamp = this_rq->clock;
4436 if (this_rq->avg_idle < sysctl_sched_migration_cost)
4439 for_each_domain(this_cpu, sd) {
4440 unsigned long interval;
4442 if (!(sd->flags & SD_LOAD_BALANCE))
4445 if (sd->flags & SD_BALANCE_NEWIDLE)
4446 /* If we've pulled tasks over stop searching: */
4447 pulled_task = load_balance_newidle(this_cpu, this_rq,
4450 interval = msecs_to_jiffies(sd->balance_interval);
4451 if (time_after(next_balance, sd->last_balance + interval))
4452 next_balance = sd->last_balance + interval;
4454 this_rq->idle_stamp = 0;
4458 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4460 * We are going idle. next_balance may be set based on
4461 * a busy processor. So reset next_balance.
4463 this_rq->next_balance = next_balance;
4468 * active_load_balance is run by migration threads. It pushes running tasks
4469 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4470 * running on each physical CPU where possible, and avoids physical /
4471 * logical imbalances.
4473 * Called with busiest_rq locked.
4475 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
4477 int target_cpu = busiest_rq->push_cpu;
4478 struct sched_domain *sd;
4479 struct rq *target_rq;
4481 /* Is there any task to move? */
4482 if (busiest_rq->nr_running <= 1)
4485 target_rq = cpu_rq(target_cpu);
4488 * This condition is "impossible", if it occurs
4489 * we need to fix it. Originally reported by
4490 * Bjorn Helgaas on a 128-cpu setup.
4492 BUG_ON(busiest_rq == target_rq);
4494 /* move a task from busiest_rq to target_rq */
4495 double_lock_balance(busiest_rq, target_rq);
4496 update_rq_clock(busiest_rq);
4497 update_rq_clock(target_rq);
4499 /* Search for an sd spanning us and the target CPU. */
4500 for_each_domain(target_cpu, sd) {
4501 if ((sd->flags & SD_LOAD_BALANCE) &&
4502 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4507 schedstat_inc(sd, alb_count);
4509 if (move_one_task(target_rq, target_cpu, busiest_rq,
4511 schedstat_inc(sd, alb_pushed);
4513 schedstat_inc(sd, alb_failed);
4515 double_unlock_balance(busiest_rq, target_rq);
4520 atomic_t load_balancer;
4521 cpumask_var_t cpu_mask;
4522 cpumask_var_t ilb_grp_nohz_mask;
4523 } nohz ____cacheline_aligned = {
4524 .load_balancer = ATOMIC_INIT(-1),
4527 int get_nohz_load_balancer(void)
4529 return atomic_read(&nohz.load_balancer);
4532 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4534 * lowest_flag_domain - Return lowest sched_domain containing flag.
4535 * @cpu: The cpu whose lowest level of sched domain is to
4537 * @flag: The flag to check for the lowest sched_domain
4538 * for the given cpu.
4540 * Returns the lowest sched_domain of a cpu which contains the given flag.
4542 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4544 struct sched_domain *sd;
4546 for_each_domain(cpu, sd)
4547 if (sd && (sd->flags & flag))
4554 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4555 * @cpu: The cpu whose domains we're iterating over.
4556 * @sd: variable holding the value of the power_savings_sd
4558 * @flag: The flag to filter the sched_domains to be iterated.
4560 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4561 * set, starting from the lowest sched_domain to the highest.
4563 #define for_each_flag_domain(cpu, sd, flag) \
4564 for (sd = lowest_flag_domain(cpu, flag); \
4565 (sd && (sd->flags & flag)); sd = sd->parent)
4568 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4569 * @ilb_group: group to be checked for semi-idleness
4571 * Returns: 1 if the group is semi-idle. 0 otherwise.
4573 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4574 * and atleast one non-idle CPU. This helper function checks if the given
4575 * sched_group is semi-idle or not.
4577 static inline int is_semi_idle_group(struct sched_group *ilb_group)
4579 cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask,
4580 sched_group_cpus(ilb_group));
4583 * A sched_group is semi-idle when it has atleast one busy cpu
4584 * and atleast one idle cpu.
4586 if (cpumask_empty(nohz.ilb_grp_nohz_mask))
4589 if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group)))
4595 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4596 * @cpu: The cpu which is nominating a new idle_load_balancer.
4598 * Returns: Returns the id of the idle load balancer if it exists,
4599 * Else, returns >= nr_cpu_ids.
4601 * This algorithm picks the idle load balancer such that it belongs to a
4602 * semi-idle powersavings sched_domain. The idea is to try and avoid
4603 * completely idle packages/cores just for the purpose of idle load balancing
4604 * when there are other idle cpu's which are better suited for that job.
4606 static int find_new_ilb(int cpu)
4608 struct sched_domain *sd;
4609 struct sched_group *ilb_group;
4612 * Have idle load balancer selection from semi-idle packages only
4613 * when power-aware load balancing is enabled
4615 if (!(sched_smt_power_savings || sched_mc_power_savings))
4619 * Optimize for the case when we have no idle CPUs or only one
4620 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4622 if (cpumask_weight(nohz.cpu_mask) < 2)
4625 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4626 ilb_group = sd->groups;
4629 if (is_semi_idle_group(ilb_group))
4630 return cpumask_first(nohz.ilb_grp_nohz_mask);
4632 ilb_group = ilb_group->next;
4634 } while (ilb_group != sd->groups);
4638 return cpumask_first(nohz.cpu_mask);
4640 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4641 static inline int find_new_ilb(int call_cpu)
4643 return cpumask_first(nohz.cpu_mask);
4648 * This routine will try to nominate the ilb (idle load balancing)
4649 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4650 * load balancing on behalf of all those cpus. If all the cpus in the system
4651 * go into this tickless mode, then there will be no ilb owner (as there is
4652 * no need for one) and all the cpus will sleep till the next wakeup event
4655 * For the ilb owner, tick is not stopped. And this tick will be used
4656 * for idle load balancing. ilb owner will still be part of
4659 * While stopping the tick, this cpu will become the ilb owner if there
4660 * is no other owner. And will be the owner till that cpu becomes busy
4661 * or if all cpus in the system stop their ticks at which point
4662 * there is no need for ilb owner.
4664 * When the ilb owner becomes busy, it nominates another owner, during the
4665 * next busy scheduler_tick()
4667 int select_nohz_load_balancer(int stop_tick)
4669 int cpu = smp_processor_id();
4672 cpu_rq(cpu)->in_nohz_recently = 1;
4674 if (!cpu_active(cpu)) {
4675 if (atomic_read(&nohz.load_balancer) != cpu)
4679 * If we are going offline and still the leader,
4682 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4688 cpumask_set_cpu(cpu, nohz.cpu_mask);
4690 /* time for ilb owner also to sleep */
4691 if (cpumask_weight(nohz.cpu_mask) == num_active_cpus()) {
4692 if (atomic_read(&nohz.load_balancer) == cpu)
4693 atomic_set(&nohz.load_balancer, -1);
4697 if (atomic_read(&nohz.load_balancer) == -1) {
4698 /* make me the ilb owner */
4699 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4701 } else if (atomic_read(&nohz.load_balancer) == cpu) {
4704 if (!(sched_smt_power_savings ||
4705 sched_mc_power_savings))
4708 * Check to see if there is a more power-efficient
4711 new_ilb = find_new_ilb(cpu);
4712 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4713 atomic_set(&nohz.load_balancer, -1);
4714 resched_cpu(new_ilb);
4720 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4723 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4725 if (atomic_read(&nohz.load_balancer) == cpu)
4726 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4733 static DEFINE_SPINLOCK(balancing);
4736 * It checks each scheduling domain to see if it is due to be balanced,
4737 * and initiates a balancing operation if so.
4739 * Balancing parameters are set up in arch_init_sched_domains.
4741 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4744 struct rq *rq = cpu_rq(cpu);
4745 unsigned long interval;
4746 struct sched_domain *sd;
4747 /* Earliest time when we have to do rebalance again */
4748 unsigned long next_balance = jiffies + 60*HZ;
4749 int update_next_balance = 0;
4752 for_each_domain(cpu, sd) {
4753 if (!(sd->flags & SD_LOAD_BALANCE))
4756 interval = sd->balance_interval;
4757 if (idle != CPU_IDLE)
4758 interval *= sd->busy_factor;
4760 /* scale ms to jiffies */
4761 interval = msecs_to_jiffies(interval);
4762 if (unlikely(!interval))
4764 if (interval > HZ*NR_CPUS/10)
4765 interval = HZ*NR_CPUS/10;
4767 need_serialize = sd->flags & SD_SERIALIZE;
4769 if (need_serialize) {
4770 if (!spin_trylock(&balancing))
4774 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4775 if (load_balance(cpu, rq, sd, idle, &balance)) {
4777 * We've pulled tasks over so either we're no
4778 * longer idle, or one of our SMT siblings is
4781 idle = CPU_NOT_IDLE;
4783 sd->last_balance = jiffies;
4786 spin_unlock(&balancing);
4788 if (time_after(next_balance, sd->last_balance + interval)) {
4789 next_balance = sd->last_balance + interval;
4790 update_next_balance = 1;
4794 * Stop the load balance at this level. There is another
4795 * CPU in our sched group which is doing load balancing more
4803 * next_balance will be updated only when there is a need.
4804 * When the cpu is attached to null domain for ex, it will not be
4807 if (likely(update_next_balance))
4808 rq->next_balance = next_balance;
4812 * run_rebalance_domains is triggered when needed from the scheduler tick.
4813 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4814 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4816 static void run_rebalance_domains(struct softirq_action *h)
4818 int this_cpu = smp_processor_id();
4819 struct rq *this_rq = cpu_rq(this_cpu);
4820 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4821 CPU_IDLE : CPU_NOT_IDLE;
4823 rebalance_domains(this_cpu, idle);
4827 * If this cpu is the owner for idle load balancing, then do the
4828 * balancing on behalf of the other idle cpus whose ticks are
4831 if (this_rq->idle_at_tick &&
4832 atomic_read(&nohz.load_balancer) == this_cpu) {
4836 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4837 if (balance_cpu == this_cpu)
4841 * If this cpu gets work to do, stop the load balancing
4842 * work being done for other cpus. Next load
4843 * balancing owner will pick it up.
4848 rebalance_domains(balance_cpu, CPU_IDLE);
4850 rq = cpu_rq(balance_cpu);
4851 if (time_after(this_rq->next_balance, rq->next_balance))
4852 this_rq->next_balance = rq->next_balance;
4858 static inline int on_null_domain(int cpu)
4860 return !rcu_dereference(cpu_rq(cpu)->sd);
4864 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4866 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4867 * idle load balancing owner or decide to stop the periodic load balancing,
4868 * if the whole system is idle.
4870 static inline void trigger_load_balance(struct rq *rq, int cpu)
4874 * If we were in the nohz mode recently and busy at the current
4875 * scheduler tick, then check if we need to nominate new idle
4878 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4879 rq->in_nohz_recently = 0;
4881 if (atomic_read(&nohz.load_balancer) == cpu) {
4882 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4883 atomic_set(&nohz.load_balancer, -1);
4886 if (atomic_read(&nohz.load_balancer) == -1) {
4887 int ilb = find_new_ilb(cpu);
4889 if (ilb < nr_cpu_ids)
4895 * If this cpu is idle and doing idle load balancing for all the
4896 * cpus with ticks stopped, is it time for that to stop?
4898 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4899 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4905 * If this cpu is idle and the idle load balancing is done by
4906 * someone else, then no need raise the SCHED_SOFTIRQ
4908 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4909 cpumask_test_cpu(cpu, nohz.cpu_mask))
4912 /* Don't need to rebalance while attached to NULL domain */
4913 if (time_after_eq(jiffies, rq->next_balance) &&
4914 likely(!on_null_domain(cpu)))
4915 raise_softirq(SCHED_SOFTIRQ);
4918 #else /* CONFIG_SMP */
4921 * on UP we do not need to balance between CPUs:
4923 static inline void idle_balance(int cpu, struct rq *rq)
4929 DEFINE_PER_CPU(struct kernel_stat, kstat);
4931 EXPORT_PER_CPU_SYMBOL(kstat);
4934 * Return any ns on the sched_clock that have not yet been accounted in
4935 * @p in case that task is currently running.
4937 * Called with task_rq_lock() held on @rq.
4939 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
4943 if (task_current(rq, p)) {
4944 update_rq_clock(rq);
4945 ns = rq->clock - p->se.exec_start;
4953 unsigned long long task_delta_exec(struct task_struct *p)
4955 unsigned long flags;
4959 rq = task_rq_lock(p, &flags);
4960 ns = do_task_delta_exec(p, rq);
4961 task_rq_unlock(rq, &flags);
4967 * Return accounted runtime for the task.
4968 * In case the task is currently running, return the runtime plus current's
4969 * pending runtime that have not been accounted yet.
4971 unsigned long long task_sched_runtime(struct task_struct *p)
4973 unsigned long flags;
4977 rq = task_rq_lock(p, &flags);
4978 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
4979 task_rq_unlock(rq, &flags);
4985 * Return sum_exec_runtime for the thread group.
4986 * In case the task is currently running, return the sum plus current's
4987 * pending runtime that have not been accounted yet.
4989 * Note that the thread group might have other running tasks as well,
4990 * so the return value not includes other pending runtime that other
4991 * running tasks might have.
4993 unsigned long long thread_group_sched_runtime(struct task_struct *p)
4995 struct task_cputime totals;
4996 unsigned long flags;
5000 rq = task_rq_lock(p, &flags);
5001 thread_group_cputime(p, &totals);
5002 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
5003 task_rq_unlock(rq, &flags);
5009 * Account user cpu time to a process.
5010 * @p: the process that the cpu time gets accounted to
5011 * @cputime: the cpu time spent in user space since the last update
5012 * @cputime_scaled: cputime scaled by cpu frequency
5014 void account_user_time(struct task_struct *p, cputime_t cputime,
5015 cputime_t cputime_scaled)
5017 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5020 /* Add user time to process. */
5021 p->utime = cputime_add(p->utime, cputime);
5022 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5023 account_group_user_time(p, cputime);
5025 /* Add user time to cpustat. */
5026 tmp = cputime_to_cputime64(cputime);
5027 if (TASK_NICE(p) > 0)
5028 cpustat->nice = cputime64_add(cpustat->nice, tmp);
5030 cpustat->user = cputime64_add(cpustat->user, tmp);
5032 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
5033 /* Account for user time used */
5034 acct_update_integrals(p);
5038 * Account guest cpu time to a process.
5039 * @p: the process that the cpu time gets accounted to
5040 * @cputime: the cpu time spent in virtual machine since the last update
5041 * @cputime_scaled: cputime scaled by cpu frequency
5043 static void account_guest_time(struct task_struct *p, cputime_t cputime,
5044 cputime_t cputime_scaled)
5047 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5049 tmp = cputime_to_cputime64(cputime);
5051 /* Add guest time to process. */
5052 p->utime = cputime_add(p->utime, cputime);
5053 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5054 account_group_user_time(p, cputime);
5055 p->gtime = cputime_add(p->gtime, cputime);
5057 /* Add guest time to cpustat. */
5058 if (TASK_NICE(p) > 0) {
5059 cpustat->nice = cputime64_add(cpustat->nice, tmp);
5060 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
5062 cpustat->user = cputime64_add(cpustat->user, tmp);
5063 cpustat->guest = cputime64_add(cpustat->guest, tmp);
5068 * Account system cpu time to a process.
5069 * @p: the process that the cpu time gets accounted to
5070 * @hardirq_offset: the offset to subtract from hardirq_count()
5071 * @cputime: the cpu time spent in kernel space since the last update
5072 * @cputime_scaled: cputime scaled by cpu frequency
5074 void account_system_time(struct task_struct *p, int hardirq_offset,
5075 cputime_t cputime, cputime_t cputime_scaled)
5077 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5080 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
5081 account_guest_time(p, cputime, cputime_scaled);
5085 /* Add system time to process. */
5086 p->stime = cputime_add(p->stime, cputime);
5087 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
5088 account_group_system_time(p, cputime);
5090 /* Add system time to cpustat. */
5091 tmp = cputime_to_cputime64(cputime);
5092 if (hardirq_count() - hardirq_offset)
5093 cpustat->irq = cputime64_add(cpustat->irq, tmp);
5094 else if (softirq_count())
5095 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
5097 cpustat->system = cputime64_add(cpustat->system, tmp);
5099 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
5101 /* Account for system time used */
5102 acct_update_integrals(p);
5106 * Account for involuntary wait time.
5107 * @steal: the cpu time spent in involuntary wait
5109 void account_steal_time(cputime_t cputime)
5111 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5112 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5114 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
5118 * Account for idle time.
5119 * @cputime: the cpu time spent in idle wait
5121 void account_idle_time(cputime_t cputime)
5123 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5124 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5125 struct rq *rq = this_rq();
5127 if (atomic_read(&rq->nr_iowait) > 0)
5128 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
5130 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
5133 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
5136 * Account a single tick of cpu time.
5137 * @p: the process that the cpu time gets accounted to
5138 * @user_tick: indicates if the tick is a user or a system tick
5140 void account_process_tick(struct task_struct *p, int user_tick)
5142 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
5143 struct rq *rq = this_rq();
5146 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
5147 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
5148 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
5151 account_idle_time(cputime_one_jiffy);
5155 * Account multiple ticks of steal time.
5156 * @p: the process from which the cpu time has been stolen
5157 * @ticks: number of stolen ticks
5159 void account_steal_ticks(unsigned long ticks)
5161 account_steal_time(jiffies_to_cputime(ticks));
5165 * Account multiple ticks of idle time.
5166 * @ticks: number of stolen ticks
5168 void account_idle_ticks(unsigned long ticks)
5170 account_idle_time(jiffies_to_cputime(ticks));
5176 * Use precise platform statistics if available:
5178 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
5179 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5185 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5187 struct task_cputime cputime;
5189 thread_group_cputime(p, &cputime);
5191 *ut = cputime.utime;
5192 *st = cputime.stime;
5196 #ifndef nsecs_to_cputime
5197 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
5200 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5202 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
5205 * Use CFS's precise accounting:
5207 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
5212 temp = (u64)(rtime * utime);
5213 do_div(temp, total);
5214 utime = (cputime_t)temp;
5219 * Compare with previous values, to keep monotonicity:
5221 p->prev_utime = max(p->prev_utime, utime);
5222 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
5224 *ut = p->prev_utime;
5225 *st = p->prev_stime;
5229 * Must be called with siglock held.
5231 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5233 struct signal_struct *sig = p->signal;
5234 struct task_cputime cputime;
5235 cputime_t rtime, utime, total;
5237 thread_group_cputime(p, &cputime);
5239 total = cputime_add(cputime.utime, cputime.stime);
5240 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
5245 temp = (u64)(rtime * cputime.utime);
5246 do_div(temp, total);
5247 utime = (cputime_t)temp;
5251 sig->prev_utime = max(sig->prev_utime, utime);
5252 sig->prev_stime = max(sig->prev_stime,
5253 cputime_sub(rtime, sig->prev_utime));
5255 *ut = sig->prev_utime;
5256 *st = sig->prev_stime;
5261 * This function gets called by the timer code, with HZ frequency.
5262 * We call it with interrupts disabled.
5264 * It also gets called by the fork code, when changing the parent's
5267 void scheduler_tick(void)
5269 int cpu = smp_processor_id();
5270 struct rq *rq = cpu_rq(cpu);
5271 struct task_struct *curr = rq->curr;
5275 spin_lock(&rq->lock);
5276 update_rq_clock(rq);
5277 update_cpu_load(rq);
5278 curr->sched_class->task_tick(rq, curr, 0);
5279 spin_unlock(&rq->lock);
5281 perf_event_task_tick(curr, cpu);
5284 rq->idle_at_tick = idle_cpu(cpu);
5285 trigger_load_balance(rq, cpu);
5289 notrace unsigned long get_parent_ip(unsigned long addr)
5291 if (in_lock_functions(addr)) {
5292 addr = CALLER_ADDR2;
5293 if (in_lock_functions(addr))
5294 addr = CALLER_ADDR3;
5299 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5300 defined(CONFIG_PREEMPT_TRACER))
5302 void __kprobes add_preempt_count(int val)
5304 #ifdef CONFIG_DEBUG_PREEMPT
5308 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5311 preempt_count() += val;
5312 #ifdef CONFIG_DEBUG_PREEMPT
5314 * Spinlock count overflowing soon?
5316 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5319 if (preempt_count() == val)
5320 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5322 EXPORT_SYMBOL(add_preempt_count);
5324 void __kprobes sub_preempt_count(int val)
5326 #ifdef CONFIG_DEBUG_PREEMPT
5330 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5333 * Is the spinlock portion underflowing?
5335 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5336 !(preempt_count() & PREEMPT_MASK)))
5340 if (preempt_count() == val)
5341 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5342 preempt_count() -= val;
5344 EXPORT_SYMBOL(sub_preempt_count);
5349 * Print scheduling while atomic bug:
5351 static noinline void __schedule_bug(struct task_struct *prev)
5353 struct pt_regs *regs = get_irq_regs();
5355 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5356 prev->comm, prev->pid, preempt_count());
5358 debug_show_held_locks(prev);
5360 if (irqs_disabled())
5361 print_irqtrace_events(prev);
5370 * Various schedule()-time debugging checks and statistics:
5372 static inline void schedule_debug(struct task_struct *prev)
5375 * Test if we are atomic. Since do_exit() needs to call into
5376 * schedule() atomically, we ignore that path for now.
5377 * Otherwise, whine if we are scheduling when we should not be.
5379 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
5380 __schedule_bug(prev);
5382 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5384 schedstat_inc(this_rq(), sched_count);
5385 #ifdef CONFIG_SCHEDSTATS
5386 if (unlikely(prev->lock_depth >= 0)) {
5387 schedstat_inc(this_rq(), bkl_count);
5388 schedstat_inc(prev, sched_info.bkl_count);
5393 static void put_prev_task(struct rq *rq, struct task_struct *p)
5395 u64 runtime = p->se.sum_exec_runtime - p->se.prev_sum_exec_runtime;
5397 update_avg(&p->se.avg_running, runtime);
5399 if (p->state == TASK_RUNNING) {
5401 * In order to avoid avg_overlap growing stale when we are
5402 * indeed overlapping and hence not getting put to sleep, grow
5403 * the avg_overlap on preemption.
5405 * We use the average preemption runtime because that
5406 * correlates to the amount of cache footprint a task can
5409 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
5410 update_avg(&p->se.avg_overlap, runtime);
5412 update_avg(&p->se.avg_running, 0);
5414 p->sched_class->put_prev_task(rq, p);
5418 * Pick up the highest-prio task:
5420 static inline struct task_struct *
5421 pick_next_task(struct rq *rq)
5423 const struct sched_class *class;
5424 struct task_struct *p;
5427 * Optimization: we know that if all tasks are in
5428 * the fair class we can call that function directly:
5430 if (likely(rq->nr_running == rq->cfs.nr_running)) {
5431 p = fair_sched_class.pick_next_task(rq);
5436 class = sched_class_highest;
5438 p = class->pick_next_task(rq);
5442 * Will never be NULL as the idle class always
5443 * returns a non-NULL p:
5445 class = class->next;
5450 * schedule() is the main scheduler function.
5452 asmlinkage void __sched schedule(void)
5454 struct task_struct *prev, *next;
5455 unsigned long *switch_count;
5461 cpu = smp_processor_id();
5465 switch_count = &prev->nivcsw;
5467 release_kernel_lock(prev);
5468 need_resched_nonpreemptible:
5470 schedule_debug(prev);
5472 if (sched_feat(HRTICK))
5475 spin_lock_irq(&rq->lock);
5476 update_rq_clock(rq);
5477 clear_tsk_need_resched(prev);
5479 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
5480 if (unlikely(signal_pending_state(prev->state, prev)))
5481 prev->state = TASK_RUNNING;
5483 deactivate_task(rq, prev, 1);
5484 switch_count = &prev->nvcsw;
5487 pre_schedule(rq, prev);
5489 if (unlikely(!rq->nr_running))
5490 idle_balance(cpu, rq);
5492 put_prev_task(rq, prev);
5493 next = pick_next_task(rq);
5495 if (likely(prev != next)) {
5496 sched_info_switch(prev, next);
5497 perf_event_task_sched_out(prev, next, cpu);
5503 context_switch(rq, prev, next); /* unlocks the rq */
5505 * the context switch might have flipped the stack from under
5506 * us, hence refresh the local variables.
5508 cpu = smp_processor_id();
5511 spin_unlock_irq(&rq->lock);
5515 if (unlikely(reacquire_kernel_lock(current) < 0))
5516 goto need_resched_nonpreemptible;
5518 preempt_enable_no_resched();
5522 EXPORT_SYMBOL(schedule);
5524 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
5526 * Look out! "owner" is an entirely speculative pointer
5527 * access and not reliable.
5529 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5534 if (!sched_feat(OWNER_SPIN))
5537 #ifdef CONFIG_DEBUG_PAGEALLOC
5539 * Need to access the cpu field knowing that
5540 * DEBUG_PAGEALLOC could have unmapped it if
5541 * the mutex owner just released it and exited.
5543 if (probe_kernel_address(&owner->cpu, cpu))
5550 * Even if the access succeeded (likely case),
5551 * the cpu field may no longer be valid.
5553 if (cpu >= nr_cpumask_bits)
5557 * We need to validate that we can do a
5558 * get_cpu() and that we have the percpu area.
5560 if (!cpu_online(cpu))
5567 * Owner changed, break to re-assess state.
5569 if (lock->owner != owner)
5573 * Is that owner really running on that cpu?
5575 if (task_thread_info(rq->curr) != owner || need_resched())
5585 #ifdef CONFIG_PREEMPT
5587 * this is the entry point to schedule() from in-kernel preemption
5588 * off of preempt_enable. Kernel preemptions off return from interrupt
5589 * occur there and call schedule directly.
5591 asmlinkage void __sched preempt_schedule(void)
5593 struct thread_info *ti = current_thread_info();
5596 * If there is a non-zero preempt_count or interrupts are disabled,
5597 * we do not want to preempt the current task. Just return..
5599 if (likely(ti->preempt_count || irqs_disabled()))
5603 add_preempt_count(PREEMPT_ACTIVE);
5605 sub_preempt_count(PREEMPT_ACTIVE);
5608 * Check again in case we missed a preemption opportunity
5609 * between schedule and now.
5612 } while (need_resched());
5614 EXPORT_SYMBOL(preempt_schedule);
5617 * this is the entry point to schedule() from kernel preemption
5618 * off of irq context.
5619 * Note, that this is called and return with irqs disabled. This will
5620 * protect us against recursive calling from irq.
5622 asmlinkage void __sched preempt_schedule_irq(void)
5624 struct thread_info *ti = current_thread_info();
5626 /* Catch callers which need to be fixed */
5627 BUG_ON(ti->preempt_count || !irqs_disabled());
5630 add_preempt_count(PREEMPT_ACTIVE);
5633 local_irq_disable();
5634 sub_preempt_count(PREEMPT_ACTIVE);
5637 * Check again in case we missed a preemption opportunity
5638 * between schedule and now.
5641 } while (need_resched());
5644 #endif /* CONFIG_PREEMPT */
5646 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
5649 return try_to_wake_up(curr->private, mode, wake_flags);
5651 EXPORT_SYMBOL(default_wake_function);
5654 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5655 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5656 * number) then we wake all the non-exclusive tasks and one exclusive task.
5658 * There are circumstances in which we can try to wake a task which has already
5659 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5660 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5662 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
5663 int nr_exclusive, int wake_flags, void *key)
5665 wait_queue_t *curr, *next;
5667 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
5668 unsigned flags = curr->flags;
5670 if (curr->func(curr, mode, wake_flags, key) &&
5671 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
5677 * __wake_up - wake up threads blocked on a waitqueue.
5679 * @mode: which threads
5680 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5681 * @key: is directly passed to the wakeup function
5683 * It may be assumed that this function implies a write memory barrier before
5684 * changing the task state if and only if any tasks are woken up.
5686 void __wake_up(wait_queue_head_t *q, unsigned int mode,
5687 int nr_exclusive, void *key)
5689 unsigned long flags;
5691 spin_lock_irqsave(&q->lock, flags);
5692 __wake_up_common(q, mode, nr_exclusive, 0, key);
5693 spin_unlock_irqrestore(&q->lock, flags);
5695 EXPORT_SYMBOL(__wake_up);
5698 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5700 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
5702 __wake_up_common(q, mode, 1, 0, NULL);
5705 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
5707 __wake_up_common(q, mode, 1, 0, key);
5711 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5713 * @mode: which threads
5714 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5715 * @key: opaque value to be passed to wakeup targets
5717 * The sync wakeup differs that the waker knows that it will schedule
5718 * away soon, so while the target thread will be woken up, it will not
5719 * be migrated to another CPU - ie. the two threads are 'synchronized'
5720 * with each other. This can prevent needless bouncing between CPUs.
5722 * On UP it can prevent extra preemption.
5724 * It may be assumed that this function implies a write memory barrier before
5725 * changing the task state if and only if any tasks are woken up.
5727 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
5728 int nr_exclusive, void *key)
5730 unsigned long flags;
5731 int wake_flags = WF_SYNC;
5736 if (unlikely(!nr_exclusive))
5739 spin_lock_irqsave(&q->lock, flags);
5740 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
5741 spin_unlock_irqrestore(&q->lock, flags);
5743 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
5746 * __wake_up_sync - see __wake_up_sync_key()
5748 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
5750 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
5752 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
5755 * complete: - signals a single thread waiting on this completion
5756 * @x: holds the state of this particular completion
5758 * This will wake up a single thread waiting on this completion. Threads will be
5759 * awakened in the same order in which they were queued.
5761 * See also complete_all(), wait_for_completion() and related routines.
5763 * It may be assumed that this function implies a write memory barrier before
5764 * changing the task state if and only if any tasks are woken up.
5766 void complete(struct completion *x)
5768 unsigned long flags;
5770 spin_lock_irqsave(&x->wait.lock, flags);
5772 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
5773 spin_unlock_irqrestore(&x->wait.lock, flags);
5775 EXPORT_SYMBOL(complete);
5778 * complete_all: - signals all threads waiting on this completion
5779 * @x: holds the state of this particular completion
5781 * This will wake up all threads waiting on this particular completion event.
5783 * It may be assumed that this function implies a write memory barrier before
5784 * changing the task state if and only if any tasks are woken up.
5786 void complete_all(struct completion *x)
5788 unsigned long flags;
5790 spin_lock_irqsave(&x->wait.lock, flags);
5791 x->done += UINT_MAX/2;
5792 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
5793 spin_unlock_irqrestore(&x->wait.lock, flags);
5795 EXPORT_SYMBOL(complete_all);
5797 static inline long __sched
5798 do_wait_for_common(struct completion *x, long timeout, int state)
5801 DECLARE_WAITQUEUE(wait, current);
5803 wait.flags |= WQ_FLAG_EXCLUSIVE;
5804 __add_wait_queue_tail(&x->wait, &wait);
5806 if (signal_pending_state(state, current)) {
5807 timeout = -ERESTARTSYS;
5810 __set_current_state(state);
5811 spin_unlock_irq(&x->wait.lock);
5812 timeout = schedule_timeout(timeout);
5813 spin_lock_irq(&x->wait.lock);
5814 } while (!x->done && timeout);
5815 __remove_wait_queue(&x->wait, &wait);
5820 return timeout ?: 1;
5824 wait_for_common(struct completion *x, long timeout, int state)
5828 spin_lock_irq(&x->wait.lock);
5829 timeout = do_wait_for_common(x, timeout, state);
5830 spin_unlock_irq(&x->wait.lock);
5835 * wait_for_completion: - waits for completion of a task
5836 * @x: holds the state of this particular completion
5838 * This waits to be signaled for completion of a specific task. It is NOT
5839 * interruptible and there is no timeout.
5841 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5842 * and interrupt capability. Also see complete().
5844 void __sched wait_for_completion(struct completion *x)
5846 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
5848 EXPORT_SYMBOL(wait_for_completion);
5851 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5852 * @x: holds the state of this particular completion
5853 * @timeout: timeout value in jiffies
5855 * This waits for either a completion of a specific task to be signaled or for a
5856 * specified timeout to expire. The timeout is in jiffies. It is not
5859 unsigned long __sched
5860 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5862 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5864 EXPORT_SYMBOL(wait_for_completion_timeout);
5867 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5868 * @x: holds the state of this particular completion
5870 * This waits for completion of a specific task to be signaled. It is
5873 int __sched wait_for_completion_interruptible(struct completion *x)
5875 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5876 if (t == -ERESTARTSYS)
5880 EXPORT_SYMBOL(wait_for_completion_interruptible);
5883 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5884 * @x: holds the state of this particular completion
5885 * @timeout: timeout value in jiffies
5887 * This waits for either a completion of a specific task to be signaled or for a
5888 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5890 unsigned long __sched
5891 wait_for_completion_interruptible_timeout(struct completion *x,
5892 unsigned long timeout)
5894 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
5896 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5899 * wait_for_completion_killable: - waits for completion of a task (killable)
5900 * @x: holds the state of this particular completion
5902 * This waits to be signaled for completion of a specific task. It can be
5903 * interrupted by a kill signal.
5905 int __sched wait_for_completion_killable(struct completion *x)
5907 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5908 if (t == -ERESTARTSYS)
5912 EXPORT_SYMBOL(wait_for_completion_killable);
5915 * try_wait_for_completion - try to decrement a completion without blocking
5916 * @x: completion structure
5918 * Returns: 0 if a decrement cannot be done without blocking
5919 * 1 if a decrement succeeded.
5921 * If a completion is being used as a counting completion,
5922 * attempt to decrement the counter without blocking. This
5923 * enables us to avoid waiting if the resource the completion
5924 * is protecting is not available.
5926 bool try_wait_for_completion(struct completion *x)
5930 spin_lock_irq(&x->wait.lock);
5935 spin_unlock_irq(&x->wait.lock);
5938 EXPORT_SYMBOL(try_wait_for_completion);
5941 * completion_done - Test to see if a completion has any waiters
5942 * @x: completion structure
5944 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5945 * 1 if there are no waiters.
5948 bool completion_done(struct completion *x)
5952 spin_lock_irq(&x->wait.lock);
5955 spin_unlock_irq(&x->wait.lock);
5958 EXPORT_SYMBOL(completion_done);
5961 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
5963 unsigned long flags;
5966 init_waitqueue_entry(&wait, current);
5968 __set_current_state(state);
5970 spin_lock_irqsave(&q->lock, flags);
5971 __add_wait_queue(q, &wait);
5972 spin_unlock(&q->lock);
5973 timeout = schedule_timeout(timeout);
5974 spin_lock_irq(&q->lock);
5975 __remove_wait_queue(q, &wait);
5976 spin_unlock_irqrestore(&q->lock, flags);
5981 void __sched interruptible_sleep_on(wait_queue_head_t *q)
5983 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5985 EXPORT_SYMBOL(interruptible_sleep_on);
5988 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
5990 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
5992 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
5994 void __sched sleep_on(wait_queue_head_t *q)
5996 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5998 EXPORT_SYMBOL(sleep_on);
6000 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
6002 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
6004 EXPORT_SYMBOL(sleep_on_timeout);
6006 #ifdef CONFIG_RT_MUTEXES
6009 * rt_mutex_setprio - set the current priority of a task
6011 * @prio: prio value (kernel-internal form)
6013 * This function changes the 'effective' priority of a task. It does
6014 * not touch ->normal_prio like __setscheduler().
6016 * Used by the rt_mutex code to implement priority inheritance logic.
6018 void rt_mutex_setprio(struct task_struct *p, int prio)
6020 unsigned long flags;
6021 int oldprio, on_rq, running;
6023 const struct sched_class *prev_class = p->sched_class;
6025 BUG_ON(prio < 0 || prio > MAX_PRIO);
6027 rq = task_rq_lock(p, &flags);
6028 update_rq_clock(rq);
6031 on_rq = p->se.on_rq;
6032 running = task_current(rq, p);
6034 dequeue_task(rq, p, 0);
6036 p->sched_class->put_prev_task(rq, p);
6039 p->sched_class = &rt_sched_class;
6041 p->sched_class = &fair_sched_class;
6046 p->sched_class->set_curr_task(rq);
6048 enqueue_task(rq, p, 0);
6050 check_class_changed(rq, p, prev_class, oldprio, running);
6052 task_rq_unlock(rq, &flags);
6057 void set_user_nice(struct task_struct *p, long nice)
6059 int old_prio, delta, on_rq;
6060 unsigned long flags;
6063 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
6066 * We have to be careful, if called from sys_setpriority(),
6067 * the task might be in the middle of scheduling on another CPU.
6069 rq = task_rq_lock(p, &flags);
6070 update_rq_clock(rq);
6072 * The RT priorities are set via sched_setscheduler(), but we still
6073 * allow the 'normal' nice value to be set - but as expected
6074 * it wont have any effect on scheduling until the task is
6075 * SCHED_FIFO/SCHED_RR:
6077 if (task_has_rt_policy(p)) {
6078 p->static_prio = NICE_TO_PRIO(nice);
6081 on_rq = p->se.on_rq;
6083 dequeue_task(rq, p, 0);
6085 p->static_prio = NICE_TO_PRIO(nice);
6088 p->prio = effective_prio(p);
6089 delta = p->prio - old_prio;
6092 enqueue_task(rq, p, 0);
6094 * If the task increased its priority or is running and
6095 * lowered its priority, then reschedule its CPU:
6097 if (delta < 0 || (delta > 0 && task_running(rq, p)))
6098 resched_task(rq->curr);
6101 task_rq_unlock(rq, &flags);
6103 EXPORT_SYMBOL(set_user_nice);
6106 * can_nice - check if a task can reduce its nice value
6110 int can_nice(const struct task_struct *p, const int nice)
6112 /* convert nice value [19,-20] to rlimit style value [1,40] */
6113 int nice_rlim = 20 - nice;
6115 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
6116 capable(CAP_SYS_NICE));
6119 #ifdef __ARCH_WANT_SYS_NICE
6122 * sys_nice - change the priority of the current process.
6123 * @increment: priority increment
6125 * sys_setpriority is a more generic, but much slower function that
6126 * does similar things.
6128 SYSCALL_DEFINE1(nice, int, increment)
6133 * Setpriority might change our priority at the same moment.
6134 * We don't have to worry. Conceptually one call occurs first
6135 * and we have a single winner.
6137 if (increment < -40)
6142 nice = TASK_NICE(current) + increment;
6148 if (increment < 0 && !can_nice(current, nice))
6151 retval = security_task_setnice(current, nice);
6155 set_user_nice(current, nice);
6162 * task_prio - return the priority value of a given task.
6163 * @p: the task in question.
6165 * This is the priority value as seen by users in /proc.
6166 * RT tasks are offset by -200. Normal tasks are centered
6167 * around 0, value goes from -16 to +15.
6169 int task_prio(const struct task_struct *p)
6171 return p->prio - MAX_RT_PRIO;
6175 * task_nice - return the nice value of a given task.
6176 * @p: the task in question.
6178 int task_nice(const struct task_struct *p)
6180 return TASK_NICE(p);
6182 EXPORT_SYMBOL(task_nice);
6185 * idle_cpu - is a given cpu idle currently?
6186 * @cpu: the processor in question.
6188 int idle_cpu(int cpu)
6190 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
6194 * idle_task - return the idle task for a given cpu.
6195 * @cpu: the processor in question.
6197 struct task_struct *idle_task(int cpu)
6199 return cpu_rq(cpu)->idle;
6203 * find_process_by_pid - find a process with a matching PID value.
6204 * @pid: the pid in question.
6206 static struct task_struct *find_process_by_pid(pid_t pid)
6208 return pid ? find_task_by_vpid(pid) : current;
6211 /* Actually do priority change: must hold rq lock. */
6213 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
6215 BUG_ON(p->se.on_rq);
6218 p->rt_priority = prio;
6219 p->normal_prio = normal_prio(p);
6220 /* we are holding p->pi_lock already */
6221 p->prio = rt_mutex_getprio(p);
6222 if (rt_prio(p->prio))
6223 p->sched_class = &rt_sched_class;
6225 p->sched_class = &fair_sched_class;
6230 * check the target process has a UID that matches the current process's
6232 static bool check_same_owner(struct task_struct *p)
6234 const struct cred *cred = current_cred(), *pcred;
6238 pcred = __task_cred(p);
6239 match = (cred->euid == pcred->euid ||
6240 cred->euid == pcred->uid);
6245 static int __sched_setscheduler(struct task_struct *p, int policy,
6246 struct sched_param *param, bool user)
6248 int retval, oldprio, oldpolicy = -1, on_rq, running;
6249 unsigned long flags;
6250 const struct sched_class *prev_class = p->sched_class;
6254 /* may grab non-irq protected spin_locks */
6255 BUG_ON(in_interrupt());
6257 /* double check policy once rq lock held */
6259 reset_on_fork = p->sched_reset_on_fork;
6260 policy = oldpolicy = p->policy;
6262 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
6263 policy &= ~SCHED_RESET_ON_FORK;
6265 if (policy != SCHED_FIFO && policy != SCHED_RR &&
6266 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
6267 policy != SCHED_IDLE)
6272 * Valid priorities for SCHED_FIFO and SCHED_RR are
6273 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6274 * SCHED_BATCH and SCHED_IDLE is 0.
6276 if (param->sched_priority < 0 ||
6277 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
6278 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
6280 if (rt_policy(policy) != (param->sched_priority != 0))
6284 * Allow unprivileged RT tasks to decrease priority:
6286 if (user && !capable(CAP_SYS_NICE)) {
6287 if (rt_policy(policy)) {
6288 unsigned long rlim_rtprio;
6290 if (!lock_task_sighand(p, &flags))
6292 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
6293 unlock_task_sighand(p, &flags);
6295 /* can't set/change the rt policy */
6296 if (policy != p->policy && !rlim_rtprio)
6299 /* can't increase priority */
6300 if (param->sched_priority > p->rt_priority &&
6301 param->sched_priority > rlim_rtprio)
6305 * Like positive nice levels, dont allow tasks to
6306 * move out of SCHED_IDLE either:
6308 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
6311 /* can't change other user's priorities */
6312 if (!check_same_owner(p))
6315 /* Normal users shall not reset the sched_reset_on_fork flag */
6316 if (p->sched_reset_on_fork && !reset_on_fork)
6321 #ifdef CONFIG_RT_GROUP_SCHED
6323 * Do not allow realtime tasks into groups that have no runtime
6326 if (rt_bandwidth_enabled() && rt_policy(policy) &&
6327 task_group(p)->rt_bandwidth.rt_runtime == 0)
6331 retval = security_task_setscheduler(p, policy, param);
6337 * make sure no PI-waiters arrive (or leave) while we are
6338 * changing the priority of the task:
6340 spin_lock_irqsave(&p->pi_lock, flags);
6342 * To be able to change p->policy safely, the apropriate
6343 * runqueue lock must be held.
6345 rq = __task_rq_lock(p);
6346 /* recheck policy now with rq lock held */
6347 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
6348 policy = oldpolicy = -1;
6349 __task_rq_unlock(rq);
6350 spin_unlock_irqrestore(&p->pi_lock, flags);
6353 update_rq_clock(rq);
6354 on_rq = p->se.on_rq;
6355 running = task_current(rq, p);
6357 deactivate_task(rq, p, 0);
6359 p->sched_class->put_prev_task(rq, p);
6361 p->sched_reset_on_fork = reset_on_fork;
6364 __setscheduler(rq, p, policy, param->sched_priority);
6367 p->sched_class->set_curr_task(rq);
6369 activate_task(rq, p, 0);
6371 check_class_changed(rq, p, prev_class, oldprio, running);
6373 __task_rq_unlock(rq);
6374 spin_unlock_irqrestore(&p->pi_lock, flags);
6376 rt_mutex_adjust_pi(p);
6382 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6383 * @p: the task in question.
6384 * @policy: new policy.
6385 * @param: structure containing the new RT priority.
6387 * NOTE that the task may be already dead.
6389 int sched_setscheduler(struct task_struct *p, int policy,
6390 struct sched_param *param)
6392 return __sched_setscheduler(p, policy, param, true);
6394 EXPORT_SYMBOL_GPL(sched_setscheduler);
6397 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6398 * @p: the task in question.
6399 * @policy: new policy.
6400 * @param: structure containing the new RT priority.
6402 * Just like sched_setscheduler, only don't bother checking if the
6403 * current context has permission. For example, this is needed in
6404 * stop_machine(): we create temporary high priority worker threads,
6405 * but our caller might not have that capability.
6407 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
6408 struct sched_param *param)
6410 return __sched_setscheduler(p, policy, param, false);
6414 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
6416 struct sched_param lparam;
6417 struct task_struct *p;
6420 if (!param || pid < 0)
6422 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
6427 p = find_process_by_pid(pid);
6429 retval = sched_setscheduler(p, policy, &lparam);
6436 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6437 * @pid: the pid in question.
6438 * @policy: new policy.
6439 * @param: structure containing the new RT priority.
6441 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
6442 struct sched_param __user *, param)
6444 /* negative values for policy are not valid */
6448 return do_sched_setscheduler(pid, policy, param);
6452 * sys_sched_setparam - set/change the RT priority of a thread
6453 * @pid: the pid in question.
6454 * @param: structure containing the new RT priority.
6456 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
6458 return do_sched_setscheduler(pid, -1, param);
6462 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6463 * @pid: the pid in question.
6465 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
6467 struct task_struct *p;
6474 read_lock(&tasklist_lock);
6475 p = find_process_by_pid(pid);
6477 retval = security_task_getscheduler(p);
6480 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
6482 read_unlock(&tasklist_lock);
6487 * sys_sched_getparam - get the RT priority of a thread
6488 * @pid: the pid in question.
6489 * @param: structure containing the RT priority.
6491 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
6493 struct sched_param lp;
6494 struct task_struct *p;
6497 if (!param || pid < 0)
6500 read_lock(&tasklist_lock);
6501 p = find_process_by_pid(pid);
6506 retval = security_task_getscheduler(p);
6510 lp.sched_priority = p->rt_priority;
6511 read_unlock(&tasklist_lock);
6514 * This one might sleep, we cannot do it with a spinlock held ...
6516 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
6521 read_unlock(&tasklist_lock);
6525 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
6527 cpumask_var_t cpus_allowed, new_mask;
6528 struct task_struct *p;
6532 read_lock(&tasklist_lock);
6534 p = find_process_by_pid(pid);
6536 read_unlock(&tasklist_lock);
6542 * It is not safe to call set_cpus_allowed with the
6543 * tasklist_lock held. We will bump the task_struct's
6544 * usage count and then drop tasklist_lock.
6547 read_unlock(&tasklist_lock);
6549 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6553 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
6555 goto out_free_cpus_allowed;
6558 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
6561 retval = security_task_setscheduler(p, 0, NULL);
6565 cpuset_cpus_allowed(p, cpus_allowed);
6566 cpumask_and(new_mask, in_mask, cpus_allowed);
6568 retval = set_cpus_allowed_ptr(p, new_mask);
6571 cpuset_cpus_allowed(p, cpus_allowed);
6572 if (!cpumask_subset(new_mask, cpus_allowed)) {
6574 * We must have raced with a concurrent cpuset
6575 * update. Just reset the cpus_allowed to the
6576 * cpuset's cpus_allowed
6578 cpumask_copy(new_mask, cpus_allowed);
6583 free_cpumask_var(new_mask);
6584 out_free_cpus_allowed:
6585 free_cpumask_var(cpus_allowed);
6592 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
6593 struct cpumask *new_mask)
6595 if (len < cpumask_size())
6596 cpumask_clear(new_mask);
6597 else if (len > cpumask_size())
6598 len = cpumask_size();
6600 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6604 * sys_sched_setaffinity - set the cpu affinity of a process
6605 * @pid: pid of the process
6606 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6607 * @user_mask_ptr: user-space pointer to the new cpu mask
6609 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6610 unsigned long __user *, user_mask_ptr)
6612 cpumask_var_t new_mask;
6615 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6618 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6620 retval = sched_setaffinity(pid, new_mask);
6621 free_cpumask_var(new_mask);
6625 long sched_getaffinity(pid_t pid, struct cpumask *mask)
6627 struct task_struct *p;
6628 unsigned long flags;
6633 read_lock(&tasklist_lock);
6636 p = find_process_by_pid(pid);
6640 retval = security_task_getscheduler(p);
6644 rq = task_rq_lock(p, &flags);
6645 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
6646 task_rq_unlock(rq, &flags);
6649 read_unlock(&tasklist_lock);
6656 * sys_sched_getaffinity - get the cpu affinity of a process
6657 * @pid: pid of the process
6658 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6659 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6661 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6662 unsigned long __user *, user_mask_ptr)
6667 if (len < cpumask_size())
6670 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6673 ret = sched_getaffinity(pid, mask);
6675 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
6678 ret = cpumask_size();
6680 free_cpumask_var(mask);
6686 * sys_sched_yield - yield the current processor to other threads.
6688 * This function yields the current CPU to other tasks. If there are no
6689 * other threads running on this CPU then this function will return.
6691 SYSCALL_DEFINE0(sched_yield)
6693 struct rq *rq = this_rq_lock();
6695 schedstat_inc(rq, yld_count);
6696 current->sched_class->yield_task(rq);
6699 * Since we are going to call schedule() anyway, there's
6700 * no need to preempt or enable interrupts:
6702 __release(rq->lock);
6703 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
6704 _raw_spin_unlock(&rq->lock);
6705 preempt_enable_no_resched();
6712 static inline int should_resched(void)
6714 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
6717 static void __cond_resched(void)
6719 add_preempt_count(PREEMPT_ACTIVE);
6721 sub_preempt_count(PREEMPT_ACTIVE);
6724 int __sched _cond_resched(void)
6726 if (should_resched()) {
6732 EXPORT_SYMBOL(_cond_resched);
6735 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
6736 * call schedule, and on return reacquire the lock.
6738 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6739 * operations here to prevent schedule() from being called twice (once via
6740 * spin_unlock(), once by hand).
6742 int __cond_resched_lock(spinlock_t *lock)
6744 int resched = should_resched();
6747 lockdep_assert_held(lock);
6749 if (spin_needbreak(lock) || resched) {
6760 EXPORT_SYMBOL(__cond_resched_lock);
6762 int __sched __cond_resched_softirq(void)
6764 BUG_ON(!in_softirq());
6766 if (should_resched()) {
6774 EXPORT_SYMBOL(__cond_resched_softirq);
6777 * yield - yield the current processor to other threads.
6779 * This is a shortcut for kernel-space yielding - it marks the
6780 * thread runnable and calls sys_sched_yield().
6782 void __sched yield(void)
6784 set_current_state(TASK_RUNNING);
6787 EXPORT_SYMBOL(yield);
6790 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6791 * that process accounting knows that this is a task in IO wait state.
6793 void __sched io_schedule(void)
6795 struct rq *rq = raw_rq();
6797 delayacct_blkio_start();
6798 atomic_inc(&rq->nr_iowait);
6799 current->in_iowait = 1;
6801 current->in_iowait = 0;
6802 atomic_dec(&rq->nr_iowait);
6803 delayacct_blkio_end();
6805 EXPORT_SYMBOL(io_schedule);
6807 long __sched io_schedule_timeout(long timeout)
6809 struct rq *rq = raw_rq();
6812 delayacct_blkio_start();
6813 atomic_inc(&rq->nr_iowait);
6814 current->in_iowait = 1;
6815 ret = schedule_timeout(timeout);
6816 current->in_iowait = 0;
6817 atomic_dec(&rq->nr_iowait);
6818 delayacct_blkio_end();
6823 * sys_sched_get_priority_max - return maximum RT priority.
6824 * @policy: scheduling class.
6826 * this syscall returns the maximum rt_priority that can be used
6827 * by a given scheduling class.
6829 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6836 ret = MAX_USER_RT_PRIO-1;
6848 * sys_sched_get_priority_min - return minimum RT priority.
6849 * @policy: scheduling class.
6851 * this syscall returns the minimum rt_priority that can be used
6852 * by a given scheduling class.
6854 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6872 * sys_sched_rr_get_interval - return the default timeslice of a process.
6873 * @pid: pid of the process.
6874 * @interval: userspace pointer to the timeslice value.
6876 * this syscall writes the default timeslice value of a given process
6877 * into the user-space timespec buffer. A value of '0' means infinity.
6879 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6880 struct timespec __user *, interval)
6882 struct task_struct *p;
6883 unsigned int time_slice;
6884 unsigned long flags;
6893 read_lock(&tasklist_lock);
6894 p = find_process_by_pid(pid);
6898 retval = security_task_getscheduler(p);
6902 rq = task_rq_lock(p, &flags);
6903 time_slice = p->sched_class->get_rr_interval(rq, p);
6904 task_rq_unlock(rq, &flags);
6906 read_unlock(&tasklist_lock);
6907 jiffies_to_timespec(time_slice, &t);
6908 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
6912 read_unlock(&tasklist_lock);
6916 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
6918 void sched_show_task(struct task_struct *p)
6920 unsigned long free = 0;
6923 state = p->state ? __ffs(p->state) + 1 : 0;
6924 printk(KERN_INFO "%-13.13s %c", p->comm,
6925 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
6926 #if BITS_PER_LONG == 32
6927 if (state == TASK_RUNNING)
6928 printk(KERN_CONT " running ");
6930 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
6932 if (state == TASK_RUNNING)
6933 printk(KERN_CONT " running task ");
6935 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
6937 #ifdef CONFIG_DEBUG_STACK_USAGE
6938 free = stack_not_used(p);
6940 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
6941 task_pid_nr(p), task_pid_nr(p->real_parent),
6942 (unsigned long)task_thread_info(p)->flags);
6944 show_stack(p, NULL);
6947 void show_state_filter(unsigned long state_filter)
6949 struct task_struct *g, *p;
6951 #if BITS_PER_LONG == 32
6953 " task PC stack pid father\n");
6956 " task PC stack pid father\n");
6958 read_lock(&tasklist_lock);
6959 do_each_thread(g, p) {
6961 * reset the NMI-timeout, listing all files on a slow
6962 * console might take alot of time:
6964 touch_nmi_watchdog();
6965 if (!state_filter || (p->state & state_filter))
6967 } while_each_thread(g, p);
6969 touch_all_softlockup_watchdogs();
6971 #ifdef CONFIG_SCHED_DEBUG
6972 sysrq_sched_debug_show();
6974 read_unlock(&tasklist_lock);
6976 * Only show locks if all tasks are dumped:
6979 debug_show_all_locks();
6982 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6984 idle->sched_class = &idle_sched_class;
6988 * init_idle - set up an idle thread for a given CPU
6989 * @idle: task in question
6990 * @cpu: cpu the idle task belongs to
6992 * NOTE: this function does not set the idle thread's NEED_RESCHED
6993 * flag, to make booting more robust.
6995 void __cpuinit init_idle(struct task_struct *idle, int cpu)
6997 struct rq *rq = cpu_rq(cpu);
6998 unsigned long flags;
7000 spin_lock_irqsave(&rq->lock, flags);
7003 idle->se.exec_start = sched_clock();
7005 idle->prio = idle->normal_prio = MAX_PRIO;
7006 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
7007 __set_task_cpu(idle, cpu);
7009 rq->curr = rq->idle = idle;
7010 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
7013 spin_unlock_irqrestore(&rq->lock, flags);
7015 /* Set the preempt count _outside_ the spinlocks! */
7016 #if defined(CONFIG_PREEMPT)
7017 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
7019 task_thread_info(idle)->preempt_count = 0;
7022 * The idle tasks have their own, simple scheduling class:
7024 idle->sched_class = &idle_sched_class;
7025 ftrace_graph_init_task(idle);
7029 * In a system that switches off the HZ timer nohz_cpu_mask
7030 * indicates which cpus entered this state. This is used
7031 * in the rcu update to wait only for active cpus. For system
7032 * which do not switch off the HZ timer nohz_cpu_mask should
7033 * always be CPU_BITS_NONE.
7035 cpumask_var_t nohz_cpu_mask;
7038 * Increase the granularity value when there are more CPUs,
7039 * because with more CPUs the 'effective latency' as visible
7040 * to users decreases. But the relationship is not linear,
7041 * so pick a second-best guess by going with the log2 of the
7044 * This idea comes from the SD scheduler of Con Kolivas:
7046 static inline void sched_init_granularity(void)
7048 unsigned int factor = 1 + ilog2(num_online_cpus());
7049 const unsigned long limit = 200000000;
7051 sysctl_sched_min_granularity *= factor;
7052 if (sysctl_sched_min_granularity > limit)
7053 sysctl_sched_min_granularity = limit;
7055 sysctl_sched_latency *= factor;
7056 if (sysctl_sched_latency > limit)
7057 sysctl_sched_latency = limit;
7059 sysctl_sched_wakeup_granularity *= factor;
7061 sysctl_sched_shares_ratelimit *= factor;
7066 * This is how migration works:
7068 * 1) we queue a struct migration_req structure in the source CPU's
7069 * runqueue and wake up that CPU's migration thread.
7070 * 2) we down() the locked semaphore => thread blocks.
7071 * 3) migration thread wakes up (implicitly it forces the migrated
7072 * thread off the CPU)
7073 * 4) it gets the migration request and checks whether the migrated
7074 * task is still in the wrong runqueue.
7075 * 5) if it's in the wrong runqueue then the migration thread removes
7076 * it and puts it into the right queue.
7077 * 6) migration thread up()s the semaphore.
7078 * 7) we wake up and the migration is done.
7082 * Change a given task's CPU affinity. Migrate the thread to a
7083 * proper CPU and schedule it away if the CPU it's executing on
7084 * is removed from the allowed bitmask.
7086 * NOTE: the caller must have a valid reference to the task, the
7087 * task must not exit() & deallocate itself prematurely. The
7088 * call is not atomic; no spinlocks may be held.
7090 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
7092 struct migration_req req;
7093 unsigned long flags;
7097 rq = task_rq_lock(p, &flags);
7098 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
7103 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
7104 !cpumask_equal(&p->cpus_allowed, new_mask))) {
7109 if (p->sched_class->set_cpus_allowed)
7110 p->sched_class->set_cpus_allowed(p, new_mask);
7112 cpumask_copy(&p->cpus_allowed, new_mask);
7113 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
7116 /* Can the task run on the task's current CPU? If so, we're done */
7117 if (cpumask_test_cpu(task_cpu(p), new_mask))
7120 if (migrate_task(p, cpumask_any_and(cpu_active_mask, new_mask), &req)) {
7121 /* Need help from migration thread: drop lock and wait. */
7122 struct task_struct *mt = rq->migration_thread;
7124 get_task_struct(mt);
7125 task_rq_unlock(rq, &flags);
7126 wake_up_process(rq->migration_thread);
7127 put_task_struct(mt);
7128 wait_for_completion(&req.done);
7129 tlb_migrate_finish(p->mm);
7133 task_rq_unlock(rq, &flags);
7137 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
7140 * Move (not current) task off this cpu, onto dest cpu. We're doing
7141 * this because either it can't run here any more (set_cpus_allowed()
7142 * away from this CPU, or CPU going down), or because we're
7143 * attempting to rebalance this task on exec (sched_exec).
7145 * So we race with normal scheduler movements, but that's OK, as long
7146 * as the task is no longer on this CPU.
7148 * Returns non-zero if task was successfully migrated.
7150 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
7152 struct rq *rq_dest, *rq_src;
7155 if (unlikely(!cpu_active(dest_cpu)))
7158 rq_src = cpu_rq(src_cpu);
7159 rq_dest = cpu_rq(dest_cpu);
7161 double_rq_lock(rq_src, rq_dest);
7162 /* Already moved. */
7163 if (task_cpu(p) != src_cpu)
7165 /* Affinity changed (again). */
7166 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7169 on_rq = p->se.on_rq;
7171 deactivate_task(rq_src, p, 0);
7173 set_task_cpu(p, dest_cpu);
7175 activate_task(rq_dest, p, 0);
7176 check_preempt_curr(rq_dest, p, 0);
7181 double_rq_unlock(rq_src, rq_dest);
7185 #define RCU_MIGRATION_IDLE 0
7186 #define RCU_MIGRATION_NEED_QS 1
7187 #define RCU_MIGRATION_GOT_QS 2
7188 #define RCU_MIGRATION_MUST_SYNC 3
7191 * migration_thread - this is a highprio system thread that performs
7192 * thread migration by bumping thread off CPU then 'pushing' onto
7195 static int migration_thread(void *data)
7198 int cpu = (long)data;
7202 BUG_ON(rq->migration_thread != current);
7204 set_current_state(TASK_INTERRUPTIBLE);
7205 while (!kthread_should_stop()) {
7206 struct migration_req *req;
7207 struct list_head *head;
7209 spin_lock_irq(&rq->lock);
7211 if (cpu_is_offline(cpu)) {
7212 spin_unlock_irq(&rq->lock);
7216 if (rq->active_balance) {
7217 active_load_balance(rq, cpu);
7218 rq->active_balance = 0;
7221 head = &rq->migration_queue;
7223 if (list_empty(head)) {
7224 spin_unlock_irq(&rq->lock);
7226 set_current_state(TASK_INTERRUPTIBLE);
7229 req = list_entry(head->next, struct migration_req, list);
7230 list_del_init(head->next);
7232 if (req->task != NULL) {
7233 spin_unlock(&rq->lock);
7234 __migrate_task(req->task, cpu, req->dest_cpu);
7235 } else if (likely(cpu == (badcpu = smp_processor_id()))) {
7236 req->dest_cpu = RCU_MIGRATION_GOT_QS;
7237 spin_unlock(&rq->lock);
7239 req->dest_cpu = RCU_MIGRATION_MUST_SYNC;
7240 spin_unlock(&rq->lock);
7241 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu, cpu);
7245 complete(&req->done);
7247 __set_current_state(TASK_RUNNING);
7252 #ifdef CONFIG_HOTPLUG_CPU
7254 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
7258 local_irq_disable();
7259 ret = __migrate_task(p, src_cpu, dest_cpu);
7265 * Figure out where task on dead CPU should go, use force if necessary.
7267 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
7270 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(dead_cpu));
7273 /* Look for allowed, online CPU in same node. */
7274 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
7275 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7278 /* Any allowed, online CPU? */
7279 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
7280 if (dest_cpu < nr_cpu_ids)
7283 /* No more Mr. Nice Guy. */
7284 if (dest_cpu >= nr_cpu_ids) {
7285 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
7286 dest_cpu = cpumask_any_and(cpu_active_mask, &p->cpus_allowed);
7289 * Don't tell them about moving exiting tasks or
7290 * kernel threads (both mm NULL), since they never
7293 if (p->mm && printk_ratelimit()) {
7294 printk(KERN_INFO "process %d (%s) no "
7295 "longer affine to cpu%d\n",
7296 task_pid_nr(p), p->comm, dead_cpu);
7301 /* It can have affinity changed while we were choosing. */
7302 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
7307 * While a dead CPU has no uninterruptible tasks queued at this point,
7308 * it might still have a nonzero ->nr_uninterruptible counter, because
7309 * for performance reasons the counter is not stricly tracking tasks to
7310 * their home CPUs. So we just add the counter to another CPU's counter,
7311 * to keep the global sum constant after CPU-down:
7313 static void migrate_nr_uninterruptible(struct rq *rq_src)
7315 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
7316 unsigned long flags;
7318 local_irq_save(flags);
7319 double_rq_lock(rq_src, rq_dest);
7320 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
7321 rq_src->nr_uninterruptible = 0;
7322 double_rq_unlock(rq_src, rq_dest);
7323 local_irq_restore(flags);
7326 /* Run through task list and migrate tasks from the dead cpu. */
7327 static void migrate_live_tasks(int src_cpu)
7329 struct task_struct *p, *t;
7331 read_lock(&tasklist_lock);
7333 do_each_thread(t, p) {
7337 if (task_cpu(p) == src_cpu)
7338 move_task_off_dead_cpu(src_cpu, p);
7339 } while_each_thread(t, p);
7341 read_unlock(&tasklist_lock);
7345 * Schedules idle task to be the next runnable task on current CPU.
7346 * It does so by boosting its priority to highest possible.
7347 * Used by CPU offline code.
7349 void sched_idle_next(void)
7351 int this_cpu = smp_processor_id();
7352 struct rq *rq = cpu_rq(this_cpu);
7353 struct task_struct *p = rq->idle;
7354 unsigned long flags;
7356 /* cpu has to be offline */
7357 BUG_ON(cpu_online(this_cpu));
7360 * Strictly not necessary since rest of the CPUs are stopped by now
7361 * and interrupts disabled on the current cpu.
7363 spin_lock_irqsave(&rq->lock, flags);
7365 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7367 update_rq_clock(rq);
7368 activate_task(rq, p, 0);
7370 spin_unlock_irqrestore(&rq->lock, flags);
7374 * Ensures that the idle task is using init_mm right before its cpu goes
7377 void idle_task_exit(void)
7379 struct mm_struct *mm = current->active_mm;
7381 BUG_ON(cpu_online(smp_processor_id()));
7384 switch_mm(mm, &init_mm, current);
7388 /* called under rq->lock with disabled interrupts */
7389 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
7391 struct rq *rq = cpu_rq(dead_cpu);
7393 /* Must be exiting, otherwise would be on tasklist. */
7394 BUG_ON(!p->exit_state);
7396 /* Cannot have done final schedule yet: would have vanished. */
7397 BUG_ON(p->state == TASK_DEAD);
7402 * Drop lock around migration; if someone else moves it,
7403 * that's OK. No task can be added to this CPU, so iteration is
7406 spin_unlock_irq(&rq->lock);
7407 move_task_off_dead_cpu(dead_cpu, p);
7408 spin_lock_irq(&rq->lock);
7413 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7414 static void migrate_dead_tasks(unsigned int dead_cpu)
7416 struct rq *rq = cpu_rq(dead_cpu);
7417 struct task_struct *next;
7420 if (!rq->nr_running)
7422 update_rq_clock(rq);
7423 next = pick_next_task(rq);
7426 next->sched_class->put_prev_task(rq, next);
7427 migrate_dead(dead_cpu, next);
7433 * remove the tasks which were accounted by rq from calc_load_tasks.
7435 static void calc_global_load_remove(struct rq *rq)
7437 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
7438 rq->calc_load_active = 0;
7440 #endif /* CONFIG_HOTPLUG_CPU */
7442 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7444 static struct ctl_table sd_ctl_dir[] = {
7446 .procname = "sched_domain",
7452 static struct ctl_table sd_ctl_root[] = {
7454 .ctl_name = CTL_KERN,
7455 .procname = "kernel",
7457 .child = sd_ctl_dir,
7462 static struct ctl_table *sd_alloc_ctl_entry(int n)
7464 struct ctl_table *entry =
7465 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
7470 static void sd_free_ctl_entry(struct ctl_table **tablep)
7472 struct ctl_table *entry;
7475 * In the intermediate directories, both the child directory and
7476 * procname are dynamically allocated and could fail but the mode
7477 * will always be set. In the lowest directory the names are
7478 * static strings and all have proc handlers.
7480 for (entry = *tablep; entry->mode; entry++) {
7482 sd_free_ctl_entry(&entry->child);
7483 if (entry->proc_handler == NULL)
7484 kfree(entry->procname);
7492 set_table_entry(struct ctl_table *entry,
7493 const char *procname, void *data, int maxlen,
7494 mode_t mode, proc_handler *proc_handler)
7496 entry->procname = procname;
7498 entry->maxlen = maxlen;
7500 entry->proc_handler = proc_handler;
7503 static struct ctl_table *
7504 sd_alloc_ctl_domain_table(struct sched_domain *sd)
7506 struct ctl_table *table = sd_alloc_ctl_entry(13);
7511 set_table_entry(&table[0], "min_interval", &sd->min_interval,
7512 sizeof(long), 0644, proc_doulongvec_minmax);
7513 set_table_entry(&table[1], "max_interval", &sd->max_interval,
7514 sizeof(long), 0644, proc_doulongvec_minmax);
7515 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
7516 sizeof(int), 0644, proc_dointvec_minmax);
7517 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
7518 sizeof(int), 0644, proc_dointvec_minmax);
7519 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
7520 sizeof(int), 0644, proc_dointvec_minmax);
7521 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
7522 sizeof(int), 0644, proc_dointvec_minmax);
7523 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
7524 sizeof(int), 0644, proc_dointvec_minmax);
7525 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
7526 sizeof(int), 0644, proc_dointvec_minmax);
7527 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
7528 sizeof(int), 0644, proc_dointvec_minmax);
7529 set_table_entry(&table[9], "cache_nice_tries",
7530 &sd->cache_nice_tries,
7531 sizeof(int), 0644, proc_dointvec_minmax);
7532 set_table_entry(&table[10], "flags", &sd->flags,
7533 sizeof(int), 0644, proc_dointvec_minmax);
7534 set_table_entry(&table[11], "name", sd->name,
7535 CORENAME_MAX_SIZE, 0444, proc_dostring);
7536 /* &table[12] is terminator */
7541 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
7543 struct ctl_table *entry, *table;
7544 struct sched_domain *sd;
7545 int domain_num = 0, i;
7548 for_each_domain(cpu, sd)
7550 entry = table = sd_alloc_ctl_entry(domain_num + 1);
7555 for_each_domain(cpu, sd) {
7556 snprintf(buf, 32, "domain%d", i);
7557 entry->procname = kstrdup(buf, GFP_KERNEL);
7559 entry->child = sd_alloc_ctl_domain_table(sd);
7566 static struct ctl_table_header *sd_sysctl_header;
7567 static void register_sched_domain_sysctl(void)
7569 int i, cpu_num = num_possible_cpus();
7570 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
7573 WARN_ON(sd_ctl_dir[0].child);
7574 sd_ctl_dir[0].child = entry;
7579 for_each_possible_cpu(i) {
7580 snprintf(buf, 32, "cpu%d", i);
7581 entry->procname = kstrdup(buf, GFP_KERNEL);
7583 entry->child = sd_alloc_ctl_cpu_table(i);
7587 WARN_ON(sd_sysctl_header);
7588 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
7591 /* may be called multiple times per register */
7592 static void unregister_sched_domain_sysctl(void)
7594 if (sd_sysctl_header)
7595 unregister_sysctl_table(sd_sysctl_header);
7596 sd_sysctl_header = NULL;
7597 if (sd_ctl_dir[0].child)
7598 sd_free_ctl_entry(&sd_ctl_dir[0].child);
7601 static void register_sched_domain_sysctl(void)
7604 static void unregister_sched_domain_sysctl(void)
7609 static void set_rq_online(struct rq *rq)
7612 const struct sched_class *class;
7614 cpumask_set_cpu(rq->cpu, rq->rd->online);
7617 for_each_class(class) {
7618 if (class->rq_online)
7619 class->rq_online(rq);
7624 static void set_rq_offline(struct rq *rq)
7627 const struct sched_class *class;
7629 for_each_class(class) {
7630 if (class->rq_offline)
7631 class->rq_offline(rq);
7634 cpumask_clear_cpu(rq->cpu, rq->rd->online);
7640 * migration_call - callback that gets triggered when a CPU is added.
7641 * Here we can start up the necessary migration thread for the new CPU.
7643 static int __cpuinit
7644 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
7646 struct task_struct *p;
7647 int cpu = (long)hcpu;
7648 unsigned long flags;
7653 case CPU_UP_PREPARE:
7654 case CPU_UP_PREPARE_FROZEN:
7655 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
7658 kthread_bind(p, cpu);
7659 /* Must be high prio: stop_machine expects to yield to it. */
7660 rq = task_rq_lock(p, &flags);
7661 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7662 task_rq_unlock(rq, &flags);
7664 cpu_rq(cpu)->migration_thread = p;
7665 rq->calc_load_update = calc_load_update;
7669 case CPU_ONLINE_FROZEN:
7670 /* Strictly unnecessary, as first user will wake it. */
7671 wake_up_process(cpu_rq(cpu)->migration_thread);
7673 /* Update our root-domain */
7675 spin_lock_irqsave(&rq->lock, flags);
7677 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7681 spin_unlock_irqrestore(&rq->lock, flags);
7684 #ifdef CONFIG_HOTPLUG_CPU
7685 case CPU_UP_CANCELED:
7686 case CPU_UP_CANCELED_FROZEN:
7687 if (!cpu_rq(cpu)->migration_thread)
7689 /* Unbind it from offline cpu so it can run. Fall thru. */
7690 kthread_bind(cpu_rq(cpu)->migration_thread,
7691 cpumask_any(cpu_online_mask));
7692 kthread_stop(cpu_rq(cpu)->migration_thread);
7693 put_task_struct(cpu_rq(cpu)->migration_thread);
7694 cpu_rq(cpu)->migration_thread = NULL;
7698 case CPU_DEAD_FROZEN:
7699 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7700 migrate_live_tasks(cpu);
7702 kthread_stop(rq->migration_thread);
7703 put_task_struct(rq->migration_thread);
7704 rq->migration_thread = NULL;
7705 /* Idle task back to normal (off runqueue, low prio) */
7706 spin_lock_irq(&rq->lock);
7707 update_rq_clock(rq);
7708 deactivate_task(rq, rq->idle, 0);
7709 rq->idle->static_prio = MAX_PRIO;
7710 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
7711 rq->idle->sched_class = &idle_sched_class;
7712 migrate_dead_tasks(cpu);
7713 spin_unlock_irq(&rq->lock);
7715 migrate_nr_uninterruptible(rq);
7716 BUG_ON(rq->nr_running != 0);
7717 calc_global_load_remove(rq);
7719 * No need to migrate the tasks: it was best-effort if
7720 * they didn't take sched_hotcpu_mutex. Just wake up
7723 spin_lock_irq(&rq->lock);
7724 while (!list_empty(&rq->migration_queue)) {
7725 struct migration_req *req;
7727 req = list_entry(rq->migration_queue.next,
7728 struct migration_req, list);
7729 list_del_init(&req->list);
7730 spin_unlock_irq(&rq->lock);
7731 complete(&req->done);
7732 spin_lock_irq(&rq->lock);
7734 spin_unlock_irq(&rq->lock);
7738 case CPU_DYING_FROZEN:
7739 /* Update our root-domain */
7741 spin_lock_irqsave(&rq->lock, flags);
7743 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7746 spin_unlock_irqrestore(&rq->lock, flags);
7754 * Register at high priority so that task migration (migrate_all_tasks)
7755 * happens before everything else. This has to be lower priority than
7756 * the notifier in the perf_event subsystem, though.
7758 static struct notifier_block __cpuinitdata migration_notifier = {
7759 .notifier_call = migration_call,
7763 static int __init migration_init(void)
7765 void *cpu = (void *)(long)smp_processor_id();
7768 /* Start one for the boot CPU: */
7769 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
7770 BUG_ON(err == NOTIFY_BAD);
7771 migration_call(&migration_notifier, CPU_ONLINE, cpu);
7772 register_cpu_notifier(&migration_notifier);
7776 early_initcall(migration_init);
7781 #ifdef CONFIG_SCHED_DEBUG
7783 static __read_mostly int sched_domain_debug_enabled;
7785 static int __init sched_domain_debug_setup(char *str)
7787 sched_domain_debug_enabled = 1;
7791 early_param("sched_debug", sched_domain_debug_setup);
7793 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
7794 struct cpumask *groupmask)
7796 struct sched_group *group = sd->groups;
7799 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
7800 cpumask_clear(groupmask);
7802 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
7804 if (!(sd->flags & SD_LOAD_BALANCE)) {
7805 printk("does not load-balance\n");
7807 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
7812 printk(KERN_CONT "span %s level %s\n", str, sd->name);
7814 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
7815 printk(KERN_ERR "ERROR: domain->span does not contain "
7818 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
7819 printk(KERN_ERR "ERROR: domain->groups does not contain"
7823 printk(KERN_DEBUG "%*s groups:", level + 1, "");
7827 printk(KERN_ERR "ERROR: group is NULL\n");
7831 if (!group->cpu_power) {
7832 printk(KERN_CONT "\n");
7833 printk(KERN_ERR "ERROR: domain->cpu_power not "
7838 if (!cpumask_weight(sched_group_cpus(group))) {
7839 printk(KERN_CONT "\n");
7840 printk(KERN_ERR "ERROR: empty group\n");
7844 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
7845 printk(KERN_CONT "\n");
7846 printk(KERN_ERR "ERROR: repeated CPUs\n");
7850 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
7852 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
7854 printk(KERN_CONT " %s", str);
7855 if (group->cpu_power != SCHED_LOAD_SCALE) {
7856 printk(KERN_CONT " (cpu_power = %d)",
7860 group = group->next;
7861 } while (group != sd->groups);
7862 printk(KERN_CONT "\n");
7864 if (!cpumask_equal(sched_domain_span(sd), groupmask))
7865 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
7868 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
7869 printk(KERN_ERR "ERROR: parent span is not a superset "
7870 "of domain->span\n");
7874 static void sched_domain_debug(struct sched_domain *sd, int cpu)
7876 cpumask_var_t groupmask;
7879 if (!sched_domain_debug_enabled)
7883 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7887 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7889 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
7890 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
7895 if (sched_domain_debug_one(sd, cpu, level, groupmask))
7902 free_cpumask_var(groupmask);
7904 #else /* !CONFIG_SCHED_DEBUG */
7905 # define sched_domain_debug(sd, cpu) do { } while (0)
7906 #endif /* CONFIG_SCHED_DEBUG */
7908 static int sd_degenerate(struct sched_domain *sd)
7910 if (cpumask_weight(sched_domain_span(sd)) == 1)
7913 /* Following flags need at least 2 groups */
7914 if (sd->flags & (SD_LOAD_BALANCE |
7915 SD_BALANCE_NEWIDLE |
7919 SD_SHARE_PKG_RESOURCES)) {
7920 if (sd->groups != sd->groups->next)
7924 /* Following flags don't use groups */
7925 if (sd->flags & (SD_WAKE_AFFINE))
7932 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
7934 unsigned long cflags = sd->flags, pflags = parent->flags;
7936 if (sd_degenerate(parent))
7939 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
7942 /* Flags needing groups don't count if only 1 group in parent */
7943 if (parent->groups == parent->groups->next) {
7944 pflags &= ~(SD_LOAD_BALANCE |
7945 SD_BALANCE_NEWIDLE |
7949 SD_SHARE_PKG_RESOURCES);
7950 if (nr_node_ids == 1)
7951 pflags &= ~SD_SERIALIZE;
7953 if (~cflags & pflags)
7959 static void free_rootdomain(struct root_domain *rd)
7961 synchronize_sched();
7963 cpupri_cleanup(&rd->cpupri);
7965 free_cpumask_var(rd->rto_mask);
7966 free_cpumask_var(rd->online);
7967 free_cpumask_var(rd->span);
7971 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7973 struct root_domain *old_rd = NULL;
7974 unsigned long flags;
7976 spin_lock_irqsave(&rq->lock, flags);
7981 if (cpumask_test_cpu(rq->cpu, old_rd->online))
7984 cpumask_clear_cpu(rq->cpu, old_rd->span);
7987 * If we dont want to free the old_rt yet then
7988 * set old_rd to NULL to skip the freeing later
7991 if (!atomic_dec_and_test(&old_rd->refcount))
7995 atomic_inc(&rd->refcount);
7998 cpumask_set_cpu(rq->cpu, rd->span);
7999 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
8002 spin_unlock_irqrestore(&rq->lock, flags);
8005 free_rootdomain(old_rd);
8008 static int init_rootdomain(struct root_domain *rd, bool bootmem)
8010 gfp_t gfp = GFP_KERNEL;
8012 memset(rd, 0, sizeof(*rd));
8017 if (!alloc_cpumask_var(&rd->span, gfp))
8019 if (!alloc_cpumask_var(&rd->online, gfp))
8021 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
8024 if (cpupri_init(&rd->cpupri, bootmem) != 0)
8029 free_cpumask_var(rd->rto_mask);
8031 free_cpumask_var(rd->online);
8033 free_cpumask_var(rd->span);
8038 static void init_defrootdomain(void)
8040 init_rootdomain(&def_root_domain, true);
8042 atomic_set(&def_root_domain.refcount, 1);
8045 static struct root_domain *alloc_rootdomain(void)
8047 struct root_domain *rd;
8049 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
8053 if (init_rootdomain(rd, false) != 0) {
8062 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
8063 * hold the hotplug lock.
8066 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
8068 struct rq *rq = cpu_rq(cpu);
8069 struct sched_domain *tmp;
8071 /* Remove the sched domains which do not contribute to scheduling. */
8072 for (tmp = sd; tmp; ) {
8073 struct sched_domain *parent = tmp->parent;
8077 if (sd_parent_degenerate(tmp, parent)) {
8078 tmp->parent = parent->parent;
8080 parent->parent->child = tmp;
8085 if (sd && sd_degenerate(sd)) {
8091 sched_domain_debug(sd, cpu);
8093 rq_attach_root(rq, rd);
8094 rcu_assign_pointer(rq->sd, sd);
8097 /* cpus with isolated domains */
8098 static cpumask_var_t cpu_isolated_map;
8100 /* Setup the mask of cpus configured for isolated domains */
8101 static int __init isolated_cpu_setup(char *str)
8103 alloc_bootmem_cpumask_var(&cpu_isolated_map);
8104 cpulist_parse(str, cpu_isolated_map);
8108 __setup("isolcpus=", isolated_cpu_setup);
8111 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
8112 * to a function which identifies what group(along with sched group) a CPU
8113 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
8114 * (due to the fact that we keep track of groups covered with a struct cpumask).
8116 * init_sched_build_groups will build a circular linked list of the groups
8117 * covered by the given span, and will set each group's ->cpumask correctly,
8118 * and ->cpu_power to 0.
8121 init_sched_build_groups(const struct cpumask *span,
8122 const struct cpumask *cpu_map,
8123 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
8124 struct sched_group **sg,
8125 struct cpumask *tmpmask),
8126 struct cpumask *covered, struct cpumask *tmpmask)
8128 struct sched_group *first = NULL, *last = NULL;
8131 cpumask_clear(covered);
8133 for_each_cpu(i, span) {
8134 struct sched_group *sg;
8135 int group = group_fn(i, cpu_map, &sg, tmpmask);
8138 if (cpumask_test_cpu(i, covered))
8141 cpumask_clear(sched_group_cpus(sg));
8144 for_each_cpu(j, span) {
8145 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
8148 cpumask_set_cpu(j, covered);
8149 cpumask_set_cpu(j, sched_group_cpus(sg));
8160 #define SD_NODES_PER_DOMAIN 16
8165 * find_next_best_node - find the next node to include in a sched_domain
8166 * @node: node whose sched_domain we're building
8167 * @used_nodes: nodes already in the sched_domain
8169 * Find the next node to include in a given scheduling domain. Simply
8170 * finds the closest node not already in the @used_nodes map.
8172 * Should use nodemask_t.
8174 static int find_next_best_node(int node, nodemask_t *used_nodes)
8176 int i, n, val, min_val, best_node = 0;
8180 for (i = 0; i < nr_node_ids; i++) {
8181 /* Start at @node */
8182 n = (node + i) % nr_node_ids;
8184 if (!nr_cpus_node(n))
8187 /* Skip already used nodes */
8188 if (node_isset(n, *used_nodes))
8191 /* Simple min distance search */
8192 val = node_distance(node, n);
8194 if (val < min_val) {
8200 node_set(best_node, *used_nodes);
8205 * sched_domain_node_span - get a cpumask for a node's sched_domain
8206 * @node: node whose cpumask we're constructing
8207 * @span: resulting cpumask
8209 * Given a node, construct a good cpumask for its sched_domain to span. It
8210 * should be one that prevents unnecessary balancing, but also spreads tasks
8213 static void sched_domain_node_span(int node, struct cpumask *span)
8215 nodemask_t used_nodes;
8218 cpumask_clear(span);
8219 nodes_clear(used_nodes);
8221 cpumask_or(span, span, cpumask_of_node(node));
8222 node_set(node, used_nodes);
8224 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
8225 int next_node = find_next_best_node(node, &used_nodes);
8227 cpumask_or(span, span, cpumask_of_node(next_node));
8230 #endif /* CONFIG_NUMA */
8232 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
8235 * The cpus mask in sched_group and sched_domain hangs off the end.
8237 * ( See the the comments in include/linux/sched.h:struct sched_group
8238 * and struct sched_domain. )
8240 struct static_sched_group {
8241 struct sched_group sg;
8242 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
8245 struct static_sched_domain {
8246 struct sched_domain sd;
8247 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
8253 cpumask_var_t domainspan;
8254 cpumask_var_t covered;
8255 cpumask_var_t notcovered;
8257 cpumask_var_t nodemask;
8258 cpumask_var_t this_sibling_map;
8259 cpumask_var_t this_core_map;
8260 cpumask_var_t send_covered;
8261 cpumask_var_t tmpmask;
8262 struct sched_group **sched_group_nodes;
8263 struct root_domain *rd;
8267 sa_sched_groups = 0,
8272 sa_this_sibling_map,
8274 sa_sched_group_nodes,
8284 * SMT sched-domains:
8286 #ifdef CONFIG_SCHED_SMT
8287 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
8288 static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
8291 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
8292 struct sched_group **sg, struct cpumask *unused)
8295 *sg = &per_cpu(sched_group_cpus, cpu).sg;
8298 #endif /* CONFIG_SCHED_SMT */
8301 * multi-core sched-domains:
8303 #ifdef CONFIG_SCHED_MC
8304 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
8305 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
8306 #endif /* CONFIG_SCHED_MC */
8308 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8310 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8311 struct sched_group **sg, struct cpumask *mask)
8315 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8316 group = cpumask_first(mask);
8318 *sg = &per_cpu(sched_group_core, group).sg;
8321 #elif defined(CONFIG_SCHED_MC)
8323 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8324 struct sched_group **sg, struct cpumask *unused)
8327 *sg = &per_cpu(sched_group_core, cpu).sg;
8332 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
8333 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
8336 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
8337 struct sched_group **sg, struct cpumask *mask)
8340 #ifdef CONFIG_SCHED_MC
8341 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
8342 group = cpumask_first(mask);
8343 #elif defined(CONFIG_SCHED_SMT)
8344 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8345 group = cpumask_first(mask);
8350 *sg = &per_cpu(sched_group_phys, group).sg;
8356 * The init_sched_build_groups can't handle what we want to do with node
8357 * groups, so roll our own. Now each node has its own list of groups which
8358 * gets dynamically allocated.
8360 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
8361 static struct sched_group ***sched_group_nodes_bycpu;
8363 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
8364 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
8366 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
8367 struct sched_group **sg,
8368 struct cpumask *nodemask)
8372 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
8373 group = cpumask_first(nodemask);
8376 *sg = &per_cpu(sched_group_allnodes, group).sg;
8380 static void init_numa_sched_groups_power(struct sched_group *group_head)
8382 struct sched_group *sg = group_head;
8388 for_each_cpu(j, sched_group_cpus(sg)) {
8389 struct sched_domain *sd;
8391 sd = &per_cpu(phys_domains, j).sd;
8392 if (j != group_first_cpu(sd->groups)) {
8394 * Only add "power" once for each
8400 sg->cpu_power += sd->groups->cpu_power;
8403 } while (sg != group_head);
8406 static int build_numa_sched_groups(struct s_data *d,
8407 const struct cpumask *cpu_map, int num)
8409 struct sched_domain *sd;
8410 struct sched_group *sg, *prev;
8413 cpumask_clear(d->covered);
8414 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
8415 if (cpumask_empty(d->nodemask)) {
8416 d->sched_group_nodes[num] = NULL;
8420 sched_domain_node_span(num, d->domainspan);
8421 cpumask_and(d->domainspan, d->domainspan, cpu_map);
8423 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8426 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
8430 d->sched_group_nodes[num] = sg;
8432 for_each_cpu(j, d->nodemask) {
8433 sd = &per_cpu(node_domains, j).sd;
8438 cpumask_copy(sched_group_cpus(sg), d->nodemask);
8440 cpumask_or(d->covered, d->covered, d->nodemask);
8443 for (j = 0; j < nr_node_ids; j++) {
8444 n = (num + j) % nr_node_ids;
8445 cpumask_complement(d->notcovered, d->covered);
8446 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
8447 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
8448 if (cpumask_empty(d->tmpmask))
8450 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
8451 if (cpumask_empty(d->tmpmask))
8453 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8457 "Can not alloc domain group for node %d\n", j);
8461 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
8462 sg->next = prev->next;
8463 cpumask_or(d->covered, d->covered, d->tmpmask);
8470 #endif /* CONFIG_NUMA */
8473 /* Free memory allocated for various sched_group structures */
8474 static void free_sched_groups(const struct cpumask *cpu_map,
8475 struct cpumask *nodemask)
8479 for_each_cpu(cpu, cpu_map) {
8480 struct sched_group **sched_group_nodes
8481 = sched_group_nodes_bycpu[cpu];
8483 if (!sched_group_nodes)
8486 for (i = 0; i < nr_node_ids; i++) {
8487 struct sched_group *oldsg, *sg = sched_group_nodes[i];
8489 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8490 if (cpumask_empty(nodemask))
8500 if (oldsg != sched_group_nodes[i])
8503 kfree(sched_group_nodes);
8504 sched_group_nodes_bycpu[cpu] = NULL;
8507 #else /* !CONFIG_NUMA */
8508 static void free_sched_groups(const struct cpumask *cpu_map,
8509 struct cpumask *nodemask)
8512 #endif /* CONFIG_NUMA */
8515 * Initialize sched groups cpu_power.
8517 * cpu_power indicates the capacity of sched group, which is used while
8518 * distributing the load between different sched groups in a sched domain.
8519 * Typically cpu_power for all the groups in a sched domain will be same unless
8520 * there are asymmetries in the topology. If there are asymmetries, group
8521 * having more cpu_power will pickup more load compared to the group having
8524 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
8526 struct sched_domain *child;
8527 struct sched_group *group;
8531 WARN_ON(!sd || !sd->groups);
8533 if (cpu != group_first_cpu(sd->groups))
8538 sd->groups->cpu_power = 0;
8541 power = SCHED_LOAD_SCALE;
8542 weight = cpumask_weight(sched_domain_span(sd));
8544 * SMT siblings share the power of a single core.
8545 * Usually multiple threads get a better yield out of
8546 * that one core than a single thread would have,
8547 * reflect that in sd->smt_gain.
8549 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
8550 power *= sd->smt_gain;
8552 power >>= SCHED_LOAD_SHIFT;
8554 sd->groups->cpu_power += power;
8559 * Add cpu_power of each child group to this groups cpu_power.
8561 group = child->groups;
8563 sd->groups->cpu_power += group->cpu_power;
8564 group = group->next;
8565 } while (group != child->groups);
8569 * Initializers for schedule domains
8570 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8573 #ifdef CONFIG_SCHED_DEBUG
8574 # define SD_INIT_NAME(sd, type) sd->name = #type
8576 # define SD_INIT_NAME(sd, type) do { } while (0)
8579 #define SD_INIT(sd, type) sd_init_##type(sd)
8581 #define SD_INIT_FUNC(type) \
8582 static noinline void sd_init_##type(struct sched_domain *sd) \
8584 memset(sd, 0, sizeof(*sd)); \
8585 *sd = SD_##type##_INIT; \
8586 sd->level = SD_LV_##type; \
8587 SD_INIT_NAME(sd, type); \
8592 SD_INIT_FUNC(ALLNODES)
8595 #ifdef CONFIG_SCHED_SMT
8596 SD_INIT_FUNC(SIBLING)
8598 #ifdef CONFIG_SCHED_MC
8602 static int default_relax_domain_level = -1;
8604 static int __init setup_relax_domain_level(char *str)
8608 val = simple_strtoul(str, NULL, 0);
8609 if (val < SD_LV_MAX)
8610 default_relax_domain_level = val;
8614 __setup("relax_domain_level=", setup_relax_domain_level);
8616 static void set_domain_attribute(struct sched_domain *sd,
8617 struct sched_domain_attr *attr)
8621 if (!attr || attr->relax_domain_level < 0) {
8622 if (default_relax_domain_level < 0)
8625 request = default_relax_domain_level;
8627 request = attr->relax_domain_level;
8628 if (request < sd->level) {
8629 /* turn off idle balance on this domain */
8630 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
8632 /* turn on idle balance on this domain */
8633 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
8637 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
8638 const struct cpumask *cpu_map)
8641 case sa_sched_groups:
8642 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
8643 d->sched_group_nodes = NULL;
8645 free_rootdomain(d->rd); /* fall through */
8647 free_cpumask_var(d->tmpmask); /* fall through */
8648 case sa_send_covered:
8649 free_cpumask_var(d->send_covered); /* fall through */
8650 case sa_this_core_map:
8651 free_cpumask_var(d->this_core_map); /* fall through */
8652 case sa_this_sibling_map:
8653 free_cpumask_var(d->this_sibling_map); /* fall through */
8655 free_cpumask_var(d->nodemask); /* fall through */
8656 case sa_sched_group_nodes:
8658 kfree(d->sched_group_nodes); /* fall through */
8660 free_cpumask_var(d->notcovered); /* fall through */
8662 free_cpumask_var(d->covered); /* fall through */
8664 free_cpumask_var(d->domainspan); /* fall through */
8671 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
8672 const struct cpumask *cpu_map)
8675 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
8677 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
8678 return sa_domainspan;
8679 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
8681 /* Allocate the per-node list of sched groups */
8682 d->sched_group_nodes = kcalloc(nr_node_ids,
8683 sizeof(struct sched_group *), GFP_KERNEL);
8684 if (!d->sched_group_nodes) {
8685 printk(KERN_WARNING "Can not alloc sched group node list\n");
8686 return sa_notcovered;
8688 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
8690 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
8691 return sa_sched_group_nodes;
8692 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
8694 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
8695 return sa_this_sibling_map;
8696 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
8697 return sa_this_core_map;
8698 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
8699 return sa_send_covered;
8700 d->rd = alloc_rootdomain();
8702 printk(KERN_WARNING "Cannot alloc root domain\n");
8705 return sa_rootdomain;
8708 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
8709 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
8711 struct sched_domain *sd = NULL;
8713 struct sched_domain *parent;
8716 if (cpumask_weight(cpu_map) >
8717 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
8718 sd = &per_cpu(allnodes_domains, i).sd;
8719 SD_INIT(sd, ALLNODES);
8720 set_domain_attribute(sd, attr);
8721 cpumask_copy(sched_domain_span(sd), cpu_map);
8722 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
8727 sd = &per_cpu(node_domains, i).sd;
8729 set_domain_attribute(sd, attr);
8730 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
8731 sd->parent = parent;
8734 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
8739 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
8740 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8741 struct sched_domain *parent, int i)
8743 struct sched_domain *sd;
8744 sd = &per_cpu(phys_domains, i).sd;
8746 set_domain_attribute(sd, attr);
8747 cpumask_copy(sched_domain_span(sd), d->nodemask);
8748 sd->parent = parent;
8751 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
8755 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
8756 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8757 struct sched_domain *parent, int i)
8759 struct sched_domain *sd = parent;
8760 #ifdef CONFIG_SCHED_MC
8761 sd = &per_cpu(core_domains, i).sd;
8763 set_domain_attribute(sd, attr);
8764 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
8765 sd->parent = parent;
8767 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
8772 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
8773 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8774 struct sched_domain *parent, int i)
8776 struct sched_domain *sd = parent;
8777 #ifdef CONFIG_SCHED_SMT
8778 sd = &per_cpu(cpu_domains, i).sd;
8779 SD_INIT(sd, SIBLING);
8780 set_domain_attribute(sd, attr);
8781 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
8782 sd->parent = parent;
8784 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
8789 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
8790 const struct cpumask *cpu_map, int cpu)
8793 #ifdef CONFIG_SCHED_SMT
8794 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
8795 cpumask_and(d->this_sibling_map, cpu_map,
8796 topology_thread_cpumask(cpu));
8797 if (cpu == cpumask_first(d->this_sibling_map))
8798 init_sched_build_groups(d->this_sibling_map, cpu_map,
8800 d->send_covered, d->tmpmask);
8803 #ifdef CONFIG_SCHED_MC
8804 case SD_LV_MC: /* set up multi-core groups */
8805 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
8806 if (cpu == cpumask_first(d->this_core_map))
8807 init_sched_build_groups(d->this_core_map, cpu_map,
8809 d->send_covered, d->tmpmask);
8812 case SD_LV_CPU: /* set up physical groups */
8813 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
8814 if (!cpumask_empty(d->nodemask))
8815 init_sched_build_groups(d->nodemask, cpu_map,
8817 d->send_covered, d->tmpmask);
8820 case SD_LV_ALLNODES:
8821 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
8822 d->send_covered, d->tmpmask);
8831 * Build sched domains for a given set of cpus and attach the sched domains
8832 * to the individual cpus
8834 static int __build_sched_domains(const struct cpumask *cpu_map,
8835 struct sched_domain_attr *attr)
8837 enum s_alloc alloc_state = sa_none;
8839 struct sched_domain *sd;
8845 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
8846 if (alloc_state != sa_rootdomain)
8848 alloc_state = sa_sched_groups;
8851 * Set up domains for cpus specified by the cpu_map.
8853 for_each_cpu(i, cpu_map) {
8854 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
8857 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
8858 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
8859 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
8860 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
8863 for_each_cpu(i, cpu_map) {
8864 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
8865 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
8868 /* Set up physical groups */
8869 for (i = 0; i < nr_node_ids; i++)
8870 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
8873 /* Set up node groups */
8875 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
8877 for (i = 0; i < nr_node_ids; i++)
8878 if (build_numa_sched_groups(&d, cpu_map, i))
8882 /* Calculate CPU power for physical packages and nodes */
8883 #ifdef CONFIG_SCHED_SMT
8884 for_each_cpu(i, cpu_map) {
8885 sd = &per_cpu(cpu_domains, i).sd;
8886 init_sched_groups_power(i, sd);
8889 #ifdef CONFIG_SCHED_MC
8890 for_each_cpu(i, cpu_map) {
8891 sd = &per_cpu(core_domains, i).sd;
8892 init_sched_groups_power(i, sd);
8896 for_each_cpu(i, cpu_map) {
8897 sd = &per_cpu(phys_domains, i).sd;
8898 init_sched_groups_power(i, sd);
8902 for (i = 0; i < nr_node_ids; i++)
8903 init_numa_sched_groups_power(d.sched_group_nodes[i]);
8905 if (d.sd_allnodes) {
8906 struct sched_group *sg;
8908 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
8910 init_numa_sched_groups_power(sg);
8914 /* Attach the domains */
8915 for_each_cpu(i, cpu_map) {
8916 #ifdef CONFIG_SCHED_SMT
8917 sd = &per_cpu(cpu_domains, i).sd;
8918 #elif defined(CONFIG_SCHED_MC)
8919 sd = &per_cpu(core_domains, i).sd;
8921 sd = &per_cpu(phys_domains, i).sd;
8923 cpu_attach_domain(sd, d.rd, i);
8926 d.sched_group_nodes = NULL; /* don't free this we still need it */
8927 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
8931 __free_domain_allocs(&d, alloc_state, cpu_map);
8935 static int build_sched_domains(const struct cpumask *cpu_map)
8937 return __build_sched_domains(cpu_map, NULL);
8940 static cpumask_var_t *doms_cur; /* current sched domains */
8941 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
8942 static struct sched_domain_attr *dattr_cur;
8943 /* attribues of custom domains in 'doms_cur' */
8946 * Special case: If a kmalloc of a doms_cur partition (array of
8947 * cpumask) fails, then fallback to a single sched domain,
8948 * as determined by the single cpumask fallback_doms.
8950 static cpumask_var_t fallback_doms;
8953 * arch_update_cpu_topology lets virtualized architectures update the
8954 * cpu core maps. It is supposed to return 1 if the topology changed
8955 * or 0 if it stayed the same.
8957 int __attribute__((weak)) arch_update_cpu_topology(void)
8962 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
8965 cpumask_var_t *doms;
8967 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
8970 for (i = 0; i < ndoms; i++) {
8971 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
8972 free_sched_domains(doms, i);
8979 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
8982 for (i = 0; i < ndoms; i++)
8983 free_cpumask_var(doms[i]);
8988 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
8989 * For now this just excludes isolated cpus, but could be used to
8990 * exclude other special cases in the future.
8992 static int arch_init_sched_domains(const struct cpumask *cpu_map)
8996 arch_update_cpu_topology();
8998 doms_cur = alloc_sched_domains(ndoms_cur);
9000 doms_cur = &fallback_doms;
9001 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
9003 err = build_sched_domains(doms_cur[0]);
9004 register_sched_domain_sysctl();
9009 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
9010 struct cpumask *tmpmask)
9012 free_sched_groups(cpu_map, tmpmask);
9016 * Detach sched domains from a group of cpus specified in cpu_map
9017 * These cpus will now be attached to the NULL domain
9019 static void detach_destroy_domains(const struct cpumask *cpu_map)
9021 /* Save because hotplug lock held. */
9022 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
9025 for_each_cpu(i, cpu_map)
9026 cpu_attach_domain(NULL, &def_root_domain, i);
9027 synchronize_sched();
9028 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
9031 /* handle null as "default" */
9032 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
9033 struct sched_domain_attr *new, int idx_new)
9035 struct sched_domain_attr tmp;
9042 return !memcmp(cur ? (cur + idx_cur) : &tmp,
9043 new ? (new + idx_new) : &tmp,
9044 sizeof(struct sched_domain_attr));
9048 * Partition sched domains as specified by the 'ndoms_new'
9049 * cpumasks in the array doms_new[] of cpumasks. This compares
9050 * doms_new[] to the current sched domain partitioning, doms_cur[].
9051 * It destroys each deleted domain and builds each new domain.
9053 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
9054 * The masks don't intersect (don't overlap.) We should setup one
9055 * sched domain for each mask. CPUs not in any of the cpumasks will
9056 * not be load balanced. If the same cpumask appears both in the
9057 * current 'doms_cur' domains and in the new 'doms_new', we can leave
9060 * The passed in 'doms_new' should be allocated using
9061 * alloc_sched_domains. This routine takes ownership of it and will
9062 * free_sched_domains it when done with it. If the caller failed the
9063 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
9064 * and partition_sched_domains() will fallback to the single partition
9065 * 'fallback_doms', it also forces the domains to be rebuilt.
9067 * If doms_new == NULL it will be replaced with cpu_online_mask.
9068 * ndoms_new == 0 is a special case for destroying existing domains,
9069 * and it will not create the default domain.
9071 * Call with hotplug lock held
9073 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
9074 struct sched_domain_attr *dattr_new)
9079 mutex_lock(&sched_domains_mutex);
9081 /* always unregister in case we don't destroy any domains */
9082 unregister_sched_domain_sysctl();
9084 /* Let architecture update cpu core mappings. */
9085 new_topology = arch_update_cpu_topology();
9087 n = doms_new ? ndoms_new : 0;
9089 /* Destroy deleted domains */
9090 for (i = 0; i < ndoms_cur; i++) {
9091 for (j = 0; j < n && !new_topology; j++) {
9092 if (cpumask_equal(doms_cur[i], doms_new[j])
9093 && dattrs_equal(dattr_cur, i, dattr_new, j))
9096 /* no match - a current sched domain not in new doms_new[] */
9097 detach_destroy_domains(doms_cur[i]);
9102 if (doms_new == NULL) {
9104 doms_new = &fallback_doms;
9105 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
9106 WARN_ON_ONCE(dattr_new);
9109 /* Build new domains */
9110 for (i = 0; i < ndoms_new; i++) {
9111 for (j = 0; j < ndoms_cur && !new_topology; j++) {
9112 if (cpumask_equal(doms_new[i], doms_cur[j])
9113 && dattrs_equal(dattr_new, i, dattr_cur, j))
9116 /* no match - add a new doms_new */
9117 __build_sched_domains(doms_new[i],
9118 dattr_new ? dattr_new + i : NULL);
9123 /* Remember the new sched domains */
9124 if (doms_cur != &fallback_doms)
9125 free_sched_domains(doms_cur, ndoms_cur);
9126 kfree(dattr_cur); /* kfree(NULL) is safe */
9127 doms_cur = doms_new;
9128 dattr_cur = dattr_new;
9129 ndoms_cur = ndoms_new;
9131 register_sched_domain_sysctl();
9133 mutex_unlock(&sched_domains_mutex);
9136 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
9137 static void arch_reinit_sched_domains(void)
9141 /* Destroy domains first to force the rebuild */
9142 partition_sched_domains(0, NULL, NULL);
9144 rebuild_sched_domains();
9148 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
9150 unsigned int level = 0;
9152 if (sscanf(buf, "%u", &level) != 1)
9156 * level is always be positive so don't check for
9157 * level < POWERSAVINGS_BALANCE_NONE which is 0
9158 * What happens on 0 or 1 byte write,
9159 * need to check for count as well?
9162 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
9166 sched_smt_power_savings = level;
9168 sched_mc_power_savings = level;
9170 arch_reinit_sched_domains();
9175 #ifdef CONFIG_SCHED_MC
9176 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
9179 return sprintf(page, "%u\n", sched_mc_power_savings);
9181 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
9182 const char *buf, size_t count)
9184 return sched_power_savings_store(buf, count, 0);
9186 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
9187 sched_mc_power_savings_show,
9188 sched_mc_power_savings_store);
9191 #ifdef CONFIG_SCHED_SMT
9192 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
9195 return sprintf(page, "%u\n", sched_smt_power_savings);
9197 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
9198 const char *buf, size_t count)
9200 return sched_power_savings_store(buf, count, 1);
9202 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
9203 sched_smt_power_savings_show,
9204 sched_smt_power_savings_store);
9207 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
9211 #ifdef CONFIG_SCHED_SMT
9213 err = sysfs_create_file(&cls->kset.kobj,
9214 &attr_sched_smt_power_savings.attr);
9216 #ifdef CONFIG_SCHED_MC
9217 if (!err && mc_capable())
9218 err = sysfs_create_file(&cls->kset.kobj,
9219 &attr_sched_mc_power_savings.attr);
9223 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
9225 #ifndef CONFIG_CPUSETS
9227 * Add online and remove offline CPUs from the scheduler domains.
9228 * When cpusets are enabled they take over this function.
9230 static int update_sched_domains(struct notifier_block *nfb,
9231 unsigned long action, void *hcpu)
9235 case CPU_ONLINE_FROZEN:
9236 case CPU_DOWN_PREPARE:
9237 case CPU_DOWN_PREPARE_FROZEN:
9238 case CPU_DOWN_FAILED:
9239 case CPU_DOWN_FAILED_FROZEN:
9240 partition_sched_domains(1, NULL, NULL);
9249 static int update_runtime(struct notifier_block *nfb,
9250 unsigned long action, void *hcpu)
9252 int cpu = (int)(long)hcpu;
9255 case CPU_DOWN_PREPARE:
9256 case CPU_DOWN_PREPARE_FROZEN:
9257 disable_runtime(cpu_rq(cpu));
9260 case CPU_DOWN_FAILED:
9261 case CPU_DOWN_FAILED_FROZEN:
9263 case CPU_ONLINE_FROZEN:
9264 enable_runtime(cpu_rq(cpu));
9272 void __init sched_init_smp(void)
9274 cpumask_var_t non_isolated_cpus;
9276 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
9277 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
9279 #if defined(CONFIG_NUMA)
9280 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
9282 BUG_ON(sched_group_nodes_bycpu == NULL);
9285 mutex_lock(&sched_domains_mutex);
9286 arch_init_sched_domains(cpu_active_mask);
9287 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
9288 if (cpumask_empty(non_isolated_cpus))
9289 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
9290 mutex_unlock(&sched_domains_mutex);
9293 #ifndef CONFIG_CPUSETS
9294 /* XXX: Theoretical race here - CPU may be hotplugged now */
9295 hotcpu_notifier(update_sched_domains, 0);
9298 /* RT runtime code needs to handle some hotplug events */
9299 hotcpu_notifier(update_runtime, 0);
9303 /* Move init over to a non-isolated CPU */
9304 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
9306 sched_init_granularity();
9307 free_cpumask_var(non_isolated_cpus);
9309 init_sched_rt_class();
9312 void __init sched_init_smp(void)
9314 sched_init_granularity();
9316 #endif /* CONFIG_SMP */
9318 const_debug unsigned int sysctl_timer_migration = 1;
9320 int in_sched_functions(unsigned long addr)
9322 return in_lock_functions(addr) ||
9323 (addr >= (unsigned long)__sched_text_start
9324 && addr < (unsigned long)__sched_text_end);
9327 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
9329 cfs_rq->tasks_timeline = RB_ROOT;
9330 INIT_LIST_HEAD(&cfs_rq->tasks);
9331 #ifdef CONFIG_FAIR_GROUP_SCHED
9334 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9337 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
9339 struct rt_prio_array *array;
9342 array = &rt_rq->active;
9343 for (i = 0; i < MAX_RT_PRIO; i++) {
9344 INIT_LIST_HEAD(array->queue + i);
9345 __clear_bit(i, array->bitmap);
9347 /* delimiter for bitsearch: */
9348 __set_bit(MAX_RT_PRIO, array->bitmap);
9350 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9351 rt_rq->highest_prio.curr = MAX_RT_PRIO;
9353 rt_rq->highest_prio.next = MAX_RT_PRIO;
9357 rt_rq->rt_nr_migratory = 0;
9358 rt_rq->overloaded = 0;
9359 plist_head_init(&rt_rq->pushable_tasks, &rq->lock);
9363 rt_rq->rt_throttled = 0;
9364 rt_rq->rt_runtime = 0;
9365 spin_lock_init(&rt_rq->rt_runtime_lock);
9367 #ifdef CONFIG_RT_GROUP_SCHED
9368 rt_rq->rt_nr_boosted = 0;
9373 #ifdef CONFIG_FAIR_GROUP_SCHED
9374 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9375 struct sched_entity *se, int cpu, int add,
9376 struct sched_entity *parent)
9378 struct rq *rq = cpu_rq(cpu);
9379 tg->cfs_rq[cpu] = cfs_rq;
9380 init_cfs_rq(cfs_rq, rq);
9383 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
9386 /* se could be NULL for init_task_group */
9391 se->cfs_rq = &rq->cfs;
9393 se->cfs_rq = parent->my_q;
9396 se->load.weight = tg->shares;
9397 se->load.inv_weight = 0;
9398 se->parent = parent;
9402 #ifdef CONFIG_RT_GROUP_SCHED
9403 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
9404 struct sched_rt_entity *rt_se, int cpu, int add,
9405 struct sched_rt_entity *parent)
9407 struct rq *rq = cpu_rq(cpu);
9409 tg->rt_rq[cpu] = rt_rq;
9410 init_rt_rq(rt_rq, rq);
9412 rt_rq->rt_se = rt_se;
9413 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
9415 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
9417 tg->rt_se[cpu] = rt_se;
9422 rt_se->rt_rq = &rq->rt;
9424 rt_se->rt_rq = parent->my_q;
9426 rt_se->my_q = rt_rq;
9427 rt_se->parent = parent;
9428 INIT_LIST_HEAD(&rt_se->run_list);
9432 void __init sched_init(void)
9435 unsigned long alloc_size = 0, ptr;
9437 #ifdef CONFIG_FAIR_GROUP_SCHED
9438 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9440 #ifdef CONFIG_RT_GROUP_SCHED
9441 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9443 #ifdef CONFIG_USER_SCHED
9446 #ifdef CONFIG_CPUMASK_OFFSTACK
9447 alloc_size += num_possible_cpus() * cpumask_size();
9450 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
9452 #ifdef CONFIG_FAIR_GROUP_SCHED
9453 init_task_group.se = (struct sched_entity **)ptr;
9454 ptr += nr_cpu_ids * sizeof(void **);
9456 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
9457 ptr += nr_cpu_ids * sizeof(void **);
9459 #ifdef CONFIG_USER_SCHED
9460 root_task_group.se = (struct sched_entity **)ptr;
9461 ptr += nr_cpu_ids * sizeof(void **);
9463 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9464 ptr += nr_cpu_ids * sizeof(void **);
9465 #endif /* CONFIG_USER_SCHED */
9466 #endif /* CONFIG_FAIR_GROUP_SCHED */
9467 #ifdef CONFIG_RT_GROUP_SCHED
9468 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
9469 ptr += nr_cpu_ids * sizeof(void **);
9471 init_task_group.rt_rq = (struct rt_rq **)ptr;
9472 ptr += nr_cpu_ids * sizeof(void **);
9474 #ifdef CONFIG_USER_SCHED
9475 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9476 ptr += nr_cpu_ids * sizeof(void **);
9478 root_task_group.rt_rq = (struct rt_rq **)ptr;
9479 ptr += nr_cpu_ids * sizeof(void **);
9480 #endif /* CONFIG_USER_SCHED */
9481 #endif /* CONFIG_RT_GROUP_SCHED */
9482 #ifdef CONFIG_CPUMASK_OFFSTACK
9483 for_each_possible_cpu(i) {
9484 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
9485 ptr += cpumask_size();
9487 #endif /* CONFIG_CPUMASK_OFFSTACK */
9491 init_defrootdomain();
9494 init_rt_bandwidth(&def_rt_bandwidth,
9495 global_rt_period(), global_rt_runtime());
9497 #ifdef CONFIG_RT_GROUP_SCHED
9498 init_rt_bandwidth(&init_task_group.rt_bandwidth,
9499 global_rt_period(), global_rt_runtime());
9500 #ifdef CONFIG_USER_SCHED
9501 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9502 global_rt_period(), RUNTIME_INF);
9503 #endif /* CONFIG_USER_SCHED */
9504 #endif /* CONFIG_RT_GROUP_SCHED */
9506 #ifdef CONFIG_GROUP_SCHED
9507 list_add(&init_task_group.list, &task_groups);
9508 INIT_LIST_HEAD(&init_task_group.children);
9510 #ifdef CONFIG_USER_SCHED
9511 INIT_LIST_HEAD(&root_task_group.children);
9512 init_task_group.parent = &root_task_group;
9513 list_add(&init_task_group.siblings, &root_task_group.children);
9514 #endif /* CONFIG_USER_SCHED */
9515 #endif /* CONFIG_GROUP_SCHED */
9517 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
9518 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
9519 __alignof__(unsigned long));
9521 for_each_possible_cpu(i) {
9525 spin_lock_init(&rq->lock);
9527 rq->calc_load_active = 0;
9528 rq->calc_load_update = jiffies + LOAD_FREQ;
9529 init_cfs_rq(&rq->cfs, rq);
9530 init_rt_rq(&rq->rt, rq);
9531 #ifdef CONFIG_FAIR_GROUP_SCHED
9532 init_task_group.shares = init_task_group_load;
9533 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9534 #ifdef CONFIG_CGROUP_SCHED
9536 * How much cpu bandwidth does init_task_group get?
9538 * In case of task-groups formed thr' the cgroup filesystem, it
9539 * gets 100% of the cpu resources in the system. This overall
9540 * system cpu resource is divided among the tasks of
9541 * init_task_group and its child task-groups in a fair manner,
9542 * based on each entity's (task or task-group's) weight
9543 * (se->load.weight).
9545 * In other words, if init_task_group has 10 tasks of weight
9546 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9547 * then A0's share of the cpu resource is:
9549 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9551 * We achieve this by letting init_task_group's tasks sit
9552 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9554 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
9555 #elif defined CONFIG_USER_SCHED
9556 root_task_group.shares = NICE_0_LOAD;
9557 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
9559 * In case of task-groups formed thr' the user id of tasks,
9560 * init_task_group represents tasks belonging to root user.
9561 * Hence it forms a sibling of all subsequent groups formed.
9562 * In this case, init_task_group gets only a fraction of overall
9563 * system cpu resource, based on the weight assigned to root
9564 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9565 * by letting tasks of init_task_group sit in a separate cfs_rq
9566 * (init_tg_cfs_rq) and having one entity represent this group of
9567 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9569 init_tg_cfs_entry(&init_task_group,
9570 &per_cpu(init_tg_cfs_rq, i),
9571 &per_cpu(init_sched_entity, i), i, 1,
9572 root_task_group.se[i]);
9575 #endif /* CONFIG_FAIR_GROUP_SCHED */
9577 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9578 #ifdef CONFIG_RT_GROUP_SCHED
9579 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
9580 #ifdef CONFIG_CGROUP_SCHED
9581 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
9582 #elif defined CONFIG_USER_SCHED
9583 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
9584 init_tg_rt_entry(&init_task_group,
9585 &per_cpu(init_rt_rq, i),
9586 &per_cpu(init_sched_rt_entity, i), i, 1,
9587 root_task_group.rt_se[i]);
9591 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
9592 rq->cpu_load[j] = 0;
9596 rq->post_schedule = 0;
9597 rq->active_balance = 0;
9598 rq->next_balance = jiffies;
9602 rq->migration_thread = NULL;
9604 rq->avg_idle = 2*sysctl_sched_migration_cost;
9605 INIT_LIST_HEAD(&rq->migration_queue);
9606 rq_attach_root(rq, &def_root_domain);
9609 atomic_set(&rq->nr_iowait, 0);
9612 set_load_weight(&init_task);
9614 #ifdef CONFIG_PREEMPT_NOTIFIERS
9615 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
9619 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9622 #ifdef CONFIG_RT_MUTEXES
9623 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
9627 * The boot idle thread does lazy MMU switching as well:
9629 atomic_inc(&init_mm.mm_count);
9630 enter_lazy_tlb(&init_mm, current);
9633 * Make us the idle thread. Technically, schedule() should not be
9634 * called from this thread, however somewhere below it might be,
9635 * but because we are the idle thread, we just pick up running again
9636 * when this runqueue becomes "idle".
9638 init_idle(current, smp_processor_id());
9640 calc_load_update = jiffies + LOAD_FREQ;
9643 * During early bootup we pretend to be a normal task:
9645 current->sched_class = &fair_sched_class;
9647 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9648 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
9651 zalloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
9652 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
9654 /* May be allocated at isolcpus cmdline parse time */
9655 if (cpu_isolated_map == NULL)
9656 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
9661 scheduler_running = 1;
9664 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9665 static inline int preempt_count_equals(int preempt_offset)
9667 int nested = preempt_count() & ~PREEMPT_ACTIVE;
9669 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
9672 void __might_sleep(char *file, int line, int preempt_offset)
9675 static unsigned long prev_jiffy; /* ratelimiting */
9677 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
9678 system_state != SYSTEM_RUNNING || oops_in_progress)
9680 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9682 prev_jiffy = jiffies;
9685 "BUG: sleeping function called from invalid context at %s:%d\n",
9688 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9689 in_atomic(), irqs_disabled(),
9690 current->pid, current->comm);
9692 debug_show_held_locks(current);
9693 if (irqs_disabled())
9694 print_irqtrace_events(current);
9698 EXPORT_SYMBOL(__might_sleep);
9701 #ifdef CONFIG_MAGIC_SYSRQ
9702 static void normalize_task(struct rq *rq, struct task_struct *p)
9706 update_rq_clock(rq);
9707 on_rq = p->se.on_rq;
9709 deactivate_task(rq, p, 0);
9710 __setscheduler(rq, p, SCHED_NORMAL, 0);
9712 activate_task(rq, p, 0);
9713 resched_task(rq->curr);
9717 void normalize_rt_tasks(void)
9719 struct task_struct *g, *p;
9720 unsigned long flags;
9723 read_lock_irqsave(&tasklist_lock, flags);
9724 do_each_thread(g, p) {
9726 * Only normalize user tasks:
9731 p->se.exec_start = 0;
9732 #ifdef CONFIG_SCHEDSTATS
9733 p->se.wait_start = 0;
9734 p->se.sleep_start = 0;
9735 p->se.block_start = 0;
9740 * Renice negative nice level userspace
9743 if (TASK_NICE(p) < 0 && p->mm)
9744 set_user_nice(p, 0);
9748 spin_lock(&p->pi_lock);
9749 rq = __task_rq_lock(p);
9751 normalize_task(rq, p);
9753 __task_rq_unlock(rq);
9754 spin_unlock(&p->pi_lock);
9755 } while_each_thread(g, p);
9757 read_unlock_irqrestore(&tasklist_lock, flags);
9760 #endif /* CONFIG_MAGIC_SYSRQ */
9764 * These functions are only useful for the IA64 MCA handling.
9766 * They can only be called when the whole system has been
9767 * stopped - every CPU needs to be quiescent, and no scheduling
9768 * activity can take place. Using them for anything else would
9769 * be a serious bug, and as a result, they aren't even visible
9770 * under any other configuration.
9774 * curr_task - return the current task for a given cpu.
9775 * @cpu: the processor in question.
9777 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9779 struct task_struct *curr_task(int cpu)
9781 return cpu_curr(cpu);
9785 * set_curr_task - set the current task for a given cpu.
9786 * @cpu: the processor in question.
9787 * @p: the task pointer to set.
9789 * Description: This function must only be used when non-maskable interrupts
9790 * are serviced on a separate stack. It allows the architecture to switch the
9791 * notion of the current task on a cpu in a non-blocking manner. This function
9792 * must be called with all CPU's synchronized, and interrupts disabled, the
9793 * and caller must save the original value of the current task (see
9794 * curr_task() above) and restore that value before reenabling interrupts and
9795 * re-starting the system.
9797 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9799 void set_curr_task(int cpu, struct task_struct *p)
9806 #ifdef CONFIG_FAIR_GROUP_SCHED
9807 static void free_fair_sched_group(struct task_group *tg)
9811 for_each_possible_cpu(i) {
9813 kfree(tg->cfs_rq[i]);
9823 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9825 struct cfs_rq *cfs_rq;
9826 struct sched_entity *se;
9830 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9833 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9837 tg->shares = NICE_0_LOAD;
9839 for_each_possible_cpu(i) {
9842 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9843 GFP_KERNEL, cpu_to_node(i));
9847 se = kzalloc_node(sizeof(struct sched_entity),
9848 GFP_KERNEL, cpu_to_node(i));
9852 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
9861 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9863 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
9864 &cpu_rq(cpu)->leaf_cfs_rq_list);
9867 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9869 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
9871 #else /* !CONFG_FAIR_GROUP_SCHED */
9872 static inline void free_fair_sched_group(struct task_group *tg)
9877 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9882 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9886 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9889 #endif /* CONFIG_FAIR_GROUP_SCHED */
9891 #ifdef CONFIG_RT_GROUP_SCHED
9892 static void free_rt_sched_group(struct task_group *tg)
9896 destroy_rt_bandwidth(&tg->rt_bandwidth);
9898 for_each_possible_cpu(i) {
9900 kfree(tg->rt_rq[i]);
9902 kfree(tg->rt_se[i]);
9910 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9912 struct rt_rq *rt_rq;
9913 struct sched_rt_entity *rt_se;
9917 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
9920 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
9924 init_rt_bandwidth(&tg->rt_bandwidth,
9925 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
9927 for_each_possible_cpu(i) {
9930 rt_rq = kzalloc_node(sizeof(struct rt_rq),
9931 GFP_KERNEL, cpu_to_node(i));
9935 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
9936 GFP_KERNEL, cpu_to_node(i));
9940 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
9949 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9951 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
9952 &cpu_rq(cpu)->leaf_rt_rq_list);
9955 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9957 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
9959 #else /* !CONFIG_RT_GROUP_SCHED */
9960 static inline void free_rt_sched_group(struct task_group *tg)
9965 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9970 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9974 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9977 #endif /* CONFIG_RT_GROUP_SCHED */
9979 #ifdef CONFIG_GROUP_SCHED
9980 static void free_sched_group(struct task_group *tg)
9982 free_fair_sched_group(tg);
9983 free_rt_sched_group(tg);
9987 /* allocate runqueue etc for a new task group */
9988 struct task_group *sched_create_group(struct task_group *parent)
9990 struct task_group *tg;
9991 unsigned long flags;
9994 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
9996 return ERR_PTR(-ENOMEM);
9998 if (!alloc_fair_sched_group(tg, parent))
10001 if (!alloc_rt_sched_group(tg, parent))
10004 spin_lock_irqsave(&task_group_lock, flags);
10005 for_each_possible_cpu(i) {
10006 register_fair_sched_group(tg, i);
10007 register_rt_sched_group(tg, i);
10009 list_add_rcu(&tg->list, &task_groups);
10011 WARN_ON(!parent); /* root should already exist */
10013 tg->parent = parent;
10014 INIT_LIST_HEAD(&tg->children);
10015 list_add_rcu(&tg->siblings, &parent->children);
10016 spin_unlock_irqrestore(&task_group_lock, flags);
10021 free_sched_group(tg);
10022 return ERR_PTR(-ENOMEM);
10025 /* rcu callback to free various structures associated with a task group */
10026 static void free_sched_group_rcu(struct rcu_head *rhp)
10028 /* now it should be safe to free those cfs_rqs */
10029 free_sched_group(container_of(rhp, struct task_group, rcu));
10032 /* Destroy runqueue etc associated with a task group */
10033 void sched_destroy_group(struct task_group *tg)
10035 unsigned long flags;
10038 spin_lock_irqsave(&task_group_lock, flags);
10039 for_each_possible_cpu(i) {
10040 unregister_fair_sched_group(tg, i);
10041 unregister_rt_sched_group(tg, i);
10043 list_del_rcu(&tg->list);
10044 list_del_rcu(&tg->siblings);
10045 spin_unlock_irqrestore(&task_group_lock, flags);
10047 /* wait for possible concurrent references to cfs_rqs complete */
10048 call_rcu(&tg->rcu, free_sched_group_rcu);
10051 /* change task's runqueue when it moves between groups.
10052 * The caller of this function should have put the task in its new group
10053 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
10054 * reflect its new group.
10056 void sched_move_task(struct task_struct *tsk)
10058 int on_rq, running;
10059 unsigned long flags;
10062 rq = task_rq_lock(tsk, &flags);
10064 update_rq_clock(rq);
10066 running = task_current(rq, tsk);
10067 on_rq = tsk->se.on_rq;
10070 dequeue_task(rq, tsk, 0);
10071 if (unlikely(running))
10072 tsk->sched_class->put_prev_task(rq, tsk);
10074 set_task_rq(tsk, task_cpu(tsk));
10076 #ifdef CONFIG_FAIR_GROUP_SCHED
10077 if (tsk->sched_class->moved_group)
10078 tsk->sched_class->moved_group(tsk);
10081 if (unlikely(running))
10082 tsk->sched_class->set_curr_task(rq);
10084 enqueue_task(rq, tsk, 0);
10086 task_rq_unlock(rq, &flags);
10088 #endif /* CONFIG_GROUP_SCHED */
10090 #ifdef CONFIG_FAIR_GROUP_SCHED
10091 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
10093 struct cfs_rq *cfs_rq = se->cfs_rq;
10098 dequeue_entity(cfs_rq, se, 0);
10100 se->load.weight = shares;
10101 se->load.inv_weight = 0;
10104 enqueue_entity(cfs_rq, se, 0);
10107 static void set_se_shares(struct sched_entity *se, unsigned long shares)
10109 struct cfs_rq *cfs_rq = se->cfs_rq;
10110 struct rq *rq = cfs_rq->rq;
10111 unsigned long flags;
10113 spin_lock_irqsave(&rq->lock, flags);
10114 __set_se_shares(se, shares);
10115 spin_unlock_irqrestore(&rq->lock, flags);
10118 static DEFINE_MUTEX(shares_mutex);
10120 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10123 unsigned long flags;
10126 * We can't change the weight of the root cgroup.
10131 if (shares < MIN_SHARES)
10132 shares = MIN_SHARES;
10133 else if (shares > MAX_SHARES)
10134 shares = MAX_SHARES;
10136 mutex_lock(&shares_mutex);
10137 if (tg->shares == shares)
10140 spin_lock_irqsave(&task_group_lock, flags);
10141 for_each_possible_cpu(i)
10142 unregister_fair_sched_group(tg, i);
10143 list_del_rcu(&tg->siblings);
10144 spin_unlock_irqrestore(&task_group_lock, flags);
10146 /* wait for any ongoing reference to this group to finish */
10147 synchronize_sched();
10150 * Now we are free to modify the group's share on each cpu
10151 * w/o tripping rebalance_share or load_balance_fair.
10153 tg->shares = shares;
10154 for_each_possible_cpu(i) {
10156 * force a rebalance
10158 cfs_rq_set_shares(tg->cfs_rq[i], 0);
10159 set_se_shares(tg->se[i], shares);
10163 * Enable load balance activity on this group, by inserting it back on
10164 * each cpu's rq->leaf_cfs_rq_list.
10166 spin_lock_irqsave(&task_group_lock, flags);
10167 for_each_possible_cpu(i)
10168 register_fair_sched_group(tg, i);
10169 list_add_rcu(&tg->siblings, &tg->parent->children);
10170 spin_unlock_irqrestore(&task_group_lock, flags);
10172 mutex_unlock(&shares_mutex);
10176 unsigned long sched_group_shares(struct task_group *tg)
10182 #ifdef CONFIG_RT_GROUP_SCHED
10184 * Ensure that the real time constraints are schedulable.
10186 static DEFINE_MUTEX(rt_constraints_mutex);
10188 static unsigned long to_ratio(u64 period, u64 runtime)
10190 if (runtime == RUNTIME_INF)
10193 return div64_u64(runtime << 20, period);
10196 /* Must be called with tasklist_lock held */
10197 static inline int tg_has_rt_tasks(struct task_group *tg)
10199 struct task_struct *g, *p;
10201 do_each_thread(g, p) {
10202 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
10204 } while_each_thread(g, p);
10209 struct rt_schedulable_data {
10210 struct task_group *tg;
10215 static int tg_schedulable(struct task_group *tg, void *data)
10217 struct rt_schedulable_data *d = data;
10218 struct task_group *child;
10219 unsigned long total, sum = 0;
10220 u64 period, runtime;
10222 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10223 runtime = tg->rt_bandwidth.rt_runtime;
10226 period = d->rt_period;
10227 runtime = d->rt_runtime;
10230 #ifdef CONFIG_USER_SCHED
10231 if (tg == &root_task_group) {
10232 period = global_rt_period();
10233 runtime = global_rt_runtime();
10238 * Cannot have more runtime than the period.
10240 if (runtime > period && runtime != RUNTIME_INF)
10244 * Ensure we don't starve existing RT tasks.
10246 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
10249 total = to_ratio(period, runtime);
10252 * Nobody can have more than the global setting allows.
10254 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
10258 * The sum of our children's runtime should not exceed our own.
10260 list_for_each_entry_rcu(child, &tg->children, siblings) {
10261 period = ktime_to_ns(child->rt_bandwidth.rt_period);
10262 runtime = child->rt_bandwidth.rt_runtime;
10264 if (child == d->tg) {
10265 period = d->rt_period;
10266 runtime = d->rt_runtime;
10269 sum += to_ratio(period, runtime);
10278 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
10280 struct rt_schedulable_data data = {
10282 .rt_period = period,
10283 .rt_runtime = runtime,
10286 return walk_tg_tree(tg_schedulable, tg_nop, &data);
10289 static int tg_set_bandwidth(struct task_group *tg,
10290 u64 rt_period, u64 rt_runtime)
10294 mutex_lock(&rt_constraints_mutex);
10295 read_lock(&tasklist_lock);
10296 err = __rt_schedulable(tg, rt_period, rt_runtime);
10300 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10301 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
10302 tg->rt_bandwidth.rt_runtime = rt_runtime;
10304 for_each_possible_cpu(i) {
10305 struct rt_rq *rt_rq = tg->rt_rq[i];
10307 spin_lock(&rt_rq->rt_runtime_lock);
10308 rt_rq->rt_runtime = rt_runtime;
10309 spin_unlock(&rt_rq->rt_runtime_lock);
10311 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10313 read_unlock(&tasklist_lock);
10314 mutex_unlock(&rt_constraints_mutex);
10319 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
10321 u64 rt_runtime, rt_period;
10323 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10324 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
10325 if (rt_runtime_us < 0)
10326 rt_runtime = RUNTIME_INF;
10328 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10331 long sched_group_rt_runtime(struct task_group *tg)
10335 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
10338 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
10339 do_div(rt_runtime_us, NSEC_PER_USEC);
10340 return rt_runtime_us;
10343 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
10345 u64 rt_runtime, rt_period;
10347 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
10348 rt_runtime = tg->rt_bandwidth.rt_runtime;
10350 if (rt_period == 0)
10353 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10356 long sched_group_rt_period(struct task_group *tg)
10360 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
10361 do_div(rt_period_us, NSEC_PER_USEC);
10362 return rt_period_us;
10365 static int sched_rt_global_constraints(void)
10367 u64 runtime, period;
10370 if (sysctl_sched_rt_period <= 0)
10373 runtime = global_rt_runtime();
10374 period = global_rt_period();
10377 * Sanity check on the sysctl variables.
10379 if (runtime > period && runtime != RUNTIME_INF)
10382 mutex_lock(&rt_constraints_mutex);
10383 read_lock(&tasklist_lock);
10384 ret = __rt_schedulable(NULL, 0, 0);
10385 read_unlock(&tasklist_lock);
10386 mutex_unlock(&rt_constraints_mutex);
10391 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
10393 /* Don't accept realtime tasks when there is no way for them to run */
10394 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
10400 #else /* !CONFIG_RT_GROUP_SCHED */
10401 static int sched_rt_global_constraints(void)
10403 unsigned long flags;
10406 if (sysctl_sched_rt_period <= 0)
10410 * There's always some RT tasks in the root group
10411 * -- migration, kstopmachine etc..
10413 if (sysctl_sched_rt_runtime == 0)
10416 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
10417 for_each_possible_cpu(i) {
10418 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
10420 spin_lock(&rt_rq->rt_runtime_lock);
10421 rt_rq->rt_runtime = global_rt_runtime();
10422 spin_unlock(&rt_rq->rt_runtime_lock);
10424 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
10428 #endif /* CONFIG_RT_GROUP_SCHED */
10430 int sched_rt_handler(struct ctl_table *table, int write,
10431 void __user *buffer, size_t *lenp,
10435 int old_period, old_runtime;
10436 static DEFINE_MUTEX(mutex);
10438 mutex_lock(&mutex);
10439 old_period = sysctl_sched_rt_period;
10440 old_runtime = sysctl_sched_rt_runtime;
10442 ret = proc_dointvec(table, write, buffer, lenp, ppos);
10444 if (!ret && write) {
10445 ret = sched_rt_global_constraints();
10447 sysctl_sched_rt_period = old_period;
10448 sysctl_sched_rt_runtime = old_runtime;
10450 def_rt_bandwidth.rt_runtime = global_rt_runtime();
10451 def_rt_bandwidth.rt_period =
10452 ns_to_ktime(global_rt_period());
10455 mutex_unlock(&mutex);
10460 #ifdef CONFIG_CGROUP_SCHED
10462 /* return corresponding task_group object of a cgroup */
10463 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
10465 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
10466 struct task_group, css);
10469 static struct cgroup_subsys_state *
10470 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
10472 struct task_group *tg, *parent;
10474 if (!cgrp->parent) {
10475 /* This is early initialization for the top cgroup */
10476 return &init_task_group.css;
10479 parent = cgroup_tg(cgrp->parent);
10480 tg = sched_create_group(parent);
10482 return ERR_PTR(-ENOMEM);
10488 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10490 struct task_group *tg = cgroup_tg(cgrp);
10492 sched_destroy_group(tg);
10496 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
10498 #ifdef CONFIG_RT_GROUP_SCHED
10499 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
10502 /* We don't support RT-tasks being in separate groups */
10503 if (tsk->sched_class != &fair_sched_class)
10510 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10511 struct task_struct *tsk, bool threadgroup)
10513 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
10517 struct task_struct *c;
10519 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
10520 retval = cpu_cgroup_can_attach_task(cgrp, c);
10532 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10533 struct cgroup *old_cont, struct task_struct *tsk,
10536 sched_move_task(tsk);
10538 struct task_struct *c;
10540 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
10541 sched_move_task(c);
10547 #ifdef CONFIG_FAIR_GROUP_SCHED
10548 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
10551 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
10554 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
10556 struct task_group *tg = cgroup_tg(cgrp);
10558 return (u64) tg->shares;
10560 #endif /* CONFIG_FAIR_GROUP_SCHED */
10562 #ifdef CONFIG_RT_GROUP_SCHED
10563 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
10566 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
10569 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
10571 return sched_group_rt_runtime(cgroup_tg(cgrp));
10574 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
10577 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
10580 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
10582 return sched_group_rt_period(cgroup_tg(cgrp));
10584 #endif /* CONFIG_RT_GROUP_SCHED */
10586 static struct cftype cpu_files[] = {
10587 #ifdef CONFIG_FAIR_GROUP_SCHED
10590 .read_u64 = cpu_shares_read_u64,
10591 .write_u64 = cpu_shares_write_u64,
10594 #ifdef CONFIG_RT_GROUP_SCHED
10596 .name = "rt_runtime_us",
10597 .read_s64 = cpu_rt_runtime_read,
10598 .write_s64 = cpu_rt_runtime_write,
10601 .name = "rt_period_us",
10602 .read_u64 = cpu_rt_period_read_uint,
10603 .write_u64 = cpu_rt_period_write_uint,
10608 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
10610 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
10613 struct cgroup_subsys cpu_cgroup_subsys = {
10615 .create = cpu_cgroup_create,
10616 .destroy = cpu_cgroup_destroy,
10617 .can_attach = cpu_cgroup_can_attach,
10618 .attach = cpu_cgroup_attach,
10619 .populate = cpu_cgroup_populate,
10620 .subsys_id = cpu_cgroup_subsys_id,
10624 #endif /* CONFIG_CGROUP_SCHED */
10626 #ifdef CONFIG_CGROUP_CPUACCT
10629 * CPU accounting code for task groups.
10631 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10632 * (balbir@in.ibm.com).
10635 /* track cpu usage of a group of tasks and its child groups */
10637 struct cgroup_subsys_state css;
10638 /* cpuusage holds pointer to a u64-type object on every cpu */
10640 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
10641 struct cpuacct *parent;
10644 struct cgroup_subsys cpuacct_subsys;
10646 /* return cpu accounting group corresponding to this container */
10647 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
10649 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
10650 struct cpuacct, css);
10653 /* return cpu accounting group to which this task belongs */
10654 static inline struct cpuacct *task_ca(struct task_struct *tsk)
10656 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
10657 struct cpuacct, css);
10660 /* create a new cpu accounting group */
10661 static struct cgroup_subsys_state *cpuacct_create(
10662 struct cgroup_subsys *ss, struct cgroup *cgrp)
10664 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
10670 ca->cpuusage = alloc_percpu(u64);
10674 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10675 if (percpu_counter_init(&ca->cpustat[i], 0))
10676 goto out_free_counters;
10679 ca->parent = cgroup_ca(cgrp->parent);
10685 percpu_counter_destroy(&ca->cpustat[i]);
10686 free_percpu(ca->cpuusage);
10690 return ERR_PTR(-ENOMEM);
10693 /* destroy an existing cpu accounting group */
10695 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10697 struct cpuacct *ca = cgroup_ca(cgrp);
10700 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10701 percpu_counter_destroy(&ca->cpustat[i]);
10702 free_percpu(ca->cpuusage);
10706 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
10708 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10711 #ifndef CONFIG_64BIT
10713 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10715 spin_lock_irq(&cpu_rq(cpu)->lock);
10717 spin_unlock_irq(&cpu_rq(cpu)->lock);
10725 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
10727 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10729 #ifndef CONFIG_64BIT
10731 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10733 spin_lock_irq(&cpu_rq(cpu)->lock);
10735 spin_unlock_irq(&cpu_rq(cpu)->lock);
10741 /* return total cpu usage (in nanoseconds) of a group */
10742 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
10744 struct cpuacct *ca = cgroup_ca(cgrp);
10745 u64 totalcpuusage = 0;
10748 for_each_present_cpu(i)
10749 totalcpuusage += cpuacct_cpuusage_read(ca, i);
10751 return totalcpuusage;
10754 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
10757 struct cpuacct *ca = cgroup_ca(cgrp);
10766 for_each_present_cpu(i)
10767 cpuacct_cpuusage_write(ca, i, 0);
10773 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
10774 struct seq_file *m)
10776 struct cpuacct *ca = cgroup_ca(cgroup);
10780 for_each_present_cpu(i) {
10781 percpu = cpuacct_cpuusage_read(ca, i);
10782 seq_printf(m, "%llu ", (unsigned long long) percpu);
10784 seq_printf(m, "\n");
10788 static const char *cpuacct_stat_desc[] = {
10789 [CPUACCT_STAT_USER] = "user",
10790 [CPUACCT_STAT_SYSTEM] = "system",
10793 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
10794 struct cgroup_map_cb *cb)
10796 struct cpuacct *ca = cgroup_ca(cgrp);
10799 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
10800 s64 val = percpu_counter_read(&ca->cpustat[i]);
10801 val = cputime64_to_clock_t(val);
10802 cb->fill(cb, cpuacct_stat_desc[i], val);
10807 static struct cftype files[] = {
10810 .read_u64 = cpuusage_read,
10811 .write_u64 = cpuusage_write,
10814 .name = "usage_percpu",
10815 .read_seq_string = cpuacct_percpu_seq_read,
10819 .read_map = cpuacct_stats_show,
10823 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
10825 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
10829 * charge this task's execution time to its accounting group.
10831 * called with rq->lock held.
10833 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
10835 struct cpuacct *ca;
10838 if (unlikely(!cpuacct_subsys.active))
10841 cpu = task_cpu(tsk);
10847 for (; ca; ca = ca->parent) {
10848 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10849 *cpuusage += cputime;
10856 * Charge the system/user time to the task's accounting group.
10858 static void cpuacct_update_stats(struct task_struct *tsk,
10859 enum cpuacct_stat_index idx, cputime_t val)
10861 struct cpuacct *ca;
10863 if (unlikely(!cpuacct_subsys.active))
10870 percpu_counter_add(&ca->cpustat[idx], val);
10876 struct cgroup_subsys cpuacct_subsys = {
10878 .create = cpuacct_create,
10879 .destroy = cpuacct_destroy,
10880 .populate = cpuacct_populate,
10881 .subsys_id = cpuacct_subsys_id,
10883 #endif /* CONFIG_CGROUP_CPUACCT */
10887 int rcu_expedited_torture_stats(char *page)
10891 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
10893 void synchronize_sched_expedited(void)
10896 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
10898 #else /* #ifndef CONFIG_SMP */
10900 static DEFINE_PER_CPU(struct migration_req, rcu_migration_req);
10901 static DEFINE_MUTEX(rcu_sched_expedited_mutex);
10903 #define RCU_EXPEDITED_STATE_POST -2
10904 #define RCU_EXPEDITED_STATE_IDLE -1
10906 static int rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
10908 int rcu_expedited_torture_stats(char *page)
10913 cnt += sprintf(&page[cnt], "state: %d /", rcu_expedited_state);
10914 for_each_online_cpu(cpu) {
10915 cnt += sprintf(&page[cnt], " %d:%d",
10916 cpu, per_cpu(rcu_migration_req, cpu).dest_cpu);
10918 cnt += sprintf(&page[cnt], "\n");
10921 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
10923 static long synchronize_sched_expedited_count;
10926 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
10927 * approach to force grace period to end quickly. This consumes
10928 * significant time on all CPUs, and is thus not recommended for
10929 * any sort of common-case code.
10931 * Note that it is illegal to call this function while holding any
10932 * lock that is acquired by a CPU-hotplug notifier. Failing to
10933 * observe this restriction will result in deadlock.
10935 void synchronize_sched_expedited(void)
10938 unsigned long flags;
10939 bool need_full_sync = 0;
10941 struct migration_req *req;
10945 smp_mb(); /* ensure prior mod happens before capturing snap. */
10946 snap = ACCESS_ONCE(synchronize_sched_expedited_count) + 1;
10948 while (!mutex_trylock(&rcu_sched_expedited_mutex)) {
10950 if (trycount++ < 10)
10951 udelay(trycount * num_online_cpus());
10953 synchronize_sched();
10956 if (ACCESS_ONCE(synchronize_sched_expedited_count) - snap > 0) {
10957 smp_mb(); /* ensure test happens before caller kfree */
10962 rcu_expedited_state = RCU_EXPEDITED_STATE_POST;
10963 for_each_online_cpu(cpu) {
10965 req = &per_cpu(rcu_migration_req, cpu);
10966 init_completion(&req->done);
10968 req->dest_cpu = RCU_MIGRATION_NEED_QS;
10969 spin_lock_irqsave(&rq->lock, flags);
10970 list_add(&req->list, &rq->migration_queue);
10971 spin_unlock_irqrestore(&rq->lock, flags);
10972 wake_up_process(rq->migration_thread);
10974 for_each_online_cpu(cpu) {
10975 rcu_expedited_state = cpu;
10976 req = &per_cpu(rcu_migration_req, cpu);
10978 wait_for_completion(&req->done);
10979 spin_lock_irqsave(&rq->lock, flags);
10980 if (unlikely(req->dest_cpu == RCU_MIGRATION_MUST_SYNC))
10981 need_full_sync = 1;
10982 req->dest_cpu = RCU_MIGRATION_IDLE;
10983 spin_unlock_irqrestore(&rq->lock, flags);
10985 rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
10986 synchronize_sched_expedited_count++;
10987 mutex_unlock(&rcu_sched_expedited_mutex);
10989 if (need_full_sync)
10990 synchronize_sched();
10992 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
10994 #endif /* #else #ifndef CONFIG_SMP */