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 raw_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 raw_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 raw_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 raw_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_var);
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 raw_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)
648 #define rcu_dereference_check_sched_domain(p) \
649 rcu_dereference_check((p), \
650 rcu_read_lock_sched_held() || \
651 lockdep_is_held(&sched_domains_mutex))
654 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
655 * See detach_destroy_domains: synchronize_sched for details.
657 * The domain tree of any CPU may only be accessed from within
658 * preempt-disabled sections.
660 #define for_each_domain(cpu, __sd) \
661 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
663 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
664 #define this_rq() (&__get_cpu_var(runqueues))
665 #define task_rq(p) cpu_rq(task_cpu(p))
666 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
667 #define raw_rq() (&__raw_get_cpu_var(runqueues))
669 inline void update_rq_clock(struct rq *rq)
671 rq->clock = sched_clock_cpu(cpu_of(rq));
675 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
677 #ifdef CONFIG_SCHED_DEBUG
678 # define const_debug __read_mostly
680 # define const_debug static const
685 * @cpu: the processor in question.
687 * Returns true if the current cpu runqueue is locked.
688 * This interface allows printk to be called with the runqueue lock
689 * held and know whether or not it is OK to wake up the klogd.
691 int runqueue_is_locked(int cpu)
693 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
697 * Debugging: various feature bits
700 #define SCHED_FEAT(name, enabled) \
701 __SCHED_FEAT_##name ,
704 #include "sched_features.h"
709 #define SCHED_FEAT(name, enabled) \
710 (1UL << __SCHED_FEAT_##name) * enabled |
712 const_debug unsigned int sysctl_sched_features =
713 #include "sched_features.h"
718 #ifdef CONFIG_SCHED_DEBUG
719 #define SCHED_FEAT(name, enabled) \
722 static __read_mostly char *sched_feat_names[] = {
723 #include "sched_features.h"
729 static int sched_feat_show(struct seq_file *m, void *v)
733 for (i = 0; sched_feat_names[i]; i++) {
734 if (!(sysctl_sched_features & (1UL << i)))
736 seq_printf(m, "%s ", sched_feat_names[i]);
744 sched_feat_write(struct file *filp, const char __user *ubuf,
745 size_t cnt, loff_t *ppos)
755 if (copy_from_user(&buf, ubuf, cnt))
760 if (strncmp(buf, "NO_", 3) == 0) {
765 for (i = 0; sched_feat_names[i]; i++) {
766 int len = strlen(sched_feat_names[i]);
768 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
770 sysctl_sched_features &= ~(1UL << i);
772 sysctl_sched_features |= (1UL << i);
777 if (!sched_feat_names[i])
785 static int sched_feat_open(struct inode *inode, struct file *filp)
787 return single_open(filp, sched_feat_show, NULL);
790 static const struct file_operations sched_feat_fops = {
791 .open = sched_feat_open,
792 .write = sched_feat_write,
795 .release = single_release,
798 static __init int sched_init_debug(void)
800 debugfs_create_file("sched_features", 0644, NULL, NULL,
805 late_initcall(sched_init_debug);
809 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
812 * Number of tasks to iterate in a single balance run.
813 * Limited because this is done with IRQs disabled.
815 const_debug unsigned int sysctl_sched_nr_migrate = 32;
818 * ratelimit for updating the group shares.
821 unsigned int sysctl_sched_shares_ratelimit = 250000;
822 unsigned int normalized_sysctl_sched_shares_ratelimit = 250000;
825 * Inject some fuzzyness into changing the per-cpu group shares
826 * this avoids remote rq-locks at the expense of fairness.
829 unsigned int sysctl_sched_shares_thresh = 4;
832 * period over which we average the RT time consumption, measured
837 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
840 * period over which we measure -rt task cpu usage in us.
843 unsigned int sysctl_sched_rt_period = 1000000;
845 static __read_mostly int scheduler_running;
848 * part of the period that we allow rt tasks to run in us.
851 int sysctl_sched_rt_runtime = 950000;
853 static inline u64 global_rt_period(void)
855 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
858 static inline u64 global_rt_runtime(void)
860 if (sysctl_sched_rt_runtime < 0)
863 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
866 #ifndef prepare_arch_switch
867 # define prepare_arch_switch(next) do { } while (0)
869 #ifndef finish_arch_switch
870 # define finish_arch_switch(prev) do { } while (0)
873 static inline int task_current(struct rq *rq, struct task_struct *p)
875 return rq->curr == p;
878 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
879 static inline int task_running(struct rq *rq, struct task_struct *p)
881 return task_current(rq, p);
884 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
888 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
890 #ifdef CONFIG_DEBUG_SPINLOCK
891 /* this is a valid case when another task releases the spinlock */
892 rq->lock.owner = current;
895 * If we are tracking spinlock dependencies then we have to
896 * fix up the runqueue lock - which gets 'carried over' from
899 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
901 raw_spin_unlock_irq(&rq->lock);
904 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
905 static inline int task_running(struct rq *rq, struct task_struct *p)
910 return task_current(rq, p);
914 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
918 * We can optimise this out completely for !SMP, because the
919 * SMP rebalancing from interrupt is the only thing that cares
924 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
925 raw_spin_unlock_irq(&rq->lock);
927 raw_spin_unlock(&rq->lock);
931 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
935 * After ->oncpu is cleared, the task can be moved to a different CPU.
936 * We must ensure this doesn't happen until the switch is completely
942 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
946 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
949 * __task_rq_lock - lock the runqueue a given task resides on.
950 * Must be called interrupts disabled.
952 static inline struct rq *__task_rq_lock(struct task_struct *p)
956 struct rq *rq = task_rq(p);
957 raw_spin_lock(&rq->lock);
958 if (likely(rq == task_rq(p)))
960 raw_spin_unlock(&rq->lock);
965 * task_rq_lock - lock the runqueue a given task resides on and disable
966 * interrupts. Note the ordering: we can safely lookup the task_rq without
967 * explicitly disabling preemption.
969 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
975 local_irq_save(*flags);
977 raw_spin_lock(&rq->lock);
978 if (likely(rq == task_rq(p)))
980 raw_spin_unlock_irqrestore(&rq->lock, *flags);
984 void task_rq_unlock_wait(struct task_struct *p)
986 struct rq *rq = task_rq(p);
988 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
989 raw_spin_unlock_wait(&rq->lock);
992 static void __task_rq_unlock(struct rq *rq)
995 raw_spin_unlock(&rq->lock);
998 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
1001 raw_spin_unlock_irqrestore(&rq->lock, *flags);
1005 * this_rq_lock - lock this runqueue and disable interrupts.
1007 static struct rq *this_rq_lock(void)
1008 __acquires(rq->lock)
1012 local_irq_disable();
1014 raw_spin_lock(&rq->lock);
1019 #ifdef CONFIG_SCHED_HRTICK
1021 * Use HR-timers to deliver accurate preemption points.
1023 * Its all a bit involved since we cannot program an hrt while holding the
1024 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1027 * When we get rescheduled we reprogram the hrtick_timer outside of the
1033 * - enabled by features
1034 * - hrtimer is actually high res
1036 static inline int hrtick_enabled(struct rq *rq)
1038 if (!sched_feat(HRTICK))
1040 if (!cpu_active(cpu_of(rq)))
1042 return hrtimer_is_hres_active(&rq->hrtick_timer);
1045 static void hrtick_clear(struct rq *rq)
1047 if (hrtimer_active(&rq->hrtick_timer))
1048 hrtimer_cancel(&rq->hrtick_timer);
1052 * High-resolution timer tick.
1053 * Runs from hardirq context with interrupts disabled.
1055 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1057 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1059 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1061 raw_spin_lock(&rq->lock);
1062 update_rq_clock(rq);
1063 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1064 raw_spin_unlock(&rq->lock);
1066 return HRTIMER_NORESTART;
1071 * called from hardirq (IPI) context
1073 static void __hrtick_start(void *arg)
1075 struct rq *rq = arg;
1077 raw_spin_lock(&rq->lock);
1078 hrtimer_restart(&rq->hrtick_timer);
1079 rq->hrtick_csd_pending = 0;
1080 raw_spin_unlock(&rq->lock);
1084 * Called to set the hrtick timer state.
1086 * called with rq->lock held and irqs disabled
1088 static void hrtick_start(struct rq *rq, u64 delay)
1090 struct hrtimer *timer = &rq->hrtick_timer;
1091 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1093 hrtimer_set_expires(timer, time);
1095 if (rq == this_rq()) {
1096 hrtimer_restart(timer);
1097 } else if (!rq->hrtick_csd_pending) {
1098 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1099 rq->hrtick_csd_pending = 1;
1104 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1106 int cpu = (int)(long)hcpu;
1109 case CPU_UP_CANCELED:
1110 case CPU_UP_CANCELED_FROZEN:
1111 case CPU_DOWN_PREPARE:
1112 case CPU_DOWN_PREPARE_FROZEN:
1114 case CPU_DEAD_FROZEN:
1115 hrtick_clear(cpu_rq(cpu));
1122 static __init void init_hrtick(void)
1124 hotcpu_notifier(hotplug_hrtick, 0);
1128 * Called to set the hrtick timer state.
1130 * called with rq->lock held and irqs disabled
1132 static void hrtick_start(struct rq *rq, u64 delay)
1134 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1135 HRTIMER_MODE_REL_PINNED, 0);
1138 static inline void init_hrtick(void)
1141 #endif /* CONFIG_SMP */
1143 static void init_rq_hrtick(struct rq *rq)
1146 rq->hrtick_csd_pending = 0;
1148 rq->hrtick_csd.flags = 0;
1149 rq->hrtick_csd.func = __hrtick_start;
1150 rq->hrtick_csd.info = rq;
1153 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1154 rq->hrtick_timer.function = hrtick;
1156 #else /* CONFIG_SCHED_HRTICK */
1157 static inline void hrtick_clear(struct rq *rq)
1161 static inline void init_rq_hrtick(struct rq *rq)
1165 static inline void init_hrtick(void)
1168 #endif /* CONFIG_SCHED_HRTICK */
1171 * resched_task - mark a task 'to be rescheduled now'.
1173 * On UP this means the setting of the need_resched flag, on SMP it
1174 * might also involve a cross-CPU call to trigger the scheduler on
1179 #ifndef tsk_is_polling
1180 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1183 static void resched_task(struct task_struct *p)
1187 assert_raw_spin_locked(&task_rq(p)->lock);
1189 if (test_tsk_need_resched(p))
1192 set_tsk_need_resched(p);
1195 if (cpu == smp_processor_id())
1198 /* NEED_RESCHED must be visible before we test polling */
1200 if (!tsk_is_polling(p))
1201 smp_send_reschedule(cpu);
1204 static void resched_cpu(int cpu)
1206 struct rq *rq = cpu_rq(cpu);
1207 unsigned long flags;
1209 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1211 resched_task(cpu_curr(cpu));
1212 raw_spin_unlock_irqrestore(&rq->lock, flags);
1217 * When add_timer_on() enqueues a timer into the timer wheel of an
1218 * idle CPU then this timer might expire before the next timer event
1219 * which is scheduled to wake up that CPU. In case of a completely
1220 * idle system the next event might even be infinite time into the
1221 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1222 * leaves the inner idle loop so the newly added timer is taken into
1223 * account when the CPU goes back to idle and evaluates the timer
1224 * wheel for the next timer event.
1226 void wake_up_idle_cpu(int cpu)
1228 struct rq *rq = cpu_rq(cpu);
1230 if (cpu == smp_processor_id())
1234 * This is safe, as this function is called with the timer
1235 * wheel base lock of (cpu) held. When the CPU is on the way
1236 * to idle and has not yet set rq->curr to idle then it will
1237 * be serialized on the timer wheel base lock and take the new
1238 * timer into account automatically.
1240 if (rq->curr != rq->idle)
1244 * We can set TIF_RESCHED on the idle task of the other CPU
1245 * lockless. The worst case is that the other CPU runs the
1246 * idle task through an additional NOOP schedule()
1248 set_tsk_need_resched(rq->idle);
1250 /* NEED_RESCHED must be visible before we test polling */
1252 if (!tsk_is_polling(rq->idle))
1253 smp_send_reschedule(cpu);
1255 #endif /* CONFIG_NO_HZ */
1257 static u64 sched_avg_period(void)
1259 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1262 static void sched_avg_update(struct rq *rq)
1264 s64 period = sched_avg_period();
1266 while ((s64)(rq->clock - rq->age_stamp) > period) {
1267 rq->age_stamp += period;
1272 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1274 rq->rt_avg += rt_delta;
1275 sched_avg_update(rq);
1278 #else /* !CONFIG_SMP */
1279 static void resched_task(struct task_struct *p)
1281 assert_raw_spin_locked(&task_rq(p)->lock);
1282 set_tsk_need_resched(p);
1285 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1288 #endif /* CONFIG_SMP */
1290 #if BITS_PER_LONG == 32
1291 # define WMULT_CONST (~0UL)
1293 # define WMULT_CONST (1UL << 32)
1296 #define WMULT_SHIFT 32
1299 * Shift right and round:
1301 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1304 * delta *= weight / lw
1306 static unsigned long
1307 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1308 struct load_weight *lw)
1312 if (!lw->inv_weight) {
1313 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1316 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1320 tmp = (u64)delta_exec * weight;
1322 * Check whether we'd overflow the 64-bit multiplication:
1324 if (unlikely(tmp > WMULT_CONST))
1325 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1328 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1330 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1333 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1339 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1346 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1347 * of tasks with abnormal "nice" values across CPUs the contribution that
1348 * each task makes to its run queue's load is weighted according to its
1349 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1350 * scaled version of the new time slice allocation that they receive on time
1354 #define WEIGHT_IDLEPRIO 3
1355 #define WMULT_IDLEPRIO 1431655765
1358 * Nice levels are multiplicative, with a gentle 10% change for every
1359 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1360 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1361 * that remained on nice 0.
1363 * The "10% effect" is relative and cumulative: from _any_ nice level,
1364 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1365 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1366 * If a task goes up by ~10% and another task goes down by ~10% then
1367 * the relative distance between them is ~25%.)
1369 static const int prio_to_weight[40] = {
1370 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1371 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1372 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1373 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1374 /* 0 */ 1024, 820, 655, 526, 423,
1375 /* 5 */ 335, 272, 215, 172, 137,
1376 /* 10 */ 110, 87, 70, 56, 45,
1377 /* 15 */ 36, 29, 23, 18, 15,
1381 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1383 * In cases where the weight does not change often, we can use the
1384 * precalculated inverse to speed up arithmetics by turning divisions
1385 * into multiplications:
1387 static const u32 prio_to_wmult[40] = {
1388 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1389 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1390 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1391 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1392 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1393 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1394 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1395 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1398 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1401 * runqueue iterator, to support SMP load-balancing between different
1402 * scheduling classes, without having to expose their internal data
1403 * structures to the load-balancing proper:
1405 struct rq_iterator {
1407 struct task_struct *(*start)(void *);
1408 struct task_struct *(*next)(void *);
1412 static unsigned long
1413 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1414 unsigned long max_load_move, struct sched_domain *sd,
1415 enum cpu_idle_type idle, int *all_pinned,
1416 int *this_best_prio, struct rq_iterator *iterator);
1419 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1420 struct sched_domain *sd, enum cpu_idle_type idle,
1421 struct rq_iterator *iterator);
1424 /* Time spent by the tasks of the cpu accounting group executing in ... */
1425 enum cpuacct_stat_index {
1426 CPUACCT_STAT_USER, /* ... user mode */
1427 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1429 CPUACCT_STAT_NSTATS,
1432 #ifdef CONFIG_CGROUP_CPUACCT
1433 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1434 static void cpuacct_update_stats(struct task_struct *tsk,
1435 enum cpuacct_stat_index idx, cputime_t val);
1437 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1438 static inline void cpuacct_update_stats(struct task_struct *tsk,
1439 enum cpuacct_stat_index idx, cputime_t val) {}
1442 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1444 update_load_add(&rq->load, load);
1447 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1449 update_load_sub(&rq->load, load);
1452 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1453 typedef int (*tg_visitor)(struct task_group *, void *);
1456 * Iterate the full tree, calling @down when first entering a node and @up when
1457 * leaving it for the final time.
1459 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1461 struct task_group *parent, *child;
1465 parent = &root_task_group;
1467 ret = (*down)(parent, data);
1470 list_for_each_entry_rcu(child, &parent->children, siblings) {
1477 ret = (*up)(parent, data);
1482 parent = parent->parent;
1491 static int tg_nop(struct task_group *tg, void *data)
1498 /* Used instead of source_load when we know the type == 0 */
1499 static unsigned long weighted_cpuload(const int cpu)
1501 return cpu_rq(cpu)->load.weight;
1505 * Return a low guess at the load of a migration-source cpu weighted
1506 * according to the scheduling class and "nice" value.
1508 * We want to under-estimate the load of migration sources, to
1509 * balance conservatively.
1511 static unsigned long source_load(int cpu, int type)
1513 struct rq *rq = cpu_rq(cpu);
1514 unsigned long total = weighted_cpuload(cpu);
1516 if (type == 0 || !sched_feat(LB_BIAS))
1519 return min(rq->cpu_load[type-1], total);
1523 * Return a high guess at the load of a migration-target cpu weighted
1524 * according to the scheduling class and "nice" value.
1526 static unsigned long target_load(int cpu, int type)
1528 struct rq *rq = cpu_rq(cpu);
1529 unsigned long total = weighted_cpuload(cpu);
1531 if (type == 0 || !sched_feat(LB_BIAS))
1534 return max(rq->cpu_load[type-1], total);
1537 static struct sched_group *group_of(int cpu)
1539 struct sched_domain *sd = rcu_dereference_sched(cpu_rq(cpu)->sd);
1547 static unsigned long power_of(int cpu)
1549 struct sched_group *group = group_of(cpu);
1552 return SCHED_LOAD_SCALE;
1554 return group->cpu_power;
1557 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1559 static unsigned long cpu_avg_load_per_task(int cpu)
1561 struct rq *rq = cpu_rq(cpu);
1562 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1565 rq->avg_load_per_task = rq->load.weight / nr_running;
1567 rq->avg_load_per_task = 0;
1569 return rq->avg_load_per_task;
1572 #ifdef CONFIG_FAIR_GROUP_SCHED
1574 static __read_mostly unsigned long *update_shares_data;
1576 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1579 * Calculate and set the cpu's group shares.
1581 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1582 unsigned long sd_shares,
1583 unsigned long sd_rq_weight,
1584 unsigned long *usd_rq_weight)
1586 unsigned long shares, rq_weight;
1589 rq_weight = usd_rq_weight[cpu];
1592 rq_weight = NICE_0_LOAD;
1596 * \Sum_j shares_j * rq_weight_i
1597 * shares_i = -----------------------------
1598 * \Sum_j rq_weight_j
1600 shares = (sd_shares * rq_weight) / sd_rq_weight;
1601 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1603 if (abs(shares - tg->se[cpu]->load.weight) >
1604 sysctl_sched_shares_thresh) {
1605 struct rq *rq = cpu_rq(cpu);
1606 unsigned long flags;
1608 raw_spin_lock_irqsave(&rq->lock, flags);
1609 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1610 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1611 __set_se_shares(tg->se[cpu], shares);
1612 raw_spin_unlock_irqrestore(&rq->lock, flags);
1617 * Re-compute the task group their per cpu shares over the given domain.
1618 * This needs to be done in a bottom-up fashion because the rq weight of a
1619 * parent group depends on the shares of its child groups.
1621 static int tg_shares_up(struct task_group *tg, void *data)
1623 unsigned long weight, rq_weight = 0, sum_weight = 0, shares = 0;
1624 unsigned long *usd_rq_weight;
1625 struct sched_domain *sd = data;
1626 unsigned long flags;
1632 local_irq_save(flags);
1633 usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id());
1635 for_each_cpu(i, sched_domain_span(sd)) {
1636 weight = tg->cfs_rq[i]->load.weight;
1637 usd_rq_weight[i] = weight;
1639 rq_weight += weight;
1641 * If there are currently no tasks on the cpu pretend there
1642 * is one of average load so that when a new task gets to
1643 * run here it will not get delayed by group starvation.
1646 weight = NICE_0_LOAD;
1648 sum_weight += weight;
1649 shares += tg->cfs_rq[i]->shares;
1653 rq_weight = sum_weight;
1655 if ((!shares && rq_weight) || shares > tg->shares)
1656 shares = tg->shares;
1658 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1659 shares = tg->shares;
1661 for_each_cpu(i, sched_domain_span(sd))
1662 update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight);
1664 local_irq_restore(flags);
1670 * Compute the cpu's hierarchical load factor for each task group.
1671 * This needs to be done in a top-down fashion because the load of a child
1672 * group is a fraction of its parents load.
1674 static int tg_load_down(struct task_group *tg, void *data)
1677 long cpu = (long)data;
1680 load = cpu_rq(cpu)->load.weight;
1682 load = tg->parent->cfs_rq[cpu]->h_load;
1683 load *= tg->cfs_rq[cpu]->shares;
1684 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1687 tg->cfs_rq[cpu]->h_load = load;
1692 static void update_shares(struct sched_domain *sd)
1697 if (root_task_group_empty())
1700 now = cpu_clock(raw_smp_processor_id());
1701 elapsed = now - sd->last_update;
1703 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1704 sd->last_update = now;
1705 walk_tg_tree(tg_nop, tg_shares_up, sd);
1709 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1711 if (root_task_group_empty())
1714 raw_spin_unlock(&rq->lock);
1716 raw_spin_lock(&rq->lock);
1719 static void update_h_load(long cpu)
1721 if (root_task_group_empty())
1724 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1729 static inline void update_shares(struct sched_domain *sd)
1733 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1739 #ifdef CONFIG_PREEMPT
1741 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1744 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1745 * way at the expense of forcing extra atomic operations in all
1746 * invocations. This assures that the double_lock is acquired using the
1747 * same underlying policy as the spinlock_t on this architecture, which
1748 * reduces latency compared to the unfair variant below. However, it
1749 * also adds more overhead and therefore may reduce throughput.
1751 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1752 __releases(this_rq->lock)
1753 __acquires(busiest->lock)
1754 __acquires(this_rq->lock)
1756 raw_spin_unlock(&this_rq->lock);
1757 double_rq_lock(this_rq, busiest);
1764 * Unfair double_lock_balance: Optimizes throughput at the expense of
1765 * latency by eliminating extra atomic operations when the locks are
1766 * already in proper order on entry. This favors lower cpu-ids and will
1767 * grant the double lock to lower cpus over higher ids under contention,
1768 * regardless of entry order into the function.
1770 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1771 __releases(this_rq->lock)
1772 __acquires(busiest->lock)
1773 __acquires(this_rq->lock)
1777 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1778 if (busiest < this_rq) {
1779 raw_spin_unlock(&this_rq->lock);
1780 raw_spin_lock(&busiest->lock);
1781 raw_spin_lock_nested(&this_rq->lock,
1782 SINGLE_DEPTH_NESTING);
1785 raw_spin_lock_nested(&busiest->lock,
1786 SINGLE_DEPTH_NESTING);
1791 #endif /* CONFIG_PREEMPT */
1794 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1796 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1798 if (unlikely(!irqs_disabled())) {
1799 /* printk() doesn't work good under rq->lock */
1800 raw_spin_unlock(&this_rq->lock);
1804 return _double_lock_balance(this_rq, busiest);
1807 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1808 __releases(busiest->lock)
1810 raw_spin_unlock(&busiest->lock);
1811 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1815 #ifdef CONFIG_FAIR_GROUP_SCHED
1816 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1819 cfs_rq->shares = shares;
1824 static void calc_load_account_active(struct rq *this_rq);
1825 static void update_sysctl(void);
1826 static int get_update_sysctl_factor(void);
1828 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1830 set_task_rq(p, cpu);
1833 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1834 * successfuly executed on another CPU. We must ensure that updates of
1835 * per-task data have been completed by this moment.
1838 task_thread_info(p)->cpu = cpu;
1842 #include "sched_stats.h"
1843 #include "sched_idletask.c"
1844 #include "sched_fair.c"
1845 #include "sched_rt.c"
1846 #ifdef CONFIG_SCHED_DEBUG
1847 # include "sched_debug.c"
1850 #define sched_class_highest (&rt_sched_class)
1851 #define for_each_class(class) \
1852 for (class = sched_class_highest; class; class = class->next)
1854 static void inc_nr_running(struct rq *rq)
1859 static void dec_nr_running(struct rq *rq)
1864 static void set_load_weight(struct task_struct *p)
1866 if (task_has_rt_policy(p)) {
1867 p->se.load.weight = prio_to_weight[0] * 2;
1868 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1873 * SCHED_IDLE tasks get minimal weight:
1875 if (p->policy == SCHED_IDLE) {
1876 p->se.load.weight = WEIGHT_IDLEPRIO;
1877 p->se.load.inv_weight = WMULT_IDLEPRIO;
1881 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1882 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1885 static void update_avg(u64 *avg, u64 sample)
1887 s64 diff = sample - *avg;
1891 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1894 p->se.start_runtime = p->se.sum_exec_runtime;
1896 sched_info_queued(p);
1897 p->sched_class->enqueue_task(rq, p, wakeup);
1901 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1904 if (p->se.last_wakeup) {
1905 update_avg(&p->se.avg_overlap,
1906 p->se.sum_exec_runtime - p->se.last_wakeup);
1907 p->se.last_wakeup = 0;
1909 update_avg(&p->se.avg_wakeup,
1910 sysctl_sched_wakeup_granularity);
1914 sched_info_dequeued(p);
1915 p->sched_class->dequeue_task(rq, p, sleep);
1920 * __normal_prio - return the priority that is based on the static prio
1922 static inline int __normal_prio(struct task_struct *p)
1924 return p->static_prio;
1928 * Calculate the expected normal priority: i.e. priority
1929 * without taking RT-inheritance into account. Might be
1930 * boosted by interactivity modifiers. Changes upon fork,
1931 * setprio syscalls, and whenever the interactivity
1932 * estimator recalculates.
1934 static inline int normal_prio(struct task_struct *p)
1938 if (task_has_rt_policy(p))
1939 prio = MAX_RT_PRIO-1 - p->rt_priority;
1941 prio = __normal_prio(p);
1946 * Calculate the current priority, i.e. the priority
1947 * taken into account by the scheduler. This value might
1948 * be boosted by RT tasks, or might be boosted by
1949 * interactivity modifiers. Will be RT if the task got
1950 * RT-boosted. If not then it returns p->normal_prio.
1952 static int effective_prio(struct task_struct *p)
1954 p->normal_prio = normal_prio(p);
1956 * If we are RT tasks or we were boosted to RT priority,
1957 * keep the priority unchanged. Otherwise, update priority
1958 * to the normal priority:
1960 if (!rt_prio(p->prio))
1961 return p->normal_prio;
1966 * activate_task - move a task to the runqueue.
1968 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1970 if (task_contributes_to_load(p))
1971 rq->nr_uninterruptible--;
1973 enqueue_task(rq, p, wakeup);
1978 * deactivate_task - remove a task from the runqueue.
1980 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1982 if (task_contributes_to_load(p))
1983 rq->nr_uninterruptible++;
1985 dequeue_task(rq, p, sleep);
1990 * task_curr - is this task currently executing on a CPU?
1991 * @p: the task in question.
1993 inline int task_curr(const struct task_struct *p)
1995 return cpu_curr(task_cpu(p)) == p;
1998 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1999 const struct sched_class *prev_class,
2000 int oldprio, int running)
2002 if (prev_class != p->sched_class) {
2003 if (prev_class->switched_from)
2004 prev_class->switched_from(rq, p, running);
2005 p->sched_class->switched_to(rq, p, running);
2007 p->sched_class->prio_changed(rq, p, oldprio, running);
2012 * Is this task likely cache-hot:
2015 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2019 if (p->sched_class != &fair_sched_class)
2023 * Buddy candidates are cache hot:
2025 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2026 (&p->se == cfs_rq_of(&p->se)->next ||
2027 &p->se == cfs_rq_of(&p->se)->last))
2030 if (sysctl_sched_migration_cost == -1)
2032 if (sysctl_sched_migration_cost == 0)
2035 delta = now - p->se.exec_start;
2037 return delta < (s64)sysctl_sched_migration_cost;
2040 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2042 #ifdef CONFIG_SCHED_DEBUG
2044 * We should never call set_task_cpu() on a blocked task,
2045 * ttwu() will sort out the placement.
2047 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2048 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2051 trace_sched_migrate_task(p, new_cpu);
2053 if (task_cpu(p) != new_cpu) {
2054 p->se.nr_migrations++;
2055 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2058 __set_task_cpu(p, new_cpu);
2061 struct migration_req {
2062 struct list_head list;
2064 struct task_struct *task;
2067 struct completion done;
2071 * The task's runqueue lock must be held.
2072 * Returns true if you have to wait for migration thread.
2075 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2077 struct rq *rq = task_rq(p);
2080 * If the task is not on a runqueue (and not running), then
2081 * the next wake-up will properly place the task.
2083 if (!p->se.on_rq && !task_running(rq, p))
2086 init_completion(&req->done);
2088 req->dest_cpu = dest_cpu;
2089 list_add(&req->list, &rq->migration_queue);
2095 * wait_task_context_switch - wait for a thread to complete at least one
2098 * @p must not be current.
2100 void wait_task_context_switch(struct task_struct *p)
2102 unsigned long nvcsw, nivcsw, flags;
2110 * The runqueue is assigned before the actual context
2111 * switch. We need to take the runqueue lock.
2113 * We could check initially without the lock but it is
2114 * very likely that we need to take the lock in every
2117 rq = task_rq_lock(p, &flags);
2118 running = task_running(rq, p);
2119 task_rq_unlock(rq, &flags);
2121 if (likely(!running))
2124 * The switch count is incremented before the actual
2125 * context switch. We thus wait for two switches to be
2126 * sure at least one completed.
2128 if ((p->nvcsw - nvcsw) > 1)
2130 if ((p->nivcsw - nivcsw) > 1)
2138 * wait_task_inactive - wait for a thread to unschedule.
2140 * If @match_state is nonzero, it's the @p->state value just checked and
2141 * not expected to change. If it changes, i.e. @p might have woken up,
2142 * then return zero. When we succeed in waiting for @p to be off its CPU,
2143 * we return a positive number (its total switch count). If a second call
2144 * a short while later returns the same number, the caller can be sure that
2145 * @p has remained unscheduled the whole time.
2147 * The caller must ensure that the task *will* unschedule sometime soon,
2148 * else this function might spin for a *long* time. This function can't
2149 * be called with interrupts off, or it may introduce deadlock with
2150 * smp_call_function() if an IPI is sent by the same process we are
2151 * waiting to become inactive.
2153 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2155 unsigned long flags;
2162 * We do the initial early heuristics without holding
2163 * any task-queue locks at all. We'll only try to get
2164 * the runqueue lock when things look like they will
2170 * If the task is actively running on another CPU
2171 * still, just relax and busy-wait without holding
2174 * NOTE! Since we don't hold any locks, it's not
2175 * even sure that "rq" stays as the right runqueue!
2176 * But we don't care, since "task_running()" will
2177 * return false if the runqueue has changed and p
2178 * is actually now running somewhere else!
2180 while (task_running(rq, p)) {
2181 if (match_state && unlikely(p->state != match_state))
2187 * Ok, time to look more closely! We need the rq
2188 * lock now, to be *sure*. If we're wrong, we'll
2189 * just go back and repeat.
2191 rq = task_rq_lock(p, &flags);
2192 trace_sched_wait_task(rq, p);
2193 running = task_running(rq, p);
2194 on_rq = p->se.on_rq;
2196 if (!match_state || p->state == match_state)
2197 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2198 task_rq_unlock(rq, &flags);
2201 * If it changed from the expected state, bail out now.
2203 if (unlikely(!ncsw))
2207 * Was it really running after all now that we
2208 * checked with the proper locks actually held?
2210 * Oops. Go back and try again..
2212 if (unlikely(running)) {
2218 * It's not enough that it's not actively running,
2219 * it must be off the runqueue _entirely_, and not
2222 * So if it was still runnable (but just not actively
2223 * running right now), it's preempted, and we should
2224 * yield - it could be a while.
2226 if (unlikely(on_rq)) {
2227 schedule_timeout_uninterruptible(1);
2232 * Ahh, all good. It wasn't running, and it wasn't
2233 * runnable, which means that it will never become
2234 * running in the future either. We're all done!
2243 * kick_process - kick a running thread to enter/exit the kernel
2244 * @p: the to-be-kicked thread
2246 * Cause a process which is running on another CPU to enter
2247 * kernel-mode, without any delay. (to get signals handled.)
2249 * NOTE: this function doesnt have to take the runqueue lock,
2250 * because all it wants to ensure is that the remote task enters
2251 * the kernel. If the IPI races and the task has been migrated
2252 * to another CPU then no harm is done and the purpose has been
2255 void kick_process(struct task_struct *p)
2261 if ((cpu != smp_processor_id()) && task_curr(p))
2262 smp_send_reschedule(cpu);
2265 EXPORT_SYMBOL_GPL(kick_process);
2266 #endif /* CONFIG_SMP */
2269 * task_oncpu_function_call - call a function on the cpu on which a task runs
2270 * @p: the task to evaluate
2271 * @func: the function to be called
2272 * @info: the function call argument
2274 * Calls the function @func when the task is currently running. This might
2275 * be on the current CPU, which just calls the function directly
2277 void task_oncpu_function_call(struct task_struct *p,
2278 void (*func) (void *info), void *info)
2285 smp_call_function_single(cpu, func, info, 1);
2290 static int select_fallback_rq(int cpu, struct task_struct *p)
2293 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2295 /* Look for allowed, online CPU in same node. */
2296 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2297 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2300 /* Any allowed, online CPU? */
2301 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2302 if (dest_cpu < nr_cpu_ids)
2305 /* No more Mr. Nice Guy. */
2306 if (dest_cpu >= nr_cpu_ids) {
2308 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
2310 dest_cpu = cpumask_any_and(cpu_active_mask, &p->cpus_allowed);
2313 * Don't tell them about moving exiting tasks or
2314 * kernel threads (both mm NULL), since they never
2317 if (p->mm && printk_ratelimit()) {
2318 printk(KERN_INFO "process %d (%s) no "
2319 "longer affine to cpu%d\n",
2320 task_pid_nr(p), p->comm, cpu);
2328 * Gets called from 3 sites (exec, fork, wakeup), since it is called without
2329 * holding rq->lock we need to ensure ->cpus_allowed is stable, this is done
2332 * exec: is unstable, retry loop
2333 * fork & wake-up: serialize ->cpus_allowed against TASK_WAKING
2336 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2338 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2341 * In order not to call set_task_cpu() on a blocking task we need
2342 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2345 * Since this is common to all placement strategies, this lives here.
2347 * [ this allows ->select_task() to simply return task_cpu(p) and
2348 * not worry about this generic constraint ]
2350 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2352 cpu = select_fallback_rq(task_cpu(p), p);
2359 * try_to_wake_up - wake up a thread
2360 * @p: the to-be-woken-up thread
2361 * @state: the mask of task states that can be woken
2362 * @sync: do a synchronous wakeup?
2364 * Put it on the run-queue if it's not already there. The "current"
2365 * thread is always on the run-queue (except when the actual
2366 * re-schedule is in progress), and as such you're allowed to do
2367 * the simpler "current->state = TASK_RUNNING" to mark yourself
2368 * runnable without the overhead of this.
2370 * returns failure only if the task is already active.
2372 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2375 int cpu, orig_cpu, this_cpu, success = 0;
2376 unsigned long flags;
2377 struct rq *rq, *orig_rq;
2379 if (!sched_feat(SYNC_WAKEUPS))
2380 wake_flags &= ~WF_SYNC;
2382 this_cpu = get_cpu();
2385 rq = orig_rq = task_rq_lock(p, &flags);
2386 update_rq_clock(rq);
2387 if (!(p->state & state))
2397 if (unlikely(task_running(rq, p)))
2401 * In order to handle concurrent wakeups and release the rq->lock
2402 * we put the task in TASK_WAKING state.
2404 * First fix up the nr_uninterruptible count:
2406 if (task_contributes_to_load(p))
2407 rq->nr_uninterruptible--;
2408 p->state = TASK_WAKING;
2410 if (p->sched_class->task_waking)
2411 p->sched_class->task_waking(rq, p);
2413 __task_rq_unlock(rq);
2415 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2416 if (cpu != orig_cpu)
2417 set_task_cpu(p, cpu);
2419 rq = __task_rq_lock(p);
2420 update_rq_clock(rq);
2422 WARN_ON(p->state != TASK_WAKING);
2425 #ifdef CONFIG_SCHEDSTATS
2426 schedstat_inc(rq, ttwu_count);
2427 if (cpu == this_cpu)
2428 schedstat_inc(rq, ttwu_local);
2430 struct sched_domain *sd;
2431 for_each_domain(this_cpu, sd) {
2432 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2433 schedstat_inc(sd, ttwu_wake_remote);
2438 #endif /* CONFIG_SCHEDSTATS */
2441 #endif /* CONFIG_SMP */
2442 schedstat_inc(p, se.nr_wakeups);
2443 if (wake_flags & WF_SYNC)
2444 schedstat_inc(p, se.nr_wakeups_sync);
2445 if (orig_cpu != cpu)
2446 schedstat_inc(p, se.nr_wakeups_migrate);
2447 if (cpu == this_cpu)
2448 schedstat_inc(p, se.nr_wakeups_local);
2450 schedstat_inc(p, se.nr_wakeups_remote);
2451 activate_task(rq, p, 1);
2455 * Only attribute actual wakeups done by this task.
2457 if (!in_interrupt()) {
2458 struct sched_entity *se = ¤t->se;
2459 u64 sample = se->sum_exec_runtime;
2461 if (se->last_wakeup)
2462 sample -= se->last_wakeup;
2464 sample -= se->start_runtime;
2465 update_avg(&se->avg_wakeup, sample);
2467 se->last_wakeup = se->sum_exec_runtime;
2471 trace_sched_wakeup(rq, p, success);
2472 check_preempt_curr(rq, p, wake_flags);
2474 p->state = TASK_RUNNING;
2476 if (p->sched_class->task_woken)
2477 p->sched_class->task_woken(rq, p);
2479 if (unlikely(rq->idle_stamp)) {
2480 u64 delta = rq->clock - rq->idle_stamp;
2481 u64 max = 2*sysctl_sched_migration_cost;
2486 update_avg(&rq->avg_idle, delta);
2491 task_rq_unlock(rq, &flags);
2498 * wake_up_process - Wake up a specific process
2499 * @p: The process to be woken up.
2501 * Attempt to wake up the nominated process and move it to the set of runnable
2502 * processes. Returns 1 if the process was woken up, 0 if it was already
2505 * It may be assumed that this function implies a write memory barrier before
2506 * changing the task state if and only if any tasks are woken up.
2508 int wake_up_process(struct task_struct *p)
2510 return try_to_wake_up(p, TASK_ALL, 0);
2512 EXPORT_SYMBOL(wake_up_process);
2514 int wake_up_state(struct task_struct *p, unsigned int state)
2516 return try_to_wake_up(p, state, 0);
2520 * Perform scheduler related setup for a newly forked process p.
2521 * p is forked by current.
2523 * __sched_fork() is basic setup used by init_idle() too:
2525 static void __sched_fork(struct task_struct *p)
2527 p->se.exec_start = 0;
2528 p->se.sum_exec_runtime = 0;
2529 p->se.prev_sum_exec_runtime = 0;
2530 p->se.nr_migrations = 0;
2531 p->se.last_wakeup = 0;
2532 p->se.avg_overlap = 0;
2533 p->se.start_runtime = 0;
2534 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2536 #ifdef CONFIG_SCHEDSTATS
2537 p->se.wait_start = 0;
2539 p->se.wait_count = 0;
2542 p->se.sleep_start = 0;
2543 p->se.sleep_max = 0;
2544 p->se.sum_sleep_runtime = 0;
2546 p->se.block_start = 0;
2547 p->se.block_max = 0;
2549 p->se.slice_max = 0;
2551 p->se.nr_migrations_cold = 0;
2552 p->se.nr_failed_migrations_affine = 0;
2553 p->se.nr_failed_migrations_running = 0;
2554 p->se.nr_failed_migrations_hot = 0;
2555 p->se.nr_forced_migrations = 0;
2557 p->se.nr_wakeups = 0;
2558 p->se.nr_wakeups_sync = 0;
2559 p->se.nr_wakeups_migrate = 0;
2560 p->se.nr_wakeups_local = 0;
2561 p->se.nr_wakeups_remote = 0;
2562 p->se.nr_wakeups_affine = 0;
2563 p->se.nr_wakeups_affine_attempts = 0;
2564 p->se.nr_wakeups_passive = 0;
2565 p->se.nr_wakeups_idle = 0;
2569 INIT_LIST_HEAD(&p->rt.run_list);
2571 INIT_LIST_HEAD(&p->se.group_node);
2573 #ifdef CONFIG_PREEMPT_NOTIFIERS
2574 INIT_HLIST_HEAD(&p->preempt_notifiers);
2579 * fork()/clone()-time setup:
2581 void sched_fork(struct task_struct *p, int clone_flags)
2583 int cpu = get_cpu();
2587 * We mark the process as waking here. This guarantees that
2588 * nobody will actually run it, and a signal or other external
2589 * event cannot wake it up and insert it on the runqueue either.
2591 p->state = TASK_WAKING;
2594 * Revert to default priority/policy on fork if requested.
2596 if (unlikely(p->sched_reset_on_fork)) {
2597 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2598 p->policy = SCHED_NORMAL;
2599 p->normal_prio = p->static_prio;
2602 if (PRIO_TO_NICE(p->static_prio) < 0) {
2603 p->static_prio = NICE_TO_PRIO(0);
2604 p->normal_prio = p->static_prio;
2609 * We don't need the reset flag anymore after the fork. It has
2610 * fulfilled its duty:
2612 p->sched_reset_on_fork = 0;
2616 * Make sure we do not leak PI boosting priority to the child.
2618 p->prio = current->normal_prio;
2620 if (!rt_prio(p->prio))
2621 p->sched_class = &fair_sched_class;
2623 if (p->sched_class->task_fork)
2624 p->sched_class->task_fork(p);
2626 set_task_cpu(p, cpu);
2628 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2629 if (likely(sched_info_on()))
2630 memset(&p->sched_info, 0, sizeof(p->sched_info));
2632 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2635 #ifdef CONFIG_PREEMPT
2636 /* Want to start with kernel preemption disabled. */
2637 task_thread_info(p)->preempt_count = 1;
2639 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2645 * wake_up_new_task - wake up a newly created task for the first time.
2647 * This function will do some initial scheduler statistics housekeeping
2648 * that must be done for every newly created context, then puts the task
2649 * on the runqueue and wakes it.
2651 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2653 unsigned long flags;
2655 int cpu = get_cpu();
2659 * Fork balancing, do it here and not earlier because:
2660 * - cpus_allowed can change in the fork path
2661 * - any previously selected cpu might disappear through hotplug
2663 * We still have TASK_WAKING but PF_STARTING is gone now, meaning
2664 * ->cpus_allowed is stable, we have preemption disabled, meaning
2665 * cpu_online_mask is stable.
2667 cpu = select_task_rq(p, SD_BALANCE_FORK, 0);
2668 set_task_cpu(p, cpu);
2671 rq = task_rq_lock(p, &flags);
2672 BUG_ON(p->state != TASK_WAKING);
2673 p->state = TASK_RUNNING;
2674 update_rq_clock(rq);
2675 activate_task(rq, p, 0);
2676 trace_sched_wakeup_new(rq, p, 1);
2677 check_preempt_curr(rq, p, WF_FORK);
2679 if (p->sched_class->task_woken)
2680 p->sched_class->task_woken(rq, p);
2682 task_rq_unlock(rq, &flags);
2686 #ifdef CONFIG_PREEMPT_NOTIFIERS
2689 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2690 * @notifier: notifier struct to register
2692 void preempt_notifier_register(struct preempt_notifier *notifier)
2694 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2696 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2699 * preempt_notifier_unregister - no longer interested in preemption notifications
2700 * @notifier: notifier struct to unregister
2702 * This is safe to call from within a preemption notifier.
2704 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2706 hlist_del(¬ifier->link);
2708 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2710 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2712 struct preempt_notifier *notifier;
2713 struct hlist_node *node;
2715 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2716 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2720 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2721 struct task_struct *next)
2723 struct preempt_notifier *notifier;
2724 struct hlist_node *node;
2726 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2727 notifier->ops->sched_out(notifier, next);
2730 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2732 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2737 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2738 struct task_struct *next)
2742 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2745 * prepare_task_switch - prepare to switch tasks
2746 * @rq: the runqueue preparing to switch
2747 * @prev: the current task that is being switched out
2748 * @next: the task we are going to switch to.
2750 * This is called with the rq lock held and interrupts off. It must
2751 * be paired with a subsequent finish_task_switch after the context
2754 * prepare_task_switch sets up locking and calls architecture specific
2758 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2759 struct task_struct *next)
2761 fire_sched_out_preempt_notifiers(prev, next);
2762 prepare_lock_switch(rq, next);
2763 prepare_arch_switch(next);
2767 * finish_task_switch - clean up after a task-switch
2768 * @rq: runqueue associated with task-switch
2769 * @prev: the thread we just switched away from.
2771 * finish_task_switch must be called after the context switch, paired
2772 * with a prepare_task_switch call before the context switch.
2773 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2774 * and do any other architecture-specific cleanup actions.
2776 * Note that we may have delayed dropping an mm in context_switch(). If
2777 * so, we finish that here outside of the runqueue lock. (Doing it
2778 * with the lock held can cause deadlocks; see schedule() for
2781 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2782 __releases(rq->lock)
2784 struct mm_struct *mm = rq->prev_mm;
2790 * A task struct has one reference for the use as "current".
2791 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2792 * schedule one last time. The schedule call will never return, and
2793 * the scheduled task must drop that reference.
2794 * The test for TASK_DEAD must occur while the runqueue locks are
2795 * still held, otherwise prev could be scheduled on another cpu, die
2796 * there before we look at prev->state, and then the reference would
2798 * Manfred Spraul <manfred@colorfullife.com>
2800 prev_state = prev->state;
2801 finish_arch_switch(prev);
2802 perf_event_task_sched_in(current, cpu_of(rq));
2803 finish_lock_switch(rq, prev);
2805 fire_sched_in_preempt_notifiers(current);
2808 if (unlikely(prev_state == TASK_DEAD)) {
2810 * Remove function-return probe instances associated with this
2811 * task and put them back on the free list.
2813 kprobe_flush_task(prev);
2814 put_task_struct(prev);
2820 /* assumes rq->lock is held */
2821 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2823 if (prev->sched_class->pre_schedule)
2824 prev->sched_class->pre_schedule(rq, prev);
2827 /* rq->lock is NOT held, but preemption is disabled */
2828 static inline void post_schedule(struct rq *rq)
2830 if (rq->post_schedule) {
2831 unsigned long flags;
2833 raw_spin_lock_irqsave(&rq->lock, flags);
2834 if (rq->curr->sched_class->post_schedule)
2835 rq->curr->sched_class->post_schedule(rq);
2836 raw_spin_unlock_irqrestore(&rq->lock, flags);
2838 rq->post_schedule = 0;
2844 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2848 static inline void post_schedule(struct rq *rq)
2855 * schedule_tail - first thing a freshly forked thread must call.
2856 * @prev: the thread we just switched away from.
2858 asmlinkage void schedule_tail(struct task_struct *prev)
2859 __releases(rq->lock)
2861 struct rq *rq = this_rq();
2863 finish_task_switch(rq, prev);
2866 * FIXME: do we need to worry about rq being invalidated by the
2871 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2872 /* In this case, finish_task_switch does not reenable preemption */
2875 if (current->set_child_tid)
2876 put_user(task_pid_vnr(current), current->set_child_tid);
2880 * context_switch - switch to the new MM and the new
2881 * thread's register state.
2884 context_switch(struct rq *rq, struct task_struct *prev,
2885 struct task_struct *next)
2887 struct mm_struct *mm, *oldmm;
2889 prepare_task_switch(rq, prev, next);
2890 trace_sched_switch(rq, prev, next);
2892 oldmm = prev->active_mm;
2894 * For paravirt, this is coupled with an exit in switch_to to
2895 * combine the page table reload and the switch backend into
2898 arch_start_context_switch(prev);
2901 next->active_mm = oldmm;
2902 atomic_inc(&oldmm->mm_count);
2903 enter_lazy_tlb(oldmm, next);
2905 switch_mm(oldmm, mm, next);
2907 if (likely(!prev->mm)) {
2908 prev->active_mm = NULL;
2909 rq->prev_mm = oldmm;
2912 * Since the runqueue lock will be released by the next
2913 * task (which is an invalid locking op but in the case
2914 * of the scheduler it's an obvious special-case), so we
2915 * do an early lockdep release here:
2917 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2918 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2921 /* Here we just switch the register state and the stack. */
2922 switch_to(prev, next, prev);
2926 * this_rq must be evaluated again because prev may have moved
2927 * CPUs since it called schedule(), thus the 'rq' on its stack
2928 * frame will be invalid.
2930 finish_task_switch(this_rq(), prev);
2934 * nr_running, nr_uninterruptible and nr_context_switches:
2936 * externally visible scheduler statistics: current number of runnable
2937 * threads, current number of uninterruptible-sleeping threads, total
2938 * number of context switches performed since bootup.
2940 unsigned long nr_running(void)
2942 unsigned long i, sum = 0;
2944 for_each_online_cpu(i)
2945 sum += cpu_rq(i)->nr_running;
2950 unsigned long nr_uninterruptible(void)
2952 unsigned long i, sum = 0;
2954 for_each_possible_cpu(i)
2955 sum += cpu_rq(i)->nr_uninterruptible;
2958 * Since we read the counters lockless, it might be slightly
2959 * inaccurate. Do not allow it to go below zero though:
2961 if (unlikely((long)sum < 0))
2967 unsigned long long nr_context_switches(void)
2970 unsigned long long sum = 0;
2972 for_each_possible_cpu(i)
2973 sum += cpu_rq(i)->nr_switches;
2978 unsigned long nr_iowait(void)
2980 unsigned long i, sum = 0;
2982 for_each_possible_cpu(i)
2983 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2988 unsigned long nr_iowait_cpu(void)
2990 struct rq *this = this_rq();
2991 return atomic_read(&this->nr_iowait);
2994 unsigned long this_cpu_load(void)
2996 struct rq *this = this_rq();
2997 return this->cpu_load[0];
3001 /* Variables and functions for calc_load */
3002 static atomic_long_t calc_load_tasks;
3003 static unsigned long calc_load_update;
3004 unsigned long avenrun[3];
3005 EXPORT_SYMBOL(avenrun);
3008 * get_avenrun - get the load average array
3009 * @loads: pointer to dest load array
3010 * @offset: offset to add
3011 * @shift: shift count to shift the result left
3013 * These values are estimates at best, so no need for locking.
3015 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3017 loads[0] = (avenrun[0] + offset) << shift;
3018 loads[1] = (avenrun[1] + offset) << shift;
3019 loads[2] = (avenrun[2] + offset) << shift;
3022 static unsigned long
3023 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3026 load += active * (FIXED_1 - exp);
3027 return load >> FSHIFT;
3031 * calc_load - update the avenrun load estimates 10 ticks after the
3032 * CPUs have updated calc_load_tasks.
3034 void calc_global_load(void)
3036 unsigned long upd = calc_load_update + 10;
3039 if (time_before(jiffies, upd))
3042 active = atomic_long_read(&calc_load_tasks);
3043 active = active > 0 ? active * FIXED_1 : 0;
3045 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3046 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3047 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3049 calc_load_update += LOAD_FREQ;
3053 * Either called from update_cpu_load() or from a cpu going idle
3055 static void calc_load_account_active(struct rq *this_rq)
3057 long nr_active, delta;
3059 nr_active = this_rq->nr_running;
3060 nr_active += (long) this_rq->nr_uninterruptible;
3062 if (nr_active != this_rq->calc_load_active) {
3063 delta = nr_active - this_rq->calc_load_active;
3064 this_rq->calc_load_active = nr_active;
3065 atomic_long_add(delta, &calc_load_tasks);
3070 * Update rq->cpu_load[] statistics. This function is usually called every
3071 * scheduler tick (TICK_NSEC).
3073 static void update_cpu_load(struct rq *this_rq)
3075 unsigned long this_load = this_rq->load.weight;
3078 this_rq->nr_load_updates++;
3080 /* Update our load: */
3081 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3082 unsigned long old_load, new_load;
3084 /* scale is effectively 1 << i now, and >> i divides by scale */
3086 old_load = this_rq->cpu_load[i];
3087 new_load = this_load;
3089 * Round up the averaging division if load is increasing. This
3090 * prevents us from getting stuck on 9 if the load is 10, for
3093 if (new_load > old_load)
3094 new_load += scale-1;
3095 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3098 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3099 this_rq->calc_load_update += LOAD_FREQ;
3100 calc_load_account_active(this_rq);
3107 * double_rq_lock - safely lock two runqueues
3109 * Note this does not disable interrupts like task_rq_lock,
3110 * you need to do so manually before calling.
3112 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
3113 __acquires(rq1->lock)
3114 __acquires(rq2->lock)
3116 BUG_ON(!irqs_disabled());
3118 raw_spin_lock(&rq1->lock);
3119 __acquire(rq2->lock); /* Fake it out ;) */
3122 raw_spin_lock(&rq1->lock);
3123 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
3125 raw_spin_lock(&rq2->lock);
3126 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
3129 update_rq_clock(rq1);
3130 update_rq_clock(rq2);
3134 * double_rq_unlock - safely unlock two runqueues
3136 * Note this does not restore interrupts like task_rq_unlock,
3137 * you need to do so manually after calling.
3139 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
3140 __releases(rq1->lock)
3141 __releases(rq2->lock)
3143 raw_spin_unlock(&rq1->lock);
3145 raw_spin_unlock(&rq2->lock);
3147 __release(rq2->lock);
3151 * sched_exec - execve() is a valuable balancing opportunity, because at
3152 * this point the task has the smallest effective memory and cache footprint.
3154 void sched_exec(void)
3156 struct task_struct *p = current;
3157 struct migration_req req;
3158 int dest_cpu, this_cpu;
3159 unsigned long flags;
3163 this_cpu = get_cpu();
3164 dest_cpu = select_task_rq(p, SD_BALANCE_EXEC, 0);
3165 if (dest_cpu == this_cpu) {
3170 rq = task_rq_lock(p, &flags);
3174 * select_task_rq() can race against ->cpus_allowed
3176 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
3177 || unlikely(!cpu_active(dest_cpu))) {
3178 task_rq_unlock(rq, &flags);
3182 /* force the process onto the specified CPU */
3183 if (migrate_task(p, dest_cpu, &req)) {
3184 /* Need to wait for migration thread (might exit: take ref). */
3185 struct task_struct *mt = rq->migration_thread;
3187 get_task_struct(mt);
3188 task_rq_unlock(rq, &flags);
3189 wake_up_process(mt);
3190 put_task_struct(mt);
3191 wait_for_completion(&req.done);
3195 task_rq_unlock(rq, &flags);
3199 * pull_task - move a task from a remote runqueue to the local runqueue.
3200 * Both runqueues must be locked.
3202 static void pull_task(struct rq *src_rq, struct task_struct *p,
3203 struct rq *this_rq, int this_cpu)
3205 deactivate_task(src_rq, p, 0);
3206 set_task_cpu(p, this_cpu);
3207 activate_task(this_rq, p, 0);
3208 check_preempt_curr(this_rq, p, 0);
3212 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3215 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3216 struct sched_domain *sd, enum cpu_idle_type idle,
3219 int tsk_cache_hot = 0;
3221 * We do not migrate tasks that are:
3222 * 1) running (obviously), or
3223 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3224 * 3) are cache-hot on their current CPU.
3226 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
3227 schedstat_inc(p, se.nr_failed_migrations_affine);
3232 if (task_running(rq, p)) {
3233 schedstat_inc(p, se.nr_failed_migrations_running);
3238 * Aggressive migration if:
3239 * 1) task is cache cold, or
3240 * 2) too many balance attempts have failed.
3243 tsk_cache_hot = task_hot(p, rq->clock, sd);
3244 if (!tsk_cache_hot ||
3245 sd->nr_balance_failed > sd->cache_nice_tries) {
3246 #ifdef CONFIG_SCHEDSTATS
3247 if (tsk_cache_hot) {
3248 schedstat_inc(sd, lb_hot_gained[idle]);
3249 schedstat_inc(p, se.nr_forced_migrations);
3255 if (tsk_cache_hot) {
3256 schedstat_inc(p, se.nr_failed_migrations_hot);
3262 static unsigned long
3263 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3264 unsigned long max_load_move, struct sched_domain *sd,
3265 enum cpu_idle_type idle, int *all_pinned,
3266 int *this_best_prio, struct rq_iterator *iterator)
3268 int loops = 0, pulled = 0, pinned = 0;
3269 struct task_struct *p;
3270 long rem_load_move = max_load_move;
3272 if (max_load_move == 0)
3278 * Start the load-balancing iterator:
3280 p = iterator->start(iterator->arg);
3282 if (!p || loops++ > sysctl_sched_nr_migrate)
3285 if ((p->se.load.weight >> 1) > rem_load_move ||
3286 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3287 p = iterator->next(iterator->arg);
3291 pull_task(busiest, p, this_rq, this_cpu);
3293 rem_load_move -= p->se.load.weight;
3295 #ifdef CONFIG_PREEMPT
3297 * NEWIDLE balancing is a source of latency, so preemptible kernels
3298 * will stop after the first task is pulled to minimize the critical
3301 if (idle == CPU_NEWLY_IDLE)
3306 * We only want to steal up to the prescribed amount of weighted load.
3308 if (rem_load_move > 0) {
3309 if (p->prio < *this_best_prio)
3310 *this_best_prio = p->prio;
3311 p = iterator->next(iterator->arg);
3316 * Right now, this is one of only two places pull_task() is called,
3317 * so we can safely collect pull_task() stats here rather than
3318 * inside pull_task().
3320 schedstat_add(sd, lb_gained[idle], pulled);
3323 *all_pinned = pinned;
3325 return max_load_move - rem_load_move;
3329 * move_tasks tries to move up to max_load_move weighted load from busiest to
3330 * this_rq, as part of a balancing operation within domain "sd".
3331 * Returns 1 if successful and 0 otherwise.
3333 * Called with both runqueues locked.
3335 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3336 unsigned long max_load_move,
3337 struct sched_domain *sd, enum cpu_idle_type idle,
3340 const struct sched_class *class = sched_class_highest;
3341 unsigned long total_load_moved = 0;
3342 int this_best_prio = this_rq->curr->prio;
3346 class->load_balance(this_rq, this_cpu, busiest,
3347 max_load_move - total_load_moved,
3348 sd, idle, all_pinned, &this_best_prio);
3349 class = class->next;
3351 #ifdef CONFIG_PREEMPT
3353 * NEWIDLE balancing is a source of latency, so preemptible
3354 * kernels will stop after the first task is pulled to minimize
3355 * the critical section.
3357 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3360 } while (class && max_load_move > total_load_moved);
3362 return total_load_moved > 0;
3366 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3367 struct sched_domain *sd, enum cpu_idle_type idle,
3368 struct rq_iterator *iterator)
3370 struct task_struct *p = iterator->start(iterator->arg);
3374 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3375 pull_task(busiest, p, this_rq, this_cpu);
3377 * Right now, this is only the second place pull_task()
3378 * is called, so we can safely collect pull_task()
3379 * stats here rather than inside pull_task().
3381 schedstat_inc(sd, lb_gained[idle]);
3385 p = iterator->next(iterator->arg);
3392 * move_one_task tries to move exactly one task from busiest to this_rq, as
3393 * part of active balancing operations within "domain".
3394 * Returns 1 if successful and 0 otherwise.
3396 * Called with both runqueues locked.
3398 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3399 struct sched_domain *sd, enum cpu_idle_type idle)
3401 const struct sched_class *class;
3403 for_each_class(class) {
3404 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3410 /********** Helpers for find_busiest_group ************************/
3412 * sd_lb_stats - Structure to store the statistics of a sched_domain
3413 * during load balancing.
3415 struct sd_lb_stats {
3416 struct sched_group *busiest; /* Busiest group in this sd */
3417 struct sched_group *this; /* Local group in this sd */
3418 unsigned long total_load; /* Total load of all groups in sd */
3419 unsigned long total_pwr; /* Total power of all groups in sd */
3420 unsigned long avg_load; /* Average load across all groups in sd */
3422 /** Statistics of this group */
3423 unsigned long this_load;
3424 unsigned long this_load_per_task;
3425 unsigned long this_nr_running;
3427 /* Statistics of the busiest group */
3428 unsigned long max_load;
3429 unsigned long busiest_load_per_task;
3430 unsigned long busiest_nr_running;
3432 int group_imb; /* Is there imbalance in this sd */
3433 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3434 int power_savings_balance; /* Is powersave balance needed for this sd */
3435 struct sched_group *group_min; /* Least loaded group in sd */
3436 struct sched_group *group_leader; /* Group which relieves group_min */
3437 unsigned long min_load_per_task; /* load_per_task in group_min */
3438 unsigned long leader_nr_running; /* Nr running of group_leader */
3439 unsigned long min_nr_running; /* Nr running of group_min */
3444 * sg_lb_stats - stats of a sched_group required for load_balancing
3446 struct sg_lb_stats {
3447 unsigned long avg_load; /*Avg load across the CPUs of the group */
3448 unsigned long group_load; /* Total load over the CPUs of the group */
3449 unsigned long sum_nr_running; /* Nr tasks running in the group */
3450 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3451 unsigned long group_capacity;
3452 int group_imb; /* Is there an imbalance in the group ? */
3456 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3457 * @group: The group whose first cpu is to be returned.
3459 static inline unsigned int group_first_cpu(struct sched_group *group)
3461 return cpumask_first(sched_group_cpus(group));
3465 * get_sd_load_idx - Obtain the load index for a given sched domain.
3466 * @sd: The sched_domain whose load_idx is to be obtained.
3467 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3469 static inline int get_sd_load_idx(struct sched_domain *sd,
3470 enum cpu_idle_type idle)
3476 load_idx = sd->busy_idx;
3479 case CPU_NEWLY_IDLE:
3480 load_idx = sd->newidle_idx;
3483 load_idx = sd->idle_idx;
3491 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3493 * init_sd_power_savings_stats - Initialize power savings statistics for
3494 * the given sched_domain, during load balancing.
3496 * @sd: Sched domain whose power-savings statistics are to be initialized.
3497 * @sds: Variable containing the statistics for sd.
3498 * @idle: Idle status of the CPU at which we're performing load-balancing.
3500 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3501 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3504 * Busy processors will not participate in power savings
3507 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3508 sds->power_savings_balance = 0;
3510 sds->power_savings_balance = 1;
3511 sds->min_nr_running = ULONG_MAX;
3512 sds->leader_nr_running = 0;
3517 * update_sd_power_savings_stats - Update the power saving stats for a
3518 * sched_domain while performing load balancing.
3520 * @group: sched_group belonging to the sched_domain under consideration.
3521 * @sds: Variable containing the statistics of the sched_domain
3522 * @local_group: Does group contain the CPU for which we're performing
3524 * @sgs: Variable containing the statistics of the group.
3526 static inline void update_sd_power_savings_stats(struct sched_group *group,
3527 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3530 if (!sds->power_savings_balance)
3534 * If the local group is idle or completely loaded
3535 * no need to do power savings balance at this domain
3537 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3538 !sds->this_nr_running))
3539 sds->power_savings_balance = 0;
3542 * If a group is already running at full capacity or idle,
3543 * don't include that group in power savings calculations
3545 if (!sds->power_savings_balance ||
3546 sgs->sum_nr_running >= sgs->group_capacity ||
3547 !sgs->sum_nr_running)
3551 * Calculate the group which has the least non-idle load.
3552 * This is the group from where we need to pick up the load
3555 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3556 (sgs->sum_nr_running == sds->min_nr_running &&
3557 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3558 sds->group_min = group;
3559 sds->min_nr_running = sgs->sum_nr_running;
3560 sds->min_load_per_task = sgs->sum_weighted_load /
3561 sgs->sum_nr_running;
3565 * Calculate the group which is almost near its
3566 * capacity but still has some space to pick up some load
3567 * from other group and save more power
3569 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
3572 if (sgs->sum_nr_running > sds->leader_nr_running ||
3573 (sgs->sum_nr_running == sds->leader_nr_running &&
3574 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3575 sds->group_leader = group;
3576 sds->leader_nr_running = sgs->sum_nr_running;
3581 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3582 * @sds: Variable containing the statistics of the sched_domain
3583 * under consideration.
3584 * @this_cpu: Cpu at which we're currently performing load-balancing.
3585 * @imbalance: Variable to store the imbalance.
3588 * Check if we have potential to perform some power-savings balance.
3589 * If yes, set the busiest group to be the least loaded group in the
3590 * sched_domain, so that it's CPUs can be put to idle.
3592 * Returns 1 if there is potential to perform power-savings balance.
3595 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3596 int this_cpu, unsigned long *imbalance)
3598 if (!sds->power_savings_balance)
3601 if (sds->this != sds->group_leader ||
3602 sds->group_leader == sds->group_min)
3605 *imbalance = sds->min_load_per_task;
3606 sds->busiest = sds->group_min;
3611 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3612 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3613 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3618 static inline void update_sd_power_savings_stats(struct sched_group *group,
3619 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3624 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3625 int this_cpu, unsigned long *imbalance)
3629 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3632 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
3634 return SCHED_LOAD_SCALE;
3637 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
3639 return default_scale_freq_power(sd, cpu);
3642 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
3644 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3645 unsigned long smt_gain = sd->smt_gain;
3652 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3654 return default_scale_smt_power(sd, cpu);
3657 unsigned long scale_rt_power(int cpu)
3659 struct rq *rq = cpu_rq(cpu);
3660 u64 total, available;
3662 sched_avg_update(rq);
3664 total = sched_avg_period() + (rq->clock - rq->age_stamp);
3665 available = total - rq->rt_avg;
3667 if (unlikely((s64)total < SCHED_LOAD_SCALE))
3668 total = SCHED_LOAD_SCALE;
3670 total >>= SCHED_LOAD_SHIFT;
3672 return div_u64(available, total);
3675 static void update_cpu_power(struct sched_domain *sd, int cpu)
3677 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3678 unsigned long power = SCHED_LOAD_SCALE;
3679 struct sched_group *sdg = sd->groups;
3681 if (sched_feat(ARCH_POWER))
3682 power *= arch_scale_freq_power(sd, cpu);
3684 power *= default_scale_freq_power(sd, cpu);
3686 power >>= SCHED_LOAD_SHIFT;
3688 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
3689 if (sched_feat(ARCH_POWER))
3690 power *= arch_scale_smt_power(sd, cpu);
3692 power *= default_scale_smt_power(sd, cpu);
3694 power >>= SCHED_LOAD_SHIFT;
3697 power *= scale_rt_power(cpu);
3698 power >>= SCHED_LOAD_SHIFT;
3703 sdg->cpu_power = power;
3706 static void update_group_power(struct sched_domain *sd, int cpu)
3708 struct sched_domain *child = sd->child;
3709 struct sched_group *group, *sdg = sd->groups;
3710 unsigned long power;
3713 update_cpu_power(sd, cpu);
3719 group = child->groups;
3721 power += group->cpu_power;
3722 group = group->next;
3723 } while (group != child->groups);
3725 sdg->cpu_power = power;
3729 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3730 * @sd: The sched_domain whose statistics are to be updated.
3731 * @group: sched_group whose statistics are to be updated.
3732 * @this_cpu: Cpu for which load balance is currently performed.
3733 * @idle: Idle status of this_cpu
3734 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3735 * @sd_idle: Idle status of the sched_domain containing group.
3736 * @local_group: Does group contain this_cpu.
3737 * @cpus: Set of cpus considered for load balancing.
3738 * @balance: Should we balance.
3739 * @sgs: variable to hold the statistics for this group.
3741 static inline void update_sg_lb_stats(struct sched_domain *sd,
3742 struct sched_group *group, int this_cpu,
3743 enum cpu_idle_type idle, int load_idx, int *sd_idle,
3744 int local_group, const struct cpumask *cpus,
3745 int *balance, struct sg_lb_stats *sgs)
3747 unsigned long load, max_cpu_load, min_cpu_load;
3749 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3750 unsigned long sum_avg_load_per_task;
3751 unsigned long avg_load_per_task;
3754 balance_cpu = group_first_cpu(group);
3755 if (balance_cpu == this_cpu)
3756 update_group_power(sd, this_cpu);
3759 /* Tally up the load of all CPUs in the group */
3760 sum_avg_load_per_task = avg_load_per_task = 0;
3762 min_cpu_load = ~0UL;
3764 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3765 struct rq *rq = cpu_rq(i);
3767 if (*sd_idle && rq->nr_running)
3770 /* Bias balancing toward cpus of our domain */
3772 if (idle_cpu(i) && !first_idle_cpu) {
3777 load = target_load(i, load_idx);
3779 load = source_load(i, load_idx);
3780 if (load > max_cpu_load)
3781 max_cpu_load = load;
3782 if (min_cpu_load > load)
3783 min_cpu_load = load;
3786 sgs->group_load += load;
3787 sgs->sum_nr_running += rq->nr_running;
3788 sgs->sum_weighted_load += weighted_cpuload(i);
3790 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3794 * First idle cpu or the first cpu(busiest) in this sched group
3795 * is eligible for doing load balancing at this and above
3796 * domains. In the newly idle case, we will allow all the cpu's
3797 * to do the newly idle load balance.
3799 if (idle != CPU_NEWLY_IDLE && local_group &&
3800 balance_cpu != this_cpu && balance) {
3805 /* Adjust by relative CPU power of the group */
3806 sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power;
3810 * Consider the group unbalanced when the imbalance is larger
3811 * than the average weight of two tasks.
3813 * APZ: with cgroup the avg task weight can vary wildly and
3814 * might not be a suitable number - should we keep a
3815 * normalized nr_running number somewhere that negates
3818 avg_load_per_task = (sum_avg_load_per_task * SCHED_LOAD_SCALE) /
3821 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3824 sgs->group_capacity =
3825 DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
3829 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3830 * @sd: sched_domain whose statistics are to be updated.
3831 * @this_cpu: Cpu for which load balance is currently performed.
3832 * @idle: Idle status of this_cpu
3833 * @sd_idle: Idle status of the sched_domain containing group.
3834 * @cpus: Set of cpus considered for load balancing.
3835 * @balance: Should we balance.
3836 * @sds: variable to hold the statistics for this sched_domain.
3838 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3839 enum cpu_idle_type idle, int *sd_idle,
3840 const struct cpumask *cpus, int *balance,
3841 struct sd_lb_stats *sds)
3843 struct sched_domain *child = sd->child;
3844 struct sched_group *group = sd->groups;
3845 struct sg_lb_stats sgs;
3846 int load_idx, prefer_sibling = 0;
3848 if (child && child->flags & SD_PREFER_SIBLING)
3851 init_sd_power_savings_stats(sd, sds, idle);
3852 load_idx = get_sd_load_idx(sd, idle);
3857 local_group = cpumask_test_cpu(this_cpu,
3858 sched_group_cpus(group));
3859 memset(&sgs, 0, sizeof(sgs));
3860 update_sg_lb_stats(sd, group, this_cpu, idle, load_idx, sd_idle,
3861 local_group, cpus, balance, &sgs);
3863 if (local_group && balance && !(*balance))
3866 sds->total_load += sgs.group_load;
3867 sds->total_pwr += group->cpu_power;
3870 * In case the child domain prefers tasks go to siblings
3871 * first, lower the group capacity to one so that we'll try
3872 * and move all the excess tasks away.
3875 sgs.group_capacity = min(sgs.group_capacity, 1UL);
3878 sds->this_load = sgs.avg_load;
3880 sds->this_nr_running = sgs.sum_nr_running;
3881 sds->this_load_per_task = sgs.sum_weighted_load;
3882 } else if (sgs.avg_load > sds->max_load &&
3883 (sgs.sum_nr_running > sgs.group_capacity ||
3885 sds->max_load = sgs.avg_load;
3886 sds->busiest = group;
3887 sds->busiest_nr_running = sgs.sum_nr_running;
3888 sds->busiest_load_per_task = sgs.sum_weighted_load;
3889 sds->group_imb = sgs.group_imb;
3892 update_sd_power_savings_stats(group, sds, local_group, &sgs);
3893 group = group->next;
3894 } while (group != sd->groups);
3898 * fix_small_imbalance - Calculate the minor imbalance that exists
3899 * amongst the groups of a sched_domain, during
3901 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3902 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3903 * @imbalance: Variable to store the imbalance.
3905 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3906 int this_cpu, unsigned long *imbalance)
3908 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3909 unsigned int imbn = 2;
3911 if (sds->this_nr_running) {
3912 sds->this_load_per_task /= sds->this_nr_running;
3913 if (sds->busiest_load_per_task >
3914 sds->this_load_per_task)
3917 sds->this_load_per_task =
3918 cpu_avg_load_per_task(this_cpu);
3920 if (sds->max_load - sds->this_load + sds->busiest_load_per_task >=
3921 sds->busiest_load_per_task * imbn) {
3922 *imbalance = sds->busiest_load_per_task;
3927 * OK, we don't have enough imbalance to justify moving tasks,
3928 * however we may be able to increase total CPU power used by
3932 pwr_now += sds->busiest->cpu_power *
3933 min(sds->busiest_load_per_task, sds->max_load);
3934 pwr_now += sds->this->cpu_power *
3935 min(sds->this_load_per_task, sds->this_load);
3936 pwr_now /= SCHED_LOAD_SCALE;
3938 /* Amount of load we'd subtract */
3939 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3940 sds->busiest->cpu_power;
3941 if (sds->max_load > tmp)
3942 pwr_move += sds->busiest->cpu_power *
3943 min(sds->busiest_load_per_task, sds->max_load - tmp);
3945 /* Amount of load we'd add */
3946 if (sds->max_load * sds->busiest->cpu_power <
3947 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
3948 tmp = (sds->max_load * sds->busiest->cpu_power) /
3949 sds->this->cpu_power;
3951 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3952 sds->this->cpu_power;
3953 pwr_move += sds->this->cpu_power *
3954 min(sds->this_load_per_task, sds->this_load + tmp);
3955 pwr_move /= SCHED_LOAD_SCALE;
3957 /* Move if we gain throughput */
3958 if (pwr_move > pwr_now)
3959 *imbalance = sds->busiest_load_per_task;
3963 * calculate_imbalance - Calculate the amount of imbalance present within the
3964 * groups of a given sched_domain during load balance.
3965 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3966 * @this_cpu: Cpu for which currently load balance is being performed.
3967 * @imbalance: The variable to store the imbalance.
3969 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3970 unsigned long *imbalance)
3972 unsigned long max_pull;
3974 * In the presence of smp nice balancing, certain scenarios can have
3975 * max load less than avg load(as we skip the groups at or below
3976 * its cpu_power, while calculating max_load..)
3978 if (sds->max_load < sds->avg_load) {
3980 return fix_small_imbalance(sds, this_cpu, imbalance);
3983 /* Don't want to pull so many tasks that a group would go idle */
3984 max_pull = min(sds->max_load - sds->avg_load,
3985 sds->max_load - sds->busiest_load_per_task);
3987 /* How much load to actually move to equalise the imbalance */
3988 *imbalance = min(max_pull * sds->busiest->cpu_power,
3989 (sds->avg_load - sds->this_load) * sds->this->cpu_power)
3993 * if *imbalance is less than the average load per runnable task
3994 * there is no gaurantee that any tasks will be moved so we'll have
3995 * a think about bumping its value to force at least one task to be
3998 if (*imbalance < sds->busiest_load_per_task)
3999 return fix_small_imbalance(sds, this_cpu, imbalance);
4002 /******* find_busiest_group() helpers end here *********************/
4005 * find_busiest_group - Returns the busiest group within the sched_domain
4006 * if there is an imbalance. If there isn't an imbalance, and
4007 * the user has opted for power-savings, it returns a group whose
4008 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
4009 * such a group exists.
4011 * Also calculates the amount of weighted load which should be moved
4012 * to restore balance.
4014 * @sd: The sched_domain whose busiest group is to be returned.
4015 * @this_cpu: The cpu for which load balancing is currently being performed.
4016 * @imbalance: Variable which stores amount of weighted load which should
4017 * be moved to restore balance/put a group to idle.
4018 * @idle: The idle status of this_cpu.
4019 * @sd_idle: The idleness of sd
4020 * @cpus: The set of CPUs under consideration for load-balancing.
4021 * @balance: Pointer to a variable indicating if this_cpu
4022 * is the appropriate cpu to perform load balancing at this_level.
4024 * Returns: - the busiest group if imbalance exists.
4025 * - If no imbalance and user has opted for power-savings balance,
4026 * return the least loaded group whose CPUs can be
4027 * put to idle by rebalancing its tasks onto our group.
4029 static struct sched_group *
4030 find_busiest_group(struct sched_domain *sd, int this_cpu,
4031 unsigned long *imbalance, enum cpu_idle_type idle,
4032 int *sd_idle, const struct cpumask *cpus, int *balance)
4034 struct sd_lb_stats sds;
4036 memset(&sds, 0, sizeof(sds));
4039 * Compute the various statistics relavent for load balancing at
4042 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
4045 /* Cases where imbalance does not exist from POV of this_cpu */
4046 /* 1) this_cpu is not the appropriate cpu to perform load balancing
4048 * 2) There is no busy sibling group to pull from.
4049 * 3) This group is the busiest group.
4050 * 4) This group is more busy than the avg busieness at this
4052 * 5) The imbalance is within the specified limit.
4053 * 6) Any rebalance would lead to ping-pong
4055 if (balance && !(*balance))
4058 if (!sds.busiest || sds.busiest_nr_running == 0)
4061 if (sds.this_load >= sds.max_load)
4064 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
4066 if (sds.this_load >= sds.avg_load)
4069 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
4072 sds.busiest_load_per_task /= sds.busiest_nr_running;
4074 sds.busiest_load_per_task =
4075 min(sds.busiest_load_per_task, sds.avg_load);
4078 * We're trying to get all the cpus to the average_load, so we don't
4079 * want to push ourselves above the average load, nor do we wish to
4080 * reduce the max loaded cpu below the average load, as either of these
4081 * actions would just result in more rebalancing later, and ping-pong
4082 * tasks around. Thus we look for the minimum possible imbalance.
4083 * Negative imbalances (*we* are more loaded than anyone else) will
4084 * be counted as no imbalance for these purposes -- we can't fix that
4085 * by pulling tasks to us. Be careful of negative numbers as they'll
4086 * appear as very large values with unsigned longs.
4088 if (sds.max_load <= sds.busiest_load_per_task)
4091 /* Looks like there is an imbalance. Compute it */
4092 calculate_imbalance(&sds, this_cpu, imbalance);
4097 * There is no obvious imbalance. But check if we can do some balancing
4100 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
4108 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4111 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
4112 unsigned long imbalance, const struct cpumask *cpus)
4114 struct rq *busiest = NULL, *rq;
4115 unsigned long max_load = 0;
4118 for_each_cpu(i, sched_group_cpus(group)) {
4119 unsigned long power = power_of(i);
4120 unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
4123 if (!cpumask_test_cpu(i, cpus))
4127 wl = weighted_cpuload(i) * SCHED_LOAD_SCALE;
4130 if (capacity && rq->nr_running == 1 && wl > imbalance)
4133 if (wl > max_load) {
4143 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4144 * so long as it is large enough.
4146 #define MAX_PINNED_INTERVAL 512
4148 /* Working cpumask for load_balance and load_balance_newidle. */
4149 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4152 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4153 * tasks if there is an imbalance.
4155 static int load_balance(int this_cpu, struct rq *this_rq,
4156 struct sched_domain *sd, enum cpu_idle_type idle,
4159 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
4160 struct sched_group *group;
4161 unsigned long imbalance;
4163 unsigned long flags;
4164 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4166 cpumask_copy(cpus, cpu_active_mask);
4169 * When power savings policy is enabled for the parent domain, idle
4170 * sibling can pick up load irrespective of busy siblings. In this case,
4171 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4172 * portraying it as CPU_NOT_IDLE.
4174 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
4175 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4178 schedstat_inc(sd, lb_count[idle]);
4182 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
4189 schedstat_inc(sd, lb_nobusyg[idle]);
4193 busiest = find_busiest_queue(group, idle, imbalance, cpus);
4195 schedstat_inc(sd, lb_nobusyq[idle]);
4199 BUG_ON(busiest == this_rq);
4201 schedstat_add(sd, lb_imbalance[idle], imbalance);
4204 if (busiest->nr_running > 1) {
4206 * Attempt to move tasks. If find_busiest_group has found
4207 * an imbalance but busiest->nr_running <= 1, the group is
4208 * still unbalanced. ld_moved simply stays zero, so it is
4209 * correctly treated as an imbalance.
4211 local_irq_save(flags);
4212 double_rq_lock(this_rq, busiest);
4213 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4214 imbalance, sd, idle, &all_pinned);
4215 double_rq_unlock(this_rq, busiest);
4216 local_irq_restore(flags);
4219 * some other cpu did the load balance for us.
4221 if (ld_moved && this_cpu != smp_processor_id())
4222 resched_cpu(this_cpu);
4224 /* All tasks on this runqueue were pinned by CPU affinity */
4225 if (unlikely(all_pinned)) {
4226 cpumask_clear_cpu(cpu_of(busiest), cpus);
4227 if (!cpumask_empty(cpus))
4234 schedstat_inc(sd, lb_failed[idle]);
4235 sd->nr_balance_failed++;
4237 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
4239 raw_spin_lock_irqsave(&busiest->lock, flags);
4241 /* don't kick the migration_thread, if the curr
4242 * task on busiest cpu can't be moved to this_cpu
4244 if (!cpumask_test_cpu(this_cpu,
4245 &busiest->curr->cpus_allowed)) {
4246 raw_spin_unlock_irqrestore(&busiest->lock,
4249 goto out_one_pinned;
4252 if (!busiest->active_balance) {
4253 busiest->active_balance = 1;
4254 busiest->push_cpu = this_cpu;
4257 raw_spin_unlock_irqrestore(&busiest->lock, flags);
4259 wake_up_process(busiest->migration_thread);
4262 * We've kicked active balancing, reset the failure
4265 sd->nr_balance_failed = sd->cache_nice_tries+1;
4268 sd->nr_balance_failed = 0;
4270 if (likely(!active_balance)) {
4271 /* We were unbalanced, so reset the balancing interval */
4272 sd->balance_interval = sd->min_interval;
4275 * If we've begun active balancing, start to back off. This
4276 * case may not be covered by the all_pinned logic if there
4277 * is only 1 task on the busy runqueue (because we don't call
4280 if (sd->balance_interval < sd->max_interval)
4281 sd->balance_interval *= 2;
4284 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4285 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4291 schedstat_inc(sd, lb_balanced[idle]);
4293 sd->nr_balance_failed = 0;
4296 /* tune up the balancing interval */
4297 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4298 (sd->balance_interval < sd->max_interval))
4299 sd->balance_interval *= 2;
4301 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4302 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4313 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4314 * tasks if there is an imbalance.
4316 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4317 * this_rq is locked.
4320 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
4322 struct sched_group *group;
4323 struct rq *busiest = NULL;
4324 unsigned long imbalance;
4328 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4330 cpumask_copy(cpus, cpu_active_mask);
4333 * When power savings policy is enabled for the parent domain, idle
4334 * sibling can pick up load irrespective of busy siblings. In this case,
4335 * let the state of idle sibling percolate up as IDLE, instead of
4336 * portraying it as CPU_NOT_IDLE.
4338 if (sd->flags & SD_SHARE_CPUPOWER &&
4339 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4342 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
4344 update_shares_locked(this_rq, sd);
4345 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
4346 &sd_idle, cpus, NULL);
4348 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
4352 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
4354 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
4358 BUG_ON(busiest == this_rq);
4360 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
4363 if (busiest->nr_running > 1) {
4364 /* Attempt to move tasks */
4365 double_lock_balance(this_rq, busiest);
4366 /* this_rq->clock is already updated */
4367 update_rq_clock(busiest);
4368 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4369 imbalance, sd, CPU_NEWLY_IDLE,
4371 double_unlock_balance(this_rq, busiest);
4373 if (unlikely(all_pinned)) {
4374 cpumask_clear_cpu(cpu_of(busiest), cpus);
4375 if (!cpumask_empty(cpus))
4381 int active_balance = 0;
4383 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
4384 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4385 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4388 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4391 if (sd->nr_balance_failed++ < 2)
4395 * The only task running in a non-idle cpu can be moved to this
4396 * cpu in an attempt to completely freeup the other CPU
4397 * package. The same method used to move task in load_balance()
4398 * have been extended for load_balance_newidle() to speedup
4399 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4401 * The package power saving logic comes from
4402 * find_busiest_group(). If there are no imbalance, then
4403 * f_b_g() will return NULL. However when sched_mc={1,2} then
4404 * f_b_g() will select a group from which a running task may be
4405 * pulled to this cpu in order to make the other package idle.
4406 * If there is no opportunity to make a package idle and if
4407 * there are no imbalance, then f_b_g() will return NULL and no
4408 * action will be taken in load_balance_newidle().
4410 * Under normal task pull operation due to imbalance, there
4411 * will be more than one task in the source run queue and
4412 * move_tasks() will succeed. ld_moved will be true and this
4413 * active balance code will not be triggered.
4416 /* Lock busiest in correct order while this_rq is held */
4417 double_lock_balance(this_rq, busiest);
4420 * don't kick the migration_thread, if the curr
4421 * task on busiest cpu can't be moved to this_cpu
4423 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
4424 double_unlock_balance(this_rq, busiest);
4429 if (!busiest->active_balance) {
4430 busiest->active_balance = 1;
4431 busiest->push_cpu = this_cpu;
4435 double_unlock_balance(this_rq, busiest);
4437 * Should not call ttwu while holding a rq->lock
4439 raw_spin_unlock(&this_rq->lock);
4441 wake_up_process(busiest->migration_thread);
4442 raw_spin_lock(&this_rq->lock);
4445 sd->nr_balance_failed = 0;
4447 update_shares_locked(this_rq, sd);
4451 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
4452 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4453 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4455 sd->nr_balance_failed = 0;
4461 * idle_balance is called by schedule() if this_cpu is about to become
4462 * idle. Attempts to pull tasks from other CPUs.
4464 static void idle_balance(int this_cpu, struct rq *this_rq)
4466 struct sched_domain *sd;
4467 int pulled_task = 0;
4468 unsigned long next_balance = jiffies + HZ;
4470 this_rq->idle_stamp = this_rq->clock;
4472 if (this_rq->avg_idle < sysctl_sched_migration_cost)
4475 for_each_domain(this_cpu, sd) {
4476 unsigned long interval;
4478 if (!(sd->flags & SD_LOAD_BALANCE))
4481 if (sd->flags & SD_BALANCE_NEWIDLE)
4482 /* If we've pulled tasks over stop searching: */
4483 pulled_task = load_balance_newidle(this_cpu, this_rq,
4486 interval = msecs_to_jiffies(sd->balance_interval);
4487 if (time_after(next_balance, sd->last_balance + interval))
4488 next_balance = sd->last_balance + interval;
4490 this_rq->idle_stamp = 0;
4494 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4496 * We are going idle. next_balance may be set based on
4497 * a busy processor. So reset next_balance.
4499 this_rq->next_balance = next_balance;
4504 * active_load_balance is run by migration threads. It pushes running tasks
4505 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4506 * running on each physical CPU where possible, and avoids physical /
4507 * logical imbalances.
4509 * Called with busiest_rq locked.
4511 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
4513 int target_cpu = busiest_rq->push_cpu;
4514 struct sched_domain *sd;
4515 struct rq *target_rq;
4517 /* Is there any task to move? */
4518 if (busiest_rq->nr_running <= 1)
4521 target_rq = cpu_rq(target_cpu);
4524 * This condition is "impossible", if it occurs
4525 * we need to fix it. Originally reported by
4526 * Bjorn Helgaas on a 128-cpu setup.
4528 BUG_ON(busiest_rq == target_rq);
4530 /* move a task from busiest_rq to target_rq */
4531 double_lock_balance(busiest_rq, target_rq);
4532 update_rq_clock(busiest_rq);
4533 update_rq_clock(target_rq);
4535 /* Search for an sd spanning us and the target CPU. */
4536 for_each_domain(target_cpu, sd) {
4537 if ((sd->flags & SD_LOAD_BALANCE) &&
4538 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4543 schedstat_inc(sd, alb_count);
4545 if (move_one_task(target_rq, target_cpu, busiest_rq,
4547 schedstat_inc(sd, alb_pushed);
4549 schedstat_inc(sd, alb_failed);
4551 double_unlock_balance(busiest_rq, target_rq);
4556 atomic_t load_balancer;
4557 cpumask_var_t cpu_mask;
4558 cpumask_var_t ilb_grp_nohz_mask;
4559 } nohz ____cacheline_aligned = {
4560 .load_balancer = ATOMIC_INIT(-1),
4563 int get_nohz_load_balancer(void)
4565 return atomic_read(&nohz.load_balancer);
4568 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4570 * lowest_flag_domain - Return lowest sched_domain containing flag.
4571 * @cpu: The cpu whose lowest level of sched domain is to
4573 * @flag: The flag to check for the lowest sched_domain
4574 * for the given cpu.
4576 * Returns the lowest sched_domain of a cpu which contains the given flag.
4578 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4580 struct sched_domain *sd;
4582 for_each_domain(cpu, sd)
4583 if (sd && (sd->flags & flag))
4590 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4591 * @cpu: The cpu whose domains we're iterating over.
4592 * @sd: variable holding the value of the power_savings_sd
4594 * @flag: The flag to filter the sched_domains to be iterated.
4596 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4597 * set, starting from the lowest sched_domain to the highest.
4599 #define for_each_flag_domain(cpu, sd, flag) \
4600 for (sd = lowest_flag_domain(cpu, flag); \
4601 (sd && (sd->flags & flag)); sd = sd->parent)
4604 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4605 * @ilb_group: group to be checked for semi-idleness
4607 * Returns: 1 if the group is semi-idle. 0 otherwise.
4609 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4610 * and atleast one non-idle CPU. This helper function checks if the given
4611 * sched_group is semi-idle or not.
4613 static inline int is_semi_idle_group(struct sched_group *ilb_group)
4615 cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask,
4616 sched_group_cpus(ilb_group));
4619 * A sched_group is semi-idle when it has atleast one busy cpu
4620 * and atleast one idle cpu.
4622 if (cpumask_empty(nohz.ilb_grp_nohz_mask))
4625 if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group)))
4631 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4632 * @cpu: The cpu which is nominating a new idle_load_balancer.
4634 * Returns: Returns the id of the idle load balancer if it exists,
4635 * Else, returns >= nr_cpu_ids.
4637 * This algorithm picks the idle load balancer such that it belongs to a
4638 * semi-idle powersavings sched_domain. The idea is to try and avoid
4639 * completely idle packages/cores just for the purpose of idle load balancing
4640 * when there are other idle cpu's which are better suited for that job.
4642 static int find_new_ilb(int cpu)
4644 struct sched_domain *sd;
4645 struct sched_group *ilb_group;
4648 * Have idle load balancer selection from semi-idle packages only
4649 * when power-aware load balancing is enabled
4651 if (!(sched_smt_power_savings || sched_mc_power_savings))
4655 * Optimize for the case when we have no idle CPUs or only one
4656 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4658 if (cpumask_weight(nohz.cpu_mask) < 2)
4661 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4662 ilb_group = sd->groups;
4665 if (is_semi_idle_group(ilb_group))
4666 return cpumask_first(nohz.ilb_grp_nohz_mask);
4668 ilb_group = ilb_group->next;
4670 } while (ilb_group != sd->groups);
4674 return cpumask_first(nohz.cpu_mask);
4676 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4677 static inline int find_new_ilb(int call_cpu)
4679 return cpumask_first(nohz.cpu_mask);
4684 * This routine will try to nominate the ilb (idle load balancing)
4685 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4686 * load balancing on behalf of all those cpus. If all the cpus in the system
4687 * go into this tickless mode, then there will be no ilb owner (as there is
4688 * no need for one) and all the cpus will sleep till the next wakeup event
4691 * For the ilb owner, tick is not stopped. And this tick will be used
4692 * for idle load balancing. ilb owner will still be part of
4695 * While stopping the tick, this cpu will become the ilb owner if there
4696 * is no other owner. And will be the owner till that cpu becomes busy
4697 * or if all cpus in the system stop their ticks at which point
4698 * there is no need for ilb owner.
4700 * When the ilb owner becomes busy, it nominates another owner, during the
4701 * next busy scheduler_tick()
4703 int select_nohz_load_balancer(int stop_tick)
4705 int cpu = smp_processor_id();
4708 cpu_rq(cpu)->in_nohz_recently = 1;
4710 if (!cpu_active(cpu)) {
4711 if (atomic_read(&nohz.load_balancer) != cpu)
4715 * If we are going offline and still the leader,
4718 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4724 cpumask_set_cpu(cpu, nohz.cpu_mask);
4726 /* time for ilb owner also to sleep */
4727 if (cpumask_weight(nohz.cpu_mask) == num_active_cpus()) {
4728 if (atomic_read(&nohz.load_balancer) == cpu)
4729 atomic_set(&nohz.load_balancer, -1);
4733 if (atomic_read(&nohz.load_balancer) == -1) {
4734 /* make me the ilb owner */
4735 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4737 } else if (atomic_read(&nohz.load_balancer) == cpu) {
4740 if (!(sched_smt_power_savings ||
4741 sched_mc_power_savings))
4744 * Check to see if there is a more power-efficient
4747 new_ilb = find_new_ilb(cpu);
4748 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4749 atomic_set(&nohz.load_balancer, -1);
4750 resched_cpu(new_ilb);
4756 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4759 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4761 if (atomic_read(&nohz.load_balancer) == cpu)
4762 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4769 static DEFINE_SPINLOCK(balancing);
4772 * It checks each scheduling domain to see if it is due to be balanced,
4773 * and initiates a balancing operation if so.
4775 * Balancing parameters are set up in arch_init_sched_domains.
4777 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4780 struct rq *rq = cpu_rq(cpu);
4781 unsigned long interval;
4782 struct sched_domain *sd;
4783 /* Earliest time when we have to do rebalance again */
4784 unsigned long next_balance = jiffies + 60*HZ;
4785 int update_next_balance = 0;
4788 for_each_domain(cpu, sd) {
4789 if (!(sd->flags & SD_LOAD_BALANCE))
4792 interval = sd->balance_interval;
4793 if (idle != CPU_IDLE)
4794 interval *= sd->busy_factor;
4796 /* scale ms to jiffies */
4797 interval = msecs_to_jiffies(interval);
4798 if (unlikely(!interval))
4800 if (interval > HZ*NR_CPUS/10)
4801 interval = HZ*NR_CPUS/10;
4803 need_serialize = sd->flags & SD_SERIALIZE;
4805 if (need_serialize) {
4806 if (!spin_trylock(&balancing))
4810 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4811 if (load_balance(cpu, rq, sd, idle, &balance)) {
4813 * We've pulled tasks over so either we're no
4814 * longer idle, or one of our SMT siblings is
4817 idle = CPU_NOT_IDLE;
4819 sd->last_balance = jiffies;
4822 spin_unlock(&balancing);
4824 if (time_after(next_balance, sd->last_balance + interval)) {
4825 next_balance = sd->last_balance + interval;
4826 update_next_balance = 1;
4830 * Stop the load balance at this level. There is another
4831 * CPU in our sched group which is doing load balancing more
4839 * next_balance will be updated only when there is a need.
4840 * When the cpu is attached to null domain for ex, it will not be
4843 if (likely(update_next_balance))
4844 rq->next_balance = next_balance;
4848 * run_rebalance_domains is triggered when needed from the scheduler tick.
4849 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4850 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4852 static void run_rebalance_domains(struct softirq_action *h)
4854 int this_cpu = smp_processor_id();
4855 struct rq *this_rq = cpu_rq(this_cpu);
4856 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4857 CPU_IDLE : CPU_NOT_IDLE;
4859 rebalance_domains(this_cpu, idle);
4863 * If this cpu is the owner for idle load balancing, then do the
4864 * balancing on behalf of the other idle cpus whose ticks are
4867 if (this_rq->idle_at_tick &&
4868 atomic_read(&nohz.load_balancer) == this_cpu) {
4872 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4873 if (balance_cpu == this_cpu)
4877 * If this cpu gets work to do, stop the load balancing
4878 * work being done for other cpus. Next load
4879 * balancing owner will pick it up.
4884 rebalance_domains(balance_cpu, CPU_IDLE);
4886 rq = cpu_rq(balance_cpu);
4887 if (time_after(this_rq->next_balance, rq->next_balance))
4888 this_rq->next_balance = rq->next_balance;
4894 static inline int on_null_domain(int cpu)
4896 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
4900 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4902 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4903 * idle load balancing owner or decide to stop the periodic load balancing,
4904 * if the whole system is idle.
4906 static inline void trigger_load_balance(struct rq *rq, int cpu)
4910 * If we were in the nohz mode recently and busy at the current
4911 * scheduler tick, then check if we need to nominate new idle
4914 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4915 rq->in_nohz_recently = 0;
4917 if (atomic_read(&nohz.load_balancer) == cpu) {
4918 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4919 atomic_set(&nohz.load_balancer, -1);
4922 if (atomic_read(&nohz.load_balancer) == -1) {
4923 int ilb = find_new_ilb(cpu);
4925 if (ilb < nr_cpu_ids)
4931 * If this cpu is idle and doing idle load balancing for all the
4932 * cpus with ticks stopped, is it time for that to stop?
4934 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4935 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4941 * If this cpu is idle and the idle load balancing is done by
4942 * someone else, then no need raise the SCHED_SOFTIRQ
4944 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4945 cpumask_test_cpu(cpu, nohz.cpu_mask))
4948 /* Don't need to rebalance while attached to NULL domain */
4949 if (time_after_eq(jiffies, rq->next_balance) &&
4950 likely(!on_null_domain(cpu)))
4951 raise_softirq(SCHED_SOFTIRQ);
4954 #else /* CONFIG_SMP */
4957 * on UP we do not need to balance between CPUs:
4959 static inline void idle_balance(int cpu, struct rq *rq)
4965 DEFINE_PER_CPU(struct kernel_stat, kstat);
4967 EXPORT_PER_CPU_SYMBOL(kstat);
4970 * Return any ns on the sched_clock that have not yet been accounted in
4971 * @p in case that task is currently running.
4973 * Called with task_rq_lock() held on @rq.
4975 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
4979 if (task_current(rq, p)) {
4980 update_rq_clock(rq);
4981 ns = rq->clock - p->se.exec_start;
4989 unsigned long long task_delta_exec(struct task_struct *p)
4991 unsigned long flags;
4995 rq = task_rq_lock(p, &flags);
4996 ns = do_task_delta_exec(p, rq);
4997 task_rq_unlock(rq, &flags);
5003 * Return accounted runtime for the task.
5004 * In case the task is currently running, return the runtime plus current's
5005 * pending runtime that have not been accounted yet.
5007 unsigned long long task_sched_runtime(struct task_struct *p)
5009 unsigned long flags;
5013 rq = task_rq_lock(p, &flags);
5014 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
5015 task_rq_unlock(rq, &flags);
5021 * Return sum_exec_runtime for the thread group.
5022 * In case the task is currently running, return the sum plus current's
5023 * pending runtime that have not been accounted yet.
5025 * Note that the thread group might have other running tasks as well,
5026 * so the return value not includes other pending runtime that other
5027 * running tasks might have.
5029 unsigned long long thread_group_sched_runtime(struct task_struct *p)
5031 struct task_cputime totals;
5032 unsigned long flags;
5036 rq = task_rq_lock(p, &flags);
5037 thread_group_cputime(p, &totals);
5038 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
5039 task_rq_unlock(rq, &flags);
5045 * Account user cpu time to a process.
5046 * @p: the process that the cpu time gets accounted to
5047 * @cputime: the cpu time spent in user space since the last update
5048 * @cputime_scaled: cputime scaled by cpu frequency
5050 void account_user_time(struct task_struct *p, cputime_t cputime,
5051 cputime_t cputime_scaled)
5053 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5056 /* Add user time to process. */
5057 p->utime = cputime_add(p->utime, cputime);
5058 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5059 account_group_user_time(p, cputime);
5061 /* Add user time to cpustat. */
5062 tmp = cputime_to_cputime64(cputime);
5063 if (TASK_NICE(p) > 0)
5064 cpustat->nice = cputime64_add(cpustat->nice, tmp);
5066 cpustat->user = cputime64_add(cpustat->user, tmp);
5068 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
5069 /* Account for user time used */
5070 acct_update_integrals(p);
5074 * Account guest cpu time to a process.
5075 * @p: the process that the cpu time gets accounted to
5076 * @cputime: the cpu time spent in virtual machine since the last update
5077 * @cputime_scaled: cputime scaled by cpu frequency
5079 static void account_guest_time(struct task_struct *p, cputime_t cputime,
5080 cputime_t cputime_scaled)
5083 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5085 tmp = cputime_to_cputime64(cputime);
5087 /* Add guest time to process. */
5088 p->utime = cputime_add(p->utime, cputime);
5089 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5090 account_group_user_time(p, cputime);
5091 p->gtime = cputime_add(p->gtime, cputime);
5093 /* Add guest time to cpustat. */
5094 if (TASK_NICE(p) > 0) {
5095 cpustat->nice = cputime64_add(cpustat->nice, tmp);
5096 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
5098 cpustat->user = cputime64_add(cpustat->user, tmp);
5099 cpustat->guest = cputime64_add(cpustat->guest, tmp);
5104 * Account system cpu time to a process.
5105 * @p: the process that the cpu time gets accounted to
5106 * @hardirq_offset: the offset to subtract from hardirq_count()
5107 * @cputime: the cpu time spent in kernel space since the last update
5108 * @cputime_scaled: cputime scaled by cpu frequency
5110 void account_system_time(struct task_struct *p, int hardirq_offset,
5111 cputime_t cputime, cputime_t cputime_scaled)
5113 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5116 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
5117 account_guest_time(p, cputime, cputime_scaled);
5121 /* Add system time to process. */
5122 p->stime = cputime_add(p->stime, cputime);
5123 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
5124 account_group_system_time(p, cputime);
5126 /* Add system time to cpustat. */
5127 tmp = cputime_to_cputime64(cputime);
5128 if (hardirq_count() - hardirq_offset)
5129 cpustat->irq = cputime64_add(cpustat->irq, tmp);
5130 else if (softirq_count())
5131 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
5133 cpustat->system = cputime64_add(cpustat->system, tmp);
5135 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
5137 /* Account for system time used */
5138 acct_update_integrals(p);
5142 * Account for involuntary wait time.
5143 * @steal: the cpu time spent in involuntary wait
5145 void account_steal_time(cputime_t cputime)
5147 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5148 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5150 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
5154 * Account for idle time.
5155 * @cputime: the cpu time spent in idle wait
5157 void account_idle_time(cputime_t cputime)
5159 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5160 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5161 struct rq *rq = this_rq();
5163 if (atomic_read(&rq->nr_iowait) > 0)
5164 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
5166 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
5169 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
5172 * Account a single tick of cpu time.
5173 * @p: the process that the cpu time gets accounted to
5174 * @user_tick: indicates if the tick is a user or a system tick
5176 void account_process_tick(struct task_struct *p, int user_tick)
5178 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
5179 struct rq *rq = this_rq();
5182 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
5183 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
5184 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
5187 account_idle_time(cputime_one_jiffy);
5191 * Account multiple ticks of steal time.
5192 * @p: the process from which the cpu time has been stolen
5193 * @ticks: number of stolen ticks
5195 void account_steal_ticks(unsigned long ticks)
5197 account_steal_time(jiffies_to_cputime(ticks));
5201 * Account multiple ticks of idle time.
5202 * @ticks: number of stolen ticks
5204 void account_idle_ticks(unsigned long ticks)
5206 account_idle_time(jiffies_to_cputime(ticks));
5212 * Use precise platform statistics if available:
5214 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
5215 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5221 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5223 struct task_cputime cputime;
5225 thread_group_cputime(p, &cputime);
5227 *ut = cputime.utime;
5228 *st = cputime.stime;
5232 #ifndef nsecs_to_cputime
5233 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
5236 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5238 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
5241 * Use CFS's precise accounting:
5243 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
5248 temp = (u64)(rtime * utime);
5249 do_div(temp, total);
5250 utime = (cputime_t)temp;
5255 * Compare with previous values, to keep monotonicity:
5257 p->prev_utime = max(p->prev_utime, utime);
5258 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
5260 *ut = p->prev_utime;
5261 *st = p->prev_stime;
5265 * Must be called with siglock held.
5267 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5269 struct signal_struct *sig = p->signal;
5270 struct task_cputime cputime;
5271 cputime_t rtime, utime, total;
5273 thread_group_cputime(p, &cputime);
5275 total = cputime_add(cputime.utime, cputime.stime);
5276 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
5281 temp = (u64)(rtime * cputime.utime);
5282 do_div(temp, total);
5283 utime = (cputime_t)temp;
5287 sig->prev_utime = max(sig->prev_utime, utime);
5288 sig->prev_stime = max(sig->prev_stime,
5289 cputime_sub(rtime, sig->prev_utime));
5291 *ut = sig->prev_utime;
5292 *st = sig->prev_stime;
5297 * This function gets called by the timer code, with HZ frequency.
5298 * We call it with interrupts disabled.
5300 * It also gets called by the fork code, when changing the parent's
5303 void scheduler_tick(void)
5305 int cpu = smp_processor_id();
5306 struct rq *rq = cpu_rq(cpu);
5307 struct task_struct *curr = rq->curr;
5311 raw_spin_lock(&rq->lock);
5312 update_rq_clock(rq);
5313 update_cpu_load(rq);
5314 curr->sched_class->task_tick(rq, curr, 0);
5315 raw_spin_unlock(&rq->lock);
5317 perf_event_task_tick(curr, cpu);
5320 rq->idle_at_tick = idle_cpu(cpu);
5321 trigger_load_balance(rq, cpu);
5325 notrace unsigned long get_parent_ip(unsigned long addr)
5327 if (in_lock_functions(addr)) {
5328 addr = CALLER_ADDR2;
5329 if (in_lock_functions(addr))
5330 addr = CALLER_ADDR3;
5335 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5336 defined(CONFIG_PREEMPT_TRACER))
5338 void __kprobes add_preempt_count(int val)
5340 #ifdef CONFIG_DEBUG_PREEMPT
5344 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5347 preempt_count() += val;
5348 #ifdef CONFIG_DEBUG_PREEMPT
5350 * Spinlock count overflowing soon?
5352 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5355 if (preempt_count() == val)
5356 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5358 EXPORT_SYMBOL(add_preempt_count);
5360 void __kprobes sub_preempt_count(int val)
5362 #ifdef CONFIG_DEBUG_PREEMPT
5366 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5369 * Is the spinlock portion underflowing?
5371 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5372 !(preempt_count() & PREEMPT_MASK)))
5376 if (preempt_count() == val)
5377 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5378 preempt_count() -= val;
5380 EXPORT_SYMBOL(sub_preempt_count);
5385 * Print scheduling while atomic bug:
5387 static noinline void __schedule_bug(struct task_struct *prev)
5389 struct pt_regs *regs = get_irq_regs();
5391 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5392 prev->comm, prev->pid, preempt_count());
5394 debug_show_held_locks(prev);
5396 if (irqs_disabled())
5397 print_irqtrace_events(prev);
5406 * Various schedule()-time debugging checks and statistics:
5408 static inline void schedule_debug(struct task_struct *prev)
5411 * Test if we are atomic. Since do_exit() needs to call into
5412 * schedule() atomically, we ignore that path for now.
5413 * Otherwise, whine if we are scheduling when we should not be.
5415 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
5416 __schedule_bug(prev);
5418 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5420 schedstat_inc(this_rq(), sched_count);
5421 #ifdef CONFIG_SCHEDSTATS
5422 if (unlikely(prev->lock_depth >= 0)) {
5423 schedstat_inc(this_rq(), bkl_count);
5424 schedstat_inc(prev, sched_info.bkl_count);
5429 static void put_prev_task(struct rq *rq, struct task_struct *prev)
5431 if (prev->state == TASK_RUNNING) {
5432 u64 runtime = prev->se.sum_exec_runtime;
5434 runtime -= prev->se.prev_sum_exec_runtime;
5435 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
5438 * In order to avoid avg_overlap growing stale when we are
5439 * indeed overlapping and hence not getting put to sleep, grow
5440 * the avg_overlap on preemption.
5442 * We use the average preemption runtime because that
5443 * correlates to the amount of cache footprint a task can
5446 update_avg(&prev->se.avg_overlap, runtime);
5448 prev->sched_class->put_prev_task(rq, prev);
5452 * Pick up the highest-prio task:
5454 static inline struct task_struct *
5455 pick_next_task(struct rq *rq)
5457 const struct sched_class *class;
5458 struct task_struct *p;
5461 * Optimization: we know that if all tasks are in
5462 * the fair class we can call that function directly:
5464 if (likely(rq->nr_running == rq->cfs.nr_running)) {
5465 p = fair_sched_class.pick_next_task(rq);
5470 class = sched_class_highest;
5472 p = class->pick_next_task(rq);
5476 * Will never be NULL as the idle class always
5477 * returns a non-NULL p:
5479 class = class->next;
5484 * schedule() is the main scheduler function.
5486 asmlinkage void __sched schedule(void)
5488 struct task_struct *prev, *next;
5489 unsigned long *switch_count;
5495 cpu = smp_processor_id();
5499 switch_count = &prev->nivcsw;
5501 release_kernel_lock(prev);
5502 need_resched_nonpreemptible:
5504 schedule_debug(prev);
5506 if (sched_feat(HRTICK))
5509 raw_spin_lock_irq(&rq->lock);
5510 update_rq_clock(rq);
5511 clear_tsk_need_resched(prev);
5513 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
5514 if (unlikely(signal_pending_state(prev->state, prev)))
5515 prev->state = TASK_RUNNING;
5517 deactivate_task(rq, prev, 1);
5518 switch_count = &prev->nvcsw;
5521 pre_schedule(rq, prev);
5523 if (unlikely(!rq->nr_running))
5524 idle_balance(cpu, rq);
5526 put_prev_task(rq, prev);
5527 next = pick_next_task(rq);
5529 if (likely(prev != next)) {
5530 sched_info_switch(prev, next);
5531 perf_event_task_sched_out(prev, next, cpu);
5537 context_switch(rq, prev, next); /* unlocks the rq */
5539 * the context switch might have flipped the stack from under
5540 * us, hence refresh the local variables.
5542 cpu = smp_processor_id();
5545 raw_spin_unlock_irq(&rq->lock);
5549 if (unlikely(reacquire_kernel_lock(current) < 0)) {
5551 switch_count = &prev->nivcsw;
5552 goto need_resched_nonpreemptible;
5555 preempt_enable_no_resched();
5559 EXPORT_SYMBOL(schedule);
5561 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
5563 * Look out! "owner" is an entirely speculative pointer
5564 * access and not reliable.
5566 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5571 if (!sched_feat(OWNER_SPIN))
5574 #ifdef CONFIG_DEBUG_PAGEALLOC
5576 * Need to access the cpu field knowing that
5577 * DEBUG_PAGEALLOC could have unmapped it if
5578 * the mutex owner just released it and exited.
5580 if (probe_kernel_address(&owner->cpu, cpu))
5587 * Even if the access succeeded (likely case),
5588 * the cpu field may no longer be valid.
5590 if (cpu >= nr_cpumask_bits)
5594 * We need to validate that we can do a
5595 * get_cpu() and that we have the percpu area.
5597 if (!cpu_online(cpu))
5604 * Owner changed, break to re-assess state.
5606 if (lock->owner != owner)
5610 * Is that owner really running on that cpu?
5612 if (task_thread_info(rq->curr) != owner || need_resched())
5622 #ifdef CONFIG_PREEMPT
5624 * this is the entry point to schedule() from in-kernel preemption
5625 * off of preempt_enable. Kernel preemptions off return from interrupt
5626 * occur there and call schedule directly.
5628 asmlinkage void __sched preempt_schedule(void)
5630 struct thread_info *ti = current_thread_info();
5633 * If there is a non-zero preempt_count or interrupts are disabled,
5634 * we do not want to preempt the current task. Just return..
5636 if (likely(ti->preempt_count || irqs_disabled()))
5640 add_preempt_count(PREEMPT_ACTIVE);
5642 sub_preempt_count(PREEMPT_ACTIVE);
5645 * Check again in case we missed a preemption opportunity
5646 * between schedule and now.
5649 } while (need_resched());
5651 EXPORT_SYMBOL(preempt_schedule);
5654 * this is the entry point to schedule() from kernel preemption
5655 * off of irq context.
5656 * Note, that this is called and return with irqs disabled. This will
5657 * protect us against recursive calling from irq.
5659 asmlinkage void __sched preempt_schedule_irq(void)
5661 struct thread_info *ti = current_thread_info();
5663 /* Catch callers which need to be fixed */
5664 BUG_ON(ti->preempt_count || !irqs_disabled());
5667 add_preempt_count(PREEMPT_ACTIVE);
5670 local_irq_disable();
5671 sub_preempt_count(PREEMPT_ACTIVE);
5674 * Check again in case we missed a preemption opportunity
5675 * between schedule and now.
5678 } while (need_resched());
5681 #endif /* CONFIG_PREEMPT */
5683 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
5686 return try_to_wake_up(curr->private, mode, wake_flags);
5688 EXPORT_SYMBOL(default_wake_function);
5691 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5692 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5693 * number) then we wake all the non-exclusive tasks and one exclusive task.
5695 * There are circumstances in which we can try to wake a task which has already
5696 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5697 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5699 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
5700 int nr_exclusive, int wake_flags, void *key)
5702 wait_queue_t *curr, *next;
5704 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
5705 unsigned flags = curr->flags;
5707 if (curr->func(curr, mode, wake_flags, key) &&
5708 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
5714 * __wake_up - wake up threads blocked on a waitqueue.
5716 * @mode: which threads
5717 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5718 * @key: is directly passed to the wakeup function
5720 * It may be assumed that this function implies a write memory barrier before
5721 * changing the task state if and only if any tasks are woken up.
5723 void __wake_up(wait_queue_head_t *q, unsigned int mode,
5724 int nr_exclusive, void *key)
5726 unsigned long flags;
5728 spin_lock_irqsave(&q->lock, flags);
5729 __wake_up_common(q, mode, nr_exclusive, 0, key);
5730 spin_unlock_irqrestore(&q->lock, flags);
5732 EXPORT_SYMBOL(__wake_up);
5735 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5737 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
5739 __wake_up_common(q, mode, 1, 0, NULL);
5742 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
5744 __wake_up_common(q, mode, 1, 0, key);
5748 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5750 * @mode: which threads
5751 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5752 * @key: opaque value to be passed to wakeup targets
5754 * The sync wakeup differs that the waker knows that it will schedule
5755 * away soon, so while the target thread will be woken up, it will not
5756 * be migrated to another CPU - ie. the two threads are 'synchronized'
5757 * with each other. This can prevent needless bouncing between CPUs.
5759 * On UP it can prevent extra preemption.
5761 * It may be assumed that this function implies a write memory barrier before
5762 * changing the task state if and only if any tasks are woken up.
5764 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
5765 int nr_exclusive, void *key)
5767 unsigned long flags;
5768 int wake_flags = WF_SYNC;
5773 if (unlikely(!nr_exclusive))
5776 spin_lock_irqsave(&q->lock, flags);
5777 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
5778 spin_unlock_irqrestore(&q->lock, flags);
5780 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
5783 * __wake_up_sync - see __wake_up_sync_key()
5785 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
5787 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
5789 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
5792 * complete: - signals a single thread waiting on this completion
5793 * @x: holds the state of this particular completion
5795 * This will wake up a single thread waiting on this completion. Threads will be
5796 * awakened in the same order in which they were queued.
5798 * See also complete_all(), wait_for_completion() and related routines.
5800 * It may be assumed that this function implies a write memory barrier before
5801 * changing the task state if and only if any tasks are woken up.
5803 void complete(struct completion *x)
5805 unsigned long flags;
5807 spin_lock_irqsave(&x->wait.lock, flags);
5809 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
5810 spin_unlock_irqrestore(&x->wait.lock, flags);
5812 EXPORT_SYMBOL(complete);
5815 * complete_all: - signals all threads waiting on this completion
5816 * @x: holds the state of this particular completion
5818 * This will wake up all threads waiting on this particular completion event.
5820 * It may be assumed that this function implies a write memory barrier before
5821 * changing the task state if and only if any tasks are woken up.
5823 void complete_all(struct completion *x)
5825 unsigned long flags;
5827 spin_lock_irqsave(&x->wait.lock, flags);
5828 x->done += UINT_MAX/2;
5829 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
5830 spin_unlock_irqrestore(&x->wait.lock, flags);
5832 EXPORT_SYMBOL(complete_all);
5834 static inline long __sched
5835 do_wait_for_common(struct completion *x, long timeout, int state)
5838 DECLARE_WAITQUEUE(wait, current);
5840 wait.flags |= WQ_FLAG_EXCLUSIVE;
5841 __add_wait_queue_tail(&x->wait, &wait);
5843 if (signal_pending_state(state, current)) {
5844 timeout = -ERESTARTSYS;
5847 __set_current_state(state);
5848 spin_unlock_irq(&x->wait.lock);
5849 timeout = schedule_timeout(timeout);
5850 spin_lock_irq(&x->wait.lock);
5851 } while (!x->done && timeout);
5852 __remove_wait_queue(&x->wait, &wait);
5857 return timeout ?: 1;
5861 wait_for_common(struct completion *x, long timeout, int state)
5865 spin_lock_irq(&x->wait.lock);
5866 timeout = do_wait_for_common(x, timeout, state);
5867 spin_unlock_irq(&x->wait.lock);
5872 * wait_for_completion: - waits for completion of a task
5873 * @x: holds the state of this particular completion
5875 * This waits to be signaled for completion of a specific task. It is NOT
5876 * interruptible and there is no timeout.
5878 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5879 * and interrupt capability. Also see complete().
5881 void __sched wait_for_completion(struct completion *x)
5883 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
5885 EXPORT_SYMBOL(wait_for_completion);
5888 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5889 * @x: holds the state of this particular completion
5890 * @timeout: timeout value in jiffies
5892 * This waits for either a completion of a specific task to be signaled or for a
5893 * specified timeout to expire. The timeout is in jiffies. It is not
5896 unsigned long __sched
5897 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5899 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5901 EXPORT_SYMBOL(wait_for_completion_timeout);
5904 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5905 * @x: holds the state of this particular completion
5907 * This waits for completion of a specific task to be signaled. It is
5910 int __sched wait_for_completion_interruptible(struct completion *x)
5912 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5913 if (t == -ERESTARTSYS)
5917 EXPORT_SYMBOL(wait_for_completion_interruptible);
5920 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5921 * @x: holds the state of this particular completion
5922 * @timeout: timeout value in jiffies
5924 * This waits for either a completion of a specific task to be signaled or for a
5925 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5927 unsigned long __sched
5928 wait_for_completion_interruptible_timeout(struct completion *x,
5929 unsigned long timeout)
5931 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
5933 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5936 * wait_for_completion_killable: - waits for completion of a task (killable)
5937 * @x: holds the state of this particular completion
5939 * This waits to be signaled for completion of a specific task. It can be
5940 * interrupted by a kill signal.
5942 int __sched wait_for_completion_killable(struct completion *x)
5944 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5945 if (t == -ERESTARTSYS)
5949 EXPORT_SYMBOL(wait_for_completion_killable);
5952 * try_wait_for_completion - try to decrement a completion without blocking
5953 * @x: completion structure
5955 * Returns: 0 if a decrement cannot be done without blocking
5956 * 1 if a decrement succeeded.
5958 * If a completion is being used as a counting completion,
5959 * attempt to decrement the counter without blocking. This
5960 * enables us to avoid waiting if the resource the completion
5961 * is protecting is not available.
5963 bool try_wait_for_completion(struct completion *x)
5965 unsigned long flags;
5968 spin_lock_irqsave(&x->wait.lock, flags);
5973 spin_unlock_irqrestore(&x->wait.lock, flags);
5976 EXPORT_SYMBOL(try_wait_for_completion);
5979 * completion_done - Test to see if a completion has any waiters
5980 * @x: completion structure
5982 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5983 * 1 if there are no waiters.
5986 bool completion_done(struct completion *x)
5988 unsigned long flags;
5991 spin_lock_irqsave(&x->wait.lock, flags);
5994 spin_unlock_irqrestore(&x->wait.lock, flags);
5997 EXPORT_SYMBOL(completion_done);
6000 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
6002 unsigned long flags;
6005 init_waitqueue_entry(&wait, current);
6007 __set_current_state(state);
6009 spin_lock_irqsave(&q->lock, flags);
6010 __add_wait_queue(q, &wait);
6011 spin_unlock(&q->lock);
6012 timeout = schedule_timeout(timeout);
6013 spin_lock_irq(&q->lock);
6014 __remove_wait_queue(q, &wait);
6015 spin_unlock_irqrestore(&q->lock, flags);
6020 void __sched interruptible_sleep_on(wait_queue_head_t *q)
6022 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
6024 EXPORT_SYMBOL(interruptible_sleep_on);
6027 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
6029 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
6031 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
6033 void __sched sleep_on(wait_queue_head_t *q)
6035 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
6037 EXPORT_SYMBOL(sleep_on);
6039 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
6041 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
6043 EXPORT_SYMBOL(sleep_on_timeout);
6045 #ifdef CONFIG_RT_MUTEXES
6048 * rt_mutex_setprio - set the current priority of a task
6050 * @prio: prio value (kernel-internal form)
6052 * This function changes the 'effective' priority of a task. It does
6053 * not touch ->normal_prio like __setscheduler().
6055 * Used by the rt_mutex code to implement priority inheritance logic.
6057 void rt_mutex_setprio(struct task_struct *p, int prio)
6059 unsigned long flags;
6060 int oldprio, on_rq, running;
6062 const struct sched_class *prev_class = p->sched_class;
6064 BUG_ON(prio < 0 || prio > MAX_PRIO);
6066 rq = task_rq_lock(p, &flags);
6067 update_rq_clock(rq);
6070 on_rq = p->se.on_rq;
6071 running = task_current(rq, p);
6073 dequeue_task(rq, p, 0);
6075 p->sched_class->put_prev_task(rq, p);
6078 p->sched_class = &rt_sched_class;
6080 p->sched_class = &fair_sched_class;
6085 p->sched_class->set_curr_task(rq);
6087 enqueue_task(rq, p, 0);
6089 check_class_changed(rq, p, prev_class, oldprio, running);
6091 task_rq_unlock(rq, &flags);
6096 void set_user_nice(struct task_struct *p, long nice)
6098 int old_prio, delta, on_rq;
6099 unsigned long flags;
6102 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
6105 * We have to be careful, if called from sys_setpriority(),
6106 * the task might be in the middle of scheduling on another CPU.
6108 rq = task_rq_lock(p, &flags);
6109 update_rq_clock(rq);
6111 * The RT priorities are set via sched_setscheduler(), but we still
6112 * allow the 'normal' nice value to be set - but as expected
6113 * it wont have any effect on scheduling until the task is
6114 * SCHED_FIFO/SCHED_RR:
6116 if (task_has_rt_policy(p)) {
6117 p->static_prio = NICE_TO_PRIO(nice);
6120 on_rq = p->se.on_rq;
6122 dequeue_task(rq, p, 0);
6124 p->static_prio = NICE_TO_PRIO(nice);
6127 p->prio = effective_prio(p);
6128 delta = p->prio - old_prio;
6131 enqueue_task(rq, p, 0);
6133 * If the task increased its priority or is running and
6134 * lowered its priority, then reschedule its CPU:
6136 if (delta < 0 || (delta > 0 && task_running(rq, p)))
6137 resched_task(rq->curr);
6140 task_rq_unlock(rq, &flags);
6142 EXPORT_SYMBOL(set_user_nice);
6145 * can_nice - check if a task can reduce its nice value
6149 int can_nice(const struct task_struct *p, const int nice)
6151 /* convert nice value [19,-20] to rlimit style value [1,40] */
6152 int nice_rlim = 20 - nice;
6154 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
6155 capable(CAP_SYS_NICE));
6158 #ifdef __ARCH_WANT_SYS_NICE
6161 * sys_nice - change the priority of the current process.
6162 * @increment: priority increment
6164 * sys_setpriority is a more generic, but much slower function that
6165 * does similar things.
6167 SYSCALL_DEFINE1(nice, int, increment)
6172 * Setpriority might change our priority at the same moment.
6173 * We don't have to worry. Conceptually one call occurs first
6174 * and we have a single winner.
6176 if (increment < -40)
6181 nice = TASK_NICE(current) + increment;
6187 if (increment < 0 && !can_nice(current, nice))
6190 retval = security_task_setnice(current, nice);
6194 set_user_nice(current, nice);
6201 * task_prio - return the priority value of a given task.
6202 * @p: the task in question.
6204 * This is the priority value as seen by users in /proc.
6205 * RT tasks are offset by -200. Normal tasks are centered
6206 * around 0, value goes from -16 to +15.
6208 int task_prio(const struct task_struct *p)
6210 return p->prio - MAX_RT_PRIO;
6214 * task_nice - return the nice value of a given task.
6215 * @p: the task in question.
6217 int task_nice(const struct task_struct *p)
6219 return TASK_NICE(p);
6221 EXPORT_SYMBOL(task_nice);
6224 * idle_cpu - is a given cpu idle currently?
6225 * @cpu: the processor in question.
6227 int idle_cpu(int cpu)
6229 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
6233 * idle_task - return the idle task for a given cpu.
6234 * @cpu: the processor in question.
6236 struct task_struct *idle_task(int cpu)
6238 return cpu_rq(cpu)->idle;
6242 * find_process_by_pid - find a process with a matching PID value.
6243 * @pid: the pid in question.
6245 static struct task_struct *find_process_by_pid(pid_t pid)
6247 return pid ? find_task_by_vpid(pid) : current;
6250 /* Actually do priority change: must hold rq lock. */
6252 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
6254 BUG_ON(p->se.on_rq);
6257 p->rt_priority = prio;
6258 p->normal_prio = normal_prio(p);
6259 /* we are holding p->pi_lock already */
6260 p->prio = rt_mutex_getprio(p);
6261 if (rt_prio(p->prio))
6262 p->sched_class = &rt_sched_class;
6264 p->sched_class = &fair_sched_class;
6269 * check the target process has a UID that matches the current process's
6271 static bool check_same_owner(struct task_struct *p)
6273 const struct cred *cred = current_cred(), *pcred;
6277 pcred = __task_cred(p);
6278 match = (cred->euid == pcred->euid ||
6279 cred->euid == pcred->uid);
6284 static int __sched_setscheduler(struct task_struct *p, int policy,
6285 struct sched_param *param, bool user)
6287 int retval, oldprio, oldpolicy = -1, on_rq, running;
6288 unsigned long flags;
6289 const struct sched_class *prev_class = p->sched_class;
6293 /* may grab non-irq protected spin_locks */
6294 BUG_ON(in_interrupt());
6296 /* double check policy once rq lock held */
6298 reset_on_fork = p->sched_reset_on_fork;
6299 policy = oldpolicy = p->policy;
6301 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
6302 policy &= ~SCHED_RESET_ON_FORK;
6304 if (policy != SCHED_FIFO && policy != SCHED_RR &&
6305 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
6306 policy != SCHED_IDLE)
6311 * Valid priorities for SCHED_FIFO and SCHED_RR are
6312 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6313 * SCHED_BATCH and SCHED_IDLE is 0.
6315 if (param->sched_priority < 0 ||
6316 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
6317 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
6319 if (rt_policy(policy) != (param->sched_priority != 0))
6323 * Allow unprivileged RT tasks to decrease priority:
6325 if (user && !capable(CAP_SYS_NICE)) {
6326 if (rt_policy(policy)) {
6327 unsigned long rlim_rtprio;
6329 if (!lock_task_sighand(p, &flags))
6331 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
6332 unlock_task_sighand(p, &flags);
6334 /* can't set/change the rt policy */
6335 if (policy != p->policy && !rlim_rtprio)
6338 /* can't increase priority */
6339 if (param->sched_priority > p->rt_priority &&
6340 param->sched_priority > rlim_rtprio)
6344 * Like positive nice levels, dont allow tasks to
6345 * move out of SCHED_IDLE either:
6347 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
6350 /* can't change other user's priorities */
6351 if (!check_same_owner(p))
6354 /* Normal users shall not reset the sched_reset_on_fork flag */
6355 if (p->sched_reset_on_fork && !reset_on_fork)
6360 #ifdef CONFIG_RT_GROUP_SCHED
6362 * Do not allow realtime tasks into groups that have no runtime
6365 if (rt_bandwidth_enabled() && rt_policy(policy) &&
6366 task_group(p)->rt_bandwidth.rt_runtime == 0)
6370 retval = security_task_setscheduler(p, policy, param);
6376 * make sure no PI-waiters arrive (or leave) while we are
6377 * changing the priority of the task:
6379 raw_spin_lock_irqsave(&p->pi_lock, flags);
6381 * To be able to change p->policy safely, the apropriate
6382 * runqueue lock must be held.
6384 rq = __task_rq_lock(p);
6385 /* recheck policy now with rq lock held */
6386 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
6387 policy = oldpolicy = -1;
6388 __task_rq_unlock(rq);
6389 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
6392 update_rq_clock(rq);
6393 on_rq = p->se.on_rq;
6394 running = task_current(rq, p);
6396 deactivate_task(rq, p, 0);
6398 p->sched_class->put_prev_task(rq, p);
6400 p->sched_reset_on_fork = reset_on_fork;
6403 __setscheduler(rq, p, policy, param->sched_priority);
6406 p->sched_class->set_curr_task(rq);
6408 activate_task(rq, p, 0);
6410 check_class_changed(rq, p, prev_class, oldprio, running);
6412 __task_rq_unlock(rq);
6413 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
6415 rt_mutex_adjust_pi(p);
6421 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6422 * @p: the task in question.
6423 * @policy: new policy.
6424 * @param: structure containing the new RT priority.
6426 * NOTE that the task may be already dead.
6428 int sched_setscheduler(struct task_struct *p, int policy,
6429 struct sched_param *param)
6431 return __sched_setscheduler(p, policy, param, true);
6433 EXPORT_SYMBOL_GPL(sched_setscheduler);
6436 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6437 * @p: the task in question.
6438 * @policy: new policy.
6439 * @param: structure containing the new RT priority.
6441 * Just like sched_setscheduler, only don't bother checking if the
6442 * current context has permission. For example, this is needed in
6443 * stop_machine(): we create temporary high priority worker threads,
6444 * but our caller might not have that capability.
6446 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
6447 struct sched_param *param)
6449 return __sched_setscheduler(p, policy, param, false);
6453 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
6455 struct sched_param lparam;
6456 struct task_struct *p;
6459 if (!param || pid < 0)
6461 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
6466 p = find_process_by_pid(pid);
6468 retval = sched_setscheduler(p, policy, &lparam);
6475 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6476 * @pid: the pid in question.
6477 * @policy: new policy.
6478 * @param: structure containing the new RT priority.
6480 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
6481 struct sched_param __user *, param)
6483 /* negative values for policy are not valid */
6487 return do_sched_setscheduler(pid, policy, param);
6491 * sys_sched_setparam - set/change the RT priority of a thread
6492 * @pid: the pid in question.
6493 * @param: structure containing the new RT priority.
6495 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
6497 return do_sched_setscheduler(pid, -1, param);
6501 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6502 * @pid: the pid in question.
6504 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
6506 struct task_struct *p;
6514 p = find_process_by_pid(pid);
6516 retval = security_task_getscheduler(p);
6519 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
6526 * sys_sched_getparam - get the RT priority of a thread
6527 * @pid: the pid in question.
6528 * @param: structure containing the RT priority.
6530 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
6532 struct sched_param lp;
6533 struct task_struct *p;
6536 if (!param || pid < 0)
6540 p = find_process_by_pid(pid);
6545 retval = security_task_getscheduler(p);
6549 lp.sched_priority = p->rt_priority;
6553 * This one might sleep, we cannot do it with a spinlock held ...
6555 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
6564 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
6566 cpumask_var_t cpus_allowed, new_mask;
6567 struct task_struct *p;
6573 p = find_process_by_pid(pid);
6580 /* Prevent p going away */
6584 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6588 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
6590 goto out_free_cpus_allowed;
6593 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
6596 retval = security_task_setscheduler(p, 0, NULL);
6600 cpuset_cpus_allowed(p, cpus_allowed);
6601 cpumask_and(new_mask, in_mask, cpus_allowed);
6603 retval = set_cpus_allowed_ptr(p, new_mask);
6606 cpuset_cpus_allowed(p, cpus_allowed);
6607 if (!cpumask_subset(new_mask, cpus_allowed)) {
6609 * We must have raced with a concurrent cpuset
6610 * update. Just reset the cpus_allowed to the
6611 * cpuset's cpus_allowed
6613 cpumask_copy(new_mask, cpus_allowed);
6618 free_cpumask_var(new_mask);
6619 out_free_cpus_allowed:
6620 free_cpumask_var(cpus_allowed);
6627 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
6628 struct cpumask *new_mask)
6630 if (len < cpumask_size())
6631 cpumask_clear(new_mask);
6632 else if (len > cpumask_size())
6633 len = cpumask_size();
6635 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6639 * sys_sched_setaffinity - set the cpu affinity of a process
6640 * @pid: pid of the process
6641 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6642 * @user_mask_ptr: user-space pointer to the new cpu mask
6644 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6645 unsigned long __user *, user_mask_ptr)
6647 cpumask_var_t new_mask;
6650 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6653 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6655 retval = sched_setaffinity(pid, new_mask);
6656 free_cpumask_var(new_mask);
6660 long sched_getaffinity(pid_t pid, struct cpumask *mask)
6662 struct task_struct *p;
6663 unsigned long flags;
6671 p = find_process_by_pid(pid);
6675 retval = security_task_getscheduler(p);
6679 rq = task_rq_lock(p, &flags);
6680 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
6681 task_rq_unlock(rq, &flags);
6691 * sys_sched_getaffinity - get the cpu affinity of a process
6692 * @pid: pid of the process
6693 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6694 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6696 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6697 unsigned long __user *, user_mask_ptr)
6702 if (len < cpumask_size())
6705 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6708 ret = sched_getaffinity(pid, mask);
6710 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
6713 ret = cpumask_size();
6715 free_cpumask_var(mask);
6721 * sys_sched_yield - yield the current processor to other threads.
6723 * This function yields the current CPU to other tasks. If there are no
6724 * other threads running on this CPU then this function will return.
6726 SYSCALL_DEFINE0(sched_yield)
6728 struct rq *rq = this_rq_lock();
6730 schedstat_inc(rq, yld_count);
6731 current->sched_class->yield_task(rq);
6734 * Since we are going to call schedule() anyway, there's
6735 * no need to preempt or enable interrupts:
6737 __release(rq->lock);
6738 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
6739 do_raw_spin_unlock(&rq->lock);
6740 preempt_enable_no_resched();
6747 static inline int should_resched(void)
6749 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
6752 static void __cond_resched(void)
6754 add_preempt_count(PREEMPT_ACTIVE);
6756 sub_preempt_count(PREEMPT_ACTIVE);
6759 int __sched _cond_resched(void)
6761 if (should_resched()) {
6767 EXPORT_SYMBOL(_cond_resched);
6770 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
6771 * call schedule, and on return reacquire the lock.
6773 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6774 * operations here to prevent schedule() from being called twice (once via
6775 * spin_unlock(), once by hand).
6777 int __cond_resched_lock(spinlock_t *lock)
6779 int resched = should_resched();
6782 lockdep_assert_held(lock);
6784 if (spin_needbreak(lock) || resched) {
6795 EXPORT_SYMBOL(__cond_resched_lock);
6797 int __sched __cond_resched_softirq(void)
6799 BUG_ON(!in_softirq());
6801 if (should_resched()) {
6809 EXPORT_SYMBOL(__cond_resched_softirq);
6812 * yield - yield the current processor to other threads.
6814 * This is a shortcut for kernel-space yielding - it marks the
6815 * thread runnable and calls sys_sched_yield().
6817 void __sched yield(void)
6819 set_current_state(TASK_RUNNING);
6822 EXPORT_SYMBOL(yield);
6825 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6826 * that process accounting knows that this is a task in IO wait state.
6828 void __sched io_schedule(void)
6830 struct rq *rq = raw_rq();
6832 delayacct_blkio_start();
6833 atomic_inc(&rq->nr_iowait);
6834 current->in_iowait = 1;
6836 current->in_iowait = 0;
6837 atomic_dec(&rq->nr_iowait);
6838 delayacct_blkio_end();
6840 EXPORT_SYMBOL(io_schedule);
6842 long __sched io_schedule_timeout(long timeout)
6844 struct rq *rq = raw_rq();
6847 delayacct_blkio_start();
6848 atomic_inc(&rq->nr_iowait);
6849 current->in_iowait = 1;
6850 ret = schedule_timeout(timeout);
6851 current->in_iowait = 0;
6852 atomic_dec(&rq->nr_iowait);
6853 delayacct_blkio_end();
6858 * sys_sched_get_priority_max - return maximum RT priority.
6859 * @policy: scheduling class.
6861 * this syscall returns the maximum rt_priority that can be used
6862 * by a given scheduling class.
6864 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6871 ret = MAX_USER_RT_PRIO-1;
6883 * sys_sched_get_priority_min - return minimum RT priority.
6884 * @policy: scheduling class.
6886 * this syscall returns the minimum rt_priority that can be used
6887 * by a given scheduling class.
6889 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6907 * sys_sched_rr_get_interval - return the default timeslice of a process.
6908 * @pid: pid of the process.
6909 * @interval: userspace pointer to the timeslice value.
6911 * this syscall writes the default timeslice value of a given process
6912 * into the user-space timespec buffer. A value of '0' means infinity.
6914 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6915 struct timespec __user *, interval)
6917 struct task_struct *p;
6918 unsigned int time_slice;
6919 unsigned long flags;
6929 p = find_process_by_pid(pid);
6933 retval = security_task_getscheduler(p);
6937 rq = task_rq_lock(p, &flags);
6938 time_slice = p->sched_class->get_rr_interval(rq, p);
6939 task_rq_unlock(rq, &flags);
6942 jiffies_to_timespec(time_slice, &t);
6943 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
6951 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
6953 void sched_show_task(struct task_struct *p)
6955 unsigned long free = 0;
6958 state = p->state ? __ffs(p->state) + 1 : 0;
6959 printk(KERN_INFO "%-13.13s %c", p->comm,
6960 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
6961 #if BITS_PER_LONG == 32
6962 if (state == TASK_RUNNING)
6963 printk(KERN_CONT " running ");
6965 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
6967 if (state == TASK_RUNNING)
6968 printk(KERN_CONT " running task ");
6970 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
6972 #ifdef CONFIG_DEBUG_STACK_USAGE
6973 free = stack_not_used(p);
6975 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
6976 task_pid_nr(p), task_pid_nr(p->real_parent),
6977 (unsigned long)task_thread_info(p)->flags);
6979 show_stack(p, NULL);
6982 void show_state_filter(unsigned long state_filter)
6984 struct task_struct *g, *p;
6986 #if BITS_PER_LONG == 32
6988 " task PC stack pid father\n");
6991 " task PC stack pid father\n");
6993 read_lock(&tasklist_lock);
6994 do_each_thread(g, p) {
6996 * reset the NMI-timeout, listing all files on a slow
6997 * console might take alot of time:
6999 touch_nmi_watchdog();
7000 if (!state_filter || (p->state & state_filter))
7002 } while_each_thread(g, p);
7004 touch_all_softlockup_watchdogs();
7006 #ifdef CONFIG_SCHED_DEBUG
7007 sysrq_sched_debug_show();
7009 read_unlock(&tasklist_lock);
7011 * Only show locks if all tasks are dumped:
7014 debug_show_all_locks();
7017 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
7019 idle->sched_class = &idle_sched_class;
7023 * init_idle - set up an idle thread for a given CPU
7024 * @idle: task in question
7025 * @cpu: cpu the idle task belongs to
7027 * NOTE: this function does not set the idle thread's NEED_RESCHED
7028 * flag, to make booting more robust.
7030 void __cpuinit init_idle(struct task_struct *idle, int cpu)
7032 struct rq *rq = cpu_rq(cpu);
7033 unsigned long flags;
7035 raw_spin_lock_irqsave(&rq->lock, flags);
7038 idle->state = TASK_RUNNING;
7039 idle->se.exec_start = sched_clock();
7041 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
7042 __set_task_cpu(idle, cpu);
7044 rq->curr = rq->idle = idle;
7045 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
7048 raw_spin_unlock_irqrestore(&rq->lock, flags);
7050 /* Set the preempt count _outside_ the spinlocks! */
7051 #if defined(CONFIG_PREEMPT)
7052 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
7054 task_thread_info(idle)->preempt_count = 0;
7057 * The idle tasks have their own, simple scheduling class:
7059 idle->sched_class = &idle_sched_class;
7060 ftrace_graph_init_task(idle);
7064 * In a system that switches off the HZ timer nohz_cpu_mask
7065 * indicates which cpus entered this state. This is used
7066 * in the rcu update to wait only for active cpus. For system
7067 * which do not switch off the HZ timer nohz_cpu_mask should
7068 * always be CPU_BITS_NONE.
7070 cpumask_var_t nohz_cpu_mask;
7073 * Increase the granularity value when there are more CPUs,
7074 * because with more CPUs the 'effective latency' as visible
7075 * to users decreases. But the relationship is not linear,
7076 * so pick a second-best guess by going with the log2 of the
7079 * This idea comes from the SD scheduler of Con Kolivas:
7081 static int get_update_sysctl_factor(void)
7083 unsigned int cpus = min_t(int, num_online_cpus(), 8);
7084 unsigned int factor;
7086 switch (sysctl_sched_tunable_scaling) {
7087 case SCHED_TUNABLESCALING_NONE:
7090 case SCHED_TUNABLESCALING_LINEAR:
7093 case SCHED_TUNABLESCALING_LOG:
7095 factor = 1 + ilog2(cpus);
7102 static void update_sysctl(void)
7104 unsigned int factor = get_update_sysctl_factor();
7106 #define SET_SYSCTL(name) \
7107 (sysctl_##name = (factor) * normalized_sysctl_##name)
7108 SET_SYSCTL(sched_min_granularity);
7109 SET_SYSCTL(sched_latency);
7110 SET_SYSCTL(sched_wakeup_granularity);
7111 SET_SYSCTL(sched_shares_ratelimit);
7115 static inline void sched_init_granularity(void)
7122 * This is how migration works:
7124 * 1) we queue a struct migration_req structure in the source CPU's
7125 * runqueue and wake up that CPU's migration thread.
7126 * 2) we down() the locked semaphore => thread blocks.
7127 * 3) migration thread wakes up (implicitly it forces the migrated
7128 * thread off the CPU)
7129 * 4) it gets the migration request and checks whether the migrated
7130 * task is still in the wrong runqueue.
7131 * 5) if it's in the wrong runqueue then the migration thread removes
7132 * it and puts it into the right queue.
7133 * 6) migration thread up()s the semaphore.
7134 * 7) we wake up and the migration is done.
7138 * Change a given task's CPU affinity. Migrate the thread to a
7139 * proper CPU and schedule it away if the CPU it's executing on
7140 * is removed from the allowed bitmask.
7142 * NOTE: the caller must have a valid reference to the task, the
7143 * task must not exit() & deallocate itself prematurely. The
7144 * call is not atomic; no spinlocks may be held.
7146 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
7148 struct migration_req req;
7149 unsigned long flags;
7154 * Since we rely on wake-ups to migrate sleeping tasks, don't change
7155 * the ->cpus_allowed mask from under waking tasks, which would be
7156 * possible when we change rq->lock in ttwu(), so synchronize against
7157 * TASK_WAKING to avoid that.
7159 * Make an exception for freshly cloned tasks, since cpuset namespaces
7160 * might move the task about, we have to validate the target in
7161 * wake_up_new_task() anyway since the cpu might have gone away.
7164 while (p->state == TASK_WAKING && !(p->flags & PF_STARTING))
7167 rq = task_rq_lock(p, &flags);
7169 if (p->state == TASK_WAKING && !(p->flags & PF_STARTING)) {
7170 task_rq_unlock(rq, &flags);
7174 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
7179 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
7180 !cpumask_equal(&p->cpus_allowed, new_mask))) {
7185 if (p->sched_class->set_cpus_allowed)
7186 p->sched_class->set_cpus_allowed(p, new_mask);
7188 cpumask_copy(&p->cpus_allowed, new_mask);
7189 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
7192 /* Can the task run on the task's current CPU? If so, we're done */
7193 if (cpumask_test_cpu(task_cpu(p), new_mask))
7196 if (migrate_task(p, cpumask_any_and(cpu_active_mask, new_mask), &req)) {
7197 /* Need help from migration thread: drop lock and wait. */
7198 struct task_struct *mt = rq->migration_thread;
7200 get_task_struct(mt);
7201 task_rq_unlock(rq, &flags);
7202 wake_up_process(rq->migration_thread);
7203 put_task_struct(mt);
7204 wait_for_completion(&req.done);
7205 tlb_migrate_finish(p->mm);
7209 task_rq_unlock(rq, &flags);
7213 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
7216 * Move (not current) task off this cpu, onto dest cpu. We're doing
7217 * this because either it can't run here any more (set_cpus_allowed()
7218 * away from this CPU, or CPU going down), or because we're
7219 * attempting to rebalance this task on exec (sched_exec).
7221 * So we race with normal scheduler movements, but that's OK, as long
7222 * as the task is no longer on this CPU.
7224 * Returns non-zero if task was successfully migrated.
7226 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
7228 struct rq *rq_dest, *rq_src;
7231 if (unlikely(!cpu_active(dest_cpu)))
7234 rq_src = cpu_rq(src_cpu);
7235 rq_dest = cpu_rq(dest_cpu);
7237 double_rq_lock(rq_src, rq_dest);
7238 /* Already moved. */
7239 if (task_cpu(p) != src_cpu)
7241 /* Affinity changed (again). */
7242 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7246 * If we're not on a rq, the next wake-up will ensure we're
7250 deactivate_task(rq_src, p, 0);
7251 set_task_cpu(p, dest_cpu);
7252 activate_task(rq_dest, p, 0);
7253 check_preempt_curr(rq_dest, p, 0);
7258 double_rq_unlock(rq_src, rq_dest);
7262 #define RCU_MIGRATION_IDLE 0
7263 #define RCU_MIGRATION_NEED_QS 1
7264 #define RCU_MIGRATION_GOT_QS 2
7265 #define RCU_MIGRATION_MUST_SYNC 3
7268 * migration_thread - this is a highprio system thread that performs
7269 * thread migration by bumping thread off CPU then 'pushing' onto
7272 static int migration_thread(void *data)
7275 int cpu = (long)data;
7279 BUG_ON(rq->migration_thread != current);
7281 set_current_state(TASK_INTERRUPTIBLE);
7282 while (!kthread_should_stop()) {
7283 struct migration_req *req;
7284 struct list_head *head;
7286 raw_spin_lock_irq(&rq->lock);
7288 if (cpu_is_offline(cpu)) {
7289 raw_spin_unlock_irq(&rq->lock);
7293 if (rq->active_balance) {
7294 active_load_balance(rq, cpu);
7295 rq->active_balance = 0;
7298 head = &rq->migration_queue;
7300 if (list_empty(head)) {
7301 raw_spin_unlock_irq(&rq->lock);
7303 set_current_state(TASK_INTERRUPTIBLE);
7306 req = list_entry(head->next, struct migration_req, list);
7307 list_del_init(head->next);
7309 if (req->task != NULL) {
7310 raw_spin_unlock(&rq->lock);
7311 __migrate_task(req->task, cpu, req->dest_cpu);
7312 } else if (likely(cpu == (badcpu = smp_processor_id()))) {
7313 req->dest_cpu = RCU_MIGRATION_GOT_QS;
7314 raw_spin_unlock(&rq->lock);
7316 req->dest_cpu = RCU_MIGRATION_MUST_SYNC;
7317 raw_spin_unlock(&rq->lock);
7318 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu, cpu);
7322 complete(&req->done);
7324 __set_current_state(TASK_RUNNING);
7329 #ifdef CONFIG_HOTPLUG_CPU
7331 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
7335 local_irq_disable();
7336 ret = __migrate_task(p, src_cpu, dest_cpu);
7342 * Figure out where task on dead CPU should go, use force if necessary.
7344 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
7349 dest_cpu = select_fallback_rq(dead_cpu, p);
7351 /* It can have affinity changed while we were choosing. */
7352 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
7357 * While a dead CPU has no uninterruptible tasks queued at this point,
7358 * it might still have a nonzero ->nr_uninterruptible counter, because
7359 * for performance reasons the counter is not stricly tracking tasks to
7360 * their home CPUs. So we just add the counter to another CPU's counter,
7361 * to keep the global sum constant after CPU-down:
7363 static void migrate_nr_uninterruptible(struct rq *rq_src)
7365 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
7366 unsigned long flags;
7368 local_irq_save(flags);
7369 double_rq_lock(rq_src, rq_dest);
7370 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
7371 rq_src->nr_uninterruptible = 0;
7372 double_rq_unlock(rq_src, rq_dest);
7373 local_irq_restore(flags);
7376 /* Run through task list and migrate tasks from the dead cpu. */
7377 static void migrate_live_tasks(int src_cpu)
7379 struct task_struct *p, *t;
7381 read_lock(&tasklist_lock);
7383 do_each_thread(t, p) {
7387 if (task_cpu(p) == src_cpu)
7388 move_task_off_dead_cpu(src_cpu, p);
7389 } while_each_thread(t, p);
7391 read_unlock(&tasklist_lock);
7395 * Schedules idle task to be the next runnable task on current CPU.
7396 * It does so by boosting its priority to highest possible.
7397 * Used by CPU offline code.
7399 void sched_idle_next(void)
7401 int this_cpu = smp_processor_id();
7402 struct rq *rq = cpu_rq(this_cpu);
7403 struct task_struct *p = rq->idle;
7404 unsigned long flags;
7406 /* cpu has to be offline */
7407 BUG_ON(cpu_online(this_cpu));
7410 * Strictly not necessary since rest of the CPUs are stopped by now
7411 * and interrupts disabled on the current cpu.
7413 raw_spin_lock_irqsave(&rq->lock, flags);
7415 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7417 update_rq_clock(rq);
7418 activate_task(rq, p, 0);
7420 raw_spin_unlock_irqrestore(&rq->lock, flags);
7424 * Ensures that the idle task is using init_mm right before its cpu goes
7427 void idle_task_exit(void)
7429 struct mm_struct *mm = current->active_mm;
7431 BUG_ON(cpu_online(smp_processor_id()));
7434 switch_mm(mm, &init_mm, current);
7438 /* called under rq->lock with disabled interrupts */
7439 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
7441 struct rq *rq = cpu_rq(dead_cpu);
7443 /* Must be exiting, otherwise would be on tasklist. */
7444 BUG_ON(!p->exit_state);
7446 /* Cannot have done final schedule yet: would have vanished. */
7447 BUG_ON(p->state == TASK_DEAD);
7452 * Drop lock around migration; if someone else moves it,
7453 * that's OK. No task can be added to this CPU, so iteration is
7456 raw_spin_unlock_irq(&rq->lock);
7457 move_task_off_dead_cpu(dead_cpu, p);
7458 raw_spin_lock_irq(&rq->lock);
7463 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7464 static void migrate_dead_tasks(unsigned int dead_cpu)
7466 struct rq *rq = cpu_rq(dead_cpu);
7467 struct task_struct *next;
7470 if (!rq->nr_running)
7472 update_rq_clock(rq);
7473 next = pick_next_task(rq);
7476 next->sched_class->put_prev_task(rq, next);
7477 migrate_dead(dead_cpu, next);
7483 * remove the tasks which were accounted by rq from calc_load_tasks.
7485 static void calc_global_load_remove(struct rq *rq)
7487 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
7488 rq->calc_load_active = 0;
7490 #endif /* CONFIG_HOTPLUG_CPU */
7492 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7494 static struct ctl_table sd_ctl_dir[] = {
7496 .procname = "sched_domain",
7502 static struct ctl_table sd_ctl_root[] = {
7504 .procname = "kernel",
7506 .child = sd_ctl_dir,
7511 static struct ctl_table *sd_alloc_ctl_entry(int n)
7513 struct ctl_table *entry =
7514 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
7519 static void sd_free_ctl_entry(struct ctl_table **tablep)
7521 struct ctl_table *entry;
7524 * In the intermediate directories, both the child directory and
7525 * procname are dynamically allocated and could fail but the mode
7526 * will always be set. In the lowest directory the names are
7527 * static strings and all have proc handlers.
7529 for (entry = *tablep; entry->mode; entry++) {
7531 sd_free_ctl_entry(&entry->child);
7532 if (entry->proc_handler == NULL)
7533 kfree(entry->procname);
7541 set_table_entry(struct ctl_table *entry,
7542 const char *procname, void *data, int maxlen,
7543 mode_t mode, proc_handler *proc_handler)
7545 entry->procname = procname;
7547 entry->maxlen = maxlen;
7549 entry->proc_handler = proc_handler;
7552 static struct ctl_table *
7553 sd_alloc_ctl_domain_table(struct sched_domain *sd)
7555 struct ctl_table *table = sd_alloc_ctl_entry(13);
7560 set_table_entry(&table[0], "min_interval", &sd->min_interval,
7561 sizeof(long), 0644, proc_doulongvec_minmax);
7562 set_table_entry(&table[1], "max_interval", &sd->max_interval,
7563 sizeof(long), 0644, proc_doulongvec_minmax);
7564 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
7565 sizeof(int), 0644, proc_dointvec_minmax);
7566 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
7567 sizeof(int), 0644, proc_dointvec_minmax);
7568 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
7569 sizeof(int), 0644, proc_dointvec_minmax);
7570 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
7571 sizeof(int), 0644, proc_dointvec_minmax);
7572 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
7573 sizeof(int), 0644, proc_dointvec_minmax);
7574 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
7575 sizeof(int), 0644, proc_dointvec_minmax);
7576 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
7577 sizeof(int), 0644, proc_dointvec_minmax);
7578 set_table_entry(&table[9], "cache_nice_tries",
7579 &sd->cache_nice_tries,
7580 sizeof(int), 0644, proc_dointvec_minmax);
7581 set_table_entry(&table[10], "flags", &sd->flags,
7582 sizeof(int), 0644, proc_dointvec_minmax);
7583 set_table_entry(&table[11], "name", sd->name,
7584 CORENAME_MAX_SIZE, 0444, proc_dostring);
7585 /* &table[12] is terminator */
7590 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
7592 struct ctl_table *entry, *table;
7593 struct sched_domain *sd;
7594 int domain_num = 0, i;
7597 for_each_domain(cpu, sd)
7599 entry = table = sd_alloc_ctl_entry(domain_num + 1);
7604 for_each_domain(cpu, sd) {
7605 snprintf(buf, 32, "domain%d", i);
7606 entry->procname = kstrdup(buf, GFP_KERNEL);
7608 entry->child = sd_alloc_ctl_domain_table(sd);
7615 static struct ctl_table_header *sd_sysctl_header;
7616 static void register_sched_domain_sysctl(void)
7618 int i, cpu_num = num_possible_cpus();
7619 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
7622 WARN_ON(sd_ctl_dir[0].child);
7623 sd_ctl_dir[0].child = entry;
7628 for_each_possible_cpu(i) {
7629 snprintf(buf, 32, "cpu%d", i);
7630 entry->procname = kstrdup(buf, GFP_KERNEL);
7632 entry->child = sd_alloc_ctl_cpu_table(i);
7636 WARN_ON(sd_sysctl_header);
7637 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
7640 /* may be called multiple times per register */
7641 static void unregister_sched_domain_sysctl(void)
7643 if (sd_sysctl_header)
7644 unregister_sysctl_table(sd_sysctl_header);
7645 sd_sysctl_header = NULL;
7646 if (sd_ctl_dir[0].child)
7647 sd_free_ctl_entry(&sd_ctl_dir[0].child);
7650 static void register_sched_domain_sysctl(void)
7653 static void unregister_sched_domain_sysctl(void)
7658 static void set_rq_online(struct rq *rq)
7661 const struct sched_class *class;
7663 cpumask_set_cpu(rq->cpu, rq->rd->online);
7666 for_each_class(class) {
7667 if (class->rq_online)
7668 class->rq_online(rq);
7673 static void set_rq_offline(struct rq *rq)
7676 const struct sched_class *class;
7678 for_each_class(class) {
7679 if (class->rq_offline)
7680 class->rq_offline(rq);
7683 cpumask_clear_cpu(rq->cpu, rq->rd->online);
7689 * migration_call - callback that gets triggered when a CPU is added.
7690 * Here we can start up the necessary migration thread for the new CPU.
7692 static int __cpuinit
7693 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
7695 struct task_struct *p;
7696 int cpu = (long)hcpu;
7697 unsigned long flags;
7702 case CPU_UP_PREPARE:
7703 case CPU_UP_PREPARE_FROZEN:
7704 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
7707 kthread_bind(p, cpu);
7708 /* Must be high prio: stop_machine expects to yield to it. */
7709 rq = task_rq_lock(p, &flags);
7710 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7711 task_rq_unlock(rq, &flags);
7713 cpu_rq(cpu)->migration_thread = p;
7714 rq->calc_load_update = calc_load_update;
7718 case CPU_ONLINE_FROZEN:
7719 /* Strictly unnecessary, as first user will wake it. */
7720 wake_up_process(cpu_rq(cpu)->migration_thread);
7722 /* Update our root-domain */
7724 raw_spin_lock_irqsave(&rq->lock, flags);
7726 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7730 raw_spin_unlock_irqrestore(&rq->lock, flags);
7733 #ifdef CONFIG_HOTPLUG_CPU
7734 case CPU_UP_CANCELED:
7735 case CPU_UP_CANCELED_FROZEN:
7736 if (!cpu_rq(cpu)->migration_thread)
7738 /* Unbind it from offline cpu so it can run. Fall thru. */
7739 kthread_bind(cpu_rq(cpu)->migration_thread,
7740 cpumask_any(cpu_online_mask));
7741 kthread_stop(cpu_rq(cpu)->migration_thread);
7742 put_task_struct(cpu_rq(cpu)->migration_thread);
7743 cpu_rq(cpu)->migration_thread = NULL;
7747 case CPU_DEAD_FROZEN:
7748 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7749 migrate_live_tasks(cpu);
7751 kthread_stop(rq->migration_thread);
7752 put_task_struct(rq->migration_thread);
7753 rq->migration_thread = NULL;
7754 /* Idle task back to normal (off runqueue, low prio) */
7755 raw_spin_lock_irq(&rq->lock);
7756 update_rq_clock(rq);
7757 deactivate_task(rq, rq->idle, 0);
7758 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
7759 rq->idle->sched_class = &idle_sched_class;
7760 migrate_dead_tasks(cpu);
7761 raw_spin_unlock_irq(&rq->lock);
7763 migrate_nr_uninterruptible(rq);
7764 BUG_ON(rq->nr_running != 0);
7765 calc_global_load_remove(rq);
7767 * No need to migrate the tasks: it was best-effort if
7768 * they didn't take sched_hotcpu_mutex. Just wake up
7771 raw_spin_lock_irq(&rq->lock);
7772 while (!list_empty(&rq->migration_queue)) {
7773 struct migration_req *req;
7775 req = list_entry(rq->migration_queue.next,
7776 struct migration_req, list);
7777 list_del_init(&req->list);
7778 raw_spin_unlock_irq(&rq->lock);
7779 complete(&req->done);
7780 raw_spin_lock_irq(&rq->lock);
7782 raw_spin_unlock_irq(&rq->lock);
7786 case CPU_DYING_FROZEN:
7787 /* Update our root-domain */
7789 raw_spin_lock_irqsave(&rq->lock, flags);
7791 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7794 raw_spin_unlock_irqrestore(&rq->lock, flags);
7802 * Register at high priority so that task migration (migrate_all_tasks)
7803 * happens before everything else. This has to be lower priority than
7804 * the notifier in the perf_event subsystem, though.
7806 static struct notifier_block __cpuinitdata migration_notifier = {
7807 .notifier_call = migration_call,
7811 static int __init migration_init(void)
7813 void *cpu = (void *)(long)smp_processor_id();
7816 /* Start one for the boot CPU: */
7817 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
7818 BUG_ON(err == NOTIFY_BAD);
7819 migration_call(&migration_notifier, CPU_ONLINE, cpu);
7820 register_cpu_notifier(&migration_notifier);
7824 early_initcall(migration_init);
7829 #ifdef CONFIG_SCHED_DEBUG
7831 static __read_mostly int sched_domain_debug_enabled;
7833 static int __init sched_domain_debug_setup(char *str)
7835 sched_domain_debug_enabled = 1;
7839 early_param("sched_debug", sched_domain_debug_setup);
7841 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
7842 struct cpumask *groupmask)
7844 struct sched_group *group = sd->groups;
7847 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
7848 cpumask_clear(groupmask);
7850 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
7852 if (!(sd->flags & SD_LOAD_BALANCE)) {
7853 printk("does not load-balance\n");
7855 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
7860 printk(KERN_CONT "span %s level %s\n", str, sd->name);
7862 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
7863 printk(KERN_ERR "ERROR: domain->span does not contain "
7866 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
7867 printk(KERN_ERR "ERROR: domain->groups does not contain"
7871 printk(KERN_DEBUG "%*s groups:", level + 1, "");
7875 printk(KERN_ERR "ERROR: group is NULL\n");
7879 if (!group->cpu_power) {
7880 printk(KERN_CONT "\n");
7881 printk(KERN_ERR "ERROR: domain->cpu_power not "
7886 if (!cpumask_weight(sched_group_cpus(group))) {
7887 printk(KERN_CONT "\n");
7888 printk(KERN_ERR "ERROR: empty group\n");
7892 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
7893 printk(KERN_CONT "\n");
7894 printk(KERN_ERR "ERROR: repeated CPUs\n");
7898 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
7900 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
7902 printk(KERN_CONT " %s", str);
7903 if (group->cpu_power != SCHED_LOAD_SCALE) {
7904 printk(KERN_CONT " (cpu_power = %d)",
7908 group = group->next;
7909 } while (group != sd->groups);
7910 printk(KERN_CONT "\n");
7912 if (!cpumask_equal(sched_domain_span(sd), groupmask))
7913 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
7916 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
7917 printk(KERN_ERR "ERROR: parent span is not a superset "
7918 "of domain->span\n");
7922 static void sched_domain_debug(struct sched_domain *sd, int cpu)
7924 cpumask_var_t groupmask;
7927 if (!sched_domain_debug_enabled)
7931 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7935 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7937 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
7938 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
7943 if (sched_domain_debug_one(sd, cpu, level, groupmask))
7950 free_cpumask_var(groupmask);
7952 #else /* !CONFIG_SCHED_DEBUG */
7953 # define sched_domain_debug(sd, cpu) do { } while (0)
7954 #endif /* CONFIG_SCHED_DEBUG */
7956 static int sd_degenerate(struct sched_domain *sd)
7958 if (cpumask_weight(sched_domain_span(sd)) == 1)
7961 /* Following flags need at least 2 groups */
7962 if (sd->flags & (SD_LOAD_BALANCE |
7963 SD_BALANCE_NEWIDLE |
7967 SD_SHARE_PKG_RESOURCES)) {
7968 if (sd->groups != sd->groups->next)
7972 /* Following flags don't use groups */
7973 if (sd->flags & (SD_WAKE_AFFINE))
7980 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
7982 unsigned long cflags = sd->flags, pflags = parent->flags;
7984 if (sd_degenerate(parent))
7987 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
7990 /* Flags needing groups don't count if only 1 group in parent */
7991 if (parent->groups == parent->groups->next) {
7992 pflags &= ~(SD_LOAD_BALANCE |
7993 SD_BALANCE_NEWIDLE |
7997 SD_SHARE_PKG_RESOURCES);
7998 if (nr_node_ids == 1)
7999 pflags &= ~SD_SERIALIZE;
8001 if (~cflags & pflags)
8007 static void free_rootdomain(struct root_domain *rd)
8009 synchronize_sched();
8011 cpupri_cleanup(&rd->cpupri);
8013 free_cpumask_var(rd->rto_mask);
8014 free_cpumask_var(rd->online);
8015 free_cpumask_var(rd->span);
8019 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
8021 struct root_domain *old_rd = NULL;
8022 unsigned long flags;
8024 raw_spin_lock_irqsave(&rq->lock, flags);
8029 if (cpumask_test_cpu(rq->cpu, old_rd->online))
8032 cpumask_clear_cpu(rq->cpu, old_rd->span);
8035 * If we dont want to free the old_rt yet then
8036 * set old_rd to NULL to skip the freeing later
8039 if (!atomic_dec_and_test(&old_rd->refcount))
8043 atomic_inc(&rd->refcount);
8046 cpumask_set_cpu(rq->cpu, rd->span);
8047 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
8050 raw_spin_unlock_irqrestore(&rq->lock, flags);
8053 free_rootdomain(old_rd);
8056 static int init_rootdomain(struct root_domain *rd, bool bootmem)
8058 gfp_t gfp = GFP_KERNEL;
8060 memset(rd, 0, sizeof(*rd));
8065 if (!alloc_cpumask_var(&rd->span, gfp))
8067 if (!alloc_cpumask_var(&rd->online, gfp))
8069 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
8072 if (cpupri_init(&rd->cpupri, bootmem) != 0)
8077 free_cpumask_var(rd->rto_mask);
8079 free_cpumask_var(rd->online);
8081 free_cpumask_var(rd->span);
8086 static void init_defrootdomain(void)
8088 init_rootdomain(&def_root_domain, true);
8090 atomic_set(&def_root_domain.refcount, 1);
8093 static struct root_domain *alloc_rootdomain(void)
8095 struct root_domain *rd;
8097 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
8101 if (init_rootdomain(rd, false) != 0) {
8110 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
8111 * hold the hotplug lock.
8114 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
8116 struct rq *rq = cpu_rq(cpu);
8117 struct sched_domain *tmp;
8119 /* Remove the sched domains which do not contribute to scheduling. */
8120 for (tmp = sd; tmp; ) {
8121 struct sched_domain *parent = tmp->parent;
8125 if (sd_parent_degenerate(tmp, parent)) {
8126 tmp->parent = parent->parent;
8128 parent->parent->child = tmp;
8133 if (sd && sd_degenerate(sd)) {
8139 sched_domain_debug(sd, cpu);
8141 rq_attach_root(rq, rd);
8142 rcu_assign_pointer(rq->sd, sd);
8145 /* cpus with isolated domains */
8146 static cpumask_var_t cpu_isolated_map;
8148 /* Setup the mask of cpus configured for isolated domains */
8149 static int __init isolated_cpu_setup(char *str)
8151 alloc_bootmem_cpumask_var(&cpu_isolated_map);
8152 cpulist_parse(str, cpu_isolated_map);
8156 __setup("isolcpus=", isolated_cpu_setup);
8159 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
8160 * to a function which identifies what group(along with sched group) a CPU
8161 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
8162 * (due to the fact that we keep track of groups covered with a struct cpumask).
8164 * init_sched_build_groups will build a circular linked list of the groups
8165 * covered by the given span, and will set each group's ->cpumask correctly,
8166 * and ->cpu_power to 0.
8169 init_sched_build_groups(const struct cpumask *span,
8170 const struct cpumask *cpu_map,
8171 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
8172 struct sched_group **sg,
8173 struct cpumask *tmpmask),
8174 struct cpumask *covered, struct cpumask *tmpmask)
8176 struct sched_group *first = NULL, *last = NULL;
8179 cpumask_clear(covered);
8181 for_each_cpu(i, span) {
8182 struct sched_group *sg;
8183 int group = group_fn(i, cpu_map, &sg, tmpmask);
8186 if (cpumask_test_cpu(i, covered))
8189 cpumask_clear(sched_group_cpus(sg));
8192 for_each_cpu(j, span) {
8193 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
8196 cpumask_set_cpu(j, covered);
8197 cpumask_set_cpu(j, sched_group_cpus(sg));
8208 #define SD_NODES_PER_DOMAIN 16
8213 * find_next_best_node - find the next node to include in a sched_domain
8214 * @node: node whose sched_domain we're building
8215 * @used_nodes: nodes already in the sched_domain
8217 * Find the next node to include in a given scheduling domain. Simply
8218 * finds the closest node not already in the @used_nodes map.
8220 * Should use nodemask_t.
8222 static int find_next_best_node(int node, nodemask_t *used_nodes)
8224 int i, n, val, min_val, best_node = 0;
8228 for (i = 0; i < nr_node_ids; i++) {
8229 /* Start at @node */
8230 n = (node + i) % nr_node_ids;
8232 if (!nr_cpus_node(n))
8235 /* Skip already used nodes */
8236 if (node_isset(n, *used_nodes))
8239 /* Simple min distance search */
8240 val = node_distance(node, n);
8242 if (val < min_val) {
8248 node_set(best_node, *used_nodes);
8253 * sched_domain_node_span - get a cpumask for a node's sched_domain
8254 * @node: node whose cpumask we're constructing
8255 * @span: resulting cpumask
8257 * Given a node, construct a good cpumask for its sched_domain to span. It
8258 * should be one that prevents unnecessary balancing, but also spreads tasks
8261 static void sched_domain_node_span(int node, struct cpumask *span)
8263 nodemask_t used_nodes;
8266 cpumask_clear(span);
8267 nodes_clear(used_nodes);
8269 cpumask_or(span, span, cpumask_of_node(node));
8270 node_set(node, used_nodes);
8272 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
8273 int next_node = find_next_best_node(node, &used_nodes);
8275 cpumask_or(span, span, cpumask_of_node(next_node));
8278 #endif /* CONFIG_NUMA */
8280 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
8283 * The cpus mask in sched_group and sched_domain hangs off the end.
8285 * ( See the the comments in include/linux/sched.h:struct sched_group
8286 * and struct sched_domain. )
8288 struct static_sched_group {
8289 struct sched_group sg;
8290 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
8293 struct static_sched_domain {
8294 struct sched_domain sd;
8295 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
8301 cpumask_var_t domainspan;
8302 cpumask_var_t covered;
8303 cpumask_var_t notcovered;
8305 cpumask_var_t nodemask;
8306 cpumask_var_t this_sibling_map;
8307 cpumask_var_t this_core_map;
8308 cpumask_var_t send_covered;
8309 cpumask_var_t tmpmask;
8310 struct sched_group **sched_group_nodes;
8311 struct root_domain *rd;
8315 sa_sched_groups = 0,
8320 sa_this_sibling_map,
8322 sa_sched_group_nodes,
8332 * SMT sched-domains:
8334 #ifdef CONFIG_SCHED_SMT
8335 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
8336 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
8339 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
8340 struct sched_group **sg, struct cpumask *unused)
8343 *sg = &per_cpu(sched_groups, cpu).sg;
8346 #endif /* CONFIG_SCHED_SMT */
8349 * multi-core sched-domains:
8351 #ifdef CONFIG_SCHED_MC
8352 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
8353 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
8354 #endif /* CONFIG_SCHED_MC */
8356 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8358 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8359 struct sched_group **sg, struct cpumask *mask)
8363 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8364 group = cpumask_first(mask);
8366 *sg = &per_cpu(sched_group_core, group).sg;
8369 #elif defined(CONFIG_SCHED_MC)
8371 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8372 struct sched_group **sg, struct cpumask *unused)
8375 *sg = &per_cpu(sched_group_core, cpu).sg;
8380 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
8381 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
8384 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
8385 struct sched_group **sg, struct cpumask *mask)
8388 #ifdef CONFIG_SCHED_MC
8389 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
8390 group = cpumask_first(mask);
8391 #elif defined(CONFIG_SCHED_SMT)
8392 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8393 group = cpumask_first(mask);
8398 *sg = &per_cpu(sched_group_phys, group).sg;
8404 * The init_sched_build_groups can't handle what we want to do with node
8405 * groups, so roll our own. Now each node has its own list of groups which
8406 * gets dynamically allocated.
8408 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
8409 static struct sched_group ***sched_group_nodes_bycpu;
8411 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
8412 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
8414 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
8415 struct sched_group **sg,
8416 struct cpumask *nodemask)
8420 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
8421 group = cpumask_first(nodemask);
8424 *sg = &per_cpu(sched_group_allnodes, group).sg;
8428 static void init_numa_sched_groups_power(struct sched_group *group_head)
8430 struct sched_group *sg = group_head;
8436 for_each_cpu(j, sched_group_cpus(sg)) {
8437 struct sched_domain *sd;
8439 sd = &per_cpu(phys_domains, j).sd;
8440 if (j != group_first_cpu(sd->groups)) {
8442 * Only add "power" once for each
8448 sg->cpu_power += sd->groups->cpu_power;
8451 } while (sg != group_head);
8454 static int build_numa_sched_groups(struct s_data *d,
8455 const struct cpumask *cpu_map, int num)
8457 struct sched_domain *sd;
8458 struct sched_group *sg, *prev;
8461 cpumask_clear(d->covered);
8462 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
8463 if (cpumask_empty(d->nodemask)) {
8464 d->sched_group_nodes[num] = NULL;
8468 sched_domain_node_span(num, d->domainspan);
8469 cpumask_and(d->domainspan, d->domainspan, cpu_map);
8471 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8474 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
8478 d->sched_group_nodes[num] = sg;
8480 for_each_cpu(j, d->nodemask) {
8481 sd = &per_cpu(node_domains, j).sd;
8486 cpumask_copy(sched_group_cpus(sg), d->nodemask);
8488 cpumask_or(d->covered, d->covered, d->nodemask);
8491 for (j = 0; j < nr_node_ids; j++) {
8492 n = (num + j) % nr_node_ids;
8493 cpumask_complement(d->notcovered, d->covered);
8494 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
8495 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
8496 if (cpumask_empty(d->tmpmask))
8498 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
8499 if (cpumask_empty(d->tmpmask))
8501 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8505 "Can not alloc domain group for node %d\n", j);
8509 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
8510 sg->next = prev->next;
8511 cpumask_or(d->covered, d->covered, d->tmpmask);
8518 #endif /* CONFIG_NUMA */
8521 /* Free memory allocated for various sched_group structures */
8522 static void free_sched_groups(const struct cpumask *cpu_map,
8523 struct cpumask *nodemask)
8527 for_each_cpu(cpu, cpu_map) {
8528 struct sched_group **sched_group_nodes
8529 = sched_group_nodes_bycpu[cpu];
8531 if (!sched_group_nodes)
8534 for (i = 0; i < nr_node_ids; i++) {
8535 struct sched_group *oldsg, *sg = sched_group_nodes[i];
8537 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8538 if (cpumask_empty(nodemask))
8548 if (oldsg != sched_group_nodes[i])
8551 kfree(sched_group_nodes);
8552 sched_group_nodes_bycpu[cpu] = NULL;
8555 #else /* !CONFIG_NUMA */
8556 static void free_sched_groups(const struct cpumask *cpu_map,
8557 struct cpumask *nodemask)
8560 #endif /* CONFIG_NUMA */
8563 * Initialize sched groups cpu_power.
8565 * cpu_power indicates the capacity of sched group, which is used while
8566 * distributing the load between different sched groups in a sched domain.
8567 * Typically cpu_power for all the groups in a sched domain will be same unless
8568 * there are asymmetries in the topology. If there are asymmetries, group
8569 * having more cpu_power will pickup more load compared to the group having
8572 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
8574 struct sched_domain *child;
8575 struct sched_group *group;
8579 WARN_ON(!sd || !sd->groups);
8581 if (cpu != group_first_cpu(sd->groups))
8586 sd->groups->cpu_power = 0;
8589 power = SCHED_LOAD_SCALE;
8590 weight = cpumask_weight(sched_domain_span(sd));
8592 * SMT siblings share the power of a single core.
8593 * Usually multiple threads get a better yield out of
8594 * that one core than a single thread would have,
8595 * reflect that in sd->smt_gain.
8597 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
8598 power *= sd->smt_gain;
8600 power >>= SCHED_LOAD_SHIFT;
8602 sd->groups->cpu_power += power;
8607 * Add cpu_power of each child group to this groups cpu_power.
8609 group = child->groups;
8611 sd->groups->cpu_power += group->cpu_power;
8612 group = group->next;
8613 } while (group != child->groups);
8617 * Initializers for schedule domains
8618 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8621 #ifdef CONFIG_SCHED_DEBUG
8622 # define SD_INIT_NAME(sd, type) sd->name = #type
8624 # define SD_INIT_NAME(sd, type) do { } while (0)
8627 #define SD_INIT(sd, type) sd_init_##type(sd)
8629 #define SD_INIT_FUNC(type) \
8630 static noinline void sd_init_##type(struct sched_domain *sd) \
8632 memset(sd, 0, sizeof(*sd)); \
8633 *sd = SD_##type##_INIT; \
8634 sd->level = SD_LV_##type; \
8635 SD_INIT_NAME(sd, type); \
8640 SD_INIT_FUNC(ALLNODES)
8643 #ifdef CONFIG_SCHED_SMT
8644 SD_INIT_FUNC(SIBLING)
8646 #ifdef CONFIG_SCHED_MC
8650 static int default_relax_domain_level = -1;
8652 static int __init setup_relax_domain_level(char *str)
8656 val = simple_strtoul(str, NULL, 0);
8657 if (val < SD_LV_MAX)
8658 default_relax_domain_level = val;
8662 __setup("relax_domain_level=", setup_relax_domain_level);
8664 static void set_domain_attribute(struct sched_domain *sd,
8665 struct sched_domain_attr *attr)
8669 if (!attr || attr->relax_domain_level < 0) {
8670 if (default_relax_domain_level < 0)
8673 request = default_relax_domain_level;
8675 request = attr->relax_domain_level;
8676 if (request < sd->level) {
8677 /* turn off idle balance on this domain */
8678 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
8680 /* turn on idle balance on this domain */
8681 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
8685 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
8686 const struct cpumask *cpu_map)
8689 case sa_sched_groups:
8690 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
8691 d->sched_group_nodes = NULL;
8693 free_rootdomain(d->rd); /* fall through */
8695 free_cpumask_var(d->tmpmask); /* fall through */
8696 case sa_send_covered:
8697 free_cpumask_var(d->send_covered); /* fall through */
8698 case sa_this_core_map:
8699 free_cpumask_var(d->this_core_map); /* fall through */
8700 case sa_this_sibling_map:
8701 free_cpumask_var(d->this_sibling_map); /* fall through */
8703 free_cpumask_var(d->nodemask); /* fall through */
8704 case sa_sched_group_nodes:
8706 kfree(d->sched_group_nodes); /* fall through */
8708 free_cpumask_var(d->notcovered); /* fall through */
8710 free_cpumask_var(d->covered); /* fall through */
8712 free_cpumask_var(d->domainspan); /* fall through */
8719 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
8720 const struct cpumask *cpu_map)
8723 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
8725 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
8726 return sa_domainspan;
8727 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
8729 /* Allocate the per-node list of sched groups */
8730 d->sched_group_nodes = kcalloc(nr_node_ids,
8731 sizeof(struct sched_group *), GFP_KERNEL);
8732 if (!d->sched_group_nodes) {
8733 printk(KERN_WARNING "Can not alloc sched group node list\n");
8734 return sa_notcovered;
8736 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
8738 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
8739 return sa_sched_group_nodes;
8740 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
8742 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
8743 return sa_this_sibling_map;
8744 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
8745 return sa_this_core_map;
8746 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
8747 return sa_send_covered;
8748 d->rd = alloc_rootdomain();
8750 printk(KERN_WARNING "Cannot alloc root domain\n");
8753 return sa_rootdomain;
8756 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
8757 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
8759 struct sched_domain *sd = NULL;
8761 struct sched_domain *parent;
8764 if (cpumask_weight(cpu_map) >
8765 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
8766 sd = &per_cpu(allnodes_domains, i).sd;
8767 SD_INIT(sd, ALLNODES);
8768 set_domain_attribute(sd, attr);
8769 cpumask_copy(sched_domain_span(sd), cpu_map);
8770 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
8775 sd = &per_cpu(node_domains, i).sd;
8777 set_domain_attribute(sd, attr);
8778 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
8779 sd->parent = parent;
8782 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
8787 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
8788 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8789 struct sched_domain *parent, int i)
8791 struct sched_domain *sd;
8792 sd = &per_cpu(phys_domains, i).sd;
8794 set_domain_attribute(sd, attr);
8795 cpumask_copy(sched_domain_span(sd), d->nodemask);
8796 sd->parent = parent;
8799 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
8803 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
8804 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8805 struct sched_domain *parent, int i)
8807 struct sched_domain *sd = parent;
8808 #ifdef CONFIG_SCHED_MC
8809 sd = &per_cpu(core_domains, i).sd;
8811 set_domain_attribute(sd, attr);
8812 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
8813 sd->parent = parent;
8815 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
8820 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
8821 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8822 struct sched_domain *parent, int i)
8824 struct sched_domain *sd = parent;
8825 #ifdef CONFIG_SCHED_SMT
8826 sd = &per_cpu(cpu_domains, i).sd;
8827 SD_INIT(sd, SIBLING);
8828 set_domain_attribute(sd, attr);
8829 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
8830 sd->parent = parent;
8832 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
8837 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
8838 const struct cpumask *cpu_map, int cpu)
8841 #ifdef CONFIG_SCHED_SMT
8842 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
8843 cpumask_and(d->this_sibling_map, cpu_map,
8844 topology_thread_cpumask(cpu));
8845 if (cpu == cpumask_first(d->this_sibling_map))
8846 init_sched_build_groups(d->this_sibling_map, cpu_map,
8848 d->send_covered, d->tmpmask);
8851 #ifdef CONFIG_SCHED_MC
8852 case SD_LV_MC: /* set up multi-core groups */
8853 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
8854 if (cpu == cpumask_first(d->this_core_map))
8855 init_sched_build_groups(d->this_core_map, cpu_map,
8857 d->send_covered, d->tmpmask);
8860 case SD_LV_CPU: /* set up physical groups */
8861 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
8862 if (!cpumask_empty(d->nodemask))
8863 init_sched_build_groups(d->nodemask, cpu_map,
8865 d->send_covered, d->tmpmask);
8868 case SD_LV_ALLNODES:
8869 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
8870 d->send_covered, d->tmpmask);
8879 * Build sched domains for a given set of cpus and attach the sched domains
8880 * to the individual cpus
8882 static int __build_sched_domains(const struct cpumask *cpu_map,
8883 struct sched_domain_attr *attr)
8885 enum s_alloc alloc_state = sa_none;
8887 struct sched_domain *sd;
8893 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
8894 if (alloc_state != sa_rootdomain)
8896 alloc_state = sa_sched_groups;
8899 * Set up domains for cpus specified by the cpu_map.
8901 for_each_cpu(i, cpu_map) {
8902 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
8905 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
8906 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
8907 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
8908 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
8911 for_each_cpu(i, cpu_map) {
8912 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
8913 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
8916 /* Set up physical groups */
8917 for (i = 0; i < nr_node_ids; i++)
8918 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
8921 /* Set up node groups */
8923 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
8925 for (i = 0; i < nr_node_ids; i++)
8926 if (build_numa_sched_groups(&d, cpu_map, i))
8930 /* Calculate CPU power for physical packages and nodes */
8931 #ifdef CONFIG_SCHED_SMT
8932 for_each_cpu(i, cpu_map) {
8933 sd = &per_cpu(cpu_domains, i).sd;
8934 init_sched_groups_power(i, sd);
8937 #ifdef CONFIG_SCHED_MC
8938 for_each_cpu(i, cpu_map) {
8939 sd = &per_cpu(core_domains, i).sd;
8940 init_sched_groups_power(i, sd);
8944 for_each_cpu(i, cpu_map) {
8945 sd = &per_cpu(phys_domains, i).sd;
8946 init_sched_groups_power(i, sd);
8950 for (i = 0; i < nr_node_ids; i++)
8951 init_numa_sched_groups_power(d.sched_group_nodes[i]);
8953 if (d.sd_allnodes) {
8954 struct sched_group *sg;
8956 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
8958 init_numa_sched_groups_power(sg);
8962 /* Attach the domains */
8963 for_each_cpu(i, cpu_map) {
8964 #ifdef CONFIG_SCHED_SMT
8965 sd = &per_cpu(cpu_domains, i).sd;
8966 #elif defined(CONFIG_SCHED_MC)
8967 sd = &per_cpu(core_domains, i).sd;
8969 sd = &per_cpu(phys_domains, i).sd;
8971 cpu_attach_domain(sd, d.rd, i);
8974 d.sched_group_nodes = NULL; /* don't free this we still need it */
8975 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
8979 __free_domain_allocs(&d, alloc_state, cpu_map);
8983 static int build_sched_domains(const struct cpumask *cpu_map)
8985 return __build_sched_domains(cpu_map, NULL);
8988 static cpumask_var_t *doms_cur; /* current sched domains */
8989 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
8990 static struct sched_domain_attr *dattr_cur;
8991 /* attribues of custom domains in 'doms_cur' */
8994 * Special case: If a kmalloc of a doms_cur partition (array of
8995 * cpumask) fails, then fallback to a single sched domain,
8996 * as determined by the single cpumask fallback_doms.
8998 static cpumask_var_t fallback_doms;
9001 * arch_update_cpu_topology lets virtualized architectures update the
9002 * cpu core maps. It is supposed to return 1 if the topology changed
9003 * or 0 if it stayed the same.
9005 int __attribute__((weak)) arch_update_cpu_topology(void)
9010 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
9013 cpumask_var_t *doms;
9015 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
9018 for (i = 0; i < ndoms; i++) {
9019 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
9020 free_sched_domains(doms, i);
9027 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
9030 for (i = 0; i < ndoms; i++)
9031 free_cpumask_var(doms[i]);
9036 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
9037 * For now this just excludes isolated cpus, but could be used to
9038 * exclude other special cases in the future.
9040 static int arch_init_sched_domains(const struct cpumask *cpu_map)
9044 arch_update_cpu_topology();
9046 doms_cur = alloc_sched_domains(ndoms_cur);
9048 doms_cur = &fallback_doms;
9049 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
9051 err = build_sched_domains(doms_cur[0]);
9052 register_sched_domain_sysctl();
9057 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
9058 struct cpumask *tmpmask)
9060 free_sched_groups(cpu_map, tmpmask);
9064 * Detach sched domains from a group of cpus specified in cpu_map
9065 * These cpus will now be attached to the NULL domain
9067 static void detach_destroy_domains(const struct cpumask *cpu_map)
9069 /* Save because hotplug lock held. */
9070 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
9073 for_each_cpu(i, cpu_map)
9074 cpu_attach_domain(NULL, &def_root_domain, i);
9075 synchronize_sched();
9076 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
9079 /* handle null as "default" */
9080 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
9081 struct sched_domain_attr *new, int idx_new)
9083 struct sched_domain_attr tmp;
9090 return !memcmp(cur ? (cur + idx_cur) : &tmp,
9091 new ? (new + idx_new) : &tmp,
9092 sizeof(struct sched_domain_attr));
9096 * Partition sched domains as specified by the 'ndoms_new'
9097 * cpumasks in the array doms_new[] of cpumasks. This compares
9098 * doms_new[] to the current sched domain partitioning, doms_cur[].
9099 * It destroys each deleted domain and builds each new domain.
9101 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
9102 * The masks don't intersect (don't overlap.) We should setup one
9103 * sched domain for each mask. CPUs not in any of the cpumasks will
9104 * not be load balanced. If the same cpumask appears both in the
9105 * current 'doms_cur' domains and in the new 'doms_new', we can leave
9108 * The passed in 'doms_new' should be allocated using
9109 * alloc_sched_domains. This routine takes ownership of it and will
9110 * free_sched_domains it when done with it. If the caller failed the
9111 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
9112 * and partition_sched_domains() will fallback to the single partition
9113 * 'fallback_doms', it also forces the domains to be rebuilt.
9115 * If doms_new == NULL it will be replaced with cpu_online_mask.
9116 * ndoms_new == 0 is a special case for destroying existing domains,
9117 * and it will not create the default domain.
9119 * Call with hotplug lock held
9121 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
9122 struct sched_domain_attr *dattr_new)
9127 mutex_lock(&sched_domains_mutex);
9129 /* always unregister in case we don't destroy any domains */
9130 unregister_sched_domain_sysctl();
9132 /* Let architecture update cpu core mappings. */
9133 new_topology = arch_update_cpu_topology();
9135 n = doms_new ? ndoms_new : 0;
9137 /* Destroy deleted domains */
9138 for (i = 0; i < ndoms_cur; i++) {
9139 for (j = 0; j < n && !new_topology; j++) {
9140 if (cpumask_equal(doms_cur[i], doms_new[j])
9141 && dattrs_equal(dattr_cur, i, dattr_new, j))
9144 /* no match - a current sched domain not in new doms_new[] */
9145 detach_destroy_domains(doms_cur[i]);
9150 if (doms_new == NULL) {
9152 doms_new = &fallback_doms;
9153 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
9154 WARN_ON_ONCE(dattr_new);
9157 /* Build new domains */
9158 for (i = 0; i < ndoms_new; i++) {
9159 for (j = 0; j < ndoms_cur && !new_topology; j++) {
9160 if (cpumask_equal(doms_new[i], doms_cur[j])
9161 && dattrs_equal(dattr_new, i, dattr_cur, j))
9164 /* no match - add a new doms_new */
9165 __build_sched_domains(doms_new[i],
9166 dattr_new ? dattr_new + i : NULL);
9171 /* Remember the new sched domains */
9172 if (doms_cur != &fallback_doms)
9173 free_sched_domains(doms_cur, ndoms_cur);
9174 kfree(dattr_cur); /* kfree(NULL) is safe */
9175 doms_cur = doms_new;
9176 dattr_cur = dattr_new;
9177 ndoms_cur = ndoms_new;
9179 register_sched_domain_sysctl();
9181 mutex_unlock(&sched_domains_mutex);
9184 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
9185 static void arch_reinit_sched_domains(void)
9189 /* Destroy domains first to force the rebuild */
9190 partition_sched_domains(0, NULL, NULL);
9192 rebuild_sched_domains();
9196 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
9198 unsigned int level = 0;
9200 if (sscanf(buf, "%u", &level) != 1)
9204 * level is always be positive so don't check for
9205 * level < POWERSAVINGS_BALANCE_NONE which is 0
9206 * What happens on 0 or 1 byte write,
9207 * need to check for count as well?
9210 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
9214 sched_smt_power_savings = level;
9216 sched_mc_power_savings = level;
9218 arch_reinit_sched_domains();
9223 #ifdef CONFIG_SCHED_MC
9224 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
9227 return sprintf(page, "%u\n", sched_mc_power_savings);
9229 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
9230 const char *buf, size_t count)
9232 return sched_power_savings_store(buf, count, 0);
9234 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
9235 sched_mc_power_savings_show,
9236 sched_mc_power_savings_store);
9239 #ifdef CONFIG_SCHED_SMT
9240 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
9243 return sprintf(page, "%u\n", sched_smt_power_savings);
9245 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
9246 const char *buf, size_t count)
9248 return sched_power_savings_store(buf, count, 1);
9250 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
9251 sched_smt_power_savings_show,
9252 sched_smt_power_savings_store);
9255 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
9259 #ifdef CONFIG_SCHED_SMT
9261 err = sysfs_create_file(&cls->kset.kobj,
9262 &attr_sched_smt_power_savings.attr);
9264 #ifdef CONFIG_SCHED_MC
9265 if (!err && mc_capable())
9266 err = sysfs_create_file(&cls->kset.kobj,
9267 &attr_sched_mc_power_savings.attr);
9271 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
9273 #ifndef CONFIG_CPUSETS
9275 * Add online and remove offline CPUs from the scheduler domains.
9276 * When cpusets are enabled they take over this function.
9278 static int update_sched_domains(struct notifier_block *nfb,
9279 unsigned long action, void *hcpu)
9283 case CPU_ONLINE_FROZEN:
9284 case CPU_DOWN_PREPARE:
9285 case CPU_DOWN_PREPARE_FROZEN:
9286 case CPU_DOWN_FAILED:
9287 case CPU_DOWN_FAILED_FROZEN:
9288 partition_sched_domains(1, NULL, NULL);
9297 static int update_runtime(struct notifier_block *nfb,
9298 unsigned long action, void *hcpu)
9300 int cpu = (int)(long)hcpu;
9303 case CPU_DOWN_PREPARE:
9304 case CPU_DOWN_PREPARE_FROZEN:
9305 disable_runtime(cpu_rq(cpu));
9308 case CPU_DOWN_FAILED:
9309 case CPU_DOWN_FAILED_FROZEN:
9311 case CPU_ONLINE_FROZEN:
9312 enable_runtime(cpu_rq(cpu));
9320 void __init sched_init_smp(void)
9322 cpumask_var_t non_isolated_cpus;
9324 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
9325 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
9327 #if defined(CONFIG_NUMA)
9328 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
9330 BUG_ON(sched_group_nodes_bycpu == NULL);
9333 mutex_lock(&sched_domains_mutex);
9334 arch_init_sched_domains(cpu_active_mask);
9335 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
9336 if (cpumask_empty(non_isolated_cpus))
9337 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
9338 mutex_unlock(&sched_domains_mutex);
9341 #ifndef CONFIG_CPUSETS
9342 /* XXX: Theoretical race here - CPU may be hotplugged now */
9343 hotcpu_notifier(update_sched_domains, 0);
9346 /* RT runtime code needs to handle some hotplug events */
9347 hotcpu_notifier(update_runtime, 0);
9351 /* Move init over to a non-isolated CPU */
9352 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
9354 sched_init_granularity();
9355 free_cpumask_var(non_isolated_cpus);
9357 init_sched_rt_class();
9360 void __init sched_init_smp(void)
9362 sched_init_granularity();
9364 #endif /* CONFIG_SMP */
9366 const_debug unsigned int sysctl_timer_migration = 1;
9368 int in_sched_functions(unsigned long addr)
9370 return in_lock_functions(addr) ||
9371 (addr >= (unsigned long)__sched_text_start
9372 && addr < (unsigned long)__sched_text_end);
9375 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
9377 cfs_rq->tasks_timeline = RB_ROOT;
9378 INIT_LIST_HEAD(&cfs_rq->tasks);
9379 #ifdef CONFIG_FAIR_GROUP_SCHED
9382 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9385 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
9387 struct rt_prio_array *array;
9390 array = &rt_rq->active;
9391 for (i = 0; i < MAX_RT_PRIO; i++) {
9392 INIT_LIST_HEAD(array->queue + i);
9393 __clear_bit(i, array->bitmap);
9395 /* delimiter for bitsearch: */
9396 __set_bit(MAX_RT_PRIO, array->bitmap);
9398 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9399 rt_rq->highest_prio.curr = MAX_RT_PRIO;
9401 rt_rq->highest_prio.next = MAX_RT_PRIO;
9405 rt_rq->rt_nr_migratory = 0;
9406 rt_rq->overloaded = 0;
9407 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
9411 rt_rq->rt_throttled = 0;
9412 rt_rq->rt_runtime = 0;
9413 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
9415 #ifdef CONFIG_RT_GROUP_SCHED
9416 rt_rq->rt_nr_boosted = 0;
9421 #ifdef CONFIG_FAIR_GROUP_SCHED
9422 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9423 struct sched_entity *se, int cpu, int add,
9424 struct sched_entity *parent)
9426 struct rq *rq = cpu_rq(cpu);
9427 tg->cfs_rq[cpu] = cfs_rq;
9428 init_cfs_rq(cfs_rq, rq);
9431 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
9434 /* se could be NULL for init_task_group */
9439 se->cfs_rq = &rq->cfs;
9441 se->cfs_rq = parent->my_q;
9444 se->load.weight = tg->shares;
9445 se->load.inv_weight = 0;
9446 se->parent = parent;
9450 #ifdef CONFIG_RT_GROUP_SCHED
9451 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
9452 struct sched_rt_entity *rt_se, int cpu, int add,
9453 struct sched_rt_entity *parent)
9455 struct rq *rq = cpu_rq(cpu);
9457 tg->rt_rq[cpu] = rt_rq;
9458 init_rt_rq(rt_rq, rq);
9460 rt_rq->rt_se = rt_se;
9461 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
9463 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
9465 tg->rt_se[cpu] = rt_se;
9470 rt_se->rt_rq = &rq->rt;
9472 rt_se->rt_rq = parent->my_q;
9474 rt_se->my_q = rt_rq;
9475 rt_se->parent = parent;
9476 INIT_LIST_HEAD(&rt_se->run_list);
9480 void __init sched_init(void)
9483 unsigned long alloc_size = 0, ptr;
9485 #ifdef CONFIG_FAIR_GROUP_SCHED
9486 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9488 #ifdef CONFIG_RT_GROUP_SCHED
9489 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9491 #ifdef CONFIG_USER_SCHED
9494 #ifdef CONFIG_CPUMASK_OFFSTACK
9495 alloc_size += num_possible_cpus() * cpumask_size();
9498 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
9500 #ifdef CONFIG_FAIR_GROUP_SCHED
9501 init_task_group.se = (struct sched_entity **)ptr;
9502 ptr += nr_cpu_ids * sizeof(void **);
9504 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
9505 ptr += nr_cpu_ids * sizeof(void **);
9507 #ifdef CONFIG_USER_SCHED
9508 root_task_group.se = (struct sched_entity **)ptr;
9509 ptr += nr_cpu_ids * sizeof(void **);
9511 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9512 ptr += nr_cpu_ids * sizeof(void **);
9513 #endif /* CONFIG_USER_SCHED */
9514 #endif /* CONFIG_FAIR_GROUP_SCHED */
9515 #ifdef CONFIG_RT_GROUP_SCHED
9516 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
9517 ptr += nr_cpu_ids * sizeof(void **);
9519 init_task_group.rt_rq = (struct rt_rq **)ptr;
9520 ptr += nr_cpu_ids * sizeof(void **);
9522 #ifdef CONFIG_USER_SCHED
9523 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9524 ptr += nr_cpu_ids * sizeof(void **);
9526 root_task_group.rt_rq = (struct rt_rq **)ptr;
9527 ptr += nr_cpu_ids * sizeof(void **);
9528 #endif /* CONFIG_USER_SCHED */
9529 #endif /* CONFIG_RT_GROUP_SCHED */
9530 #ifdef CONFIG_CPUMASK_OFFSTACK
9531 for_each_possible_cpu(i) {
9532 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
9533 ptr += cpumask_size();
9535 #endif /* CONFIG_CPUMASK_OFFSTACK */
9539 init_defrootdomain();
9542 init_rt_bandwidth(&def_rt_bandwidth,
9543 global_rt_period(), global_rt_runtime());
9545 #ifdef CONFIG_RT_GROUP_SCHED
9546 init_rt_bandwidth(&init_task_group.rt_bandwidth,
9547 global_rt_period(), global_rt_runtime());
9548 #ifdef CONFIG_USER_SCHED
9549 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9550 global_rt_period(), RUNTIME_INF);
9551 #endif /* CONFIG_USER_SCHED */
9552 #endif /* CONFIG_RT_GROUP_SCHED */
9554 #ifdef CONFIG_GROUP_SCHED
9555 list_add(&init_task_group.list, &task_groups);
9556 INIT_LIST_HEAD(&init_task_group.children);
9558 #ifdef CONFIG_USER_SCHED
9559 INIT_LIST_HEAD(&root_task_group.children);
9560 init_task_group.parent = &root_task_group;
9561 list_add(&init_task_group.siblings, &root_task_group.children);
9562 #endif /* CONFIG_USER_SCHED */
9563 #endif /* CONFIG_GROUP_SCHED */
9565 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
9566 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
9567 __alignof__(unsigned long));
9569 for_each_possible_cpu(i) {
9573 raw_spin_lock_init(&rq->lock);
9575 rq->calc_load_active = 0;
9576 rq->calc_load_update = jiffies + LOAD_FREQ;
9577 init_cfs_rq(&rq->cfs, rq);
9578 init_rt_rq(&rq->rt, rq);
9579 #ifdef CONFIG_FAIR_GROUP_SCHED
9580 init_task_group.shares = init_task_group_load;
9581 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9582 #ifdef CONFIG_CGROUP_SCHED
9584 * How much cpu bandwidth does init_task_group get?
9586 * In case of task-groups formed thr' the cgroup filesystem, it
9587 * gets 100% of the cpu resources in the system. This overall
9588 * system cpu resource is divided among the tasks of
9589 * init_task_group and its child task-groups in a fair manner,
9590 * based on each entity's (task or task-group's) weight
9591 * (se->load.weight).
9593 * In other words, if init_task_group has 10 tasks of weight
9594 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9595 * then A0's share of the cpu resource is:
9597 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9599 * We achieve this by letting init_task_group's tasks sit
9600 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9602 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
9603 #elif defined CONFIG_USER_SCHED
9604 root_task_group.shares = NICE_0_LOAD;
9605 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
9607 * In case of task-groups formed thr' the user id of tasks,
9608 * init_task_group represents tasks belonging to root user.
9609 * Hence it forms a sibling of all subsequent groups formed.
9610 * In this case, init_task_group gets only a fraction of overall
9611 * system cpu resource, based on the weight assigned to root
9612 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9613 * by letting tasks of init_task_group sit in a separate cfs_rq
9614 * (init_tg_cfs_rq) and having one entity represent this group of
9615 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9617 init_tg_cfs_entry(&init_task_group,
9618 &per_cpu(init_tg_cfs_rq, i),
9619 &per_cpu(init_sched_entity, i), i, 1,
9620 root_task_group.se[i]);
9623 #endif /* CONFIG_FAIR_GROUP_SCHED */
9625 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9626 #ifdef CONFIG_RT_GROUP_SCHED
9627 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
9628 #ifdef CONFIG_CGROUP_SCHED
9629 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
9630 #elif defined CONFIG_USER_SCHED
9631 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
9632 init_tg_rt_entry(&init_task_group,
9633 &per_cpu(init_rt_rq_var, i),
9634 &per_cpu(init_sched_rt_entity, i), i, 1,
9635 root_task_group.rt_se[i]);
9639 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
9640 rq->cpu_load[j] = 0;
9644 rq->post_schedule = 0;
9645 rq->active_balance = 0;
9646 rq->next_balance = jiffies;
9650 rq->migration_thread = NULL;
9652 rq->avg_idle = 2*sysctl_sched_migration_cost;
9653 INIT_LIST_HEAD(&rq->migration_queue);
9654 rq_attach_root(rq, &def_root_domain);
9657 atomic_set(&rq->nr_iowait, 0);
9660 set_load_weight(&init_task);
9662 #ifdef CONFIG_PREEMPT_NOTIFIERS
9663 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
9667 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9670 #ifdef CONFIG_RT_MUTEXES
9671 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
9675 * The boot idle thread does lazy MMU switching as well:
9677 atomic_inc(&init_mm.mm_count);
9678 enter_lazy_tlb(&init_mm, current);
9681 * Make us the idle thread. Technically, schedule() should not be
9682 * called from this thread, however somewhere below it might be,
9683 * but because we are the idle thread, we just pick up running again
9684 * when this runqueue becomes "idle".
9686 init_idle(current, smp_processor_id());
9688 calc_load_update = jiffies + LOAD_FREQ;
9691 * During early bootup we pretend to be a normal task:
9693 current->sched_class = &fair_sched_class;
9695 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9696 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
9699 zalloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
9700 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
9702 /* May be allocated at isolcpus cmdline parse time */
9703 if (cpu_isolated_map == NULL)
9704 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
9709 scheduler_running = 1;
9712 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9713 static inline int preempt_count_equals(int preempt_offset)
9715 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
9717 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
9720 void __might_sleep(char *file, int line, int preempt_offset)
9723 static unsigned long prev_jiffy; /* ratelimiting */
9725 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
9726 system_state != SYSTEM_RUNNING || oops_in_progress)
9728 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9730 prev_jiffy = jiffies;
9733 "BUG: sleeping function called from invalid context at %s:%d\n",
9736 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9737 in_atomic(), irqs_disabled(),
9738 current->pid, current->comm);
9740 debug_show_held_locks(current);
9741 if (irqs_disabled())
9742 print_irqtrace_events(current);
9746 EXPORT_SYMBOL(__might_sleep);
9749 #ifdef CONFIG_MAGIC_SYSRQ
9750 static void normalize_task(struct rq *rq, struct task_struct *p)
9754 update_rq_clock(rq);
9755 on_rq = p->se.on_rq;
9757 deactivate_task(rq, p, 0);
9758 __setscheduler(rq, p, SCHED_NORMAL, 0);
9760 activate_task(rq, p, 0);
9761 resched_task(rq->curr);
9765 void normalize_rt_tasks(void)
9767 struct task_struct *g, *p;
9768 unsigned long flags;
9771 read_lock_irqsave(&tasklist_lock, flags);
9772 do_each_thread(g, p) {
9774 * Only normalize user tasks:
9779 p->se.exec_start = 0;
9780 #ifdef CONFIG_SCHEDSTATS
9781 p->se.wait_start = 0;
9782 p->se.sleep_start = 0;
9783 p->se.block_start = 0;
9788 * Renice negative nice level userspace
9791 if (TASK_NICE(p) < 0 && p->mm)
9792 set_user_nice(p, 0);
9796 raw_spin_lock(&p->pi_lock);
9797 rq = __task_rq_lock(p);
9799 normalize_task(rq, p);
9801 __task_rq_unlock(rq);
9802 raw_spin_unlock(&p->pi_lock);
9803 } while_each_thread(g, p);
9805 read_unlock_irqrestore(&tasklist_lock, flags);
9808 #endif /* CONFIG_MAGIC_SYSRQ */
9812 * These functions are only useful for the IA64 MCA handling.
9814 * They can only be called when the whole system has been
9815 * stopped - every CPU needs to be quiescent, and no scheduling
9816 * activity can take place. Using them for anything else would
9817 * be a serious bug, and as a result, they aren't even visible
9818 * under any other configuration.
9822 * curr_task - return the current task for a given cpu.
9823 * @cpu: the processor in question.
9825 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9827 struct task_struct *curr_task(int cpu)
9829 return cpu_curr(cpu);
9833 * set_curr_task - set the current task for a given cpu.
9834 * @cpu: the processor in question.
9835 * @p: the task pointer to set.
9837 * Description: This function must only be used when non-maskable interrupts
9838 * are serviced on a separate stack. It allows the architecture to switch the
9839 * notion of the current task on a cpu in a non-blocking manner. This function
9840 * must be called with all CPU's synchronized, and interrupts disabled, the
9841 * and caller must save the original value of the current task (see
9842 * curr_task() above) and restore that value before reenabling interrupts and
9843 * re-starting the system.
9845 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9847 void set_curr_task(int cpu, struct task_struct *p)
9854 #ifdef CONFIG_FAIR_GROUP_SCHED
9855 static void free_fair_sched_group(struct task_group *tg)
9859 for_each_possible_cpu(i) {
9861 kfree(tg->cfs_rq[i]);
9871 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9873 struct cfs_rq *cfs_rq;
9874 struct sched_entity *se;
9878 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9881 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9885 tg->shares = NICE_0_LOAD;
9887 for_each_possible_cpu(i) {
9890 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9891 GFP_KERNEL, cpu_to_node(i));
9895 se = kzalloc_node(sizeof(struct sched_entity),
9896 GFP_KERNEL, cpu_to_node(i));
9900 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
9911 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9913 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
9914 &cpu_rq(cpu)->leaf_cfs_rq_list);
9917 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9919 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
9921 #else /* !CONFG_FAIR_GROUP_SCHED */
9922 static inline void free_fair_sched_group(struct task_group *tg)
9927 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9932 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9936 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9939 #endif /* CONFIG_FAIR_GROUP_SCHED */
9941 #ifdef CONFIG_RT_GROUP_SCHED
9942 static void free_rt_sched_group(struct task_group *tg)
9946 destroy_rt_bandwidth(&tg->rt_bandwidth);
9948 for_each_possible_cpu(i) {
9950 kfree(tg->rt_rq[i]);
9952 kfree(tg->rt_se[i]);
9960 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9962 struct rt_rq *rt_rq;
9963 struct sched_rt_entity *rt_se;
9967 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
9970 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
9974 init_rt_bandwidth(&tg->rt_bandwidth,
9975 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
9977 for_each_possible_cpu(i) {
9980 rt_rq = kzalloc_node(sizeof(struct rt_rq),
9981 GFP_KERNEL, cpu_to_node(i));
9985 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
9986 GFP_KERNEL, cpu_to_node(i));
9990 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
10001 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
10003 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
10004 &cpu_rq(cpu)->leaf_rt_rq_list);
10007 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
10009 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
10011 #else /* !CONFIG_RT_GROUP_SCHED */
10012 static inline void free_rt_sched_group(struct task_group *tg)
10017 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
10022 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
10026 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
10029 #endif /* CONFIG_RT_GROUP_SCHED */
10031 #ifdef CONFIG_GROUP_SCHED
10032 static void free_sched_group(struct task_group *tg)
10034 free_fair_sched_group(tg);
10035 free_rt_sched_group(tg);
10039 /* allocate runqueue etc for a new task group */
10040 struct task_group *sched_create_group(struct task_group *parent)
10042 struct task_group *tg;
10043 unsigned long flags;
10046 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
10048 return ERR_PTR(-ENOMEM);
10050 if (!alloc_fair_sched_group(tg, parent))
10053 if (!alloc_rt_sched_group(tg, parent))
10056 spin_lock_irqsave(&task_group_lock, flags);
10057 for_each_possible_cpu(i) {
10058 register_fair_sched_group(tg, i);
10059 register_rt_sched_group(tg, i);
10061 list_add_rcu(&tg->list, &task_groups);
10063 WARN_ON(!parent); /* root should already exist */
10065 tg->parent = parent;
10066 INIT_LIST_HEAD(&tg->children);
10067 list_add_rcu(&tg->siblings, &parent->children);
10068 spin_unlock_irqrestore(&task_group_lock, flags);
10073 free_sched_group(tg);
10074 return ERR_PTR(-ENOMEM);
10077 /* rcu callback to free various structures associated with a task group */
10078 static void free_sched_group_rcu(struct rcu_head *rhp)
10080 /* now it should be safe to free those cfs_rqs */
10081 free_sched_group(container_of(rhp, struct task_group, rcu));
10084 /* Destroy runqueue etc associated with a task group */
10085 void sched_destroy_group(struct task_group *tg)
10087 unsigned long flags;
10090 spin_lock_irqsave(&task_group_lock, flags);
10091 for_each_possible_cpu(i) {
10092 unregister_fair_sched_group(tg, i);
10093 unregister_rt_sched_group(tg, i);
10095 list_del_rcu(&tg->list);
10096 list_del_rcu(&tg->siblings);
10097 spin_unlock_irqrestore(&task_group_lock, flags);
10099 /* wait for possible concurrent references to cfs_rqs complete */
10100 call_rcu(&tg->rcu, free_sched_group_rcu);
10103 /* change task's runqueue when it moves between groups.
10104 * The caller of this function should have put the task in its new group
10105 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
10106 * reflect its new group.
10108 void sched_move_task(struct task_struct *tsk)
10110 int on_rq, running;
10111 unsigned long flags;
10114 rq = task_rq_lock(tsk, &flags);
10116 update_rq_clock(rq);
10118 running = task_current(rq, tsk);
10119 on_rq = tsk->se.on_rq;
10122 dequeue_task(rq, tsk, 0);
10123 if (unlikely(running))
10124 tsk->sched_class->put_prev_task(rq, tsk);
10126 set_task_rq(tsk, task_cpu(tsk));
10128 #ifdef CONFIG_FAIR_GROUP_SCHED
10129 if (tsk->sched_class->moved_group)
10130 tsk->sched_class->moved_group(tsk, on_rq);
10133 if (unlikely(running))
10134 tsk->sched_class->set_curr_task(rq);
10136 enqueue_task(rq, tsk, 0);
10138 task_rq_unlock(rq, &flags);
10140 #endif /* CONFIG_GROUP_SCHED */
10142 #ifdef CONFIG_FAIR_GROUP_SCHED
10143 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
10145 struct cfs_rq *cfs_rq = se->cfs_rq;
10150 dequeue_entity(cfs_rq, se, 0);
10152 se->load.weight = shares;
10153 se->load.inv_weight = 0;
10156 enqueue_entity(cfs_rq, se, 0);
10159 static void set_se_shares(struct sched_entity *se, unsigned long shares)
10161 struct cfs_rq *cfs_rq = se->cfs_rq;
10162 struct rq *rq = cfs_rq->rq;
10163 unsigned long flags;
10165 raw_spin_lock_irqsave(&rq->lock, flags);
10166 __set_se_shares(se, shares);
10167 raw_spin_unlock_irqrestore(&rq->lock, flags);
10170 static DEFINE_MUTEX(shares_mutex);
10172 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10175 unsigned long flags;
10178 * We can't change the weight of the root cgroup.
10183 if (shares < MIN_SHARES)
10184 shares = MIN_SHARES;
10185 else if (shares > MAX_SHARES)
10186 shares = MAX_SHARES;
10188 mutex_lock(&shares_mutex);
10189 if (tg->shares == shares)
10192 spin_lock_irqsave(&task_group_lock, flags);
10193 for_each_possible_cpu(i)
10194 unregister_fair_sched_group(tg, i);
10195 list_del_rcu(&tg->siblings);
10196 spin_unlock_irqrestore(&task_group_lock, flags);
10198 /* wait for any ongoing reference to this group to finish */
10199 synchronize_sched();
10202 * Now we are free to modify the group's share on each cpu
10203 * w/o tripping rebalance_share or load_balance_fair.
10205 tg->shares = shares;
10206 for_each_possible_cpu(i) {
10208 * force a rebalance
10210 cfs_rq_set_shares(tg->cfs_rq[i], 0);
10211 set_se_shares(tg->se[i], shares);
10215 * Enable load balance activity on this group, by inserting it back on
10216 * each cpu's rq->leaf_cfs_rq_list.
10218 spin_lock_irqsave(&task_group_lock, flags);
10219 for_each_possible_cpu(i)
10220 register_fair_sched_group(tg, i);
10221 list_add_rcu(&tg->siblings, &tg->parent->children);
10222 spin_unlock_irqrestore(&task_group_lock, flags);
10224 mutex_unlock(&shares_mutex);
10228 unsigned long sched_group_shares(struct task_group *tg)
10234 #ifdef CONFIG_RT_GROUP_SCHED
10236 * Ensure that the real time constraints are schedulable.
10238 static DEFINE_MUTEX(rt_constraints_mutex);
10240 static unsigned long to_ratio(u64 period, u64 runtime)
10242 if (runtime == RUNTIME_INF)
10245 return div64_u64(runtime << 20, period);
10248 /* Must be called with tasklist_lock held */
10249 static inline int tg_has_rt_tasks(struct task_group *tg)
10251 struct task_struct *g, *p;
10253 do_each_thread(g, p) {
10254 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
10256 } while_each_thread(g, p);
10261 struct rt_schedulable_data {
10262 struct task_group *tg;
10267 static int tg_schedulable(struct task_group *tg, void *data)
10269 struct rt_schedulable_data *d = data;
10270 struct task_group *child;
10271 unsigned long total, sum = 0;
10272 u64 period, runtime;
10274 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10275 runtime = tg->rt_bandwidth.rt_runtime;
10278 period = d->rt_period;
10279 runtime = d->rt_runtime;
10282 #ifdef CONFIG_USER_SCHED
10283 if (tg == &root_task_group) {
10284 period = global_rt_period();
10285 runtime = global_rt_runtime();
10290 * Cannot have more runtime than the period.
10292 if (runtime > period && runtime != RUNTIME_INF)
10296 * Ensure we don't starve existing RT tasks.
10298 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
10301 total = to_ratio(period, runtime);
10304 * Nobody can have more than the global setting allows.
10306 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
10310 * The sum of our children's runtime should not exceed our own.
10312 list_for_each_entry_rcu(child, &tg->children, siblings) {
10313 period = ktime_to_ns(child->rt_bandwidth.rt_period);
10314 runtime = child->rt_bandwidth.rt_runtime;
10316 if (child == d->tg) {
10317 period = d->rt_period;
10318 runtime = d->rt_runtime;
10321 sum += to_ratio(period, runtime);
10330 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
10332 struct rt_schedulable_data data = {
10334 .rt_period = period,
10335 .rt_runtime = runtime,
10338 return walk_tg_tree(tg_schedulable, tg_nop, &data);
10341 static int tg_set_bandwidth(struct task_group *tg,
10342 u64 rt_period, u64 rt_runtime)
10346 mutex_lock(&rt_constraints_mutex);
10347 read_lock(&tasklist_lock);
10348 err = __rt_schedulable(tg, rt_period, rt_runtime);
10352 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10353 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
10354 tg->rt_bandwidth.rt_runtime = rt_runtime;
10356 for_each_possible_cpu(i) {
10357 struct rt_rq *rt_rq = tg->rt_rq[i];
10359 raw_spin_lock(&rt_rq->rt_runtime_lock);
10360 rt_rq->rt_runtime = rt_runtime;
10361 raw_spin_unlock(&rt_rq->rt_runtime_lock);
10363 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10365 read_unlock(&tasklist_lock);
10366 mutex_unlock(&rt_constraints_mutex);
10371 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
10373 u64 rt_runtime, rt_period;
10375 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10376 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
10377 if (rt_runtime_us < 0)
10378 rt_runtime = RUNTIME_INF;
10380 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10383 long sched_group_rt_runtime(struct task_group *tg)
10387 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
10390 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
10391 do_div(rt_runtime_us, NSEC_PER_USEC);
10392 return rt_runtime_us;
10395 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
10397 u64 rt_runtime, rt_period;
10399 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
10400 rt_runtime = tg->rt_bandwidth.rt_runtime;
10402 if (rt_period == 0)
10405 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10408 long sched_group_rt_period(struct task_group *tg)
10412 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
10413 do_div(rt_period_us, NSEC_PER_USEC);
10414 return rt_period_us;
10417 static int sched_rt_global_constraints(void)
10419 u64 runtime, period;
10422 if (sysctl_sched_rt_period <= 0)
10425 runtime = global_rt_runtime();
10426 period = global_rt_period();
10429 * Sanity check on the sysctl variables.
10431 if (runtime > period && runtime != RUNTIME_INF)
10434 mutex_lock(&rt_constraints_mutex);
10435 read_lock(&tasklist_lock);
10436 ret = __rt_schedulable(NULL, 0, 0);
10437 read_unlock(&tasklist_lock);
10438 mutex_unlock(&rt_constraints_mutex);
10443 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
10445 /* Don't accept realtime tasks when there is no way for them to run */
10446 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
10452 #else /* !CONFIG_RT_GROUP_SCHED */
10453 static int sched_rt_global_constraints(void)
10455 unsigned long flags;
10458 if (sysctl_sched_rt_period <= 0)
10462 * There's always some RT tasks in the root group
10463 * -- migration, kstopmachine etc..
10465 if (sysctl_sched_rt_runtime == 0)
10468 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
10469 for_each_possible_cpu(i) {
10470 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
10472 raw_spin_lock(&rt_rq->rt_runtime_lock);
10473 rt_rq->rt_runtime = global_rt_runtime();
10474 raw_spin_unlock(&rt_rq->rt_runtime_lock);
10476 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
10480 #endif /* CONFIG_RT_GROUP_SCHED */
10482 int sched_rt_handler(struct ctl_table *table, int write,
10483 void __user *buffer, size_t *lenp,
10487 int old_period, old_runtime;
10488 static DEFINE_MUTEX(mutex);
10490 mutex_lock(&mutex);
10491 old_period = sysctl_sched_rt_period;
10492 old_runtime = sysctl_sched_rt_runtime;
10494 ret = proc_dointvec(table, write, buffer, lenp, ppos);
10496 if (!ret && write) {
10497 ret = sched_rt_global_constraints();
10499 sysctl_sched_rt_period = old_period;
10500 sysctl_sched_rt_runtime = old_runtime;
10502 def_rt_bandwidth.rt_runtime = global_rt_runtime();
10503 def_rt_bandwidth.rt_period =
10504 ns_to_ktime(global_rt_period());
10507 mutex_unlock(&mutex);
10512 #ifdef CONFIG_CGROUP_SCHED
10514 /* return corresponding task_group object of a cgroup */
10515 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
10517 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
10518 struct task_group, css);
10521 static struct cgroup_subsys_state *
10522 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
10524 struct task_group *tg, *parent;
10526 if (!cgrp->parent) {
10527 /* This is early initialization for the top cgroup */
10528 return &init_task_group.css;
10531 parent = cgroup_tg(cgrp->parent);
10532 tg = sched_create_group(parent);
10534 return ERR_PTR(-ENOMEM);
10540 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10542 struct task_group *tg = cgroup_tg(cgrp);
10544 sched_destroy_group(tg);
10548 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
10550 #ifdef CONFIG_RT_GROUP_SCHED
10551 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
10554 /* We don't support RT-tasks being in separate groups */
10555 if (tsk->sched_class != &fair_sched_class)
10562 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10563 struct task_struct *tsk, bool threadgroup)
10565 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
10569 struct task_struct *c;
10571 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
10572 retval = cpu_cgroup_can_attach_task(cgrp, c);
10584 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10585 struct cgroup *old_cont, struct task_struct *tsk,
10588 sched_move_task(tsk);
10590 struct task_struct *c;
10592 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
10593 sched_move_task(c);
10599 #ifdef CONFIG_FAIR_GROUP_SCHED
10600 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
10603 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
10606 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
10608 struct task_group *tg = cgroup_tg(cgrp);
10610 return (u64) tg->shares;
10612 #endif /* CONFIG_FAIR_GROUP_SCHED */
10614 #ifdef CONFIG_RT_GROUP_SCHED
10615 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
10618 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
10621 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
10623 return sched_group_rt_runtime(cgroup_tg(cgrp));
10626 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
10629 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
10632 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
10634 return sched_group_rt_period(cgroup_tg(cgrp));
10636 #endif /* CONFIG_RT_GROUP_SCHED */
10638 static struct cftype cpu_files[] = {
10639 #ifdef CONFIG_FAIR_GROUP_SCHED
10642 .read_u64 = cpu_shares_read_u64,
10643 .write_u64 = cpu_shares_write_u64,
10646 #ifdef CONFIG_RT_GROUP_SCHED
10648 .name = "rt_runtime_us",
10649 .read_s64 = cpu_rt_runtime_read,
10650 .write_s64 = cpu_rt_runtime_write,
10653 .name = "rt_period_us",
10654 .read_u64 = cpu_rt_period_read_uint,
10655 .write_u64 = cpu_rt_period_write_uint,
10660 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
10662 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
10665 struct cgroup_subsys cpu_cgroup_subsys = {
10667 .create = cpu_cgroup_create,
10668 .destroy = cpu_cgroup_destroy,
10669 .can_attach = cpu_cgroup_can_attach,
10670 .attach = cpu_cgroup_attach,
10671 .populate = cpu_cgroup_populate,
10672 .subsys_id = cpu_cgroup_subsys_id,
10676 #endif /* CONFIG_CGROUP_SCHED */
10678 #ifdef CONFIG_CGROUP_CPUACCT
10681 * CPU accounting code for task groups.
10683 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10684 * (balbir@in.ibm.com).
10687 /* track cpu usage of a group of tasks and its child groups */
10689 struct cgroup_subsys_state css;
10690 /* cpuusage holds pointer to a u64-type object on every cpu */
10692 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
10693 struct cpuacct *parent;
10696 struct cgroup_subsys cpuacct_subsys;
10698 /* return cpu accounting group corresponding to this container */
10699 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
10701 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
10702 struct cpuacct, css);
10705 /* return cpu accounting group to which this task belongs */
10706 static inline struct cpuacct *task_ca(struct task_struct *tsk)
10708 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
10709 struct cpuacct, css);
10712 /* create a new cpu accounting group */
10713 static struct cgroup_subsys_state *cpuacct_create(
10714 struct cgroup_subsys *ss, struct cgroup *cgrp)
10716 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
10722 ca->cpuusage = alloc_percpu(u64);
10726 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10727 if (percpu_counter_init(&ca->cpustat[i], 0))
10728 goto out_free_counters;
10731 ca->parent = cgroup_ca(cgrp->parent);
10737 percpu_counter_destroy(&ca->cpustat[i]);
10738 free_percpu(ca->cpuusage);
10742 return ERR_PTR(-ENOMEM);
10745 /* destroy an existing cpu accounting group */
10747 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10749 struct cpuacct *ca = cgroup_ca(cgrp);
10752 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10753 percpu_counter_destroy(&ca->cpustat[i]);
10754 free_percpu(ca->cpuusage);
10758 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
10760 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10763 #ifndef CONFIG_64BIT
10765 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10767 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
10769 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
10777 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
10779 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10781 #ifndef CONFIG_64BIT
10783 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10785 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
10787 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
10793 /* return total cpu usage (in nanoseconds) of a group */
10794 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
10796 struct cpuacct *ca = cgroup_ca(cgrp);
10797 u64 totalcpuusage = 0;
10800 for_each_present_cpu(i)
10801 totalcpuusage += cpuacct_cpuusage_read(ca, i);
10803 return totalcpuusage;
10806 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
10809 struct cpuacct *ca = cgroup_ca(cgrp);
10818 for_each_present_cpu(i)
10819 cpuacct_cpuusage_write(ca, i, 0);
10825 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
10826 struct seq_file *m)
10828 struct cpuacct *ca = cgroup_ca(cgroup);
10832 for_each_present_cpu(i) {
10833 percpu = cpuacct_cpuusage_read(ca, i);
10834 seq_printf(m, "%llu ", (unsigned long long) percpu);
10836 seq_printf(m, "\n");
10840 static const char *cpuacct_stat_desc[] = {
10841 [CPUACCT_STAT_USER] = "user",
10842 [CPUACCT_STAT_SYSTEM] = "system",
10845 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
10846 struct cgroup_map_cb *cb)
10848 struct cpuacct *ca = cgroup_ca(cgrp);
10851 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
10852 s64 val = percpu_counter_read(&ca->cpustat[i]);
10853 val = cputime64_to_clock_t(val);
10854 cb->fill(cb, cpuacct_stat_desc[i], val);
10859 static struct cftype files[] = {
10862 .read_u64 = cpuusage_read,
10863 .write_u64 = cpuusage_write,
10866 .name = "usage_percpu",
10867 .read_seq_string = cpuacct_percpu_seq_read,
10871 .read_map = cpuacct_stats_show,
10875 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
10877 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
10881 * charge this task's execution time to its accounting group.
10883 * called with rq->lock held.
10885 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
10887 struct cpuacct *ca;
10890 if (unlikely(!cpuacct_subsys.active))
10893 cpu = task_cpu(tsk);
10899 for (; ca; ca = ca->parent) {
10900 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10901 *cpuusage += cputime;
10908 * Charge the system/user time to the task's accounting group.
10910 static void cpuacct_update_stats(struct task_struct *tsk,
10911 enum cpuacct_stat_index idx, cputime_t val)
10913 struct cpuacct *ca;
10915 if (unlikely(!cpuacct_subsys.active))
10922 percpu_counter_add(&ca->cpustat[idx], val);
10928 struct cgroup_subsys cpuacct_subsys = {
10930 .create = cpuacct_create,
10931 .destroy = cpuacct_destroy,
10932 .populate = cpuacct_populate,
10933 .subsys_id = cpuacct_subsys_id,
10935 #endif /* CONFIG_CGROUP_CPUACCT */
10939 int rcu_expedited_torture_stats(char *page)
10943 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
10945 void synchronize_sched_expedited(void)
10948 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
10950 #else /* #ifndef CONFIG_SMP */
10952 static DEFINE_PER_CPU(struct migration_req, rcu_migration_req);
10953 static DEFINE_MUTEX(rcu_sched_expedited_mutex);
10955 #define RCU_EXPEDITED_STATE_POST -2
10956 #define RCU_EXPEDITED_STATE_IDLE -1
10958 static int rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
10960 int rcu_expedited_torture_stats(char *page)
10965 cnt += sprintf(&page[cnt], "state: %d /", rcu_expedited_state);
10966 for_each_online_cpu(cpu) {
10967 cnt += sprintf(&page[cnt], " %d:%d",
10968 cpu, per_cpu(rcu_migration_req, cpu).dest_cpu);
10970 cnt += sprintf(&page[cnt], "\n");
10973 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
10975 static long synchronize_sched_expedited_count;
10978 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
10979 * approach to force grace period to end quickly. This consumes
10980 * significant time on all CPUs, and is thus not recommended for
10981 * any sort of common-case code.
10983 * Note that it is illegal to call this function while holding any
10984 * lock that is acquired by a CPU-hotplug notifier. Failing to
10985 * observe this restriction will result in deadlock.
10987 void synchronize_sched_expedited(void)
10990 unsigned long flags;
10991 bool need_full_sync = 0;
10993 struct migration_req *req;
10997 smp_mb(); /* ensure prior mod happens before capturing snap. */
10998 snap = ACCESS_ONCE(synchronize_sched_expedited_count) + 1;
11000 while (!mutex_trylock(&rcu_sched_expedited_mutex)) {
11002 if (trycount++ < 10)
11003 udelay(trycount * num_online_cpus());
11005 synchronize_sched();
11008 if (ACCESS_ONCE(synchronize_sched_expedited_count) - snap > 0) {
11009 smp_mb(); /* ensure test happens before caller kfree */
11014 rcu_expedited_state = RCU_EXPEDITED_STATE_POST;
11015 for_each_online_cpu(cpu) {
11017 req = &per_cpu(rcu_migration_req, cpu);
11018 init_completion(&req->done);
11020 req->dest_cpu = RCU_MIGRATION_NEED_QS;
11021 raw_spin_lock_irqsave(&rq->lock, flags);
11022 list_add(&req->list, &rq->migration_queue);
11023 raw_spin_unlock_irqrestore(&rq->lock, flags);
11024 wake_up_process(rq->migration_thread);
11026 for_each_online_cpu(cpu) {
11027 rcu_expedited_state = cpu;
11028 req = &per_cpu(rcu_migration_req, cpu);
11030 wait_for_completion(&req->done);
11031 raw_spin_lock_irqsave(&rq->lock, flags);
11032 if (unlikely(req->dest_cpu == RCU_MIGRATION_MUST_SYNC))
11033 need_full_sync = 1;
11034 req->dest_cpu = RCU_MIGRATION_IDLE;
11035 raw_spin_unlock_irqrestore(&rq->lock, flags);
11037 rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
11038 synchronize_sched_expedited_count++;
11039 mutex_unlock(&rcu_sched_expedited_mutex);
11041 if (need_full_sync)
11042 synchronize_sched();
11044 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
11046 #endif /* #else #ifndef CONFIG_SMP */