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/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/kthread.h>
58 #include <linux/proc_fs.h>
59 #include <linux/seq_file.h>
60 #include <linux/sysctl.h>
61 #include <linux/syscalls.h>
62 #include <linux/times.h>
63 #include <linux/tsacct_kern.h>
64 #include <linux/kprobes.h>
65 #include <linux/delayacct.h>
66 #include <linux/reciprocal_div.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/bootmem.h>
72 #include <linux/debugfs.h>
73 #include <linux/ctype.h>
74 #include <linux/ftrace.h>
77 #include <asm/irq_regs.h>
79 #include "sched_cpupri.h"
81 #define CREATE_TRACE_POINTS
82 #include <trace/events/sched.h>
85 * Convert user-nice values [ -20 ... 0 ... 19 ]
86 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
89 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
90 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
91 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
94 * 'User priority' is the nice value converted to something we
95 * can work with better when scaling various scheduler parameters,
96 * it's a [ 0 ... 39 ] range.
98 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
99 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
100 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
103 * Helpers for converting nanosecond timing to jiffy resolution
105 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
107 #define NICE_0_LOAD SCHED_LOAD_SCALE
108 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
111 * These are the 'tuning knobs' of the scheduler:
113 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
114 * Timeslices get refilled after they expire.
116 #define DEF_TIMESLICE (100 * HZ / 1000)
119 * single value that denotes runtime == period, ie unlimited time.
121 #define RUNTIME_INF ((u64)~0ULL)
125 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
128 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
129 * Since cpu_power is a 'constant', we can use a reciprocal divide.
131 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
133 return reciprocal_divide(load, sg->reciprocal_cpu_power);
137 * Each time a sched group cpu_power is changed,
138 * we must compute its reciprocal value
140 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
142 sg->__cpu_power += val;
143 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
147 static inline int rt_policy(int policy)
149 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
154 static inline int task_has_rt_policy(struct task_struct *p)
156 return rt_policy(p->policy);
160 * This is the priority-queue data structure of the RT scheduling class:
162 struct rt_prio_array {
163 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
164 struct list_head queue[MAX_RT_PRIO];
167 struct rt_bandwidth {
168 /* nests inside the rq lock: */
169 spinlock_t rt_runtime_lock;
172 struct hrtimer rt_period_timer;
175 static struct rt_bandwidth def_rt_bandwidth;
177 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
179 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
181 struct rt_bandwidth *rt_b =
182 container_of(timer, struct rt_bandwidth, rt_period_timer);
188 now = hrtimer_cb_get_time(timer);
189 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
194 idle = do_sched_rt_period_timer(rt_b, overrun);
197 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
201 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
203 rt_b->rt_period = ns_to_ktime(period);
204 rt_b->rt_runtime = runtime;
206 spin_lock_init(&rt_b->rt_runtime_lock);
208 hrtimer_init(&rt_b->rt_period_timer,
209 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
210 rt_b->rt_period_timer.function = sched_rt_period_timer;
213 static inline int rt_bandwidth_enabled(void)
215 return sysctl_sched_rt_runtime >= 0;
218 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
222 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
225 if (hrtimer_active(&rt_b->rt_period_timer))
228 spin_lock(&rt_b->rt_runtime_lock);
233 if (hrtimer_active(&rt_b->rt_period_timer))
236 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
237 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
239 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
240 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
241 delta = ktime_to_ns(ktime_sub(hard, soft));
242 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
243 HRTIMER_MODE_ABS, 0);
245 spin_unlock(&rt_b->rt_runtime_lock);
248 #ifdef CONFIG_RT_GROUP_SCHED
249 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
251 hrtimer_cancel(&rt_b->rt_period_timer);
256 * sched_domains_mutex serializes calls to arch_init_sched_domains,
257 * detach_destroy_domains and partition_sched_domains.
259 static DEFINE_MUTEX(sched_domains_mutex);
261 #ifdef CONFIG_GROUP_SCHED
263 #include <linux/cgroup.h>
267 static LIST_HEAD(task_groups);
269 /* task group related information */
271 #ifdef CONFIG_CGROUP_SCHED
272 struct cgroup_subsys_state css;
275 #ifdef CONFIG_USER_SCHED
279 #ifdef CONFIG_FAIR_GROUP_SCHED
280 /* schedulable entities of this group on each cpu */
281 struct sched_entity **se;
282 /* runqueue "owned" by this group on each cpu */
283 struct cfs_rq **cfs_rq;
284 unsigned long shares;
287 #ifdef CONFIG_RT_GROUP_SCHED
288 struct sched_rt_entity **rt_se;
289 struct rt_rq **rt_rq;
291 struct rt_bandwidth rt_bandwidth;
295 struct list_head list;
297 struct task_group *parent;
298 struct list_head siblings;
299 struct list_head children;
302 #ifdef CONFIG_USER_SCHED
304 /* Helper function to pass uid information to create_sched_user() */
305 void set_tg_uid(struct user_struct *user)
307 user->tg->uid = user->uid;
312 * Every UID task group (including init_task_group aka UID-0) will
313 * be a child to this group.
315 struct task_group root_task_group;
317 #ifdef CONFIG_FAIR_GROUP_SCHED
318 /* Default task group's sched entity on each cpu */
319 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
320 /* Default task group's cfs_rq on each cpu */
321 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
322 #endif /* CONFIG_FAIR_GROUP_SCHED */
324 #ifdef CONFIG_RT_GROUP_SCHED
325 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
326 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
327 #endif /* CONFIG_RT_GROUP_SCHED */
328 #else /* !CONFIG_USER_SCHED */
329 #define root_task_group init_task_group
330 #endif /* CONFIG_USER_SCHED */
332 /* task_group_lock serializes add/remove of task groups and also changes to
333 * a task group's cpu shares.
335 static DEFINE_SPINLOCK(task_group_lock);
338 static int root_task_group_empty(void)
340 return list_empty(&root_task_group.children);
344 #ifdef CONFIG_FAIR_GROUP_SCHED
345 #ifdef CONFIG_USER_SCHED
346 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
347 #else /* !CONFIG_USER_SCHED */
348 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
349 #endif /* CONFIG_USER_SCHED */
352 * A weight of 0 or 1 can cause arithmetics problems.
353 * A weight of a cfs_rq is the sum of weights of which entities
354 * are queued on this cfs_rq, so a weight of a entity should not be
355 * too large, so as the shares value of a task group.
356 * (The default weight is 1024 - so there's no practical
357 * limitation from this.)
360 #define MAX_SHARES (1UL << 18)
362 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
365 /* Default task group.
366 * Every task in system belong to this group at bootup.
368 struct task_group init_task_group;
370 /* return group to which a task belongs */
371 static inline struct task_group *task_group(struct task_struct *p)
373 struct task_group *tg;
375 #ifdef CONFIG_USER_SCHED
377 tg = __task_cred(p)->user->tg;
379 #elif defined(CONFIG_CGROUP_SCHED)
380 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
381 struct task_group, css);
383 tg = &init_task_group;
388 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
389 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
391 #ifdef CONFIG_FAIR_GROUP_SCHED
392 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
393 p->se.parent = task_group(p)->se[cpu];
396 #ifdef CONFIG_RT_GROUP_SCHED
397 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
398 p->rt.parent = task_group(p)->rt_se[cpu];
405 static int root_task_group_empty(void)
411 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
412 static inline struct task_group *task_group(struct task_struct *p)
417 #endif /* CONFIG_GROUP_SCHED */
419 /* CFS-related fields in a runqueue */
421 struct load_weight load;
422 unsigned long nr_running;
427 struct rb_root tasks_timeline;
428 struct rb_node *rb_leftmost;
430 struct list_head tasks;
431 struct list_head *balance_iterator;
434 * 'curr' points to currently running entity on this cfs_rq.
435 * It is set to NULL otherwise (i.e when none are currently running).
437 struct sched_entity *curr, *next, *last;
439 unsigned int nr_spread_over;
441 #ifdef CONFIG_FAIR_GROUP_SCHED
442 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
445 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
446 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
447 * (like users, containers etc.)
449 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
450 * list is used during load balance.
452 struct list_head leaf_cfs_rq_list;
453 struct task_group *tg; /* group that "owns" this runqueue */
457 * the part of load.weight contributed by tasks
459 unsigned long task_weight;
462 * h_load = weight * f(tg)
464 * Where f(tg) is the recursive weight fraction assigned to
467 unsigned long h_load;
470 * this cpu's part of tg->shares
472 unsigned long shares;
475 * load.weight at the time we set shares
477 unsigned long rq_weight;
482 /* Real-Time classes' related field in a runqueue: */
484 struct rt_prio_array active;
485 unsigned long rt_nr_running;
486 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
488 int curr; /* highest queued rt task prio */
490 int next; /* next highest */
495 unsigned long rt_nr_migratory;
497 struct plist_head pushable_tasks;
502 /* Nests inside the rq lock: */
503 spinlock_t rt_runtime_lock;
505 #ifdef CONFIG_RT_GROUP_SCHED
506 unsigned long rt_nr_boosted;
509 struct list_head leaf_rt_rq_list;
510 struct task_group *tg;
511 struct sched_rt_entity *rt_se;
518 * We add the notion of a root-domain which will be used to define per-domain
519 * variables. Each exclusive cpuset essentially defines an island domain by
520 * fully partitioning the member cpus from any other cpuset. Whenever a new
521 * exclusive cpuset is created, we also create and attach a new root-domain
528 cpumask_var_t online;
531 * The "RT overload" flag: it gets set if a CPU has more than
532 * one runnable RT task.
534 cpumask_var_t rto_mask;
537 struct cpupri cpupri;
539 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
541 * Preferred wake up cpu nominated by sched_mc balance that will be
542 * used when most cpus are idle in the system indicating overall very
543 * low system utilisation. Triggered at POWERSAVINGS_BALANCE_WAKEUP(2)
545 unsigned int sched_mc_preferred_wakeup_cpu;
550 * By default the system creates a single root-domain with all cpus as
551 * members (mimicking the global state we have today).
553 static struct root_domain def_root_domain;
558 * This is the main, per-CPU runqueue data structure.
560 * Locking rule: those places that want to lock multiple runqueues
561 * (such as the load balancing or the thread migration code), lock
562 * acquire operations must be ordered by ascending &runqueue.
569 * nr_running and cpu_load should be in the same cacheline because
570 * remote CPUs use both these fields when doing load calculation.
572 unsigned long nr_running;
573 #define CPU_LOAD_IDX_MAX 5
574 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
576 unsigned long last_tick_seen;
577 unsigned char in_nohz_recently;
579 /* capture load from *all* tasks on this cpu: */
580 struct load_weight load;
581 unsigned long nr_load_updates;
587 #ifdef CONFIG_FAIR_GROUP_SCHED
588 /* list of leaf cfs_rq on this cpu: */
589 struct list_head leaf_cfs_rq_list;
591 #ifdef CONFIG_RT_GROUP_SCHED
592 struct list_head leaf_rt_rq_list;
596 * This is part of a global counter where only the total sum
597 * over all CPUs matters. A task can increase this counter on
598 * one CPU and if it got migrated afterwards it may decrease
599 * it on another CPU. Always updated under the runqueue lock:
601 unsigned long nr_uninterruptible;
603 struct task_struct *curr, *idle;
604 unsigned long next_balance;
605 struct mm_struct *prev_mm;
612 struct root_domain *rd;
613 struct sched_domain *sd;
615 unsigned char idle_at_tick;
616 /* For active balancing */
619 /* cpu of this runqueue: */
623 unsigned long avg_load_per_task;
625 struct task_struct *migration_thread;
626 struct list_head migration_queue;
629 #ifdef CONFIG_SCHED_HRTICK
631 int hrtick_csd_pending;
632 struct call_single_data hrtick_csd;
634 struct hrtimer hrtick_timer;
637 #ifdef CONFIG_SCHEDSTATS
639 struct sched_info rq_sched_info;
640 unsigned long long rq_cpu_time;
641 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
643 /* sys_sched_yield() stats */
644 unsigned int yld_count;
646 /* schedule() stats */
647 unsigned int sched_switch;
648 unsigned int sched_count;
649 unsigned int sched_goidle;
651 /* try_to_wake_up() stats */
652 unsigned int ttwu_count;
653 unsigned int ttwu_local;
656 unsigned int bkl_count;
660 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
662 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p, int sync)
664 rq->curr->sched_class->check_preempt_curr(rq, p, sync);
667 static inline int cpu_of(struct rq *rq)
677 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
678 * See detach_destroy_domains: synchronize_sched for details.
680 * The domain tree of any CPU may only be accessed from within
681 * preempt-disabled sections.
683 #define for_each_domain(cpu, __sd) \
684 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
686 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
687 #define this_rq() (&__get_cpu_var(runqueues))
688 #define task_rq(p) cpu_rq(task_cpu(p))
689 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
691 static inline void update_rq_clock(struct rq *rq)
693 rq->clock = sched_clock_cpu(cpu_of(rq));
697 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
699 #ifdef CONFIG_SCHED_DEBUG
700 # define const_debug __read_mostly
702 # define const_debug static const
708 * Returns true if the current cpu runqueue is locked.
709 * This interface allows printk to be called with the runqueue lock
710 * held and know whether or not it is OK to wake up the klogd.
712 int runqueue_is_locked(void)
715 struct rq *rq = cpu_rq(cpu);
718 ret = spin_is_locked(&rq->lock);
724 * Debugging: various feature bits
727 #define SCHED_FEAT(name, enabled) \
728 __SCHED_FEAT_##name ,
731 #include "sched_features.h"
736 #define SCHED_FEAT(name, enabled) \
737 (1UL << __SCHED_FEAT_##name) * enabled |
739 const_debug unsigned int sysctl_sched_features =
740 #include "sched_features.h"
745 #ifdef CONFIG_SCHED_DEBUG
746 #define SCHED_FEAT(name, enabled) \
749 static __read_mostly char *sched_feat_names[] = {
750 #include "sched_features.h"
756 static int sched_feat_show(struct seq_file *m, void *v)
760 for (i = 0; sched_feat_names[i]; i++) {
761 if (!(sysctl_sched_features & (1UL << i)))
763 seq_printf(m, "%s ", sched_feat_names[i]);
771 sched_feat_write(struct file *filp, const char __user *ubuf,
772 size_t cnt, loff_t *ppos)
782 if (copy_from_user(&buf, ubuf, cnt))
787 if (strncmp(buf, "NO_", 3) == 0) {
792 for (i = 0; sched_feat_names[i]; i++) {
793 int len = strlen(sched_feat_names[i]);
795 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
797 sysctl_sched_features &= ~(1UL << i);
799 sysctl_sched_features |= (1UL << i);
804 if (!sched_feat_names[i])
812 static int sched_feat_open(struct inode *inode, struct file *filp)
814 return single_open(filp, sched_feat_show, NULL);
817 static struct file_operations sched_feat_fops = {
818 .open = sched_feat_open,
819 .write = sched_feat_write,
822 .release = single_release,
825 static __init int sched_init_debug(void)
827 debugfs_create_file("sched_features", 0644, NULL, NULL,
832 late_initcall(sched_init_debug);
836 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
839 * Number of tasks to iterate in a single balance run.
840 * Limited because this is done with IRQs disabled.
842 const_debug unsigned int sysctl_sched_nr_migrate = 32;
845 * ratelimit for updating the group shares.
848 unsigned int sysctl_sched_shares_ratelimit = 250000;
851 * Inject some fuzzyness into changing the per-cpu group shares
852 * this avoids remote rq-locks at the expense of fairness.
855 unsigned int sysctl_sched_shares_thresh = 4;
858 * period over which we measure -rt task cpu usage in us.
861 unsigned int sysctl_sched_rt_period = 1000000;
863 static __read_mostly int scheduler_running;
866 * part of the period that we allow rt tasks to run in us.
869 int sysctl_sched_rt_runtime = 950000;
871 static inline u64 global_rt_period(void)
873 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
876 static inline u64 global_rt_runtime(void)
878 if (sysctl_sched_rt_runtime < 0)
881 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
884 #ifndef prepare_arch_switch
885 # define prepare_arch_switch(next) do { } while (0)
887 #ifndef finish_arch_switch
888 # define finish_arch_switch(prev) do { } while (0)
891 static inline int task_current(struct rq *rq, struct task_struct *p)
893 return rq->curr == p;
896 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
897 static inline int task_running(struct rq *rq, struct task_struct *p)
899 return task_current(rq, p);
902 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
906 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
908 #ifdef CONFIG_DEBUG_SPINLOCK
909 /* this is a valid case when another task releases the spinlock */
910 rq->lock.owner = current;
913 * If we are tracking spinlock dependencies then we have to
914 * fix up the runqueue lock - which gets 'carried over' from
917 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
919 spin_unlock_irq(&rq->lock);
922 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
923 static inline int task_running(struct rq *rq, struct task_struct *p)
928 return task_current(rq, p);
932 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
936 * We can optimise this out completely for !SMP, because the
937 * SMP rebalancing from interrupt is the only thing that cares
942 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
943 spin_unlock_irq(&rq->lock);
945 spin_unlock(&rq->lock);
949 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
953 * After ->oncpu is cleared, the task can be moved to a different CPU.
954 * We must ensure this doesn't happen until the switch is completely
960 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
964 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
967 * __task_rq_lock - lock the runqueue a given task resides on.
968 * Must be called interrupts disabled.
970 static inline struct rq *__task_rq_lock(struct task_struct *p)
974 struct rq *rq = task_rq(p);
975 spin_lock(&rq->lock);
976 if (likely(rq == task_rq(p)))
978 spin_unlock(&rq->lock);
983 * task_rq_lock - lock the runqueue a given task resides on and disable
984 * interrupts. Note the ordering: we can safely lookup the task_rq without
985 * explicitly disabling preemption.
987 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
993 local_irq_save(*flags);
995 spin_lock(&rq->lock);
996 if (likely(rq == task_rq(p)))
998 spin_unlock_irqrestore(&rq->lock, *flags);
1002 void task_rq_unlock_wait(struct task_struct *p)
1004 struct rq *rq = task_rq(p);
1006 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
1007 spin_unlock_wait(&rq->lock);
1010 static void __task_rq_unlock(struct rq *rq)
1011 __releases(rq->lock)
1013 spin_unlock(&rq->lock);
1016 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
1017 __releases(rq->lock)
1019 spin_unlock_irqrestore(&rq->lock, *flags);
1023 * this_rq_lock - lock this runqueue and disable interrupts.
1025 static struct rq *this_rq_lock(void)
1026 __acquires(rq->lock)
1030 local_irq_disable();
1032 spin_lock(&rq->lock);
1037 #ifdef CONFIG_SCHED_HRTICK
1039 * Use HR-timers to deliver accurate preemption points.
1041 * Its all a bit involved since we cannot program an hrt while holding the
1042 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1045 * When we get rescheduled we reprogram the hrtick_timer outside of the
1051 * - enabled by features
1052 * - hrtimer is actually high res
1054 static inline int hrtick_enabled(struct rq *rq)
1056 if (!sched_feat(HRTICK))
1058 if (!cpu_active(cpu_of(rq)))
1060 return hrtimer_is_hres_active(&rq->hrtick_timer);
1063 static void hrtick_clear(struct rq *rq)
1065 if (hrtimer_active(&rq->hrtick_timer))
1066 hrtimer_cancel(&rq->hrtick_timer);
1070 * High-resolution timer tick.
1071 * Runs from hardirq context with interrupts disabled.
1073 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1075 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1077 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1079 spin_lock(&rq->lock);
1080 update_rq_clock(rq);
1081 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1082 spin_unlock(&rq->lock);
1084 return HRTIMER_NORESTART;
1089 * called from hardirq (IPI) context
1091 static void __hrtick_start(void *arg)
1093 struct rq *rq = arg;
1095 spin_lock(&rq->lock);
1096 hrtimer_restart(&rq->hrtick_timer);
1097 rq->hrtick_csd_pending = 0;
1098 spin_unlock(&rq->lock);
1102 * Called to set the hrtick timer state.
1104 * called with rq->lock held and irqs disabled
1106 static void hrtick_start(struct rq *rq, u64 delay)
1108 struct hrtimer *timer = &rq->hrtick_timer;
1109 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1111 hrtimer_set_expires(timer, time);
1113 if (rq == this_rq()) {
1114 hrtimer_restart(timer);
1115 } else if (!rq->hrtick_csd_pending) {
1116 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1117 rq->hrtick_csd_pending = 1;
1122 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1124 int cpu = (int)(long)hcpu;
1127 case CPU_UP_CANCELED:
1128 case CPU_UP_CANCELED_FROZEN:
1129 case CPU_DOWN_PREPARE:
1130 case CPU_DOWN_PREPARE_FROZEN:
1132 case CPU_DEAD_FROZEN:
1133 hrtick_clear(cpu_rq(cpu));
1140 static __init void init_hrtick(void)
1142 hotcpu_notifier(hotplug_hrtick, 0);
1146 * Called to set the hrtick timer state.
1148 * called with rq->lock held and irqs disabled
1150 static void hrtick_start(struct rq *rq, u64 delay)
1152 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1153 HRTIMER_MODE_REL, 0);
1156 static inline void init_hrtick(void)
1159 #endif /* CONFIG_SMP */
1161 static void init_rq_hrtick(struct rq *rq)
1164 rq->hrtick_csd_pending = 0;
1166 rq->hrtick_csd.flags = 0;
1167 rq->hrtick_csd.func = __hrtick_start;
1168 rq->hrtick_csd.info = rq;
1171 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1172 rq->hrtick_timer.function = hrtick;
1174 #else /* CONFIG_SCHED_HRTICK */
1175 static inline void hrtick_clear(struct rq *rq)
1179 static inline void init_rq_hrtick(struct rq *rq)
1183 static inline void init_hrtick(void)
1186 #endif /* CONFIG_SCHED_HRTICK */
1189 * resched_task - mark a task 'to be rescheduled now'.
1191 * On UP this means the setting of the need_resched flag, on SMP it
1192 * might also involve a cross-CPU call to trigger the scheduler on
1197 #ifndef tsk_is_polling
1198 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1201 static void resched_task(struct task_struct *p)
1205 assert_spin_locked(&task_rq(p)->lock);
1207 if (test_tsk_need_resched(p))
1210 set_tsk_need_resched(p);
1213 if (cpu == smp_processor_id())
1216 /* NEED_RESCHED must be visible before we test polling */
1218 if (!tsk_is_polling(p))
1219 smp_send_reschedule(cpu);
1222 static void resched_cpu(int cpu)
1224 struct rq *rq = cpu_rq(cpu);
1225 unsigned long flags;
1227 if (!spin_trylock_irqsave(&rq->lock, flags))
1229 resched_task(cpu_curr(cpu));
1230 spin_unlock_irqrestore(&rq->lock, flags);
1235 * When add_timer_on() enqueues a timer into the timer wheel of an
1236 * idle CPU then this timer might expire before the next timer event
1237 * which is scheduled to wake up that CPU. In case of a completely
1238 * idle system the next event might even be infinite time into the
1239 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1240 * leaves the inner idle loop so the newly added timer is taken into
1241 * account when the CPU goes back to idle and evaluates the timer
1242 * wheel for the next timer event.
1244 void wake_up_idle_cpu(int cpu)
1246 struct rq *rq = cpu_rq(cpu);
1248 if (cpu == smp_processor_id())
1252 * This is safe, as this function is called with the timer
1253 * wheel base lock of (cpu) held. When the CPU is on the way
1254 * to idle and has not yet set rq->curr to idle then it will
1255 * be serialized on the timer wheel base lock and take the new
1256 * timer into account automatically.
1258 if (rq->curr != rq->idle)
1262 * We can set TIF_RESCHED on the idle task of the other CPU
1263 * lockless. The worst case is that the other CPU runs the
1264 * idle task through an additional NOOP schedule()
1266 set_tsk_need_resched(rq->idle);
1268 /* NEED_RESCHED must be visible before we test polling */
1270 if (!tsk_is_polling(rq->idle))
1271 smp_send_reschedule(cpu);
1273 #endif /* CONFIG_NO_HZ */
1275 #else /* !CONFIG_SMP */
1276 static void resched_task(struct task_struct *p)
1278 assert_spin_locked(&task_rq(p)->lock);
1279 set_tsk_need_resched(p);
1281 #endif /* CONFIG_SMP */
1283 #if BITS_PER_LONG == 32
1284 # define WMULT_CONST (~0UL)
1286 # define WMULT_CONST (1UL << 32)
1289 #define WMULT_SHIFT 32
1292 * Shift right and round:
1294 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1297 * delta *= weight / lw
1299 static unsigned long
1300 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1301 struct load_weight *lw)
1305 if (!lw->inv_weight) {
1306 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1309 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1313 tmp = (u64)delta_exec * weight;
1315 * Check whether we'd overflow the 64-bit multiplication:
1317 if (unlikely(tmp > WMULT_CONST))
1318 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1321 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1323 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1326 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1332 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1339 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1340 * of tasks with abnormal "nice" values across CPUs the contribution that
1341 * each task makes to its run queue's load is weighted according to its
1342 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1343 * scaled version of the new time slice allocation that they receive on time
1347 #define WEIGHT_IDLEPRIO 3
1348 #define WMULT_IDLEPRIO 1431655765
1351 * Nice levels are multiplicative, with a gentle 10% change for every
1352 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1353 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1354 * that remained on nice 0.
1356 * The "10% effect" is relative and cumulative: from _any_ nice level,
1357 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1358 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1359 * If a task goes up by ~10% and another task goes down by ~10% then
1360 * the relative distance between them is ~25%.)
1362 static const int prio_to_weight[40] = {
1363 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1364 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1365 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1366 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1367 /* 0 */ 1024, 820, 655, 526, 423,
1368 /* 5 */ 335, 272, 215, 172, 137,
1369 /* 10 */ 110, 87, 70, 56, 45,
1370 /* 15 */ 36, 29, 23, 18, 15,
1374 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1376 * In cases where the weight does not change often, we can use the
1377 * precalculated inverse to speed up arithmetics by turning divisions
1378 * into multiplications:
1380 static const u32 prio_to_wmult[40] = {
1381 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1382 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1383 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1384 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1385 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1386 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1387 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1388 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1391 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1394 * runqueue iterator, to support SMP load-balancing between different
1395 * scheduling classes, without having to expose their internal data
1396 * structures to the load-balancing proper:
1398 struct rq_iterator {
1400 struct task_struct *(*start)(void *);
1401 struct task_struct *(*next)(void *);
1405 static unsigned long
1406 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1407 unsigned long max_load_move, struct sched_domain *sd,
1408 enum cpu_idle_type idle, int *all_pinned,
1409 int *this_best_prio, struct rq_iterator *iterator);
1412 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1413 struct sched_domain *sd, enum cpu_idle_type idle,
1414 struct rq_iterator *iterator);
1417 /* Time spent by the tasks of the cpu accounting group executing in ... */
1418 enum cpuacct_stat_index {
1419 CPUACCT_STAT_USER, /* ... user mode */
1420 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1422 CPUACCT_STAT_NSTATS,
1425 #ifdef CONFIG_CGROUP_CPUACCT
1426 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1427 static void cpuacct_update_stats(struct task_struct *tsk,
1428 enum cpuacct_stat_index idx, cputime_t val);
1430 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1431 static inline void cpuacct_update_stats(struct task_struct *tsk,
1432 enum cpuacct_stat_index idx, cputime_t val) {}
1435 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1437 update_load_add(&rq->load, load);
1440 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1442 update_load_sub(&rq->load, load);
1445 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1446 typedef int (*tg_visitor)(struct task_group *, void *);
1449 * Iterate the full tree, calling @down when first entering a node and @up when
1450 * leaving it for the final time.
1452 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1454 struct task_group *parent, *child;
1458 parent = &root_task_group;
1460 ret = (*down)(parent, data);
1463 list_for_each_entry_rcu(child, &parent->children, siblings) {
1470 ret = (*up)(parent, data);
1475 parent = parent->parent;
1484 static int tg_nop(struct task_group *tg, void *data)
1491 static unsigned long source_load(int cpu, int type);
1492 static unsigned long target_load(int cpu, int type);
1493 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1495 static unsigned long cpu_avg_load_per_task(int cpu)
1497 struct rq *rq = cpu_rq(cpu);
1498 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1501 rq->avg_load_per_task = rq->load.weight / nr_running;
1503 rq->avg_load_per_task = 0;
1505 return rq->avg_load_per_task;
1508 #ifdef CONFIG_FAIR_GROUP_SCHED
1510 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1513 * Calculate and set the cpu's group shares.
1516 update_group_shares_cpu(struct task_group *tg, int cpu,
1517 unsigned long sd_shares, unsigned long sd_rq_weight)
1519 unsigned long shares;
1520 unsigned long rq_weight;
1525 rq_weight = tg->cfs_rq[cpu]->rq_weight;
1528 * \Sum shares * rq_weight
1529 * shares = -----------------------
1533 shares = (sd_shares * rq_weight) / sd_rq_weight;
1534 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1536 if (abs(shares - tg->se[cpu]->load.weight) >
1537 sysctl_sched_shares_thresh) {
1538 struct rq *rq = cpu_rq(cpu);
1539 unsigned long flags;
1541 spin_lock_irqsave(&rq->lock, flags);
1542 tg->cfs_rq[cpu]->shares = shares;
1544 __set_se_shares(tg->se[cpu], shares);
1545 spin_unlock_irqrestore(&rq->lock, flags);
1550 * Re-compute the task group their per cpu shares over the given domain.
1551 * This needs to be done in a bottom-up fashion because the rq weight of a
1552 * parent group depends on the shares of its child groups.
1554 static int tg_shares_up(struct task_group *tg, void *data)
1556 unsigned long weight, rq_weight = 0;
1557 unsigned long shares = 0;
1558 struct sched_domain *sd = data;
1561 for_each_cpu(i, sched_domain_span(sd)) {
1563 * If there are currently no tasks on the cpu pretend there
1564 * is one of average load so that when a new task gets to
1565 * run here it will not get delayed by group starvation.
1567 weight = tg->cfs_rq[i]->load.weight;
1569 weight = NICE_0_LOAD;
1571 tg->cfs_rq[i]->rq_weight = weight;
1572 rq_weight += weight;
1573 shares += tg->cfs_rq[i]->shares;
1576 if ((!shares && rq_weight) || shares > tg->shares)
1577 shares = tg->shares;
1579 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1580 shares = tg->shares;
1582 for_each_cpu(i, sched_domain_span(sd))
1583 update_group_shares_cpu(tg, i, shares, rq_weight);
1589 * Compute the cpu's hierarchical load factor for each task group.
1590 * This needs to be done in a top-down fashion because the load of a child
1591 * group is a fraction of its parents load.
1593 static int tg_load_down(struct task_group *tg, void *data)
1596 long cpu = (long)data;
1599 load = cpu_rq(cpu)->load.weight;
1601 load = tg->parent->cfs_rq[cpu]->h_load;
1602 load *= tg->cfs_rq[cpu]->shares;
1603 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1606 tg->cfs_rq[cpu]->h_load = load;
1611 static void update_shares(struct sched_domain *sd)
1613 u64 now = cpu_clock(raw_smp_processor_id());
1614 s64 elapsed = now - sd->last_update;
1616 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1617 sd->last_update = now;
1618 walk_tg_tree(tg_nop, tg_shares_up, sd);
1622 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1624 spin_unlock(&rq->lock);
1626 spin_lock(&rq->lock);
1629 static void update_h_load(long cpu)
1631 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1636 static inline void update_shares(struct sched_domain *sd)
1640 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1646 #ifdef CONFIG_PREEMPT
1649 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1650 * way at the expense of forcing extra atomic operations in all
1651 * invocations. This assures that the double_lock is acquired using the
1652 * same underlying policy as the spinlock_t on this architecture, which
1653 * reduces latency compared to the unfair variant below. However, it
1654 * also adds more overhead and therefore may reduce throughput.
1656 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1657 __releases(this_rq->lock)
1658 __acquires(busiest->lock)
1659 __acquires(this_rq->lock)
1661 spin_unlock(&this_rq->lock);
1662 double_rq_lock(this_rq, busiest);
1669 * Unfair double_lock_balance: Optimizes throughput at the expense of
1670 * latency by eliminating extra atomic operations when the locks are
1671 * already in proper order on entry. This favors lower cpu-ids and will
1672 * grant the double lock to lower cpus over higher ids under contention,
1673 * regardless of entry order into the function.
1675 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1676 __releases(this_rq->lock)
1677 __acquires(busiest->lock)
1678 __acquires(this_rq->lock)
1682 if (unlikely(!spin_trylock(&busiest->lock))) {
1683 if (busiest < this_rq) {
1684 spin_unlock(&this_rq->lock);
1685 spin_lock(&busiest->lock);
1686 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
1689 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
1694 #endif /* CONFIG_PREEMPT */
1697 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1699 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1701 if (unlikely(!irqs_disabled())) {
1702 /* printk() doesn't work good under rq->lock */
1703 spin_unlock(&this_rq->lock);
1707 return _double_lock_balance(this_rq, busiest);
1710 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1711 __releases(busiest->lock)
1713 spin_unlock(&busiest->lock);
1714 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1718 #ifdef CONFIG_FAIR_GROUP_SCHED
1719 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1722 cfs_rq->shares = shares;
1727 #include "sched_stats.h"
1728 #include "sched_idletask.c"
1729 #include "sched_fair.c"
1730 #include "sched_rt.c"
1731 #ifdef CONFIG_SCHED_DEBUG
1732 # include "sched_debug.c"
1735 #define sched_class_highest (&rt_sched_class)
1736 #define for_each_class(class) \
1737 for (class = sched_class_highest; class; class = class->next)
1739 static void inc_nr_running(struct rq *rq)
1744 static void dec_nr_running(struct rq *rq)
1749 static void set_load_weight(struct task_struct *p)
1751 if (task_has_rt_policy(p)) {
1752 p->se.load.weight = prio_to_weight[0] * 2;
1753 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1758 * SCHED_IDLE tasks get minimal weight:
1760 if (p->policy == SCHED_IDLE) {
1761 p->se.load.weight = WEIGHT_IDLEPRIO;
1762 p->se.load.inv_weight = WMULT_IDLEPRIO;
1766 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1767 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1770 static void update_avg(u64 *avg, u64 sample)
1772 s64 diff = sample - *avg;
1776 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1779 p->se.start_runtime = p->se.sum_exec_runtime;
1781 sched_info_queued(p);
1782 p->sched_class->enqueue_task(rq, p, wakeup);
1786 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1789 if (p->se.last_wakeup) {
1790 update_avg(&p->se.avg_overlap,
1791 p->se.sum_exec_runtime - p->se.last_wakeup);
1792 p->se.last_wakeup = 0;
1794 update_avg(&p->se.avg_wakeup,
1795 sysctl_sched_wakeup_granularity);
1799 sched_info_dequeued(p);
1800 p->sched_class->dequeue_task(rq, p, sleep);
1805 * __normal_prio - return the priority that is based on the static prio
1807 static inline int __normal_prio(struct task_struct *p)
1809 return p->static_prio;
1813 * Calculate the expected normal priority: i.e. priority
1814 * without taking RT-inheritance into account. Might be
1815 * boosted by interactivity modifiers. Changes upon fork,
1816 * setprio syscalls, and whenever the interactivity
1817 * estimator recalculates.
1819 static inline int normal_prio(struct task_struct *p)
1823 if (task_has_rt_policy(p))
1824 prio = MAX_RT_PRIO-1 - p->rt_priority;
1826 prio = __normal_prio(p);
1831 * Calculate the current priority, i.e. the priority
1832 * taken into account by the scheduler. This value might
1833 * be boosted by RT tasks, or might be boosted by
1834 * interactivity modifiers. Will be RT if the task got
1835 * RT-boosted. If not then it returns p->normal_prio.
1837 static int effective_prio(struct task_struct *p)
1839 p->normal_prio = normal_prio(p);
1841 * If we are RT tasks or we were boosted to RT priority,
1842 * keep the priority unchanged. Otherwise, update priority
1843 * to the normal priority:
1845 if (!rt_prio(p->prio))
1846 return p->normal_prio;
1851 * activate_task - move a task to the runqueue.
1853 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1855 if (task_contributes_to_load(p))
1856 rq->nr_uninterruptible--;
1858 enqueue_task(rq, p, wakeup);
1863 * deactivate_task - remove a task from the runqueue.
1865 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1867 if (task_contributes_to_load(p))
1868 rq->nr_uninterruptible++;
1870 dequeue_task(rq, p, sleep);
1875 * task_curr - is this task currently executing on a CPU?
1876 * @p: the task in question.
1878 inline int task_curr(const struct task_struct *p)
1880 return cpu_curr(task_cpu(p)) == p;
1883 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1885 set_task_rq(p, cpu);
1888 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1889 * successfuly executed on another CPU. We must ensure that updates of
1890 * per-task data have been completed by this moment.
1893 task_thread_info(p)->cpu = cpu;
1897 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1898 const struct sched_class *prev_class,
1899 int oldprio, int running)
1901 if (prev_class != p->sched_class) {
1902 if (prev_class->switched_from)
1903 prev_class->switched_from(rq, p, running);
1904 p->sched_class->switched_to(rq, p, running);
1906 p->sched_class->prio_changed(rq, p, oldprio, running);
1911 /* Used instead of source_load when we know the type == 0 */
1912 static unsigned long weighted_cpuload(const int cpu)
1914 return cpu_rq(cpu)->load.weight;
1918 * Is this task likely cache-hot:
1921 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1926 * Buddy candidates are cache hot:
1928 if (sched_feat(CACHE_HOT_BUDDY) &&
1929 (&p->se == cfs_rq_of(&p->se)->next ||
1930 &p->se == cfs_rq_of(&p->se)->last))
1933 if (p->sched_class != &fair_sched_class)
1936 if (sysctl_sched_migration_cost == -1)
1938 if (sysctl_sched_migration_cost == 0)
1941 delta = now - p->se.exec_start;
1943 return delta < (s64)sysctl_sched_migration_cost;
1947 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1949 int old_cpu = task_cpu(p);
1950 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1951 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1952 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1955 clock_offset = old_rq->clock - new_rq->clock;
1957 trace_sched_migrate_task(p, new_cpu);
1959 #ifdef CONFIG_SCHEDSTATS
1960 if (p->se.wait_start)
1961 p->se.wait_start -= clock_offset;
1962 if (p->se.sleep_start)
1963 p->se.sleep_start -= clock_offset;
1964 if (p->se.block_start)
1965 p->se.block_start -= clock_offset;
1966 if (old_cpu != new_cpu) {
1967 schedstat_inc(p, se.nr_migrations);
1968 if (task_hot(p, old_rq->clock, NULL))
1969 schedstat_inc(p, se.nr_forced2_migrations);
1972 p->se.vruntime -= old_cfsrq->min_vruntime -
1973 new_cfsrq->min_vruntime;
1975 __set_task_cpu(p, new_cpu);
1978 struct migration_req {
1979 struct list_head list;
1981 struct task_struct *task;
1984 struct completion done;
1988 * The task's runqueue lock must be held.
1989 * Returns true if you have to wait for migration thread.
1992 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1994 struct rq *rq = task_rq(p);
1997 * If the task is not on a runqueue (and not running), then
1998 * it is sufficient to simply update the task's cpu field.
2000 if (!p->se.on_rq && !task_running(rq, p)) {
2001 set_task_cpu(p, dest_cpu);
2005 init_completion(&req->done);
2007 req->dest_cpu = dest_cpu;
2008 list_add(&req->list, &rq->migration_queue);
2014 * wait_task_context_switch - wait for a thread to complete at least one
2017 * @p must not be current.
2019 void wait_task_context_switch(struct task_struct *p)
2021 unsigned long nvcsw, nivcsw, flags;
2029 * The runqueue is assigned before the actual context
2030 * switch. We need to take the runqueue lock.
2032 * We could check initially without the lock but it is
2033 * very likely that we need to take the lock in every
2036 rq = task_rq_lock(p, &flags);
2037 running = task_running(rq, p);
2038 task_rq_unlock(rq, &flags);
2040 if (likely(!running))
2043 * The switch count is incremented before the actual
2044 * context switch. We thus wait for two switches to be
2045 * sure at least one completed.
2047 if ((p->nvcsw - nvcsw) > 1)
2049 if ((p->nivcsw - nivcsw) > 1)
2057 * wait_task_inactive - wait for a thread to unschedule.
2059 * If @match_state is nonzero, it's the @p->state value just checked and
2060 * not expected to change. If it changes, i.e. @p might have woken up,
2061 * then return zero. When we succeed in waiting for @p to be off its CPU,
2062 * we return a positive number (its total switch count). If a second call
2063 * a short while later returns the same number, the caller can be sure that
2064 * @p has remained unscheduled the whole time.
2066 * The caller must ensure that the task *will* unschedule sometime soon,
2067 * else this function might spin for a *long* time. This function can't
2068 * be called with interrupts off, or it may introduce deadlock with
2069 * smp_call_function() if an IPI is sent by the same process we are
2070 * waiting to become inactive.
2072 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2074 unsigned long flags;
2081 * We do the initial early heuristics without holding
2082 * any task-queue locks at all. We'll only try to get
2083 * the runqueue lock when things look like they will
2089 * If the task is actively running on another CPU
2090 * still, just relax and busy-wait without holding
2093 * NOTE! Since we don't hold any locks, it's not
2094 * even sure that "rq" stays as the right runqueue!
2095 * But we don't care, since "task_running()" will
2096 * return false if the runqueue has changed and p
2097 * is actually now running somewhere else!
2099 while (task_running(rq, p)) {
2100 if (match_state && unlikely(p->state != match_state))
2106 * Ok, time to look more closely! We need the rq
2107 * lock now, to be *sure*. If we're wrong, we'll
2108 * just go back and repeat.
2110 rq = task_rq_lock(p, &flags);
2111 trace_sched_wait_task(rq, p);
2112 running = task_running(rq, p);
2113 on_rq = p->se.on_rq;
2115 if (!match_state || p->state == match_state)
2116 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2117 task_rq_unlock(rq, &flags);
2120 * If it changed from the expected state, bail out now.
2122 if (unlikely(!ncsw))
2126 * Was it really running after all now that we
2127 * checked with the proper locks actually held?
2129 * Oops. Go back and try again..
2131 if (unlikely(running)) {
2137 * It's not enough that it's not actively running,
2138 * it must be off the runqueue _entirely_, and not
2141 * So if it was still runnable (but just not actively
2142 * running right now), it's preempted, and we should
2143 * yield - it could be a while.
2145 if (unlikely(on_rq)) {
2146 schedule_timeout_uninterruptible(1);
2151 * Ahh, all good. It wasn't running, and it wasn't
2152 * runnable, which means that it will never become
2153 * running in the future either. We're all done!
2162 * kick_process - kick a running thread to enter/exit the kernel
2163 * @p: the to-be-kicked thread
2165 * Cause a process which is running on another CPU to enter
2166 * kernel-mode, without any delay. (to get signals handled.)
2168 * NOTE: this function doesnt have to take the runqueue lock,
2169 * because all it wants to ensure is that the remote task enters
2170 * the kernel. If the IPI races and the task has been migrated
2171 * to another CPU then no harm is done and the purpose has been
2174 void kick_process(struct task_struct *p)
2180 if ((cpu != smp_processor_id()) && task_curr(p))
2181 smp_send_reschedule(cpu);
2186 * Return a low guess at the load of a migration-source cpu weighted
2187 * according to the scheduling class and "nice" value.
2189 * We want to under-estimate the load of migration sources, to
2190 * balance conservatively.
2192 static unsigned long source_load(int cpu, int type)
2194 struct rq *rq = cpu_rq(cpu);
2195 unsigned long total = weighted_cpuload(cpu);
2197 if (type == 0 || !sched_feat(LB_BIAS))
2200 return min(rq->cpu_load[type-1], total);
2204 * Return a high guess at the load of a migration-target cpu weighted
2205 * according to the scheduling class and "nice" value.
2207 static unsigned long target_load(int cpu, int type)
2209 struct rq *rq = cpu_rq(cpu);
2210 unsigned long total = weighted_cpuload(cpu);
2212 if (type == 0 || !sched_feat(LB_BIAS))
2215 return max(rq->cpu_load[type-1], total);
2219 * find_idlest_group finds and returns the least busy CPU group within the
2222 static struct sched_group *
2223 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2225 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2226 unsigned long min_load = ULONG_MAX, this_load = 0;
2227 int load_idx = sd->forkexec_idx;
2228 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2231 unsigned long load, avg_load;
2235 /* Skip over this group if it has no CPUs allowed */
2236 if (!cpumask_intersects(sched_group_cpus(group),
2240 local_group = cpumask_test_cpu(this_cpu,
2241 sched_group_cpus(group));
2243 /* Tally up the load of all CPUs in the group */
2246 for_each_cpu(i, sched_group_cpus(group)) {
2247 /* Bias balancing toward cpus of our domain */
2249 load = source_load(i, load_idx);
2251 load = target_load(i, load_idx);
2256 /* Adjust by relative CPU power of the group */
2257 avg_load = sg_div_cpu_power(group,
2258 avg_load * SCHED_LOAD_SCALE);
2261 this_load = avg_load;
2263 } else if (avg_load < min_load) {
2264 min_load = avg_load;
2267 } while (group = group->next, group != sd->groups);
2269 if (!idlest || 100*this_load < imbalance*min_load)
2275 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2278 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
2280 unsigned long load, min_load = ULONG_MAX;
2284 /* Traverse only the allowed CPUs */
2285 for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
2286 load = weighted_cpuload(i);
2288 if (load < min_load || (load == min_load && i == this_cpu)) {
2298 * sched_balance_self: balance the current task (running on cpu) in domains
2299 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2302 * Balance, ie. select the least loaded group.
2304 * Returns the target CPU number, or the same CPU if no balancing is needed.
2306 * preempt must be disabled.
2308 static int sched_balance_self(int cpu, int flag)
2310 struct task_struct *t = current;
2311 struct sched_domain *tmp, *sd = NULL;
2313 for_each_domain(cpu, tmp) {
2315 * If power savings logic is enabled for a domain, stop there.
2317 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2319 if (tmp->flags & flag)
2327 struct sched_group *group;
2328 int new_cpu, weight;
2330 if (!(sd->flags & flag)) {
2335 group = find_idlest_group(sd, t, cpu);
2341 new_cpu = find_idlest_cpu(group, t, cpu);
2342 if (new_cpu == -1 || new_cpu == cpu) {
2343 /* Now try balancing at a lower domain level of cpu */
2348 /* Now try balancing at a lower domain level of new_cpu */
2350 weight = cpumask_weight(sched_domain_span(sd));
2352 for_each_domain(cpu, tmp) {
2353 if (weight <= cpumask_weight(sched_domain_span(tmp)))
2355 if (tmp->flags & flag)
2358 /* while loop will break here if sd == NULL */
2364 #endif /* CONFIG_SMP */
2367 * try_to_wake_up - wake up a thread
2368 * @p: the to-be-woken-up thread
2369 * @state: the mask of task states that can be woken
2370 * @sync: do a synchronous wakeup?
2372 * Put it on the run-queue if it's not already there. The "current"
2373 * thread is always on the run-queue (except when the actual
2374 * re-schedule is in progress), and as such you're allowed to do
2375 * the simpler "current->state = TASK_RUNNING" to mark yourself
2376 * runnable without the overhead of this.
2378 * returns failure only if the task is already active.
2380 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2382 int cpu, orig_cpu, this_cpu, success = 0;
2383 unsigned long flags;
2387 if (!sched_feat(SYNC_WAKEUPS))
2391 if (sched_feat(LB_WAKEUP_UPDATE) && !root_task_group_empty()) {
2392 struct sched_domain *sd;
2394 this_cpu = raw_smp_processor_id();
2397 for_each_domain(this_cpu, sd) {
2398 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2407 rq = task_rq_lock(p, &flags);
2408 update_rq_clock(rq);
2409 old_state = p->state;
2410 if (!(old_state & state))
2418 this_cpu = smp_processor_id();
2421 if (unlikely(task_running(rq, p)))
2424 cpu = p->sched_class->select_task_rq(p, sync);
2425 if (cpu != orig_cpu) {
2426 set_task_cpu(p, cpu);
2427 task_rq_unlock(rq, &flags);
2428 /* might preempt at this point */
2429 rq = task_rq_lock(p, &flags);
2430 old_state = p->state;
2431 if (!(old_state & state))
2436 this_cpu = smp_processor_id();
2440 #ifdef CONFIG_SCHEDSTATS
2441 schedstat_inc(rq, ttwu_count);
2442 if (cpu == this_cpu)
2443 schedstat_inc(rq, ttwu_local);
2445 struct sched_domain *sd;
2446 for_each_domain(this_cpu, sd) {
2447 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2448 schedstat_inc(sd, ttwu_wake_remote);
2453 #endif /* CONFIG_SCHEDSTATS */
2456 #endif /* CONFIG_SMP */
2457 schedstat_inc(p, se.nr_wakeups);
2459 schedstat_inc(p, se.nr_wakeups_sync);
2460 if (orig_cpu != cpu)
2461 schedstat_inc(p, se.nr_wakeups_migrate);
2462 if (cpu == this_cpu)
2463 schedstat_inc(p, se.nr_wakeups_local);
2465 schedstat_inc(p, se.nr_wakeups_remote);
2466 activate_task(rq, p, 1);
2470 * Only attribute actual wakeups done by this task.
2472 if (!in_interrupt()) {
2473 struct sched_entity *se = ¤t->se;
2474 u64 sample = se->sum_exec_runtime;
2476 if (se->last_wakeup)
2477 sample -= se->last_wakeup;
2479 sample -= se->start_runtime;
2480 update_avg(&se->avg_wakeup, sample);
2482 se->last_wakeup = se->sum_exec_runtime;
2486 trace_sched_wakeup(rq, p, success);
2487 check_preempt_curr(rq, p, sync);
2489 p->state = TASK_RUNNING;
2491 if (p->sched_class->task_wake_up)
2492 p->sched_class->task_wake_up(rq, p);
2495 task_rq_unlock(rq, &flags);
2500 int wake_up_process(struct task_struct *p)
2502 return try_to_wake_up(p, TASK_ALL, 0);
2504 EXPORT_SYMBOL(wake_up_process);
2506 int wake_up_state(struct task_struct *p, unsigned int state)
2508 return try_to_wake_up(p, state, 0);
2512 * Perform scheduler related setup for a newly forked process p.
2513 * p is forked by current.
2515 * __sched_fork() is basic setup used by init_idle() too:
2517 static void __sched_fork(struct task_struct *p)
2519 p->se.exec_start = 0;
2520 p->se.sum_exec_runtime = 0;
2521 p->se.prev_sum_exec_runtime = 0;
2522 p->se.last_wakeup = 0;
2523 p->se.avg_overlap = 0;
2524 p->se.start_runtime = 0;
2525 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2527 #ifdef CONFIG_SCHEDSTATS
2528 p->se.wait_start = 0;
2529 p->se.sum_sleep_runtime = 0;
2530 p->se.sleep_start = 0;
2531 p->se.block_start = 0;
2532 p->se.sleep_max = 0;
2533 p->se.block_max = 0;
2535 p->se.slice_max = 0;
2539 INIT_LIST_HEAD(&p->rt.run_list);
2541 INIT_LIST_HEAD(&p->se.group_node);
2543 #ifdef CONFIG_PREEMPT_NOTIFIERS
2544 INIT_HLIST_HEAD(&p->preempt_notifiers);
2548 * We mark the process as running here, but have not actually
2549 * inserted it onto the runqueue yet. This guarantees that
2550 * nobody will actually run it, and a signal or other external
2551 * event cannot wake it up and insert it on the runqueue either.
2553 p->state = TASK_RUNNING;
2557 * fork()/clone()-time setup:
2559 void sched_fork(struct task_struct *p, int clone_flags)
2561 int cpu = get_cpu();
2566 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2568 set_task_cpu(p, cpu);
2571 * Make sure we do not leak PI boosting priority to the child:
2573 p->prio = current->normal_prio;
2574 if (!rt_prio(p->prio))
2575 p->sched_class = &fair_sched_class;
2577 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2578 if (likely(sched_info_on()))
2579 memset(&p->sched_info, 0, sizeof(p->sched_info));
2581 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2584 #ifdef CONFIG_PREEMPT
2585 /* Want to start with kernel preemption disabled. */
2586 task_thread_info(p)->preempt_count = 1;
2588 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2594 * wake_up_new_task - wake up a newly created task for the first time.
2596 * This function will do some initial scheduler statistics housekeeping
2597 * that must be done for every newly created context, then puts the task
2598 * on the runqueue and wakes it.
2600 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2602 unsigned long flags;
2605 rq = task_rq_lock(p, &flags);
2606 BUG_ON(p->state != TASK_RUNNING);
2607 update_rq_clock(rq);
2609 p->prio = effective_prio(p);
2611 if (!p->sched_class->task_new || !current->se.on_rq) {
2612 activate_task(rq, p, 0);
2615 * Let the scheduling class do new task startup
2616 * management (if any):
2618 p->sched_class->task_new(rq, p);
2621 trace_sched_wakeup_new(rq, p, 1);
2622 check_preempt_curr(rq, p, 0);
2624 if (p->sched_class->task_wake_up)
2625 p->sched_class->task_wake_up(rq, p);
2627 task_rq_unlock(rq, &flags);
2630 #ifdef CONFIG_PREEMPT_NOTIFIERS
2633 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2634 * @notifier: notifier struct to register
2636 void preempt_notifier_register(struct preempt_notifier *notifier)
2638 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2640 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2643 * preempt_notifier_unregister - no longer interested in preemption notifications
2644 * @notifier: notifier struct to unregister
2646 * This is safe to call from within a preemption notifier.
2648 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2650 hlist_del(¬ifier->link);
2652 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2654 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2656 struct preempt_notifier *notifier;
2657 struct hlist_node *node;
2659 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2660 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2664 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2665 struct task_struct *next)
2667 struct preempt_notifier *notifier;
2668 struct hlist_node *node;
2670 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2671 notifier->ops->sched_out(notifier, next);
2674 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2676 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2681 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2682 struct task_struct *next)
2686 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2689 * prepare_task_switch - prepare to switch tasks
2690 * @rq: the runqueue preparing to switch
2691 * @prev: the current task that is being switched out
2692 * @next: the task we are going to switch to.
2694 * This is called with the rq lock held and interrupts off. It must
2695 * be paired with a subsequent finish_task_switch after the context
2698 * prepare_task_switch sets up locking and calls architecture specific
2702 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2703 struct task_struct *next)
2705 fire_sched_out_preempt_notifiers(prev, next);
2706 prepare_lock_switch(rq, next);
2707 prepare_arch_switch(next);
2711 * finish_task_switch - clean up after a task-switch
2712 * @rq: runqueue associated with task-switch
2713 * @prev: the thread we just switched away from.
2715 * finish_task_switch must be called after the context switch, paired
2716 * with a prepare_task_switch call before the context switch.
2717 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2718 * and do any other architecture-specific cleanup actions.
2720 * Note that we may have delayed dropping an mm in context_switch(). If
2721 * so, we finish that here outside of the runqueue lock. (Doing it
2722 * with the lock held can cause deadlocks; see schedule() for
2725 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2726 __releases(rq->lock)
2728 struct mm_struct *mm = rq->prev_mm;
2731 int post_schedule = 0;
2733 if (current->sched_class->needs_post_schedule)
2734 post_schedule = current->sched_class->needs_post_schedule(rq);
2740 * A task struct has one reference for the use as "current".
2741 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2742 * schedule one last time. The schedule call will never return, and
2743 * the scheduled task must drop that reference.
2744 * The test for TASK_DEAD must occur while the runqueue locks are
2745 * still held, otherwise prev could be scheduled on another cpu, die
2746 * there before we look at prev->state, and then the reference would
2748 * Manfred Spraul <manfred@colorfullife.com>
2750 prev_state = prev->state;
2751 finish_arch_switch(prev);
2752 finish_lock_switch(rq, prev);
2755 current->sched_class->post_schedule(rq);
2758 fire_sched_in_preempt_notifiers(current);
2761 if (unlikely(prev_state == TASK_DEAD)) {
2763 * Remove function-return probe instances associated with this
2764 * task and put them back on the free list.
2766 kprobe_flush_task(prev);
2767 put_task_struct(prev);
2772 * schedule_tail - first thing a freshly forked thread must call.
2773 * @prev: the thread we just switched away from.
2775 asmlinkage void schedule_tail(struct task_struct *prev)
2776 __releases(rq->lock)
2778 struct rq *rq = this_rq();
2780 finish_task_switch(rq, prev);
2781 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2782 /* In this case, finish_task_switch does not reenable preemption */
2785 if (current->set_child_tid)
2786 put_user(task_pid_vnr(current), current->set_child_tid);
2790 * context_switch - switch to the new MM and the new
2791 * thread's register state.
2794 context_switch(struct rq *rq, struct task_struct *prev,
2795 struct task_struct *next)
2797 struct mm_struct *mm, *oldmm;
2799 prepare_task_switch(rq, prev, next);
2800 trace_sched_switch(rq, prev, next);
2802 oldmm = prev->active_mm;
2804 * For paravirt, this is coupled with an exit in switch_to to
2805 * combine the page table reload and the switch backend into
2808 arch_enter_lazy_cpu_mode();
2810 if (unlikely(!mm)) {
2811 next->active_mm = oldmm;
2812 atomic_inc(&oldmm->mm_count);
2813 enter_lazy_tlb(oldmm, next);
2815 switch_mm(oldmm, mm, next);
2817 if (unlikely(!prev->mm)) {
2818 prev->active_mm = NULL;
2819 rq->prev_mm = oldmm;
2822 * Since the runqueue lock will be released by the next
2823 * task (which is an invalid locking op but in the case
2824 * of the scheduler it's an obvious special-case), so we
2825 * do an early lockdep release here:
2827 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2828 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2831 /* Here we just switch the register state and the stack. */
2832 switch_to(prev, next, prev);
2836 * this_rq must be evaluated again because prev may have moved
2837 * CPUs since it called schedule(), thus the 'rq' on its stack
2838 * frame will be invalid.
2840 finish_task_switch(this_rq(), prev);
2844 * nr_running, nr_uninterruptible and nr_context_switches:
2846 * externally visible scheduler statistics: current number of runnable
2847 * threads, current number of uninterruptible-sleeping threads, total
2848 * number of context switches performed since bootup.
2850 unsigned long nr_running(void)
2852 unsigned long i, sum = 0;
2854 for_each_online_cpu(i)
2855 sum += cpu_rq(i)->nr_running;
2860 unsigned long nr_uninterruptible(void)
2862 unsigned long i, sum = 0;
2864 for_each_possible_cpu(i)
2865 sum += cpu_rq(i)->nr_uninterruptible;
2868 * Since we read the counters lockless, it might be slightly
2869 * inaccurate. Do not allow it to go below zero though:
2871 if (unlikely((long)sum < 0))
2877 unsigned long long nr_context_switches(void)
2880 unsigned long long sum = 0;
2882 for_each_possible_cpu(i)
2883 sum += cpu_rq(i)->nr_switches;
2888 unsigned long nr_iowait(void)
2890 unsigned long i, sum = 0;
2892 for_each_possible_cpu(i)
2893 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2898 unsigned long nr_active(void)
2900 unsigned long i, running = 0, uninterruptible = 0;
2902 for_each_online_cpu(i) {
2903 running += cpu_rq(i)->nr_running;
2904 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2907 if (unlikely((long)uninterruptible < 0))
2908 uninterruptible = 0;
2910 return running + uninterruptible;
2914 * Update rq->cpu_load[] statistics. This function is usually called every
2915 * scheduler tick (TICK_NSEC).
2917 static void update_cpu_load(struct rq *this_rq)
2919 unsigned long this_load = this_rq->load.weight;
2922 this_rq->nr_load_updates++;
2924 /* Update our load: */
2925 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2926 unsigned long old_load, new_load;
2928 /* scale is effectively 1 << i now, and >> i divides by scale */
2930 old_load = this_rq->cpu_load[i];
2931 new_load = this_load;
2933 * Round up the averaging division if load is increasing. This
2934 * prevents us from getting stuck on 9 if the load is 10, for
2937 if (new_load > old_load)
2938 new_load += scale-1;
2939 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2946 * double_rq_lock - safely lock two runqueues
2948 * Note this does not disable interrupts like task_rq_lock,
2949 * you need to do so manually before calling.
2951 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2952 __acquires(rq1->lock)
2953 __acquires(rq2->lock)
2955 BUG_ON(!irqs_disabled());
2957 spin_lock(&rq1->lock);
2958 __acquire(rq2->lock); /* Fake it out ;) */
2961 spin_lock(&rq1->lock);
2962 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
2964 spin_lock(&rq2->lock);
2965 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
2968 update_rq_clock(rq1);
2969 update_rq_clock(rq2);
2973 * double_rq_unlock - safely unlock two runqueues
2975 * Note this does not restore interrupts like task_rq_unlock,
2976 * you need to do so manually after calling.
2978 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2979 __releases(rq1->lock)
2980 __releases(rq2->lock)
2982 spin_unlock(&rq1->lock);
2984 spin_unlock(&rq2->lock);
2986 __release(rq2->lock);
2990 * If dest_cpu is allowed for this process, migrate the task to it.
2991 * This is accomplished by forcing the cpu_allowed mask to only
2992 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2993 * the cpu_allowed mask is restored.
2995 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2997 struct migration_req req;
2998 unsigned long flags;
3001 rq = task_rq_lock(p, &flags);
3002 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
3003 || unlikely(!cpu_active(dest_cpu)))
3006 /* force the process onto the specified CPU */
3007 if (migrate_task(p, dest_cpu, &req)) {
3008 /* Need to wait for migration thread (might exit: take ref). */
3009 struct task_struct *mt = rq->migration_thread;
3011 get_task_struct(mt);
3012 task_rq_unlock(rq, &flags);
3013 wake_up_process(mt);
3014 put_task_struct(mt);
3015 wait_for_completion(&req.done);
3020 task_rq_unlock(rq, &flags);
3024 * sched_exec - execve() is a valuable balancing opportunity, because at
3025 * this point the task has the smallest effective memory and cache footprint.
3027 void sched_exec(void)
3029 int new_cpu, this_cpu = get_cpu();
3030 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
3032 if (new_cpu != this_cpu)
3033 sched_migrate_task(current, new_cpu);
3037 * pull_task - move a task from a remote runqueue to the local runqueue.
3038 * Both runqueues must be locked.
3040 static void pull_task(struct rq *src_rq, struct task_struct *p,
3041 struct rq *this_rq, int this_cpu)
3043 deactivate_task(src_rq, p, 0);
3044 set_task_cpu(p, this_cpu);
3045 activate_task(this_rq, p, 0);
3047 * Note that idle threads have a prio of MAX_PRIO, for this test
3048 * to be always true for them.
3050 check_preempt_curr(this_rq, p, 0);
3054 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3057 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3058 struct sched_domain *sd, enum cpu_idle_type idle,
3061 int tsk_cache_hot = 0;
3063 * We do not migrate tasks that are:
3064 * 1) running (obviously), or
3065 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3066 * 3) are cache-hot on their current CPU.
3068 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
3069 schedstat_inc(p, se.nr_failed_migrations_affine);
3074 if (task_running(rq, p)) {
3075 schedstat_inc(p, se.nr_failed_migrations_running);
3080 * Aggressive migration if:
3081 * 1) task is cache cold, or
3082 * 2) too many balance attempts have failed.
3085 tsk_cache_hot = task_hot(p, rq->clock, sd);
3086 if (!tsk_cache_hot ||
3087 sd->nr_balance_failed > sd->cache_nice_tries) {
3088 #ifdef CONFIG_SCHEDSTATS
3089 if (tsk_cache_hot) {
3090 schedstat_inc(sd, lb_hot_gained[idle]);
3091 schedstat_inc(p, se.nr_forced_migrations);
3097 if (tsk_cache_hot) {
3098 schedstat_inc(p, se.nr_failed_migrations_hot);
3104 static unsigned long
3105 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3106 unsigned long max_load_move, struct sched_domain *sd,
3107 enum cpu_idle_type idle, int *all_pinned,
3108 int *this_best_prio, struct rq_iterator *iterator)
3110 int loops = 0, pulled = 0, pinned = 0;
3111 struct task_struct *p;
3112 long rem_load_move = max_load_move;
3114 if (max_load_move == 0)
3120 * Start the load-balancing iterator:
3122 p = iterator->start(iterator->arg);
3124 if (!p || loops++ > sysctl_sched_nr_migrate)
3127 if ((p->se.load.weight >> 1) > rem_load_move ||
3128 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3129 p = iterator->next(iterator->arg);
3133 pull_task(busiest, p, this_rq, this_cpu);
3135 rem_load_move -= p->se.load.weight;
3137 #ifdef CONFIG_PREEMPT
3139 * NEWIDLE balancing is a source of latency, so preemptible kernels
3140 * will stop after the first task is pulled to minimize the critical
3143 if (idle == CPU_NEWLY_IDLE)
3148 * We only want to steal up to the prescribed amount of weighted load.
3150 if (rem_load_move > 0) {
3151 if (p->prio < *this_best_prio)
3152 *this_best_prio = p->prio;
3153 p = iterator->next(iterator->arg);
3158 * Right now, this is one of only two places pull_task() is called,
3159 * so we can safely collect pull_task() stats here rather than
3160 * inside pull_task().
3162 schedstat_add(sd, lb_gained[idle], pulled);
3165 *all_pinned = pinned;
3167 return max_load_move - rem_load_move;
3171 * move_tasks tries to move up to max_load_move weighted load from busiest to
3172 * this_rq, as part of a balancing operation within domain "sd".
3173 * Returns 1 if successful and 0 otherwise.
3175 * Called with both runqueues locked.
3177 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3178 unsigned long max_load_move,
3179 struct sched_domain *sd, enum cpu_idle_type idle,
3182 const struct sched_class *class = sched_class_highest;
3183 unsigned long total_load_moved = 0;
3184 int this_best_prio = this_rq->curr->prio;
3188 class->load_balance(this_rq, this_cpu, busiest,
3189 max_load_move - total_load_moved,
3190 sd, idle, all_pinned, &this_best_prio);
3191 class = class->next;
3193 #ifdef CONFIG_PREEMPT
3195 * NEWIDLE balancing is a source of latency, so preemptible
3196 * kernels will stop after the first task is pulled to minimize
3197 * the critical section.
3199 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3202 } while (class && max_load_move > total_load_moved);
3204 return total_load_moved > 0;
3208 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3209 struct sched_domain *sd, enum cpu_idle_type idle,
3210 struct rq_iterator *iterator)
3212 struct task_struct *p = iterator->start(iterator->arg);
3216 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3217 pull_task(busiest, p, this_rq, this_cpu);
3219 * Right now, this is only the second place pull_task()
3220 * is called, so we can safely collect pull_task()
3221 * stats here rather than inside pull_task().
3223 schedstat_inc(sd, lb_gained[idle]);
3227 p = iterator->next(iterator->arg);
3234 * move_one_task tries to move exactly one task from busiest to this_rq, as
3235 * part of active balancing operations within "domain".
3236 * Returns 1 if successful and 0 otherwise.
3238 * Called with both runqueues locked.
3240 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3241 struct sched_domain *sd, enum cpu_idle_type idle)
3243 const struct sched_class *class;
3245 for (class = sched_class_highest; class; class = class->next)
3246 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3251 /********** Helpers for find_busiest_group ************************/
3253 * sd_lb_stats - Structure to store the statistics of a sched_domain
3254 * during load balancing.
3256 struct sd_lb_stats {
3257 struct sched_group *busiest; /* Busiest group in this sd */
3258 struct sched_group *this; /* Local group in this sd */
3259 unsigned long total_load; /* Total load of all groups in sd */
3260 unsigned long total_pwr; /* Total power of all groups in sd */
3261 unsigned long avg_load; /* Average load across all groups in sd */
3263 /** Statistics of this group */
3264 unsigned long this_load;
3265 unsigned long this_load_per_task;
3266 unsigned long this_nr_running;
3268 /* Statistics of the busiest group */
3269 unsigned long max_load;
3270 unsigned long busiest_load_per_task;
3271 unsigned long busiest_nr_running;
3273 int group_imb; /* Is there imbalance in this sd */
3274 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3275 int power_savings_balance; /* Is powersave balance needed for this sd */
3276 struct sched_group *group_min; /* Least loaded group in sd */
3277 struct sched_group *group_leader; /* Group which relieves group_min */
3278 unsigned long min_load_per_task; /* load_per_task in group_min */
3279 unsigned long leader_nr_running; /* Nr running of group_leader */
3280 unsigned long min_nr_running; /* Nr running of group_min */
3285 * sg_lb_stats - stats of a sched_group required for load_balancing
3287 struct sg_lb_stats {
3288 unsigned long avg_load; /*Avg load across the CPUs of the group */
3289 unsigned long group_load; /* Total load over the CPUs of the group */
3290 unsigned long sum_nr_running; /* Nr tasks running in the group */
3291 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3292 unsigned long group_capacity;
3293 int group_imb; /* Is there an imbalance in the group ? */
3297 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3298 * @group: The group whose first cpu is to be returned.
3300 static inline unsigned int group_first_cpu(struct sched_group *group)
3302 return cpumask_first(sched_group_cpus(group));
3306 * get_sd_load_idx - Obtain the load index for a given sched domain.
3307 * @sd: The sched_domain whose load_idx is to be obtained.
3308 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3310 static inline int get_sd_load_idx(struct sched_domain *sd,
3311 enum cpu_idle_type idle)
3317 load_idx = sd->busy_idx;
3320 case CPU_NEWLY_IDLE:
3321 load_idx = sd->newidle_idx;
3324 load_idx = sd->idle_idx;
3332 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3334 * init_sd_power_savings_stats - Initialize power savings statistics for
3335 * the given sched_domain, during load balancing.
3337 * @sd: Sched domain whose power-savings statistics are to be initialized.
3338 * @sds: Variable containing the statistics for sd.
3339 * @idle: Idle status of the CPU at which we're performing load-balancing.
3341 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3342 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3345 * Busy processors will not participate in power savings
3348 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3349 sds->power_savings_balance = 0;
3351 sds->power_savings_balance = 1;
3352 sds->min_nr_running = ULONG_MAX;
3353 sds->leader_nr_running = 0;
3358 * update_sd_power_savings_stats - Update the power saving stats for a
3359 * sched_domain while performing load balancing.
3361 * @group: sched_group belonging to the sched_domain under consideration.
3362 * @sds: Variable containing the statistics of the sched_domain
3363 * @local_group: Does group contain the CPU for which we're performing
3365 * @sgs: Variable containing the statistics of the group.
3367 static inline void update_sd_power_savings_stats(struct sched_group *group,
3368 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3371 if (!sds->power_savings_balance)
3375 * If the local group is idle or completely loaded
3376 * no need to do power savings balance at this domain
3378 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3379 !sds->this_nr_running))
3380 sds->power_savings_balance = 0;
3383 * If a group is already running at full capacity or idle,
3384 * don't include that group in power savings calculations
3386 if (!sds->power_savings_balance ||
3387 sgs->sum_nr_running >= sgs->group_capacity ||
3388 !sgs->sum_nr_running)
3392 * Calculate the group which has the least non-idle load.
3393 * This is the group from where we need to pick up the load
3396 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3397 (sgs->sum_nr_running == sds->min_nr_running &&
3398 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3399 sds->group_min = group;
3400 sds->min_nr_running = sgs->sum_nr_running;
3401 sds->min_load_per_task = sgs->sum_weighted_load /
3402 sgs->sum_nr_running;
3406 * Calculate the group which is almost near its
3407 * capacity but still has some space to pick up some load
3408 * from other group and save more power
3410 if (sgs->sum_nr_running > sgs->group_capacity - 1)
3413 if (sgs->sum_nr_running > sds->leader_nr_running ||
3414 (sgs->sum_nr_running == sds->leader_nr_running &&
3415 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3416 sds->group_leader = group;
3417 sds->leader_nr_running = sgs->sum_nr_running;
3422 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3423 * @sds: Variable containing the statistics of the sched_domain
3424 * under consideration.
3425 * @this_cpu: Cpu at which we're currently performing load-balancing.
3426 * @imbalance: Variable to store the imbalance.
3429 * Check if we have potential to perform some power-savings balance.
3430 * If yes, set the busiest group to be the least loaded group in the
3431 * sched_domain, so that it's CPUs can be put to idle.
3433 * Returns 1 if there is potential to perform power-savings balance.
3436 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3437 int this_cpu, unsigned long *imbalance)
3439 if (!sds->power_savings_balance)
3442 if (sds->this != sds->group_leader ||
3443 sds->group_leader == sds->group_min)
3446 *imbalance = sds->min_load_per_task;
3447 sds->busiest = sds->group_min;
3449 if (sched_mc_power_savings >= POWERSAVINGS_BALANCE_WAKEUP) {
3450 cpu_rq(this_cpu)->rd->sched_mc_preferred_wakeup_cpu =
3451 group_first_cpu(sds->group_leader);
3457 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3458 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3459 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3464 static inline void update_sd_power_savings_stats(struct sched_group *group,
3465 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3470 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3471 int this_cpu, unsigned long *imbalance)
3475 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3479 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3480 * @group: sched_group whose statistics are to be updated.
3481 * @this_cpu: Cpu for which load balance is currently performed.
3482 * @idle: Idle status of this_cpu
3483 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3484 * @sd_idle: Idle status of the sched_domain containing group.
3485 * @local_group: Does group contain this_cpu.
3486 * @cpus: Set of cpus considered for load balancing.
3487 * @balance: Should we balance.
3488 * @sgs: variable to hold the statistics for this group.
3490 static inline void update_sg_lb_stats(struct sched_group *group, int this_cpu,
3491 enum cpu_idle_type idle, int load_idx, int *sd_idle,
3492 int local_group, const struct cpumask *cpus,
3493 int *balance, struct sg_lb_stats *sgs)
3495 unsigned long load, max_cpu_load, min_cpu_load;
3497 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3498 unsigned long sum_avg_load_per_task;
3499 unsigned long avg_load_per_task;
3502 balance_cpu = group_first_cpu(group);
3504 /* Tally up the load of all CPUs in the group */
3505 sum_avg_load_per_task = avg_load_per_task = 0;
3507 min_cpu_load = ~0UL;
3509 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3510 struct rq *rq = cpu_rq(i);
3512 if (*sd_idle && rq->nr_running)
3515 /* Bias balancing toward cpus of our domain */
3517 if (idle_cpu(i) && !first_idle_cpu) {
3522 load = target_load(i, load_idx);
3524 load = source_load(i, load_idx);
3525 if (load > max_cpu_load)
3526 max_cpu_load = load;
3527 if (min_cpu_load > load)
3528 min_cpu_load = load;
3531 sgs->group_load += load;
3532 sgs->sum_nr_running += rq->nr_running;
3533 sgs->sum_weighted_load += weighted_cpuload(i);
3535 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3539 * First idle cpu or the first cpu(busiest) in this sched group
3540 * is eligible for doing load balancing at this and above
3541 * domains. In the newly idle case, we will allow all the cpu's
3542 * to do the newly idle load balance.
3544 if (idle != CPU_NEWLY_IDLE && local_group &&
3545 balance_cpu != this_cpu && balance) {
3550 /* Adjust by relative CPU power of the group */
3551 sgs->avg_load = sg_div_cpu_power(group,
3552 sgs->group_load * SCHED_LOAD_SCALE);
3556 * Consider the group unbalanced when the imbalance is larger
3557 * than the average weight of two tasks.
3559 * APZ: with cgroup the avg task weight can vary wildly and
3560 * might not be a suitable number - should we keep a
3561 * normalized nr_running number somewhere that negates
3564 avg_load_per_task = sg_div_cpu_power(group,
3565 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3567 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3570 sgs->group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3575 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3576 * @sd: sched_domain whose statistics are to be updated.
3577 * @this_cpu: Cpu for which load balance is currently performed.
3578 * @idle: Idle status of this_cpu
3579 * @sd_idle: Idle status of the sched_domain containing group.
3580 * @cpus: Set of cpus considered for load balancing.
3581 * @balance: Should we balance.
3582 * @sds: variable to hold the statistics for this sched_domain.
3584 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3585 enum cpu_idle_type idle, int *sd_idle,
3586 const struct cpumask *cpus, int *balance,
3587 struct sd_lb_stats *sds)
3589 struct sched_group *group = sd->groups;
3590 struct sg_lb_stats sgs;
3593 init_sd_power_savings_stats(sd, sds, idle);
3594 load_idx = get_sd_load_idx(sd, idle);
3599 local_group = cpumask_test_cpu(this_cpu,
3600 sched_group_cpus(group));
3601 memset(&sgs, 0, sizeof(sgs));
3602 update_sg_lb_stats(group, this_cpu, idle, load_idx, sd_idle,
3603 local_group, cpus, balance, &sgs);
3605 if (local_group && balance && !(*balance))
3608 sds->total_load += sgs.group_load;
3609 sds->total_pwr += group->__cpu_power;
3612 sds->this_load = sgs.avg_load;
3614 sds->this_nr_running = sgs.sum_nr_running;
3615 sds->this_load_per_task = sgs.sum_weighted_load;
3616 } else if (sgs.avg_load > sds->max_load &&
3617 (sgs.sum_nr_running > sgs.group_capacity ||
3619 sds->max_load = sgs.avg_load;
3620 sds->busiest = group;
3621 sds->busiest_nr_running = sgs.sum_nr_running;
3622 sds->busiest_load_per_task = sgs.sum_weighted_load;
3623 sds->group_imb = sgs.group_imb;
3626 update_sd_power_savings_stats(group, sds, local_group, &sgs);
3627 group = group->next;
3628 } while (group != sd->groups);
3633 * fix_small_imbalance - Calculate the minor imbalance that exists
3634 * amongst the groups of a sched_domain, during
3636 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3637 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3638 * @imbalance: Variable to store the imbalance.
3640 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3641 int this_cpu, unsigned long *imbalance)
3643 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3644 unsigned int imbn = 2;
3646 if (sds->this_nr_running) {
3647 sds->this_load_per_task /= sds->this_nr_running;
3648 if (sds->busiest_load_per_task >
3649 sds->this_load_per_task)
3652 sds->this_load_per_task =
3653 cpu_avg_load_per_task(this_cpu);
3655 if (sds->max_load - sds->this_load + sds->busiest_load_per_task >=
3656 sds->busiest_load_per_task * imbn) {
3657 *imbalance = sds->busiest_load_per_task;
3662 * OK, we don't have enough imbalance to justify moving tasks,
3663 * however we may be able to increase total CPU power used by
3667 pwr_now += sds->busiest->__cpu_power *
3668 min(sds->busiest_load_per_task, sds->max_load);
3669 pwr_now += sds->this->__cpu_power *
3670 min(sds->this_load_per_task, sds->this_load);
3671 pwr_now /= SCHED_LOAD_SCALE;
3673 /* Amount of load we'd subtract */
3674 tmp = sg_div_cpu_power(sds->busiest,
3675 sds->busiest_load_per_task * SCHED_LOAD_SCALE);
3676 if (sds->max_load > tmp)
3677 pwr_move += sds->busiest->__cpu_power *
3678 min(sds->busiest_load_per_task, sds->max_load - tmp);
3680 /* Amount of load we'd add */
3681 if (sds->max_load * sds->busiest->__cpu_power <
3682 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
3683 tmp = sg_div_cpu_power(sds->this,
3684 sds->max_load * sds->busiest->__cpu_power);
3686 tmp = sg_div_cpu_power(sds->this,
3687 sds->busiest_load_per_task * SCHED_LOAD_SCALE);
3688 pwr_move += sds->this->__cpu_power *
3689 min(sds->this_load_per_task, sds->this_load + tmp);
3690 pwr_move /= SCHED_LOAD_SCALE;
3692 /* Move if we gain throughput */
3693 if (pwr_move > pwr_now)
3694 *imbalance = sds->busiest_load_per_task;
3698 * calculate_imbalance - Calculate the amount of imbalance present within the
3699 * groups of a given sched_domain during load balance.
3700 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3701 * @this_cpu: Cpu for which currently load balance is being performed.
3702 * @imbalance: The variable to store the imbalance.
3704 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3705 unsigned long *imbalance)
3707 unsigned long max_pull;
3709 * In the presence of smp nice balancing, certain scenarios can have
3710 * max load less than avg load(as we skip the groups at or below
3711 * its cpu_power, while calculating max_load..)
3713 if (sds->max_load < sds->avg_load) {
3715 return fix_small_imbalance(sds, this_cpu, imbalance);
3718 /* Don't want to pull so many tasks that a group would go idle */
3719 max_pull = min(sds->max_load - sds->avg_load,
3720 sds->max_load - sds->busiest_load_per_task);
3722 /* How much load to actually move to equalise the imbalance */
3723 *imbalance = min(max_pull * sds->busiest->__cpu_power,
3724 (sds->avg_load - sds->this_load) * sds->this->__cpu_power)
3728 * if *imbalance is less than the average load per runnable task
3729 * there is no gaurantee that any tasks will be moved so we'll have
3730 * a think about bumping its value to force at least one task to be
3733 if (*imbalance < sds->busiest_load_per_task)
3734 return fix_small_imbalance(sds, this_cpu, imbalance);
3737 /******* find_busiest_group() helpers end here *********************/
3740 * find_busiest_group - Returns the busiest group within the sched_domain
3741 * if there is an imbalance. If there isn't an imbalance, and
3742 * the user has opted for power-savings, it returns a group whose
3743 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3744 * such a group exists.
3746 * Also calculates the amount of weighted load which should be moved
3747 * to restore balance.
3749 * @sd: The sched_domain whose busiest group is to be returned.
3750 * @this_cpu: The cpu for which load balancing is currently being performed.
3751 * @imbalance: Variable which stores amount of weighted load which should
3752 * be moved to restore balance/put a group to idle.
3753 * @idle: The idle status of this_cpu.
3754 * @sd_idle: The idleness of sd
3755 * @cpus: The set of CPUs under consideration for load-balancing.
3756 * @balance: Pointer to a variable indicating if this_cpu
3757 * is the appropriate cpu to perform load balancing at this_level.
3759 * Returns: - the busiest group if imbalance exists.
3760 * - If no imbalance and user has opted for power-savings balance,
3761 * return the least loaded group whose CPUs can be
3762 * put to idle by rebalancing its tasks onto our group.
3764 static struct sched_group *
3765 find_busiest_group(struct sched_domain *sd, int this_cpu,
3766 unsigned long *imbalance, enum cpu_idle_type idle,
3767 int *sd_idle, const struct cpumask *cpus, int *balance)
3769 struct sd_lb_stats sds;
3771 memset(&sds, 0, sizeof(sds));
3774 * Compute the various statistics relavent for load balancing at
3777 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
3780 /* Cases where imbalance does not exist from POV of this_cpu */
3781 /* 1) this_cpu is not the appropriate cpu to perform load balancing
3783 * 2) There is no busy sibling group to pull from.
3784 * 3) This group is the busiest group.
3785 * 4) This group is more busy than the avg busieness at this
3787 * 5) The imbalance is within the specified limit.
3788 * 6) Any rebalance would lead to ping-pong
3790 if (balance && !(*balance))
3793 if (!sds.busiest || sds.busiest_nr_running == 0)
3796 if (sds.this_load >= sds.max_load)
3799 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
3801 if (sds.this_load >= sds.avg_load)
3804 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
3807 sds.busiest_load_per_task /= sds.busiest_nr_running;
3809 sds.busiest_load_per_task =
3810 min(sds.busiest_load_per_task, sds.avg_load);
3813 * We're trying to get all the cpus to the average_load, so we don't
3814 * want to push ourselves above the average load, nor do we wish to
3815 * reduce the max loaded cpu below the average load, as either of these
3816 * actions would just result in more rebalancing later, and ping-pong
3817 * tasks around. Thus we look for the minimum possible imbalance.
3818 * Negative imbalances (*we* are more loaded than anyone else) will
3819 * be counted as no imbalance for these purposes -- we can't fix that
3820 * by pulling tasks to us. Be careful of negative numbers as they'll
3821 * appear as very large values with unsigned longs.
3823 if (sds.max_load <= sds.busiest_load_per_task)
3826 /* Looks like there is an imbalance. Compute it */
3827 calculate_imbalance(&sds, this_cpu, imbalance);
3832 * There is no obvious imbalance. But check if we can do some balancing
3835 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
3843 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3846 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3847 unsigned long imbalance, const struct cpumask *cpus)
3849 struct rq *busiest = NULL, *rq;
3850 unsigned long max_load = 0;
3853 for_each_cpu(i, sched_group_cpus(group)) {
3856 if (!cpumask_test_cpu(i, cpus))
3860 wl = weighted_cpuload(i);
3862 if (rq->nr_running == 1 && wl > imbalance)
3865 if (wl > max_load) {
3875 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3876 * so long as it is large enough.
3878 #define MAX_PINNED_INTERVAL 512
3880 /* Working cpumask for load_balance and load_balance_newidle. */
3881 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
3884 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3885 * tasks if there is an imbalance.
3887 static int load_balance(int this_cpu, struct rq *this_rq,
3888 struct sched_domain *sd, enum cpu_idle_type idle,
3891 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3892 struct sched_group *group;
3893 unsigned long imbalance;
3895 unsigned long flags;
3896 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
3898 cpumask_setall(cpus);
3901 * When power savings policy is enabled for the parent domain, idle
3902 * sibling can pick up load irrespective of busy siblings. In this case,
3903 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3904 * portraying it as CPU_NOT_IDLE.
3906 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3907 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3910 schedstat_inc(sd, lb_count[idle]);
3914 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3921 schedstat_inc(sd, lb_nobusyg[idle]);
3925 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3927 schedstat_inc(sd, lb_nobusyq[idle]);
3931 BUG_ON(busiest == this_rq);
3933 schedstat_add(sd, lb_imbalance[idle], imbalance);
3936 if (busiest->nr_running > 1) {
3938 * Attempt to move tasks. If find_busiest_group has found
3939 * an imbalance but busiest->nr_running <= 1, the group is
3940 * still unbalanced. ld_moved simply stays zero, so it is
3941 * correctly treated as an imbalance.
3943 local_irq_save(flags);
3944 double_rq_lock(this_rq, busiest);
3945 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3946 imbalance, sd, idle, &all_pinned);
3947 double_rq_unlock(this_rq, busiest);
3948 local_irq_restore(flags);
3951 * some other cpu did the load balance for us.
3953 if (ld_moved && this_cpu != smp_processor_id())
3954 resched_cpu(this_cpu);
3956 /* All tasks on this runqueue were pinned by CPU affinity */
3957 if (unlikely(all_pinned)) {
3958 cpumask_clear_cpu(cpu_of(busiest), cpus);
3959 if (!cpumask_empty(cpus))
3966 schedstat_inc(sd, lb_failed[idle]);
3967 sd->nr_balance_failed++;
3969 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3971 spin_lock_irqsave(&busiest->lock, flags);
3973 /* don't kick the migration_thread, if the curr
3974 * task on busiest cpu can't be moved to this_cpu
3976 if (!cpumask_test_cpu(this_cpu,
3977 &busiest->curr->cpus_allowed)) {
3978 spin_unlock_irqrestore(&busiest->lock, flags);
3980 goto out_one_pinned;
3983 if (!busiest->active_balance) {
3984 busiest->active_balance = 1;
3985 busiest->push_cpu = this_cpu;
3988 spin_unlock_irqrestore(&busiest->lock, flags);
3990 wake_up_process(busiest->migration_thread);
3993 * We've kicked active balancing, reset the failure
3996 sd->nr_balance_failed = sd->cache_nice_tries+1;
3999 sd->nr_balance_failed = 0;
4001 if (likely(!active_balance)) {
4002 /* We were unbalanced, so reset the balancing interval */
4003 sd->balance_interval = sd->min_interval;
4006 * If we've begun active balancing, start to back off. This
4007 * case may not be covered by the all_pinned logic if there
4008 * is only 1 task on the busy runqueue (because we don't call
4011 if (sd->balance_interval < sd->max_interval)
4012 sd->balance_interval *= 2;
4015 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4016 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4022 schedstat_inc(sd, lb_balanced[idle]);
4024 sd->nr_balance_failed = 0;
4027 /* tune up the balancing interval */
4028 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4029 (sd->balance_interval < sd->max_interval))
4030 sd->balance_interval *= 2;
4032 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4033 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4044 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4045 * tasks if there is an imbalance.
4047 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4048 * this_rq is locked.
4051 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
4053 struct sched_group *group;
4054 struct rq *busiest = NULL;
4055 unsigned long imbalance;
4059 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4061 cpumask_setall(cpus);
4064 * When power savings policy is enabled for the parent domain, idle
4065 * sibling can pick up load irrespective of busy siblings. In this case,
4066 * let the state of idle sibling percolate up as IDLE, instead of
4067 * portraying it as CPU_NOT_IDLE.
4069 if (sd->flags & SD_SHARE_CPUPOWER &&
4070 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4073 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
4075 update_shares_locked(this_rq, sd);
4076 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
4077 &sd_idle, cpus, NULL);
4079 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
4083 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
4085 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
4089 BUG_ON(busiest == this_rq);
4091 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
4094 if (busiest->nr_running > 1) {
4095 /* Attempt to move tasks */
4096 double_lock_balance(this_rq, busiest);
4097 /* this_rq->clock is already updated */
4098 update_rq_clock(busiest);
4099 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4100 imbalance, sd, CPU_NEWLY_IDLE,
4102 double_unlock_balance(this_rq, busiest);
4104 if (unlikely(all_pinned)) {
4105 cpumask_clear_cpu(cpu_of(busiest), cpus);
4106 if (!cpumask_empty(cpus))
4112 int active_balance = 0;
4114 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
4115 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4116 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4119 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4122 if (sd->nr_balance_failed++ < 2)
4126 * The only task running in a non-idle cpu can be moved to this
4127 * cpu in an attempt to completely freeup the other CPU
4128 * package. The same method used to move task in load_balance()
4129 * have been extended for load_balance_newidle() to speedup
4130 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4132 * The package power saving logic comes from
4133 * find_busiest_group(). If there are no imbalance, then
4134 * f_b_g() will return NULL. However when sched_mc={1,2} then
4135 * f_b_g() will select a group from which a running task may be
4136 * pulled to this cpu in order to make the other package idle.
4137 * If there is no opportunity to make a package idle and if
4138 * there are no imbalance, then f_b_g() will return NULL and no
4139 * action will be taken in load_balance_newidle().
4141 * Under normal task pull operation due to imbalance, there
4142 * will be more than one task in the source run queue and
4143 * move_tasks() will succeed. ld_moved will be true and this
4144 * active balance code will not be triggered.
4147 /* Lock busiest in correct order while this_rq is held */
4148 double_lock_balance(this_rq, busiest);
4151 * don't kick the migration_thread, if the curr
4152 * task on busiest cpu can't be moved to this_cpu
4154 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
4155 double_unlock_balance(this_rq, busiest);
4160 if (!busiest->active_balance) {
4161 busiest->active_balance = 1;
4162 busiest->push_cpu = this_cpu;
4166 double_unlock_balance(this_rq, busiest);
4168 * Should not call ttwu while holding a rq->lock
4170 spin_unlock(&this_rq->lock);
4172 wake_up_process(busiest->migration_thread);
4173 spin_lock(&this_rq->lock);
4176 sd->nr_balance_failed = 0;
4178 update_shares_locked(this_rq, sd);
4182 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
4183 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4184 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4186 sd->nr_balance_failed = 0;
4192 * idle_balance is called by schedule() if this_cpu is about to become
4193 * idle. Attempts to pull tasks from other CPUs.
4195 static void idle_balance(int this_cpu, struct rq *this_rq)
4197 struct sched_domain *sd;
4198 int pulled_task = 0;
4199 unsigned long next_balance = jiffies + HZ;
4201 for_each_domain(this_cpu, sd) {
4202 unsigned long interval;
4204 if (!(sd->flags & SD_LOAD_BALANCE))
4207 if (sd->flags & SD_BALANCE_NEWIDLE)
4208 /* If we've pulled tasks over stop searching: */
4209 pulled_task = load_balance_newidle(this_cpu, this_rq,
4212 interval = msecs_to_jiffies(sd->balance_interval);
4213 if (time_after(next_balance, sd->last_balance + interval))
4214 next_balance = sd->last_balance + interval;
4218 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4220 * We are going idle. next_balance may be set based on
4221 * a busy processor. So reset next_balance.
4223 this_rq->next_balance = next_balance;
4228 * active_load_balance is run by migration threads. It pushes running tasks
4229 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4230 * running on each physical CPU where possible, and avoids physical /
4231 * logical imbalances.
4233 * Called with busiest_rq locked.
4235 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
4237 int target_cpu = busiest_rq->push_cpu;
4238 struct sched_domain *sd;
4239 struct rq *target_rq;
4241 /* Is there any task to move? */
4242 if (busiest_rq->nr_running <= 1)
4245 target_rq = cpu_rq(target_cpu);
4248 * This condition is "impossible", if it occurs
4249 * we need to fix it. Originally reported by
4250 * Bjorn Helgaas on a 128-cpu setup.
4252 BUG_ON(busiest_rq == target_rq);
4254 /* move a task from busiest_rq to target_rq */
4255 double_lock_balance(busiest_rq, target_rq);
4256 update_rq_clock(busiest_rq);
4257 update_rq_clock(target_rq);
4259 /* Search for an sd spanning us and the target CPU. */
4260 for_each_domain(target_cpu, sd) {
4261 if ((sd->flags & SD_LOAD_BALANCE) &&
4262 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4267 schedstat_inc(sd, alb_count);
4269 if (move_one_task(target_rq, target_cpu, busiest_rq,
4271 schedstat_inc(sd, alb_pushed);
4273 schedstat_inc(sd, alb_failed);
4275 double_unlock_balance(busiest_rq, target_rq);
4280 atomic_t load_balancer;
4281 cpumask_var_t cpu_mask;
4282 } nohz ____cacheline_aligned = {
4283 .load_balancer = ATOMIC_INIT(-1),
4287 * This routine will try to nominate the ilb (idle load balancing)
4288 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4289 * load balancing on behalf of all those cpus. If all the cpus in the system
4290 * go into this tickless mode, then there will be no ilb owner (as there is
4291 * no need for one) and all the cpus will sleep till the next wakeup event
4294 * For the ilb owner, tick is not stopped. And this tick will be used
4295 * for idle load balancing. ilb owner will still be part of
4298 * While stopping the tick, this cpu will become the ilb owner if there
4299 * is no other owner. And will be the owner till that cpu becomes busy
4300 * or if all cpus in the system stop their ticks at which point
4301 * there is no need for ilb owner.
4303 * When the ilb owner becomes busy, it nominates another owner, during the
4304 * next busy scheduler_tick()
4306 int select_nohz_load_balancer(int stop_tick)
4308 int cpu = smp_processor_id();
4311 cpu_rq(cpu)->in_nohz_recently = 1;
4313 if (!cpu_active(cpu)) {
4314 if (atomic_read(&nohz.load_balancer) != cpu)
4318 * If we are going offline and still the leader,
4321 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4327 cpumask_set_cpu(cpu, nohz.cpu_mask);
4329 /* time for ilb owner also to sleep */
4330 if (cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4331 if (atomic_read(&nohz.load_balancer) == cpu)
4332 atomic_set(&nohz.load_balancer, -1);
4336 if (atomic_read(&nohz.load_balancer) == -1) {
4337 /* make me the ilb owner */
4338 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4340 } else if (atomic_read(&nohz.load_balancer) == cpu)
4343 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4346 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4348 if (atomic_read(&nohz.load_balancer) == cpu)
4349 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4356 static DEFINE_SPINLOCK(balancing);
4359 * It checks each scheduling domain to see if it is due to be balanced,
4360 * and initiates a balancing operation if so.
4362 * Balancing parameters are set up in arch_init_sched_domains.
4364 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4367 struct rq *rq = cpu_rq(cpu);
4368 unsigned long interval;
4369 struct sched_domain *sd;
4370 /* Earliest time when we have to do rebalance again */
4371 unsigned long next_balance = jiffies + 60*HZ;
4372 int update_next_balance = 0;
4375 for_each_domain(cpu, sd) {
4376 if (!(sd->flags & SD_LOAD_BALANCE))
4379 interval = sd->balance_interval;
4380 if (idle != CPU_IDLE)
4381 interval *= sd->busy_factor;
4383 /* scale ms to jiffies */
4384 interval = msecs_to_jiffies(interval);
4385 if (unlikely(!interval))
4387 if (interval > HZ*NR_CPUS/10)
4388 interval = HZ*NR_CPUS/10;
4390 need_serialize = sd->flags & SD_SERIALIZE;
4392 if (need_serialize) {
4393 if (!spin_trylock(&balancing))
4397 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4398 if (load_balance(cpu, rq, sd, idle, &balance)) {
4400 * We've pulled tasks over so either we're no
4401 * longer idle, or one of our SMT siblings is
4404 idle = CPU_NOT_IDLE;
4406 sd->last_balance = jiffies;
4409 spin_unlock(&balancing);
4411 if (time_after(next_balance, sd->last_balance + interval)) {
4412 next_balance = sd->last_balance + interval;
4413 update_next_balance = 1;
4417 * Stop the load balance at this level. There is another
4418 * CPU in our sched group which is doing load balancing more
4426 * next_balance will be updated only when there is a need.
4427 * When the cpu is attached to null domain for ex, it will not be
4430 if (likely(update_next_balance))
4431 rq->next_balance = next_balance;
4435 * run_rebalance_domains is triggered when needed from the scheduler tick.
4436 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4437 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4439 static void run_rebalance_domains(struct softirq_action *h)
4441 int this_cpu = smp_processor_id();
4442 struct rq *this_rq = cpu_rq(this_cpu);
4443 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4444 CPU_IDLE : CPU_NOT_IDLE;
4446 rebalance_domains(this_cpu, idle);
4450 * If this cpu is the owner for idle load balancing, then do the
4451 * balancing on behalf of the other idle cpus whose ticks are
4454 if (this_rq->idle_at_tick &&
4455 atomic_read(&nohz.load_balancer) == this_cpu) {
4459 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4460 if (balance_cpu == this_cpu)
4464 * If this cpu gets work to do, stop the load balancing
4465 * work being done for other cpus. Next load
4466 * balancing owner will pick it up.
4471 rebalance_domains(balance_cpu, CPU_IDLE);
4473 rq = cpu_rq(balance_cpu);
4474 if (time_after(this_rq->next_balance, rq->next_balance))
4475 this_rq->next_balance = rq->next_balance;
4481 static inline int on_null_domain(int cpu)
4483 return !rcu_dereference(cpu_rq(cpu)->sd);
4487 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4489 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4490 * idle load balancing owner or decide to stop the periodic load balancing,
4491 * if the whole system is idle.
4493 static inline void trigger_load_balance(struct rq *rq, int cpu)
4497 * If we were in the nohz mode recently and busy at the current
4498 * scheduler tick, then check if we need to nominate new idle
4501 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4502 rq->in_nohz_recently = 0;
4504 if (atomic_read(&nohz.load_balancer) == cpu) {
4505 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4506 atomic_set(&nohz.load_balancer, -1);
4509 if (atomic_read(&nohz.load_balancer) == -1) {
4511 * simple selection for now: Nominate the
4512 * first cpu in the nohz list to be the next
4515 * TBD: Traverse the sched domains and nominate
4516 * the nearest cpu in the nohz.cpu_mask.
4518 int ilb = cpumask_first(nohz.cpu_mask);
4520 if (ilb < nr_cpu_ids)
4526 * If this cpu is idle and doing idle load balancing for all the
4527 * cpus with ticks stopped, is it time for that to stop?
4529 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4530 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4536 * If this cpu is idle and the idle load balancing is done by
4537 * someone else, then no need raise the SCHED_SOFTIRQ
4539 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4540 cpumask_test_cpu(cpu, nohz.cpu_mask))
4543 /* Don't need to rebalance while attached to NULL domain */
4544 if (time_after_eq(jiffies, rq->next_balance) &&
4545 likely(!on_null_domain(cpu)))
4546 raise_softirq(SCHED_SOFTIRQ);
4549 #else /* CONFIG_SMP */
4552 * on UP we do not need to balance between CPUs:
4554 static inline void idle_balance(int cpu, struct rq *rq)
4560 DEFINE_PER_CPU(struct kernel_stat, kstat);
4562 EXPORT_PER_CPU_SYMBOL(kstat);
4565 * Return any ns on the sched_clock that have not yet been accounted in
4566 * @p in case that task is currently running.
4568 * Called with task_rq_lock() held on @rq.
4570 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
4574 if (task_current(rq, p)) {
4575 update_rq_clock(rq);
4576 ns = rq->clock - p->se.exec_start;
4584 unsigned long long task_delta_exec(struct task_struct *p)
4586 unsigned long flags;
4590 rq = task_rq_lock(p, &flags);
4591 ns = do_task_delta_exec(p, rq);
4592 task_rq_unlock(rq, &flags);
4598 * Return accounted runtime for the task.
4599 * In case the task is currently running, return the runtime plus current's
4600 * pending runtime that have not been accounted yet.
4602 unsigned long long task_sched_runtime(struct task_struct *p)
4604 unsigned long flags;
4608 rq = task_rq_lock(p, &flags);
4609 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
4610 task_rq_unlock(rq, &flags);
4616 * Return sum_exec_runtime for the thread group.
4617 * In case the task is currently running, return the sum plus current's
4618 * pending runtime that have not been accounted yet.
4620 * Note that the thread group might have other running tasks as well,
4621 * so the return value not includes other pending runtime that other
4622 * running tasks might have.
4624 unsigned long long thread_group_sched_runtime(struct task_struct *p)
4626 struct task_cputime totals;
4627 unsigned long flags;
4631 rq = task_rq_lock(p, &flags);
4632 thread_group_cputime(p, &totals);
4633 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
4634 task_rq_unlock(rq, &flags);
4640 * Account user cpu time to a process.
4641 * @p: the process that the cpu time gets accounted to
4642 * @cputime: the cpu time spent in user space since the last update
4643 * @cputime_scaled: cputime scaled by cpu frequency
4645 void account_user_time(struct task_struct *p, cputime_t cputime,
4646 cputime_t cputime_scaled)
4648 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4651 /* Add user time to process. */
4652 p->utime = cputime_add(p->utime, cputime);
4653 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4654 account_group_user_time(p, cputime);
4656 /* Add user time to cpustat. */
4657 tmp = cputime_to_cputime64(cputime);
4658 if (TASK_NICE(p) > 0)
4659 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4661 cpustat->user = cputime64_add(cpustat->user, tmp);
4663 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
4664 /* Account for user time used */
4665 acct_update_integrals(p);
4669 * Account guest cpu time to a process.
4670 * @p: the process that the cpu time gets accounted to
4671 * @cputime: the cpu time spent in virtual machine since the last update
4672 * @cputime_scaled: cputime scaled by cpu frequency
4674 static void account_guest_time(struct task_struct *p, cputime_t cputime,
4675 cputime_t cputime_scaled)
4678 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4680 tmp = cputime_to_cputime64(cputime);
4682 /* Add guest time to process. */
4683 p->utime = cputime_add(p->utime, cputime);
4684 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4685 account_group_user_time(p, cputime);
4686 p->gtime = cputime_add(p->gtime, cputime);
4688 /* Add guest time to cpustat. */
4689 cpustat->user = cputime64_add(cpustat->user, tmp);
4690 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4694 * Account system cpu time to a process.
4695 * @p: the process that the cpu time gets accounted to
4696 * @hardirq_offset: the offset to subtract from hardirq_count()
4697 * @cputime: the cpu time spent in kernel space since the last update
4698 * @cputime_scaled: cputime scaled by cpu frequency
4700 void account_system_time(struct task_struct *p, int hardirq_offset,
4701 cputime_t cputime, cputime_t cputime_scaled)
4703 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4706 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4707 account_guest_time(p, cputime, cputime_scaled);
4711 /* Add system time to process. */
4712 p->stime = cputime_add(p->stime, cputime);
4713 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
4714 account_group_system_time(p, cputime);
4716 /* Add system time to cpustat. */
4717 tmp = cputime_to_cputime64(cputime);
4718 if (hardirq_count() - hardirq_offset)
4719 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4720 else if (softirq_count())
4721 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4723 cpustat->system = cputime64_add(cpustat->system, tmp);
4725 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
4727 /* Account for system time used */
4728 acct_update_integrals(p);
4732 * Account for involuntary wait time.
4733 * @steal: the cpu time spent in involuntary wait
4735 void account_steal_time(cputime_t cputime)
4737 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4738 cputime64_t cputime64 = cputime_to_cputime64(cputime);
4740 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
4744 * Account for idle time.
4745 * @cputime: the cpu time spent in idle wait
4747 void account_idle_time(cputime_t cputime)
4749 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4750 cputime64_t cputime64 = cputime_to_cputime64(cputime);
4751 struct rq *rq = this_rq();
4753 if (atomic_read(&rq->nr_iowait) > 0)
4754 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
4756 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
4759 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
4762 * Account a single tick of cpu time.
4763 * @p: the process that the cpu time gets accounted to
4764 * @user_tick: indicates if the tick is a user or a system tick
4766 void account_process_tick(struct task_struct *p, int user_tick)
4768 cputime_t one_jiffy = jiffies_to_cputime(1);
4769 cputime_t one_jiffy_scaled = cputime_to_scaled(one_jiffy);
4770 struct rq *rq = this_rq();
4773 account_user_time(p, one_jiffy, one_jiffy_scaled);
4774 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
4775 account_system_time(p, HARDIRQ_OFFSET, one_jiffy,
4778 account_idle_time(one_jiffy);
4782 * Account multiple ticks of steal time.
4783 * @p: the process from which the cpu time has been stolen
4784 * @ticks: number of stolen ticks
4786 void account_steal_ticks(unsigned long ticks)
4788 account_steal_time(jiffies_to_cputime(ticks));
4792 * Account multiple ticks of idle time.
4793 * @ticks: number of stolen ticks
4795 void account_idle_ticks(unsigned long ticks)
4797 account_idle_time(jiffies_to_cputime(ticks));
4803 * Use precise platform statistics if available:
4805 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4806 cputime_t task_utime(struct task_struct *p)
4811 cputime_t task_stime(struct task_struct *p)
4816 cputime_t task_utime(struct task_struct *p)
4818 clock_t utime = cputime_to_clock_t(p->utime),
4819 total = utime + cputime_to_clock_t(p->stime);
4823 * Use CFS's precise accounting:
4825 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
4829 do_div(temp, total);
4831 utime = (clock_t)temp;
4833 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
4834 return p->prev_utime;
4837 cputime_t task_stime(struct task_struct *p)
4842 * Use CFS's precise accounting. (we subtract utime from
4843 * the total, to make sure the total observed by userspace
4844 * grows monotonically - apps rely on that):
4846 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
4847 cputime_to_clock_t(task_utime(p));
4850 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
4852 return p->prev_stime;
4856 inline cputime_t task_gtime(struct task_struct *p)
4862 * This function gets called by the timer code, with HZ frequency.
4863 * We call it with interrupts disabled.
4865 * It also gets called by the fork code, when changing the parent's
4868 void scheduler_tick(void)
4870 int cpu = smp_processor_id();
4871 struct rq *rq = cpu_rq(cpu);
4872 struct task_struct *curr = rq->curr;
4876 spin_lock(&rq->lock);
4877 update_rq_clock(rq);
4878 update_cpu_load(rq);
4879 curr->sched_class->task_tick(rq, curr, 0);
4880 spin_unlock(&rq->lock);
4883 rq->idle_at_tick = idle_cpu(cpu);
4884 trigger_load_balance(rq, cpu);
4888 notrace unsigned long get_parent_ip(unsigned long addr)
4890 if (in_lock_functions(addr)) {
4891 addr = CALLER_ADDR2;
4892 if (in_lock_functions(addr))
4893 addr = CALLER_ADDR3;
4898 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4899 defined(CONFIG_PREEMPT_TRACER))
4901 void __kprobes add_preempt_count(int val)
4903 #ifdef CONFIG_DEBUG_PREEMPT
4907 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4910 preempt_count() += val;
4911 #ifdef CONFIG_DEBUG_PREEMPT
4913 * Spinlock count overflowing soon?
4915 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4918 if (preempt_count() == val)
4919 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4921 EXPORT_SYMBOL(add_preempt_count);
4923 void __kprobes sub_preempt_count(int val)
4925 #ifdef CONFIG_DEBUG_PREEMPT
4929 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4932 * Is the spinlock portion underflowing?
4934 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4935 !(preempt_count() & PREEMPT_MASK)))
4939 if (preempt_count() == val)
4940 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4941 preempt_count() -= val;
4943 EXPORT_SYMBOL(sub_preempt_count);
4948 * Print scheduling while atomic bug:
4950 static noinline void __schedule_bug(struct task_struct *prev)
4952 struct pt_regs *regs = get_irq_regs();
4954 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4955 prev->comm, prev->pid, preempt_count());
4957 debug_show_held_locks(prev);
4959 if (irqs_disabled())
4960 print_irqtrace_events(prev);
4969 * Various schedule()-time debugging checks and statistics:
4971 static inline void schedule_debug(struct task_struct *prev)
4974 * Test if we are atomic. Since do_exit() needs to call into
4975 * schedule() atomically, we ignore that path for now.
4976 * Otherwise, whine if we are scheduling when we should not be.
4978 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4979 __schedule_bug(prev);
4981 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4983 schedstat_inc(this_rq(), sched_count);
4984 #ifdef CONFIG_SCHEDSTATS
4985 if (unlikely(prev->lock_depth >= 0)) {
4986 schedstat_inc(this_rq(), bkl_count);
4987 schedstat_inc(prev, sched_info.bkl_count);
4992 static void put_prev_task(struct rq *rq, struct task_struct *prev)
4994 if (prev->state == TASK_RUNNING) {
4995 u64 runtime = prev->se.sum_exec_runtime;
4997 runtime -= prev->se.prev_sum_exec_runtime;
4998 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
5001 * In order to avoid avg_overlap growing stale when we are
5002 * indeed overlapping and hence not getting put to sleep, grow
5003 * the avg_overlap on preemption.
5005 * We use the average preemption runtime because that
5006 * correlates to the amount of cache footprint a task can
5009 update_avg(&prev->se.avg_overlap, runtime);
5011 prev->sched_class->put_prev_task(rq, prev);
5015 * Pick up the highest-prio task:
5017 static inline struct task_struct *
5018 pick_next_task(struct rq *rq)
5020 const struct sched_class *class;
5021 struct task_struct *p;
5024 * Optimization: we know that if all tasks are in
5025 * the fair class we can call that function directly:
5027 if (likely(rq->nr_running == rq->cfs.nr_running)) {
5028 p = fair_sched_class.pick_next_task(rq);
5033 class = sched_class_highest;
5035 p = class->pick_next_task(rq);
5039 * Will never be NULL as the idle class always
5040 * returns a non-NULL p:
5042 class = class->next;
5047 * schedule() is the main scheduler function.
5049 asmlinkage void __sched __schedule(void)
5051 struct task_struct *prev, *next;
5052 unsigned long *switch_count;
5056 cpu = smp_processor_id();
5060 switch_count = &prev->nivcsw;
5062 release_kernel_lock(prev);
5063 need_resched_nonpreemptible:
5065 schedule_debug(prev);
5067 if (sched_feat(HRTICK))
5070 spin_lock_irq(&rq->lock);
5071 update_rq_clock(rq);
5072 clear_tsk_need_resched(prev);
5074 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
5075 if (unlikely(signal_pending_state(prev->state, prev)))
5076 prev->state = TASK_RUNNING;
5078 deactivate_task(rq, prev, 1);
5079 switch_count = &prev->nvcsw;
5083 if (prev->sched_class->pre_schedule)
5084 prev->sched_class->pre_schedule(rq, prev);
5087 if (unlikely(!rq->nr_running))
5088 idle_balance(cpu, rq);
5090 put_prev_task(rq, prev);
5091 next = pick_next_task(rq);
5093 if (likely(prev != next)) {
5094 sched_info_switch(prev, next);
5100 context_switch(rq, prev, next); /* unlocks the rq */
5102 * the context switch might have flipped the stack from under
5103 * us, hence refresh the local variables.
5105 cpu = smp_processor_id();
5108 spin_unlock_irq(&rq->lock);
5110 if (unlikely(reacquire_kernel_lock(current) < 0))
5111 goto need_resched_nonpreemptible;
5114 asmlinkage void __sched schedule(void)
5119 preempt_enable_no_resched();
5120 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
5123 EXPORT_SYMBOL(schedule);
5127 * Look out! "owner" is an entirely speculative pointer
5128 * access and not reliable.
5130 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5135 if (!sched_feat(OWNER_SPIN))
5138 #ifdef CONFIG_DEBUG_PAGEALLOC
5140 * Need to access the cpu field knowing that
5141 * DEBUG_PAGEALLOC could have unmapped it if
5142 * the mutex owner just released it and exited.
5144 if (probe_kernel_address(&owner->cpu, cpu))
5151 * Even if the access succeeded (likely case),
5152 * the cpu field may no longer be valid.
5154 if (cpu >= nr_cpumask_bits)
5158 * We need to validate that we can do a
5159 * get_cpu() and that we have the percpu area.
5161 if (!cpu_online(cpu))
5168 * Owner changed, break to re-assess state.
5170 if (lock->owner != owner)
5174 * Is that owner really running on that cpu?
5176 if (task_thread_info(rq->curr) != owner || need_resched())
5186 #ifdef CONFIG_PREEMPT
5188 * this is the entry point to schedule() from in-kernel preemption
5189 * off of preempt_enable. Kernel preemptions off return from interrupt
5190 * occur there and call schedule directly.
5192 asmlinkage void __sched preempt_schedule(void)
5194 struct thread_info *ti = current_thread_info();
5197 * If there is a non-zero preempt_count or interrupts are disabled,
5198 * we do not want to preempt the current task. Just return..
5200 if (likely(ti->preempt_count || irqs_disabled()))
5204 add_preempt_count(PREEMPT_ACTIVE);
5206 sub_preempt_count(PREEMPT_ACTIVE);
5209 * Check again in case we missed a preemption opportunity
5210 * between schedule and now.
5213 } while (need_resched());
5215 EXPORT_SYMBOL(preempt_schedule);
5218 * this is the entry point to schedule() from kernel preemption
5219 * off of irq context.
5220 * Note, that this is called and return with irqs disabled. This will
5221 * protect us against recursive calling from irq.
5223 asmlinkage void __sched preempt_schedule_irq(void)
5225 struct thread_info *ti = current_thread_info();
5227 /* Catch callers which need to be fixed */
5228 BUG_ON(ti->preempt_count || !irqs_disabled());
5231 add_preempt_count(PREEMPT_ACTIVE);
5234 local_irq_disable();
5235 sub_preempt_count(PREEMPT_ACTIVE);
5238 * Check again in case we missed a preemption opportunity
5239 * between schedule and now.
5242 } while (need_resched());
5245 #endif /* CONFIG_PREEMPT */
5247 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
5250 return try_to_wake_up(curr->private, mode, sync);
5252 EXPORT_SYMBOL(default_wake_function);
5255 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5256 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5257 * number) then we wake all the non-exclusive tasks and one exclusive task.
5259 * There are circumstances in which we can try to wake a task which has already
5260 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5261 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5263 void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
5264 int nr_exclusive, int sync, void *key)
5266 wait_queue_t *curr, *next;
5268 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
5269 unsigned flags = curr->flags;
5271 if (curr->func(curr, mode, sync, key) &&
5272 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
5278 * __wake_up - wake up threads blocked on a waitqueue.
5280 * @mode: which threads
5281 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5282 * @key: is directly passed to the wakeup function
5284 void __wake_up(wait_queue_head_t *q, unsigned int mode,
5285 int nr_exclusive, void *key)
5287 unsigned long flags;
5289 spin_lock_irqsave(&q->lock, flags);
5290 __wake_up_common(q, mode, nr_exclusive, 0, key);
5291 spin_unlock_irqrestore(&q->lock, flags);
5293 EXPORT_SYMBOL(__wake_up);
5296 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5298 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
5300 __wake_up_common(q, mode, 1, 0, NULL);
5303 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
5305 __wake_up_common(q, mode, 1, 0, key);
5309 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5311 * @mode: which threads
5312 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5313 * @key: opaque value to be passed to wakeup targets
5315 * The sync wakeup differs that the waker knows that it will schedule
5316 * away soon, so while the target thread will be woken up, it will not
5317 * be migrated to another CPU - ie. the two threads are 'synchronized'
5318 * with each other. This can prevent needless bouncing between CPUs.
5320 * On UP it can prevent extra preemption.
5322 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
5323 int nr_exclusive, void *key)
5325 unsigned long flags;
5331 if (unlikely(!nr_exclusive))
5334 spin_lock_irqsave(&q->lock, flags);
5335 __wake_up_common(q, mode, nr_exclusive, sync, key);
5336 spin_unlock_irqrestore(&q->lock, flags);
5338 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
5341 * __wake_up_sync - see __wake_up_sync_key()
5343 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
5345 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
5347 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
5350 * complete: - signals a single thread waiting on this completion
5351 * @x: holds the state of this particular completion
5353 * This will wake up a single thread waiting on this completion. Threads will be
5354 * awakened in the same order in which they were queued.
5356 * See also complete_all(), wait_for_completion() and related routines.
5358 void complete(struct completion *x)
5360 unsigned long flags;
5362 spin_lock_irqsave(&x->wait.lock, flags);
5364 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
5365 spin_unlock_irqrestore(&x->wait.lock, flags);
5367 EXPORT_SYMBOL(complete);
5370 * complete_all: - signals all threads waiting on this completion
5371 * @x: holds the state of this particular completion
5373 * This will wake up all threads waiting on this particular completion event.
5375 void complete_all(struct completion *x)
5377 unsigned long flags;
5379 spin_lock_irqsave(&x->wait.lock, flags);
5380 x->done += UINT_MAX/2;
5381 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
5382 spin_unlock_irqrestore(&x->wait.lock, flags);
5384 EXPORT_SYMBOL(complete_all);
5386 static inline long __sched
5387 do_wait_for_common(struct completion *x, long timeout, int state)
5390 DECLARE_WAITQUEUE(wait, current);
5392 wait.flags |= WQ_FLAG_EXCLUSIVE;
5393 __add_wait_queue_tail(&x->wait, &wait);
5395 if (signal_pending_state(state, current)) {
5396 timeout = -ERESTARTSYS;
5399 __set_current_state(state);
5400 spin_unlock_irq(&x->wait.lock);
5401 timeout = schedule_timeout(timeout);
5402 spin_lock_irq(&x->wait.lock);
5403 } while (!x->done && timeout);
5404 __remove_wait_queue(&x->wait, &wait);
5409 return timeout ?: 1;
5413 wait_for_common(struct completion *x, long timeout, int state)
5417 spin_lock_irq(&x->wait.lock);
5418 timeout = do_wait_for_common(x, timeout, state);
5419 spin_unlock_irq(&x->wait.lock);
5424 * wait_for_completion: - waits for completion of a task
5425 * @x: holds the state of this particular completion
5427 * This waits to be signaled for completion of a specific task. It is NOT
5428 * interruptible and there is no timeout.
5430 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5431 * and interrupt capability. Also see complete().
5433 void __sched wait_for_completion(struct completion *x)
5435 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
5437 EXPORT_SYMBOL(wait_for_completion);
5440 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5441 * @x: holds the state of this particular completion
5442 * @timeout: timeout value in jiffies
5444 * This waits for either a completion of a specific task to be signaled or for a
5445 * specified timeout to expire. The timeout is in jiffies. It is not
5448 unsigned long __sched
5449 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5451 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5453 EXPORT_SYMBOL(wait_for_completion_timeout);
5456 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5457 * @x: holds the state of this particular completion
5459 * This waits for completion of a specific task to be signaled. It is
5462 int __sched wait_for_completion_interruptible(struct completion *x)
5464 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5465 if (t == -ERESTARTSYS)
5469 EXPORT_SYMBOL(wait_for_completion_interruptible);
5472 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5473 * @x: holds the state of this particular completion
5474 * @timeout: timeout value in jiffies
5476 * This waits for either a completion of a specific task to be signaled or for a
5477 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5479 unsigned long __sched
5480 wait_for_completion_interruptible_timeout(struct completion *x,
5481 unsigned long timeout)
5483 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
5485 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5488 * wait_for_completion_killable: - waits for completion of a task (killable)
5489 * @x: holds the state of this particular completion
5491 * This waits to be signaled for completion of a specific task. It can be
5492 * interrupted by a kill signal.
5494 int __sched wait_for_completion_killable(struct completion *x)
5496 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5497 if (t == -ERESTARTSYS)
5501 EXPORT_SYMBOL(wait_for_completion_killable);
5504 * try_wait_for_completion - try to decrement a completion without blocking
5505 * @x: completion structure
5507 * Returns: 0 if a decrement cannot be done without blocking
5508 * 1 if a decrement succeeded.
5510 * If a completion is being used as a counting completion,
5511 * attempt to decrement the counter without blocking. This
5512 * enables us to avoid waiting if the resource the completion
5513 * is protecting is not available.
5515 bool try_wait_for_completion(struct completion *x)
5519 spin_lock_irq(&x->wait.lock);
5524 spin_unlock_irq(&x->wait.lock);
5527 EXPORT_SYMBOL(try_wait_for_completion);
5530 * completion_done - Test to see if a completion has any waiters
5531 * @x: completion structure
5533 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5534 * 1 if there are no waiters.
5537 bool completion_done(struct completion *x)
5541 spin_lock_irq(&x->wait.lock);
5544 spin_unlock_irq(&x->wait.lock);
5547 EXPORT_SYMBOL(completion_done);
5550 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
5552 unsigned long flags;
5555 init_waitqueue_entry(&wait, current);
5557 __set_current_state(state);
5559 spin_lock_irqsave(&q->lock, flags);
5560 __add_wait_queue(q, &wait);
5561 spin_unlock(&q->lock);
5562 timeout = schedule_timeout(timeout);
5563 spin_lock_irq(&q->lock);
5564 __remove_wait_queue(q, &wait);
5565 spin_unlock_irqrestore(&q->lock, flags);
5570 void __sched interruptible_sleep_on(wait_queue_head_t *q)
5572 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5574 EXPORT_SYMBOL(interruptible_sleep_on);
5577 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
5579 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
5581 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
5583 void __sched sleep_on(wait_queue_head_t *q)
5585 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5587 EXPORT_SYMBOL(sleep_on);
5589 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
5591 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
5593 EXPORT_SYMBOL(sleep_on_timeout);
5595 #ifdef CONFIG_RT_MUTEXES
5598 * rt_mutex_setprio - set the current priority of a task
5600 * @prio: prio value (kernel-internal form)
5602 * This function changes the 'effective' priority of a task. It does
5603 * not touch ->normal_prio like __setscheduler().
5605 * Used by the rt_mutex code to implement priority inheritance logic.
5607 void rt_mutex_setprio(struct task_struct *p, int prio)
5609 unsigned long flags;
5610 int oldprio, on_rq, running;
5612 const struct sched_class *prev_class = p->sched_class;
5614 BUG_ON(prio < 0 || prio > MAX_PRIO);
5616 rq = task_rq_lock(p, &flags);
5617 update_rq_clock(rq);
5620 on_rq = p->se.on_rq;
5621 running = task_current(rq, p);
5623 dequeue_task(rq, p, 0);
5625 p->sched_class->put_prev_task(rq, p);
5628 p->sched_class = &rt_sched_class;
5630 p->sched_class = &fair_sched_class;
5635 p->sched_class->set_curr_task(rq);
5637 enqueue_task(rq, p, 0);
5639 check_class_changed(rq, p, prev_class, oldprio, running);
5641 task_rq_unlock(rq, &flags);
5646 void set_user_nice(struct task_struct *p, long nice)
5648 int old_prio, delta, on_rq;
5649 unsigned long flags;
5652 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
5655 * We have to be careful, if called from sys_setpriority(),
5656 * the task might be in the middle of scheduling on another CPU.
5658 rq = task_rq_lock(p, &flags);
5659 update_rq_clock(rq);
5661 * The RT priorities are set via sched_setscheduler(), but we still
5662 * allow the 'normal' nice value to be set - but as expected
5663 * it wont have any effect on scheduling until the task is
5664 * SCHED_FIFO/SCHED_RR:
5666 if (task_has_rt_policy(p)) {
5667 p->static_prio = NICE_TO_PRIO(nice);
5670 on_rq = p->se.on_rq;
5672 dequeue_task(rq, p, 0);
5674 p->static_prio = NICE_TO_PRIO(nice);
5677 p->prio = effective_prio(p);
5678 delta = p->prio - old_prio;
5681 enqueue_task(rq, p, 0);
5683 * If the task increased its priority or is running and
5684 * lowered its priority, then reschedule its CPU:
5686 if (delta < 0 || (delta > 0 && task_running(rq, p)))
5687 resched_task(rq->curr);
5690 task_rq_unlock(rq, &flags);
5692 EXPORT_SYMBOL(set_user_nice);
5695 * can_nice - check if a task can reduce its nice value
5699 int can_nice(const struct task_struct *p, const int nice)
5701 /* convert nice value [19,-20] to rlimit style value [1,40] */
5702 int nice_rlim = 20 - nice;
5704 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
5705 capable(CAP_SYS_NICE));
5708 #ifdef __ARCH_WANT_SYS_NICE
5711 * sys_nice - change the priority of the current process.
5712 * @increment: priority increment
5714 * sys_setpriority is a more generic, but much slower function that
5715 * does similar things.
5717 SYSCALL_DEFINE1(nice, int, increment)
5722 * Setpriority might change our priority at the same moment.
5723 * We don't have to worry. Conceptually one call occurs first
5724 * and we have a single winner.
5726 if (increment < -40)
5731 nice = TASK_NICE(current) + increment;
5737 if (increment < 0 && !can_nice(current, nice))
5740 retval = security_task_setnice(current, nice);
5744 set_user_nice(current, nice);
5751 * task_prio - return the priority value of a given task.
5752 * @p: the task in question.
5754 * This is the priority value as seen by users in /proc.
5755 * RT tasks are offset by -200. Normal tasks are centered
5756 * around 0, value goes from -16 to +15.
5758 int task_prio(const struct task_struct *p)
5760 return p->prio - MAX_RT_PRIO;
5764 * task_nice - return the nice value of a given task.
5765 * @p: the task in question.
5767 int task_nice(const struct task_struct *p)
5769 return TASK_NICE(p);
5771 EXPORT_SYMBOL(task_nice);
5774 * idle_cpu - is a given cpu idle currently?
5775 * @cpu: the processor in question.
5777 int idle_cpu(int cpu)
5779 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5783 * idle_task - return the idle task for a given cpu.
5784 * @cpu: the processor in question.
5786 struct task_struct *idle_task(int cpu)
5788 return cpu_rq(cpu)->idle;
5792 * find_process_by_pid - find a process with a matching PID value.
5793 * @pid: the pid in question.
5795 static struct task_struct *find_process_by_pid(pid_t pid)
5797 return pid ? find_task_by_vpid(pid) : current;
5800 /* Actually do priority change: must hold rq lock. */
5802 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5804 BUG_ON(p->se.on_rq);
5807 switch (p->policy) {
5811 p->sched_class = &fair_sched_class;
5815 p->sched_class = &rt_sched_class;
5819 p->rt_priority = prio;
5820 p->normal_prio = normal_prio(p);
5821 /* we are holding p->pi_lock already */
5822 p->prio = rt_mutex_getprio(p);
5827 * check the target process has a UID that matches the current process's
5829 static bool check_same_owner(struct task_struct *p)
5831 const struct cred *cred = current_cred(), *pcred;
5835 pcred = __task_cred(p);
5836 match = (cred->euid == pcred->euid ||
5837 cred->euid == pcred->uid);
5842 static int __sched_setscheduler(struct task_struct *p, int policy,
5843 struct sched_param *param, bool user)
5845 int retval, oldprio, oldpolicy = -1, on_rq, running;
5846 unsigned long flags;
5847 const struct sched_class *prev_class = p->sched_class;
5850 /* may grab non-irq protected spin_locks */
5851 BUG_ON(in_interrupt());
5853 /* double check policy once rq lock held */
5855 policy = oldpolicy = p->policy;
5856 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
5857 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5858 policy != SCHED_IDLE)
5861 * Valid priorities for SCHED_FIFO and SCHED_RR are
5862 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5863 * SCHED_BATCH and SCHED_IDLE is 0.
5865 if (param->sched_priority < 0 ||
5866 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5867 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5869 if (rt_policy(policy) != (param->sched_priority != 0))
5873 * Allow unprivileged RT tasks to decrease priority:
5875 if (user && !capable(CAP_SYS_NICE)) {
5876 if (rt_policy(policy)) {
5877 unsigned long rlim_rtprio;
5879 if (!lock_task_sighand(p, &flags))
5881 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
5882 unlock_task_sighand(p, &flags);
5884 /* can't set/change the rt policy */
5885 if (policy != p->policy && !rlim_rtprio)
5888 /* can't increase priority */
5889 if (param->sched_priority > p->rt_priority &&
5890 param->sched_priority > rlim_rtprio)
5894 * Like positive nice levels, dont allow tasks to
5895 * move out of SCHED_IDLE either:
5897 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
5900 /* can't change other user's priorities */
5901 if (!check_same_owner(p))
5906 #ifdef CONFIG_RT_GROUP_SCHED
5908 * Do not allow realtime tasks into groups that have no runtime
5911 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5912 task_group(p)->rt_bandwidth.rt_runtime == 0)
5916 retval = security_task_setscheduler(p, policy, param);
5922 * make sure no PI-waiters arrive (or leave) while we are
5923 * changing the priority of the task:
5925 spin_lock_irqsave(&p->pi_lock, flags);
5927 * To be able to change p->policy safely, the apropriate
5928 * runqueue lock must be held.
5930 rq = __task_rq_lock(p);
5931 /* recheck policy now with rq lock held */
5932 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5933 policy = oldpolicy = -1;
5934 __task_rq_unlock(rq);
5935 spin_unlock_irqrestore(&p->pi_lock, flags);
5938 update_rq_clock(rq);
5939 on_rq = p->se.on_rq;
5940 running = task_current(rq, p);
5942 deactivate_task(rq, p, 0);
5944 p->sched_class->put_prev_task(rq, p);
5947 __setscheduler(rq, p, policy, param->sched_priority);
5950 p->sched_class->set_curr_task(rq);
5952 activate_task(rq, p, 0);
5954 check_class_changed(rq, p, prev_class, oldprio, running);
5956 __task_rq_unlock(rq);
5957 spin_unlock_irqrestore(&p->pi_lock, flags);
5959 rt_mutex_adjust_pi(p);
5965 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5966 * @p: the task in question.
5967 * @policy: new policy.
5968 * @param: structure containing the new RT priority.
5970 * NOTE that the task may be already dead.
5972 int sched_setscheduler(struct task_struct *p, int policy,
5973 struct sched_param *param)
5975 return __sched_setscheduler(p, policy, param, true);
5977 EXPORT_SYMBOL_GPL(sched_setscheduler);
5980 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5981 * @p: the task in question.
5982 * @policy: new policy.
5983 * @param: structure containing the new RT priority.
5985 * Just like sched_setscheduler, only don't bother checking if the
5986 * current context has permission. For example, this is needed in
5987 * stop_machine(): we create temporary high priority worker threads,
5988 * but our caller might not have that capability.
5990 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5991 struct sched_param *param)
5993 return __sched_setscheduler(p, policy, param, false);
5997 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5999 struct sched_param lparam;
6000 struct task_struct *p;
6003 if (!param || pid < 0)
6005 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
6010 p = find_process_by_pid(pid);
6012 retval = sched_setscheduler(p, policy, &lparam);
6019 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6020 * @pid: the pid in question.
6021 * @policy: new policy.
6022 * @param: structure containing the new RT priority.
6024 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
6025 struct sched_param __user *, param)
6027 /* negative values for policy are not valid */
6031 return do_sched_setscheduler(pid, policy, param);
6035 * sys_sched_setparam - set/change the RT priority of a thread
6036 * @pid: the pid in question.
6037 * @param: structure containing the new RT priority.
6039 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
6041 return do_sched_setscheduler(pid, -1, param);
6045 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6046 * @pid: the pid in question.
6048 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
6050 struct task_struct *p;
6057 read_lock(&tasklist_lock);
6058 p = find_process_by_pid(pid);
6060 retval = security_task_getscheduler(p);
6064 read_unlock(&tasklist_lock);
6069 * sys_sched_getscheduler - get the RT priority of a thread
6070 * @pid: the pid in question.
6071 * @param: structure containing the RT priority.
6073 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
6075 struct sched_param lp;
6076 struct task_struct *p;
6079 if (!param || pid < 0)
6082 read_lock(&tasklist_lock);
6083 p = find_process_by_pid(pid);
6088 retval = security_task_getscheduler(p);
6092 lp.sched_priority = p->rt_priority;
6093 read_unlock(&tasklist_lock);
6096 * This one might sleep, we cannot do it with a spinlock held ...
6098 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
6103 read_unlock(&tasklist_lock);
6107 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
6109 cpumask_var_t cpus_allowed, new_mask;
6110 struct task_struct *p;
6114 read_lock(&tasklist_lock);
6116 p = find_process_by_pid(pid);
6118 read_unlock(&tasklist_lock);
6124 * It is not safe to call set_cpus_allowed with the
6125 * tasklist_lock held. We will bump the task_struct's
6126 * usage count and then drop tasklist_lock.
6129 read_unlock(&tasklist_lock);
6131 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6135 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
6137 goto out_free_cpus_allowed;
6140 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
6143 retval = security_task_setscheduler(p, 0, NULL);
6147 cpuset_cpus_allowed(p, cpus_allowed);
6148 cpumask_and(new_mask, in_mask, cpus_allowed);
6150 retval = set_cpus_allowed_ptr(p, new_mask);
6153 cpuset_cpus_allowed(p, cpus_allowed);
6154 if (!cpumask_subset(new_mask, cpus_allowed)) {
6156 * We must have raced with a concurrent cpuset
6157 * update. Just reset the cpus_allowed to the
6158 * cpuset's cpus_allowed
6160 cpumask_copy(new_mask, cpus_allowed);
6165 free_cpumask_var(new_mask);
6166 out_free_cpus_allowed:
6167 free_cpumask_var(cpus_allowed);
6174 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
6175 struct cpumask *new_mask)
6177 if (len < cpumask_size())
6178 cpumask_clear(new_mask);
6179 else if (len > cpumask_size())
6180 len = cpumask_size();
6182 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6186 * sys_sched_setaffinity - set the cpu affinity of a process
6187 * @pid: pid of the process
6188 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6189 * @user_mask_ptr: user-space pointer to the new cpu mask
6191 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6192 unsigned long __user *, user_mask_ptr)
6194 cpumask_var_t new_mask;
6197 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6200 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6202 retval = sched_setaffinity(pid, new_mask);
6203 free_cpumask_var(new_mask);
6207 long sched_getaffinity(pid_t pid, struct cpumask *mask)
6209 struct task_struct *p;
6213 read_lock(&tasklist_lock);
6216 p = find_process_by_pid(pid);
6220 retval = security_task_getscheduler(p);
6224 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
6227 read_unlock(&tasklist_lock);
6234 * sys_sched_getaffinity - get the cpu affinity of a process
6235 * @pid: pid of the process
6236 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6237 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6239 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6240 unsigned long __user *, user_mask_ptr)
6245 if (len < cpumask_size())
6248 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6251 ret = sched_getaffinity(pid, mask);
6253 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
6256 ret = cpumask_size();
6258 free_cpumask_var(mask);
6264 * sys_sched_yield - yield the current processor to other threads.
6266 * This function yields the current CPU to other tasks. If there are no
6267 * other threads running on this CPU then this function will return.
6269 SYSCALL_DEFINE0(sched_yield)
6271 struct rq *rq = this_rq_lock();
6273 schedstat_inc(rq, yld_count);
6274 current->sched_class->yield_task(rq);
6277 * Since we are going to call schedule() anyway, there's
6278 * no need to preempt or enable interrupts:
6280 __release(rq->lock);
6281 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
6282 _raw_spin_unlock(&rq->lock);
6283 preempt_enable_no_resched();
6290 static void __cond_resched(void)
6292 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6293 __might_sleep(__FILE__, __LINE__);
6296 * The BKS might be reacquired before we have dropped
6297 * PREEMPT_ACTIVE, which could trigger a second
6298 * cond_resched() call.
6301 add_preempt_count(PREEMPT_ACTIVE);
6303 sub_preempt_count(PREEMPT_ACTIVE);
6304 } while (need_resched());
6307 int __sched _cond_resched(void)
6309 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
6310 system_state == SYSTEM_RUNNING) {
6316 EXPORT_SYMBOL(_cond_resched);
6319 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
6320 * call schedule, and on return reacquire the lock.
6322 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6323 * operations here to prevent schedule() from being called twice (once via
6324 * spin_unlock(), once by hand).
6326 int cond_resched_lock(spinlock_t *lock)
6328 int resched = need_resched() && system_state == SYSTEM_RUNNING;
6331 if (spin_needbreak(lock) || resched) {
6333 if (resched && need_resched())
6342 EXPORT_SYMBOL(cond_resched_lock);
6344 int __sched cond_resched_softirq(void)
6346 BUG_ON(!in_softirq());
6348 if (need_resched() && system_state == SYSTEM_RUNNING) {
6356 EXPORT_SYMBOL(cond_resched_softirq);
6359 * yield - yield the current processor to other threads.
6361 * This is a shortcut for kernel-space yielding - it marks the
6362 * thread runnable and calls sys_sched_yield().
6364 void __sched yield(void)
6366 set_current_state(TASK_RUNNING);
6369 EXPORT_SYMBOL(yield);
6372 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6373 * that process accounting knows that this is a task in IO wait state.
6375 * But don't do that if it is a deliberate, throttling IO wait (this task
6376 * has set its backing_dev_info: the queue against which it should throttle)
6378 void __sched io_schedule(void)
6380 struct rq *rq = &__raw_get_cpu_var(runqueues);
6382 delayacct_blkio_start();
6383 atomic_inc(&rq->nr_iowait);
6385 atomic_dec(&rq->nr_iowait);
6386 delayacct_blkio_end();
6388 EXPORT_SYMBOL(io_schedule);
6390 long __sched io_schedule_timeout(long timeout)
6392 struct rq *rq = &__raw_get_cpu_var(runqueues);
6395 delayacct_blkio_start();
6396 atomic_inc(&rq->nr_iowait);
6397 ret = schedule_timeout(timeout);
6398 atomic_dec(&rq->nr_iowait);
6399 delayacct_blkio_end();
6404 * sys_sched_get_priority_max - return maximum RT priority.
6405 * @policy: scheduling class.
6407 * this syscall returns the maximum rt_priority that can be used
6408 * by a given scheduling class.
6410 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6417 ret = MAX_USER_RT_PRIO-1;
6429 * sys_sched_get_priority_min - return minimum RT priority.
6430 * @policy: scheduling class.
6432 * this syscall returns the minimum rt_priority that can be used
6433 * by a given scheduling class.
6435 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6453 * sys_sched_rr_get_interval - return the default timeslice of a process.
6454 * @pid: pid of the process.
6455 * @interval: userspace pointer to the timeslice value.
6457 * this syscall writes the default timeslice value of a given process
6458 * into the user-space timespec buffer. A value of '0' means infinity.
6460 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6461 struct timespec __user *, interval)
6463 struct task_struct *p;
6464 unsigned int time_slice;
6472 read_lock(&tasklist_lock);
6473 p = find_process_by_pid(pid);
6477 retval = security_task_getscheduler(p);
6482 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
6483 * tasks that are on an otherwise idle runqueue:
6486 if (p->policy == SCHED_RR) {
6487 time_slice = DEF_TIMESLICE;
6488 } else if (p->policy != SCHED_FIFO) {
6489 struct sched_entity *se = &p->se;
6490 unsigned long flags;
6493 rq = task_rq_lock(p, &flags);
6494 if (rq->cfs.load.weight)
6495 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
6496 task_rq_unlock(rq, &flags);
6498 read_unlock(&tasklist_lock);
6499 jiffies_to_timespec(time_slice, &t);
6500 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
6504 read_unlock(&tasklist_lock);
6508 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
6510 void sched_show_task(struct task_struct *p)
6512 unsigned long free = 0;
6515 state = p->state ? __ffs(p->state) + 1 : 0;
6516 printk(KERN_INFO "%-13.13s %c", p->comm,
6517 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
6518 #if BITS_PER_LONG == 32
6519 if (state == TASK_RUNNING)
6520 printk(KERN_CONT " running ");
6522 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
6524 if (state == TASK_RUNNING)
6525 printk(KERN_CONT " running task ");
6527 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
6529 #ifdef CONFIG_DEBUG_STACK_USAGE
6530 free = stack_not_used(p);
6532 printk(KERN_CONT "%5lu %5d %6d\n", free,
6533 task_pid_nr(p), task_pid_nr(p->real_parent));
6535 show_stack(p, NULL);
6538 void show_state_filter(unsigned long state_filter)
6540 struct task_struct *g, *p;
6542 #if BITS_PER_LONG == 32
6544 " task PC stack pid father\n");
6547 " task PC stack pid father\n");
6549 read_lock(&tasklist_lock);
6550 do_each_thread(g, p) {
6552 * reset the NMI-timeout, listing all files on a slow
6553 * console might take alot of time:
6555 touch_nmi_watchdog();
6556 if (!state_filter || (p->state & state_filter))
6558 } while_each_thread(g, p);
6560 touch_all_softlockup_watchdogs();
6562 #ifdef CONFIG_SCHED_DEBUG
6563 sysrq_sched_debug_show();
6565 read_unlock(&tasklist_lock);
6567 * Only show locks if all tasks are dumped:
6569 if (state_filter == -1)
6570 debug_show_all_locks();
6573 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6575 idle->sched_class = &idle_sched_class;
6579 * init_idle - set up an idle thread for a given CPU
6580 * @idle: task in question
6581 * @cpu: cpu the idle task belongs to
6583 * NOTE: this function does not set the idle thread's NEED_RESCHED
6584 * flag, to make booting more robust.
6586 void __cpuinit init_idle(struct task_struct *idle, int cpu)
6588 struct rq *rq = cpu_rq(cpu);
6589 unsigned long flags;
6591 spin_lock_irqsave(&rq->lock, flags);
6594 idle->se.exec_start = sched_clock();
6596 idle->prio = idle->normal_prio = MAX_PRIO;
6597 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
6598 __set_task_cpu(idle, cpu);
6600 rq->curr = rq->idle = idle;
6601 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6604 spin_unlock_irqrestore(&rq->lock, flags);
6606 /* Set the preempt count _outside_ the spinlocks! */
6607 #if defined(CONFIG_PREEMPT)
6608 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
6610 task_thread_info(idle)->preempt_count = 0;
6613 * The idle tasks have their own, simple scheduling class:
6615 idle->sched_class = &idle_sched_class;
6616 ftrace_graph_init_task(idle);
6620 * In a system that switches off the HZ timer nohz_cpu_mask
6621 * indicates which cpus entered this state. This is used
6622 * in the rcu update to wait only for active cpus. For system
6623 * which do not switch off the HZ timer nohz_cpu_mask should
6624 * always be CPU_BITS_NONE.
6626 cpumask_var_t nohz_cpu_mask;
6629 * Increase the granularity value when there are more CPUs,
6630 * because with more CPUs the 'effective latency' as visible
6631 * to users decreases. But the relationship is not linear,
6632 * so pick a second-best guess by going with the log2 of the
6635 * This idea comes from the SD scheduler of Con Kolivas:
6637 static inline void sched_init_granularity(void)
6639 unsigned int factor = 1 + ilog2(num_online_cpus());
6640 const unsigned long limit = 200000000;
6642 sysctl_sched_min_granularity *= factor;
6643 if (sysctl_sched_min_granularity > limit)
6644 sysctl_sched_min_granularity = limit;
6646 sysctl_sched_latency *= factor;
6647 if (sysctl_sched_latency > limit)
6648 sysctl_sched_latency = limit;
6650 sysctl_sched_wakeup_granularity *= factor;
6652 sysctl_sched_shares_ratelimit *= factor;
6657 * This is how migration works:
6659 * 1) we queue a struct migration_req structure in the source CPU's
6660 * runqueue and wake up that CPU's migration thread.
6661 * 2) we down() the locked semaphore => thread blocks.
6662 * 3) migration thread wakes up (implicitly it forces the migrated
6663 * thread off the CPU)
6664 * 4) it gets the migration request and checks whether the migrated
6665 * task is still in the wrong runqueue.
6666 * 5) if it's in the wrong runqueue then the migration thread removes
6667 * it and puts it into the right queue.
6668 * 6) migration thread up()s the semaphore.
6669 * 7) we wake up and the migration is done.
6673 * Change a given task's CPU affinity. Migrate the thread to a
6674 * proper CPU and schedule it away if the CPU it's executing on
6675 * is removed from the allowed bitmask.
6677 * NOTE: the caller must have a valid reference to the task, the
6678 * task must not exit() & deallocate itself prematurely. The
6679 * call is not atomic; no spinlocks may be held.
6681 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
6683 struct migration_req req;
6684 unsigned long flags;
6688 rq = task_rq_lock(p, &flags);
6689 if (!cpumask_intersects(new_mask, cpu_online_mask)) {
6694 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
6695 !cpumask_equal(&p->cpus_allowed, new_mask))) {
6700 if (p->sched_class->set_cpus_allowed)
6701 p->sched_class->set_cpus_allowed(p, new_mask);
6703 cpumask_copy(&p->cpus_allowed, new_mask);
6704 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
6707 /* Can the task run on the task's current CPU? If so, we're done */
6708 if (cpumask_test_cpu(task_cpu(p), new_mask))
6711 if (migrate_task(p, cpumask_any_and(cpu_online_mask, new_mask), &req)) {
6712 /* Need help from migration thread: drop lock and wait. */
6713 task_rq_unlock(rq, &flags);
6714 wake_up_process(rq->migration_thread);
6715 wait_for_completion(&req.done);
6716 tlb_migrate_finish(p->mm);
6720 task_rq_unlock(rq, &flags);
6724 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
6727 * Move (not current) task off this cpu, onto dest cpu. We're doing
6728 * this because either it can't run here any more (set_cpus_allowed()
6729 * away from this CPU, or CPU going down), or because we're
6730 * attempting to rebalance this task on exec (sched_exec).
6732 * So we race with normal scheduler movements, but that's OK, as long
6733 * as the task is no longer on this CPU.
6735 * Returns non-zero if task was successfully migrated.
6737 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6739 struct rq *rq_dest, *rq_src;
6742 if (unlikely(!cpu_active(dest_cpu)))
6745 rq_src = cpu_rq(src_cpu);
6746 rq_dest = cpu_rq(dest_cpu);
6748 double_rq_lock(rq_src, rq_dest);
6749 /* Already moved. */
6750 if (task_cpu(p) != src_cpu)
6752 /* Affinity changed (again). */
6753 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6756 on_rq = p->se.on_rq;
6758 deactivate_task(rq_src, p, 0);
6760 set_task_cpu(p, dest_cpu);
6762 activate_task(rq_dest, p, 0);
6763 check_preempt_curr(rq_dest, p, 0);
6768 double_rq_unlock(rq_src, rq_dest);
6773 * migration_thread - this is a highprio system thread that performs
6774 * thread migration by bumping thread off CPU then 'pushing' onto
6777 static int migration_thread(void *data)
6779 int cpu = (long)data;
6783 BUG_ON(rq->migration_thread != current);
6785 set_current_state(TASK_INTERRUPTIBLE);
6786 while (!kthread_should_stop()) {
6787 struct migration_req *req;
6788 struct list_head *head;
6790 spin_lock_irq(&rq->lock);
6792 if (cpu_is_offline(cpu)) {
6793 spin_unlock_irq(&rq->lock);
6797 if (rq->active_balance) {
6798 active_load_balance(rq, cpu);
6799 rq->active_balance = 0;
6802 head = &rq->migration_queue;
6804 if (list_empty(head)) {
6805 spin_unlock_irq(&rq->lock);
6807 set_current_state(TASK_INTERRUPTIBLE);
6810 req = list_entry(head->next, struct migration_req, list);
6811 list_del_init(head->next);
6813 spin_unlock(&rq->lock);
6814 __migrate_task(req->task, cpu, req->dest_cpu);
6817 complete(&req->done);
6819 __set_current_state(TASK_RUNNING);
6823 /* Wait for kthread_stop */
6824 set_current_state(TASK_INTERRUPTIBLE);
6825 while (!kthread_should_stop()) {
6827 set_current_state(TASK_INTERRUPTIBLE);
6829 __set_current_state(TASK_RUNNING);
6833 #ifdef CONFIG_HOTPLUG_CPU
6835 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
6839 local_irq_disable();
6840 ret = __migrate_task(p, src_cpu, dest_cpu);
6846 * Figure out where task on dead CPU should go, use force if necessary.
6848 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
6851 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(dead_cpu));
6854 /* Look for allowed, online CPU in same node. */
6855 for_each_cpu_and(dest_cpu, nodemask, cpu_online_mask)
6856 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6859 /* Any allowed, online CPU? */
6860 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_online_mask);
6861 if (dest_cpu < nr_cpu_ids)
6864 /* No more Mr. Nice Guy. */
6865 if (dest_cpu >= nr_cpu_ids) {
6866 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
6867 dest_cpu = cpumask_any_and(cpu_online_mask, &p->cpus_allowed);
6870 * Don't tell them about moving exiting tasks or
6871 * kernel threads (both mm NULL), since they never
6874 if (p->mm && printk_ratelimit()) {
6875 printk(KERN_INFO "process %d (%s) no "
6876 "longer affine to cpu%d\n",
6877 task_pid_nr(p), p->comm, dead_cpu);
6882 /* It can have affinity changed while we were choosing. */
6883 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
6888 * While a dead CPU has no uninterruptible tasks queued at this point,
6889 * it might still have a nonzero ->nr_uninterruptible counter, because
6890 * for performance reasons the counter is not stricly tracking tasks to
6891 * their home CPUs. So we just add the counter to another CPU's counter,
6892 * to keep the global sum constant after CPU-down:
6894 static void migrate_nr_uninterruptible(struct rq *rq_src)
6896 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_online_mask));
6897 unsigned long flags;
6899 local_irq_save(flags);
6900 double_rq_lock(rq_src, rq_dest);
6901 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6902 rq_src->nr_uninterruptible = 0;
6903 double_rq_unlock(rq_src, rq_dest);
6904 local_irq_restore(flags);
6907 /* Run through task list and migrate tasks from the dead cpu. */
6908 static void migrate_live_tasks(int src_cpu)
6910 struct task_struct *p, *t;
6912 read_lock(&tasklist_lock);
6914 do_each_thread(t, p) {
6918 if (task_cpu(p) == src_cpu)
6919 move_task_off_dead_cpu(src_cpu, p);
6920 } while_each_thread(t, p);
6922 read_unlock(&tasklist_lock);
6926 * Schedules idle task to be the next runnable task on current CPU.
6927 * It does so by boosting its priority to highest possible.
6928 * Used by CPU offline code.
6930 void sched_idle_next(void)
6932 int this_cpu = smp_processor_id();
6933 struct rq *rq = cpu_rq(this_cpu);
6934 struct task_struct *p = rq->idle;
6935 unsigned long flags;
6937 /* cpu has to be offline */
6938 BUG_ON(cpu_online(this_cpu));
6941 * Strictly not necessary since rest of the CPUs are stopped by now
6942 * and interrupts disabled on the current cpu.
6944 spin_lock_irqsave(&rq->lock, flags);
6946 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6948 update_rq_clock(rq);
6949 activate_task(rq, p, 0);
6951 spin_unlock_irqrestore(&rq->lock, flags);
6955 * Ensures that the idle task is using init_mm right before its cpu goes
6958 void idle_task_exit(void)
6960 struct mm_struct *mm = current->active_mm;
6962 BUG_ON(cpu_online(smp_processor_id()));
6965 switch_mm(mm, &init_mm, current);
6969 /* called under rq->lock with disabled interrupts */
6970 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
6972 struct rq *rq = cpu_rq(dead_cpu);
6974 /* Must be exiting, otherwise would be on tasklist. */
6975 BUG_ON(!p->exit_state);
6977 /* Cannot have done final schedule yet: would have vanished. */
6978 BUG_ON(p->state == TASK_DEAD);
6983 * Drop lock around migration; if someone else moves it,
6984 * that's OK. No task can be added to this CPU, so iteration is
6987 spin_unlock_irq(&rq->lock);
6988 move_task_off_dead_cpu(dead_cpu, p);
6989 spin_lock_irq(&rq->lock);
6994 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6995 static void migrate_dead_tasks(unsigned int dead_cpu)
6997 struct rq *rq = cpu_rq(dead_cpu);
6998 struct task_struct *next;
7001 if (!rq->nr_running)
7003 update_rq_clock(rq);
7004 next = pick_next_task(rq);
7007 next->sched_class->put_prev_task(rq, next);
7008 migrate_dead(dead_cpu, next);
7012 #endif /* CONFIG_HOTPLUG_CPU */
7014 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7016 static struct ctl_table sd_ctl_dir[] = {
7018 .procname = "sched_domain",
7024 static struct ctl_table sd_ctl_root[] = {
7026 .ctl_name = CTL_KERN,
7027 .procname = "kernel",
7029 .child = sd_ctl_dir,
7034 static struct ctl_table *sd_alloc_ctl_entry(int n)
7036 struct ctl_table *entry =
7037 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
7042 static void sd_free_ctl_entry(struct ctl_table **tablep)
7044 struct ctl_table *entry;
7047 * In the intermediate directories, both the child directory and
7048 * procname are dynamically allocated and could fail but the mode
7049 * will always be set. In the lowest directory the names are
7050 * static strings and all have proc handlers.
7052 for (entry = *tablep; entry->mode; entry++) {
7054 sd_free_ctl_entry(&entry->child);
7055 if (entry->proc_handler == NULL)
7056 kfree(entry->procname);
7064 set_table_entry(struct ctl_table *entry,
7065 const char *procname, void *data, int maxlen,
7066 mode_t mode, proc_handler *proc_handler)
7068 entry->procname = procname;
7070 entry->maxlen = maxlen;
7072 entry->proc_handler = proc_handler;
7075 static struct ctl_table *
7076 sd_alloc_ctl_domain_table(struct sched_domain *sd)
7078 struct ctl_table *table = sd_alloc_ctl_entry(13);
7083 set_table_entry(&table[0], "min_interval", &sd->min_interval,
7084 sizeof(long), 0644, proc_doulongvec_minmax);
7085 set_table_entry(&table[1], "max_interval", &sd->max_interval,
7086 sizeof(long), 0644, proc_doulongvec_minmax);
7087 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
7088 sizeof(int), 0644, proc_dointvec_minmax);
7089 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
7090 sizeof(int), 0644, proc_dointvec_minmax);
7091 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
7092 sizeof(int), 0644, proc_dointvec_minmax);
7093 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
7094 sizeof(int), 0644, proc_dointvec_minmax);
7095 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
7096 sizeof(int), 0644, proc_dointvec_minmax);
7097 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
7098 sizeof(int), 0644, proc_dointvec_minmax);
7099 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
7100 sizeof(int), 0644, proc_dointvec_minmax);
7101 set_table_entry(&table[9], "cache_nice_tries",
7102 &sd->cache_nice_tries,
7103 sizeof(int), 0644, proc_dointvec_minmax);
7104 set_table_entry(&table[10], "flags", &sd->flags,
7105 sizeof(int), 0644, proc_dointvec_minmax);
7106 set_table_entry(&table[11], "name", sd->name,
7107 CORENAME_MAX_SIZE, 0444, proc_dostring);
7108 /* &table[12] is terminator */
7113 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
7115 struct ctl_table *entry, *table;
7116 struct sched_domain *sd;
7117 int domain_num = 0, i;
7120 for_each_domain(cpu, sd)
7122 entry = table = sd_alloc_ctl_entry(domain_num + 1);
7127 for_each_domain(cpu, sd) {
7128 snprintf(buf, 32, "domain%d", i);
7129 entry->procname = kstrdup(buf, GFP_KERNEL);
7131 entry->child = sd_alloc_ctl_domain_table(sd);
7138 static struct ctl_table_header *sd_sysctl_header;
7139 static void register_sched_domain_sysctl(void)
7141 int i, cpu_num = num_online_cpus();
7142 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
7145 WARN_ON(sd_ctl_dir[0].child);
7146 sd_ctl_dir[0].child = entry;
7151 for_each_online_cpu(i) {
7152 snprintf(buf, 32, "cpu%d", i);
7153 entry->procname = kstrdup(buf, GFP_KERNEL);
7155 entry->child = sd_alloc_ctl_cpu_table(i);
7159 WARN_ON(sd_sysctl_header);
7160 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
7163 /* may be called multiple times per register */
7164 static void unregister_sched_domain_sysctl(void)
7166 if (sd_sysctl_header)
7167 unregister_sysctl_table(sd_sysctl_header);
7168 sd_sysctl_header = NULL;
7169 if (sd_ctl_dir[0].child)
7170 sd_free_ctl_entry(&sd_ctl_dir[0].child);
7173 static void register_sched_domain_sysctl(void)
7176 static void unregister_sched_domain_sysctl(void)
7181 static void set_rq_online(struct rq *rq)
7184 const struct sched_class *class;
7186 cpumask_set_cpu(rq->cpu, rq->rd->online);
7189 for_each_class(class) {
7190 if (class->rq_online)
7191 class->rq_online(rq);
7196 static void set_rq_offline(struct rq *rq)
7199 const struct sched_class *class;
7201 for_each_class(class) {
7202 if (class->rq_offline)
7203 class->rq_offline(rq);
7206 cpumask_clear_cpu(rq->cpu, rq->rd->online);
7212 * migration_call - callback that gets triggered when a CPU is added.
7213 * Here we can start up the necessary migration thread for the new CPU.
7215 static int __cpuinit
7216 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
7218 struct task_struct *p;
7219 int cpu = (long)hcpu;
7220 unsigned long flags;
7225 case CPU_UP_PREPARE:
7226 case CPU_UP_PREPARE_FROZEN:
7227 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
7230 kthread_bind(p, cpu);
7231 /* Must be high prio: stop_machine expects to yield to it. */
7232 rq = task_rq_lock(p, &flags);
7233 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7234 task_rq_unlock(rq, &flags);
7235 cpu_rq(cpu)->migration_thread = p;
7239 case CPU_ONLINE_FROZEN:
7240 /* Strictly unnecessary, as first user will wake it. */
7241 wake_up_process(cpu_rq(cpu)->migration_thread);
7243 /* Update our root-domain */
7245 spin_lock_irqsave(&rq->lock, flags);
7247 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7251 spin_unlock_irqrestore(&rq->lock, flags);
7254 #ifdef CONFIG_HOTPLUG_CPU
7255 case CPU_UP_CANCELED:
7256 case CPU_UP_CANCELED_FROZEN:
7257 if (!cpu_rq(cpu)->migration_thread)
7259 /* Unbind it from offline cpu so it can run. Fall thru. */
7260 kthread_bind(cpu_rq(cpu)->migration_thread,
7261 cpumask_any(cpu_online_mask));
7262 kthread_stop(cpu_rq(cpu)->migration_thread);
7263 cpu_rq(cpu)->migration_thread = NULL;
7267 case CPU_DEAD_FROZEN:
7268 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7269 migrate_live_tasks(cpu);
7271 kthread_stop(rq->migration_thread);
7272 rq->migration_thread = NULL;
7273 /* Idle task back to normal (off runqueue, low prio) */
7274 spin_lock_irq(&rq->lock);
7275 update_rq_clock(rq);
7276 deactivate_task(rq, rq->idle, 0);
7277 rq->idle->static_prio = MAX_PRIO;
7278 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
7279 rq->idle->sched_class = &idle_sched_class;
7280 migrate_dead_tasks(cpu);
7281 spin_unlock_irq(&rq->lock);
7283 migrate_nr_uninterruptible(rq);
7284 BUG_ON(rq->nr_running != 0);
7287 * No need to migrate the tasks: it was best-effort if
7288 * they didn't take sched_hotcpu_mutex. Just wake up
7291 spin_lock_irq(&rq->lock);
7292 while (!list_empty(&rq->migration_queue)) {
7293 struct migration_req *req;
7295 req = list_entry(rq->migration_queue.next,
7296 struct migration_req, list);
7297 list_del_init(&req->list);
7298 spin_unlock_irq(&rq->lock);
7299 complete(&req->done);
7300 spin_lock_irq(&rq->lock);
7302 spin_unlock_irq(&rq->lock);
7306 case CPU_DYING_FROZEN:
7307 /* Update our root-domain */
7309 spin_lock_irqsave(&rq->lock, flags);
7311 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7314 spin_unlock_irqrestore(&rq->lock, flags);
7321 /* Register at highest priority so that task migration (migrate_all_tasks)
7322 * happens before everything else.
7324 static struct notifier_block __cpuinitdata migration_notifier = {
7325 .notifier_call = migration_call,
7329 static int __init migration_init(void)
7331 void *cpu = (void *)(long)smp_processor_id();
7334 /* Start one for the boot CPU: */
7335 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
7336 BUG_ON(err == NOTIFY_BAD);
7337 migration_call(&migration_notifier, CPU_ONLINE, cpu);
7338 register_cpu_notifier(&migration_notifier);
7342 early_initcall(migration_init);
7347 #ifdef CONFIG_SCHED_DEBUG
7349 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
7350 struct cpumask *groupmask)
7352 struct sched_group *group = sd->groups;
7355 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
7356 cpumask_clear(groupmask);
7358 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
7360 if (!(sd->flags & SD_LOAD_BALANCE)) {
7361 printk("does not load-balance\n");
7363 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
7368 printk(KERN_CONT "span %s level %s\n", str, sd->name);
7370 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
7371 printk(KERN_ERR "ERROR: domain->span does not contain "
7374 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
7375 printk(KERN_ERR "ERROR: domain->groups does not contain"
7379 printk(KERN_DEBUG "%*s groups:", level + 1, "");
7383 printk(KERN_ERR "ERROR: group is NULL\n");
7387 if (!group->__cpu_power) {
7388 printk(KERN_CONT "\n");
7389 printk(KERN_ERR "ERROR: domain->cpu_power not "
7394 if (!cpumask_weight(sched_group_cpus(group))) {
7395 printk(KERN_CONT "\n");
7396 printk(KERN_ERR "ERROR: empty group\n");
7400 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
7401 printk(KERN_CONT "\n");
7402 printk(KERN_ERR "ERROR: repeated CPUs\n");
7406 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
7408 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
7410 printk(KERN_CONT " %s", str);
7411 if (group->__cpu_power != SCHED_LOAD_SCALE) {
7412 printk(KERN_CONT " (__cpu_power = %d)",
7413 group->__cpu_power);
7416 group = group->next;
7417 } while (group != sd->groups);
7418 printk(KERN_CONT "\n");
7420 if (!cpumask_equal(sched_domain_span(sd), groupmask))
7421 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
7424 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
7425 printk(KERN_ERR "ERROR: parent span is not a superset "
7426 "of domain->span\n");
7430 static void sched_domain_debug(struct sched_domain *sd, int cpu)
7432 cpumask_var_t groupmask;
7436 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7440 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7442 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
7443 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
7448 if (sched_domain_debug_one(sd, cpu, level, groupmask))
7455 free_cpumask_var(groupmask);
7457 #else /* !CONFIG_SCHED_DEBUG */
7458 # define sched_domain_debug(sd, cpu) do { } while (0)
7459 #endif /* CONFIG_SCHED_DEBUG */
7461 static int sd_degenerate(struct sched_domain *sd)
7463 if (cpumask_weight(sched_domain_span(sd)) == 1)
7466 /* Following flags need at least 2 groups */
7467 if (sd->flags & (SD_LOAD_BALANCE |
7468 SD_BALANCE_NEWIDLE |
7472 SD_SHARE_PKG_RESOURCES)) {
7473 if (sd->groups != sd->groups->next)
7477 /* Following flags don't use groups */
7478 if (sd->flags & (SD_WAKE_IDLE |
7487 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
7489 unsigned long cflags = sd->flags, pflags = parent->flags;
7491 if (sd_degenerate(parent))
7494 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
7497 /* Does parent contain flags not in child? */
7498 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
7499 if (cflags & SD_WAKE_AFFINE)
7500 pflags &= ~SD_WAKE_BALANCE;
7501 /* Flags needing groups don't count if only 1 group in parent */
7502 if (parent->groups == parent->groups->next) {
7503 pflags &= ~(SD_LOAD_BALANCE |
7504 SD_BALANCE_NEWIDLE |
7508 SD_SHARE_PKG_RESOURCES);
7509 if (nr_node_ids == 1)
7510 pflags &= ~SD_SERIALIZE;
7512 if (~cflags & pflags)
7518 static void free_rootdomain(struct root_domain *rd)
7520 cpupri_cleanup(&rd->cpupri);
7522 free_cpumask_var(rd->rto_mask);
7523 free_cpumask_var(rd->online);
7524 free_cpumask_var(rd->span);
7528 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7530 struct root_domain *old_rd = NULL;
7531 unsigned long flags;
7533 spin_lock_irqsave(&rq->lock, flags);
7538 if (cpumask_test_cpu(rq->cpu, old_rd->online))
7541 cpumask_clear_cpu(rq->cpu, old_rd->span);
7544 * If we dont want to free the old_rt yet then
7545 * set old_rd to NULL to skip the freeing later
7548 if (!atomic_dec_and_test(&old_rd->refcount))
7552 atomic_inc(&rd->refcount);
7555 cpumask_set_cpu(rq->cpu, rd->span);
7556 if (cpumask_test_cpu(rq->cpu, cpu_online_mask))
7559 spin_unlock_irqrestore(&rq->lock, flags);
7562 free_rootdomain(old_rd);
7565 static int __init_refok init_rootdomain(struct root_domain *rd, bool bootmem)
7567 memset(rd, 0, sizeof(*rd));
7570 alloc_bootmem_cpumask_var(&def_root_domain.span);
7571 alloc_bootmem_cpumask_var(&def_root_domain.online);
7572 alloc_bootmem_cpumask_var(&def_root_domain.rto_mask);
7573 cpupri_init(&rd->cpupri, true);
7577 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
7579 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
7581 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
7584 if (cpupri_init(&rd->cpupri, false) != 0)
7589 free_cpumask_var(rd->rto_mask);
7591 free_cpumask_var(rd->online);
7593 free_cpumask_var(rd->span);
7598 static void init_defrootdomain(void)
7600 init_rootdomain(&def_root_domain, true);
7602 atomic_set(&def_root_domain.refcount, 1);
7605 static struct root_domain *alloc_rootdomain(void)
7607 struct root_domain *rd;
7609 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
7613 if (init_rootdomain(rd, false) != 0) {
7622 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7623 * hold the hotplug lock.
7626 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
7628 struct rq *rq = cpu_rq(cpu);
7629 struct sched_domain *tmp;
7631 /* Remove the sched domains which do not contribute to scheduling. */
7632 for (tmp = sd; tmp; ) {
7633 struct sched_domain *parent = tmp->parent;
7637 if (sd_parent_degenerate(tmp, parent)) {
7638 tmp->parent = parent->parent;
7640 parent->parent->child = tmp;
7645 if (sd && sd_degenerate(sd)) {
7651 sched_domain_debug(sd, cpu);
7653 rq_attach_root(rq, rd);
7654 rcu_assign_pointer(rq->sd, sd);
7657 /* cpus with isolated domains */
7658 static cpumask_var_t cpu_isolated_map;
7660 /* Setup the mask of cpus configured for isolated domains */
7661 static int __init isolated_cpu_setup(char *str)
7663 cpulist_parse(str, cpu_isolated_map);
7667 __setup("isolcpus=", isolated_cpu_setup);
7670 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
7671 * to a function which identifies what group(along with sched group) a CPU
7672 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
7673 * (due to the fact that we keep track of groups covered with a struct cpumask).
7675 * init_sched_build_groups will build a circular linked list of the groups
7676 * covered by the given span, and will set each group's ->cpumask correctly,
7677 * and ->cpu_power to 0.
7680 init_sched_build_groups(const struct cpumask *span,
7681 const struct cpumask *cpu_map,
7682 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
7683 struct sched_group **sg,
7684 struct cpumask *tmpmask),
7685 struct cpumask *covered, struct cpumask *tmpmask)
7687 struct sched_group *first = NULL, *last = NULL;
7690 cpumask_clear(covered);
7692 for_each_cpu(i, span) {
7693 struct sched_group *sg;
7694 int group = group_fn(i, cpu_map, &sg, tmpmask);
7697 if (cpumask_test_cpu(i, covered))
7700 cpumask_clear(sched_group_cpus(sg));
7701 sg->__cpu_power = 0;
7703 for_each_cpu(j, span) {
7704 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
7707 cpumask_set_cpu(j, covered);
7708 cpumask_set_cpu(j, sched_group_cpus(sg));
7719 #define SD_NODES_PER_DOMAIN 16
7724 * find_next_best_node - find the next node to include in a sched_domain
7725 * @node: node whose sched_domain we're building
7726 * @used_nodes: nodes already in the sched_domain
7728 * Find the next node to include in a given scheduling domain. Simply
7729 * finds the closest node not already in the @used_nodes map.
7731 * Should use nodemask_t.
7733 static int find_next_best_node(int node, nodemask_t *used_nodes)
7735 int i, n, val, min_val, best_node = 0;
7739 for (i = 0; i < nr_node_ids; i++) {
7740 /* Start at @node */
7741 n = (node + i) % nr_node_ids;
7743 if (!nr_cpus_node(n))
7746 /* Skip already used nodes */
7747 if (node_isset(n, *used_nodes))
7750 /* Simple min distance search */
7751 val = node_distance(node, n);
7753 if (val < min_val) {
7759 node_set(best_node, *used_nodes);
7764 * sched_domain_node_span - get a cpumask for a node's sched_domain
7765 * @node: node whose cpumask we're constructing
7766 * @span: resulting cpumask
7768 * Given a node, construct a good cpumask for its sched_domain to span. It
7769 * should be one that prevents unnecessary balancing, but also spreads tasks
7772 static void sched_domain_node_span(int node, struct cpumask *span)
7774 nodemask_t used_nodes;
7777 cpumask_clear(span);
7778 nodes_clear(used_nodes);
7780 cpumask_or(span, span, cpumask_of_node(node));
7781 node_set(node, used_nodes);
7783 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
7784 int next_node = find_next_best_node(node, &used_nodes);
7786 cpumask_or(span, span, cpumask_of_node(next_node));
7789 #endif /* CONFIG_NUMA */
7791 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
7794 * The cpus mask in sched_group and sched_domain hangs off the end.
7795 * FIXME: use cpumask_var_t or dynamic percpu alloc to avoid wasting space
7796 * for nr_cpu_ids < CONFIG_NR_CPUS.
7798 struct static_sched_group {
7799 struct sched_group sg;
7800 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
7803 struct static_sched_domain {
7804 struct sched_domain sd;
7805 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
7809 * SMT sched-domains:
7811 #ifdef CONFIG_SCHED_SMT
7812 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
7813 static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
7816 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
7817 struct sched_group **sg, struct cpumask *unused)
7820 *sg = &per_cpu(sched_group_cpus, cpu).sg;
7823 #endif /* CONFIG_SCHED_SMT */
7826 * multi-core sched-domains:
7828 #ifdef CONFIG_SCHED_MC
7829 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
7830 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
7831 #endif /* CONFIG_SCHED_MC */
7833 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7835 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
7836 struct sched_group **sg, struct cpumask *mask)
7840 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
7841 group = cpumask_first(mask);
7843 *sg = &per_cpu(sched_group_core, group).sg;
7846 #elif defined(CONFIG_SCHED_MC)
7848 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
7849 struct sched_group **sg, struct cpumask *unused)
7852 *sg = &per_cpu(sched_group_core, cpu).sg;
7857 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
7858 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
7861 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
7862 struct sched_group **sg, struct cpumask *mask)
7865 #ifdef CONFIG_SCHED_MC
7866 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
7867 group = cpumask_first(mask);
7868 #elif defined(CONFIG_SCHED_SMT)
7869 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
7870 group = cpumask_first(mask);
7875 *sg = &per_cpu(sched_group_phys, group).sg;
7881 * The init_sched_build_groups can't handle what we want to do with node
7882 * groups, so roll our own. Now each node has its own list of groups which
7883 * gets dynamically allocated.
7885 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
7886 static struct sched_group ***sched_group_nodes_bycpu;
7888 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
7889 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
7891 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
7892 struct sched_group **sg,
7893 struct cpumask *nodemask)
7897 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
7898 group = cpumask_first(nodemask);
7901 *sg = &per_cpu(sched_group_allnodes, group).sg;
7905 static void init_numa_sched_groups_power(struct sched_group *group_head)
7907 struct sched_group *sg = group_head;
7913 for_each_cpu(j, sched_group_cpus(sg)) {
7914 struct sched_domain *sd;
7916 sd = &per_cpu(phys_domains, j).sd;
7917 if (j != cpumask_first(sched_group_cpus(sd->groups))) {
7919 * Only add "power" once for each
7925 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
7928 } while (sg != group_head);
7930 #endif /* CONFIG_NUMA */
7933 /* Free memory allocated for various sched_group structures */
7934 static void free_sched_groups(const struct cpumask *cpu_map,
7935 struct cpumask *nodemask)
7939 for_each_cpu(cpu, cpu_map) {
7940 struct sched_group **sched_group_nodes
7941 = sched_group_nodes_bycpu[cpu];
7943 if (!sched_group_nodes)
7946 for (i = 0; i < nr_node_ids; i++) {
7947 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7949 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
7950 if (cpumask_empty(nodemask))
7960 if (oldsg != sched_group_nodes[i])
7963 kfree(sched_group_nodes);
7964 sched_group_nodes_bycpu[cpu] = NULL;
7967 #else /* !CONFIG_NUMA */
7968 static void free_sched_groups(const struct cpumask *cpu_map,
7969 struct cpumask *nodemask)
7972 #endif /* CONFIG_NUMA */
7975 * Initialize sched groups cpu_power.
7977 * cpu_power indicates the capacity of sched group, which is used while
7978 * distributing the load between different sched groups in a sched domain.
7979 * Typically cpu_power for all the groups in a sched domain will be same unless
7980 * there are asymmetries in the topology. If there are asymmetries, group
7981 * having more cpu_power will pickup more load compared to the group having
7984 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7985 * the maximum number of tasks a group can handle in the presence of other idle
7986 * or lightly loaded groups in the same sched domain.
7988 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7990 struct sched_domain *child;
7991 struct sched_group *group;
7993 WARN_ON(!sd || !sd->groups);
7995 if (cpu != cpumask_first(sched_group_cpus(sd->groups)))
8000 sd->groups->__cpu_power = 0;
8003 * For perf policy, if the groups in child domain share resources
8004 * (for example cores sharing some portions of the cache hierarchy
8005 * or SMT), then set this domain groups cpu_power such that each group
8006 * can handle only one task, when there are other idle groups in the
8007 * same sched domain.
8009 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
8011 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
8012 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
8017 * add cpu_power of each child group to this groups cpu_power
8019 group = child->groups;
8021 sg_inc_cpu_power(sd->groups, group->__cpu_power);
8022 group = group->next;
8023 } while (group != child->groups);
8027 * Initializers for schedule domains
8028 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8031 #ifdef CONFIG_SCHED_DEBUG
8032 # define SD_INIT_NAME(sd, type) sd->name = #type
8034 # define SD_INIT_NAME(sd, type) do { } while (0)
8037 #define SD_INIT(sd, type) sd_init_##type(sd)
8039 #define SD_INIT_FUNC(type) \
8040 static noinline void sd_init_##type(struct sched_domain *sd) \
8042 memset(sd, 0, sizeof(*sd)); \
8043 *sd = SD_##type##_INIT; \
8044 sd->level = SD_LV_##type; \
8045 SD_INIT_NAME(sd, type); \
8050 SD_INIT_FUNC(ALLNODES)
8053 #ifdef CONFIG_SCHED_SMT
8054 SD_INIT_FUNC(SIBLING)
8056 #ifdef CONFIG_SCHED_MC
8060 static int default_relax_domain_level = -1;
8062 static int __init setup_relax_domain_level(char *str)
8066 val = simple_strtoul(str, NULL, 0);
8067 if (val < SD_LV_MAX)
8068 default_relax_domain_level = val;
8072 __setup("relax_domain_level=", setup_relax_domain_level);
8074 static void set_domain_attribute(struct sched_domain *sd,
8075 struct sched_domain_attr *attr)
8079 if (!attr || attr->relax_domain_level < 0) {
8080 if (default_relax_domain_level < 0)
8083 request = default_relax_domain_level;
8085 request = attr->relax_domain_level;
8086 if (request < sd->level) {
8087 /* turn off idle balance on this domain */
8088 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
8090 /* turn on idle balance on this domain */
8091 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
8096 * Build sched domains for a given set of cpus and attach the sched domains
8097 * to the individual cpus
8099 static int __build_sched_domains(const struct cpumask *cpu_map,
8100 struct sched_domain_attr *attr)
8102 int i, err = -ENOMEM;
8103 struct root_domain *rd;
8104 cpumask_var_t nodemask, this_sibling_map, this_core_map, send_covered,
8107 cpumask_var_t domainspan, covered, notcovered;
8108 struct sched_group **sched_group_nodes = NULL;
8109 int sd_allnodes = 0;
8111 if (!alloc_cpumask_var(&domainspan, GFP_KERNEL))
8113 if (!alloc_cpumask_var(&covered, GFP_KERNEL))
8114 goto free_domainspan;
8115 if (!alloc_cpumask_var(¬covered, GFP_KERNEL))
8119 if (!alloc_cpumask_var(&nodemask, GFP_KERNEL))
8120 goto free_notcovered;
8121 if (!alloc_cpumask_var(&this_sibling_map, GFP_KERNEL))
8123 if (!alloc_cpumask_var(&this_core_map, GFP_KERNEL))
8124 goto free_this_sibling_map;
8125 if (!alloc_cpumask_var(&send_covered, GFP_KERNEL))
8126 goto free_this_core_map;
8127 if (!alloc_cpumask_var(&tmpmask, GFP_KERNEL))
8128 goto free_send_covered;
8132 * Allocate the per-node list of sched groups
8134 sched_group_nodes = kcalloc(nr_node_ids, sizeof(struct sched_group *),
8136 if (!sched_group_nodes) {
8137 printk(KERN_WARNING "Can not alloc sched group node list\n");
8142 rd = alloc_rootdomain();
8144 printk(KERN_WARNING "Cannot alloc root domain\n");
8145 goto free_sched_groups;
8149 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = sched_group_nodes;
8153 * Set up domains for cpus specified by the cpu_map.
8155 for_each_cpu(i, cpu_map) {
8156 struct sched_domain *sd = NULL, *p;
8158 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(i)), cpu_map);
8161 if (cpumask_weight(cpu_map) >
8162 SD_NODES_PER_DOMAIN*cpumask_weight(nodemask)) {
8163 sd = &per_cpu(allnodes_domains, i).sd;
8164 SD_INIT(sd, ALLNODES);
8165 set_domain_attribute(sd, attr);
8166 cpumask_copy(sched_domain_span(sd), cpu_map);
8167 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
8173 sd = &per_cpu(node_domains, i).sd;
8175 set_domain_attribute(sd, attr);
8176 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
8180 cpumask_and(sched_domain_span(sd),
8181 sched_domain_span(sd), cpu_map);
8185 sd = &per_cpu(phys_domains, i).sd;
8187 set_domain_attribute(sd, attr);
8188 cpumask_copy(sched_domain_span(sd), nodemask);
8192 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
8194 #ifdef CONFIG_SCHED_MC
8196 sd = &per_cpu(core_domains, i).sd;
8198 set_domain_attribute(sd, attr);
8199 cpumask_and(sched_domain_span(sd), cpu_map,
8200 cpu_coregroup_mask(i));
8203 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
8206 #ifdef CONFIG_SCHED_SMT
8208 sd = &per_cpu(cpu_domains, i).sd;
8209 SD_INIT(sd, SIBLING);
8210 set_domain_attribute(sd, attr);
8211 cpumask_and(sched_domain_span(sd),
8212 topology_thread_cpumask(i), cpu_map);
8215 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
8219 #ifdef CONFIG_SCHED_SMT
8220 /* Set up CPU (sibling) groups */
8221 for_each_cpu(i, cpu_map) {
8222 cpumask_and(this_sibling_map,
8223 topology_thread_cpumask(i), cpu_map);
8224 if (i != cpumask_first(this_sibling_map))
8227 init_sched_build_groups(this_sibling_map, cpu_map,
8229 send_covered, tmpmask);
8233 #ifdef CONFIG_SCHED_MC
8234 /* Set up multi-core groups */
8235 for_each_cpu(i, cpu_map) {
8236 cpumask_and(this_core_map, cpu_coregroup_mask(i), cpu_map);
8237 if (i != cpumask_first(this_core_map))
8240 init_sched_build_groups(this_core_map, cpu_map,
8242 send_covered, tmpmask);
8246 /* Set up physical groups */
8247 for (i = 0; i < nr_node_ids; i++) {
8248 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8249 if (cpumask_empty(nodemask))
8252 init_sched_build_groups(nodemask, cpu_map,
8254 send_covered, tmpmask);
8258 /* Set up node groups */
8260 init_sched_build_groups(cpu_map, cpu_map,
8261 &cpu_to_allnodes_group,
8262 send_covered, tmpmask);
8265 for (i = 0; i < nr_node_ids; i++) {
8266 /* Set up node groups */
8267 struct sched_group *sg, *prev;
8270 cpumask_clear(covered);
8271 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8272 if (cpumask_empty(nodemask)) {
8273 sched_group_nodes[i] = NULL;
8277 sched_domain_node_span(i, domainspan);
8278 cpumask_and(domainspan, domainspan, cpu_map);
8280 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8283 printk(KERN_WARNING "Can not alloc domain group for "
8287 sched_group_nodes[i] = sg;
8288 for_each_cpu(j, nodemask) {
8289 struct sched_domain *sd;
8291 sd = &per_cpu(node_domains, j).sd;
8294 sg->__cpu_power = 0;
8295 cpumask_copy(sched_group_cpus(sg), nodemask);
8297 cpumask_or(covered, covered, nodemask);
8300 for (j = 0; j < nr_node_ids; j++) {
8301 int n = (i + j) % nr_node_ids;
8303 cpumask_complement(notcovered, covered);
8304 cpumask_and(tmpmask, notcovered, cpu_map);
8305 cpumask_and(tmpmask, tmpmask, domainspan);
8306 if (cpumask_empty(tmpmask))
8309 cpumask_and(tmpmask, tmpmask, cpumask_of_node(n));
8310 if (cpumask_empty(tmpmask))
8313 sg = kmalloc_node(sizeof(struct sched_group) +
8318 "Can not alloc domain group for node %d\n", j);
8321 sg->__cpu_power = 0;
8322 cpumask_copy(sched_group_cpus(sg), tmpmask);
8323 sg->next = prev->next;
8324 cpumask_or(covered, covered, tmpmask);
8331 /* Calculate CPU power for physical packages and nodes */
8332 #ifdef CONFIG_SCHED_SMT
8333 for_each_cpu(i, cpu_map) {
8334 struct sched_domain *sd = &per_cpu(cpu_domains, i).sd;
8336 init_sched_groups_power(i, sd);
8339 #ifdef CONFIG_SCHED_MC
8340 for_each_cpu(i, cpu_map) {
8341 struct sched_domain *sd = &per_cpu(core_domains, i).sd;
8343 init_sched_groups_power(i, sd);
8347 for_each_cpu(i, cpu_map) {
8348 struct sched_domain *sd = &per_cpu(phys_domains, i).sd;
8350 init_sched_groups_power(i, sd);
8354 for (i = 0; i < nr_node_ids; i++)
8355 init_numa_sched_groups_power(sched_group_nodes[i]);
8358 struct sched_group *sg;
8360 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
8362 init_numa_sched_groups_power(sg);
8366 /* Attach the domains */
8367 for_each_cpu(i, cpu_map) {
8368 struct sched_domain *sd;
8369 #ifdef CONFIG_SCHED_SMT
8370 sd = &per_cpu(cpu_domains, i).sd;
8371 #elif defined(CONFIG_SCHED_MC)
8372 sd = &per_cpu(core_domains, i).sd;
8374 sd = &per_cpu(phys_domains, i).sd;
8376 cpu_attach_domain(sd, rd, i);
8382 free_cpumask_var(tmpmask);
8384 free_cpumask_var(send_covered);
8386 free_cpumask_var(this_core_map);
8387 free_this_sibling_map:
8388 free_cpumask_var(this_sibling_map);
8390 free_cpumask_var(nodemask);
8393 free_cpumask_var(notcovered);
8395 free_cpumask_var(covered);
8397 free_cpumask_var(domainspan);
8404 kfree(sched_group_nodes);
8410 free_sched_groups(cpu_map, tmpmask);
8411 free_rootdomain(rd);
8416 static int build_sched_domains(const struct cpumask *cpu_map)
8418 return __build_sched_domains(cpu_map, NULL);
8421 static struct cpumask *doms_cur; /* current sched domains */
8422 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
8423 static struct sched_domain_attr *dattr_cur;
8424 /* attribues of custom domains in 'doms_cur' */
8427 * Special case: If a kmalloc of a doms_cur partition (array of
8428 * cpumask) fails, then fallback to a single sched domain,
8429 * as determined by the single cpumask fallback_doms.
8431 static cpumask_var_t fallback_doms;
8434 * arch_update_cpu_topology lets virtualized architectures update the
8435 * cpu core maps. It is supposed to return 1 if the topology changed
8436 * or 0 if it stayed the same.
8438 int __attribute__((weak)) arch_update_cpu_topology(void)
8444 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
8445 * For now this just excludes isolated cpus, but could be used to
8446 * exclude other special cases in the future.
8448 static int arch_init_sched_domains(const struct cpumask *cpu_map)
8452 arch_update_cpu_topology();
8454 doms_cur = kmalloc(cpumask_size(), GFP_KERNEL);
8456 doms_cur = fallback_doms;
8457 cpumask_andnot(doms_cur, cpu_map, cpu_isolated_map);
8459 err = build_sched_domains(doms_cur);
8460 register_sched_domain_sysctl();
8465 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
8466 struct cpumask *tmpmask)
8468 free_sched_groups(cpu_map, tmpmask);
8472 * Detach sched domains from a group of cpus specified in cpu_map
8473 * These cpus will now be attached to the NULL domain
8475 static void detach_destroy_domains(const struct cpumask *cpu_map)
8477 /* Save because hotplug lock held. */
8478 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
8481 for_each_cpu(i, cpu_map)
8482 cpu_attach_domain(NULL, &def_root_domain, i);
8483 synchronize_sched();
8484 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
8487 /* handle null as "default" */
8488 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
8489 struct sched_domain_attr *new, int idx_new)
8491 struct sched_domain_attr tmp;
8498 return !memcmp(cur ? (cur + idx_cur) : &tmp,
8499 new ? (new + idx_new) : &tmp,
8500 sizeof(struct sched_domain_attr));
8504 * Partition sched domains as specified by the 'ndoms_new'
8505 * cpumasks in the array doms_new[] of cpumasks. This compares
8506 * doms_new[] to the current sched domain partitioning, doms_cur[].
8507 * It destroys each deleted domain and builds each new domain.
8509 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
8510 * The masks don't intersect (don't overlap.) We should setup one
8511 * sched domain for each mask. CPUs not in any of the cpumasks will
8512 * not be load balanced. If the same cpumask appears both in the
8513 * current 'doms_cur' domains and in the new 'doms_new', we can leave
8516 * The passed in 'doms_new' should be kmalloc'd. This routine takes
8517 * ownership of it and will kfree it when done with it. If the caller
8518 * failed the kmalloc call, then it can pass in doms_new == NULL &&
8519 * ndoms_new == 1, and partition_sched_domains() will fallback to
8520 * the single partition 'fallback_doms', it also forces the domains
8523 * If doms_new == NULL it will be replaced with cpu_online_mask.
8524 * ndoms_new == 0 is a special case for destroying existing domains,
8525 * and it will not create the default domain.
8527 * Call with hotplug lock held
8529 /* FIXME: Change to struct cpumask *doms_new[] */
8530 void partition_sched_domains(int ndoms_new, struct cpumask *doms_new,
8531 struct sched_domain_attr *dattr_new)
8536 mutex_lock(&sched_domains_mutex);
8538 /* always unregister in case we don't destroy any domains */
8539 unregister_sched_domain_sysctl();
8541 /* Let architecture update cpu core mappings. */
8542 new_topology = arch_update_cpu_topology();
8544 n = doms_new ? ndoms_new : 0;
8546 /* Destroy deleted domains */
8547 for (i = 0; i < ndoms_cur; i++) {
8548 for (j = 0; j < n && !new_topology; j++) {
8549 if (cpumask_equal(&doms_cur[i], &doms_new[j])
8550 && dattrs_equal(dattr_cur, i, dattr_new, j))
8553 /* no match - a current sched domain not in new doms_new[] */
8554 detach_destroy_domains(doms_cur + i);
8559 if (doms_new == NULL) {
8561 doms_new = fallback_doms;
8562 cpumask_andnot(&doms_new[0], cpu_online_mask, cpu_isolated_map);
8563 WARN_ON_ONCE(dattr_new);
8566 /* Build new domains */
8567 for (i = 0; i < ndoms_new; i++) {
8568 for (j = 0; j < ndoms_cur && !new_topology; j++) {
8569 if (cpumask_equal(&doms_new[i], &doms_cur[j])
8570 && dattrs_equal(dattr_new, i, dattr_cur, j))
8573 /* no match - add a new doms_new */
8574 __build_sched_domains(doms_new + i,
8575 dattr_new ? dattr_new + i : NULL);
8580 /* Remember the new sched domains */
8581 if (doms_cur != fallback_doms)
8583 kfree(dattr_cur); /* kfree(NULL) is safe */
8584 doms_cur = doms_new;
8585 dattr_cur = dattr_new;
8586 ndoms_cur = ndoms_new;
8588 register_sched_domain_sysctl();
8590 mutex_unlock(&sched_domains_mutex);
8593 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
8594 static void arch_reinit_sched_domains(void)
8598 /* Destroy domains first to force the rebuild */
8599 partition_sched_domains(0, NULL, NULL);
8601 rebuild_sched_domains();
8605 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
8607 unsigned int level = 0;
8609 if (sscanf(buf, "%u", &level) != 1)
8613 * level is always be positive so don't check for
8614 * level < POWERSAVINGS_BALANCE_NONE which is 0
8615 * What happens on 0 or 1 byte write,
8616 * need to check for count as well?
8619 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
8623 sched_smt_power_savings = level;
8625 sched_mc_power_savings = level;
8627 arch_reinit_sched_domains();
8632 #ifdef CONFIG_SCHED_MC
8633 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
8636 return sprintf(page, "%u\n", sched_mc_power_savings);
8638 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
8639 const char *buf, size_t count)
8641 return sched_power_savings_store(buf, count, 0);
8643 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
8644 sched_mc_power_savings_show,
8645 sched_mc_power_savings_store);
8648 #ifdef CONFIG_SCHED_SMT
8649 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
8652 return sprintf(page, "%u\n", sched_smt_power_savings);
8654 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
8655 const char *buf, size_t count)
8657 return sched_power_savings_store(buf, count, 1);
8659 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
8660 sched_smt_power_savings_show,
8661 sched_smt_power_savings_store);
8664 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
8668 #ifdef CONFIG_SCHED_SMT
8670 err = sysfs_create_file(&cls->kset.kobj,
8671 &attr_sched_smt_power_savings.attr);
8673 #ifdef CONFIG_SCHED_MC
8674 if (!err && mc_capable())
8675 err = sysfs_create_file(&cls->kset.kobj,
8676 &attr_sched_mc_power_savings.attr);
8680 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
8682 #ifndef CONFIG_CPUSETS
8684 * Add online and remove offline CPUs from the scheduler domains.
8685 * When cpusets are enabled they take over this function.
8687 static int update_sched_domains(struct notifier_block *nfb,
8688 unsigned long action, void *hcpu)
8692 case CPU_ONLINE_FROZEN:
8694 case CPU_DEAD_FROZEN:
8695 partition_sched_domains(1, NULL, NULL);
8704 static int update_runtime(struct notifier_block *nfb,
8705 unsigned long action, void *hcpu)
8707 int cpu = (int)(long)hcpu;
8710 case CPU_DOWN_PREPARE:
8711 case CPU_DOWN_PREPARE_FROZEN:
8712 disable_runtime(cpu_rq(cpu));
8715 case CPU_DOWN_FAILED:
8716 case CPU_DOWN_FAILED_FROZEN:
8718 case CPU_ONLINE_FROZEN:
8719 enable_runtime(cpu_rq(cpu));
8727 void __init sched_init_smp(void)
8729 cpumask_var_t non_isolated_cpus;
8731 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
8733 #if defined(CONFIG_NUMA)
8734 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
8736 BUG_ON(sched_group_nodes_bycpu == NULL);
8739 mutex_lock(&sched_domains_mutex);
8740 arch_init_sched_domains(cpu_online_mask);
8741 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
8742 if (cpumask_empty(non_isolated_cpus))
8743 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
8744 mutex_unlock(&sched_domains_mutex);
8747 #ifndef CONFIG_CPUSETS
8748 /* XXX: Theoretical race here - CPU may be hotplugged now */
8749 hotcpu_notifier(update_sched_domains, 0);
8752 /* RT runtime code needs to handle some hotplug events */
8753 hotcpu_notifier(update_runtime, 0);
8757 /* Move init over to a non-isolated CPU */
8758 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
8760 sched_init_granularity();
8761 free_cpumask_var(non_isolated_cpus);
8763 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
8764 init_sched_rt_class();
8767 void __init sched_init_smp(void)
8769 sched_init_granularity();
8771 #endif /* CONFIG_SMP */
8773 int in_sched_functions(unsigned long addr)
8775 return in_lock_functions(addr) ||
8776 (addr >= (unsigned long)__sched_text_start
8777 && addr < (unsigned long)__sched_text_end);
8780 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
8782 cfs_rq->tasks_timeline = RB_ROOT;
8783 INIT_LIST_HEAD(&cfs_rq->tasks);
8784 #ifdef CONFIG_FAIR_GROUP_SCHED
8787 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8790 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
8792 struct rt_prio_array *array;
8795 array = &rt_rq->active;
8796 for (i = 0; i < MAX_RT_PRIO; i++) {
8797 INIT_LIST_HEAD(array->queue + i);
8798 __clear_bit(i, array->bitmap);
8800 /* delimiter for bitsearch: */
8801 __set_bit(MAX_RT_PRIO, array->bitmap);
8803 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8804 rt_rq->highest_prio.curr = MAX_RT_PRIO;
8806 rt_rq->highest_prio.next = MAX_RT_PRIO;
8810 rt_rq->rt_nr_migratory = 0;
8811 rt_rq->overloaded = 0;
8812 plist_head_init(&rq->rt.pushable_tasks, &rq->lock);
8816 rt_rq->rt_throttled = 0;
8817 rt_rq->rt_runtime = 0;
8818 spin_lock_init(&rt_rq->rt_runtime_lock);
8820 #ifdef CONFIG_RT_GROUP_SCHED
8821 rt_rq->rt_nr_boosted = 0;
8826 #ifdef CONFIG_FAIR_GROUP_SCHED
8827 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8828 struct sched_entity *se, int cpu, int add,
8829 struct sched_entity *parent)
8831 struct rq *rq = cpu_rq(cpu);
8832 tg->cfs_rq[cpu] = cfs_rq;
8833 init_cfs_rq(cfs_rq, rq);
8836 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
8839 /* se could be NULL for init_task_group */
8844 se->cfs_rq = &rq->cfs;
8846 se->cfs_rq = parent->my_q;
8849 se->load.weight = tg->shares;
8850 se->load.inv_weight = 0;
8851 se->parent = parent;
8855 #ifdef CONFIG_RT_GROUP_SCHED
8856 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8857 struct sched_rt_entity *rt_se, int cpu, int add,
8858 struct sched_rt_entity *parent)
8860 struct rq *rq = cpu_rq(cpu);
8862 tg->rt_rq[cpu] = rt_rq;
8863 init_rt_rq(rt_rq, rq);
8865 rt_rq->rt_se = rt_se;
8866 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8868 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
8870 tg->rt_se[cpu] = rt_se;
8875 rt_se->rt_rq = &rq->rt;
8877 rt_se->rt_rq = parent->my_q;
8879 rt_se->my_q = rt_rq;
8880 rt_se->parent = parent;
8881 INIT_LIST_HEAD(&rt_se->run_list);
8885 void __init sched_init(void)
8888 unsigned long alloc_size = 0, ptr;
8890 #ifdef CONFIG_FAIR_GROUP_SCHED
8891 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8893 #ifdef CONFIG_RT_GROUP_SCHED
8894 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8896 #ifdef CONFIG_USER_SCHED
8899 #ifdef CONFIG_CPUMASK_OFFSTACK
8900 alloc_size += num_possible_cpus() * cpumask_size();
8903 * As sched_init() is called before page_alloc is setup,
8904 * we use alloc_bootmem().
8907 ptr = (unsigned long)alloc_bootmem(alloc_size);
8909 #ifdef CONFIG_FAIR_GROUP_SCHED
8910 init_task_group.se = (struct sched_entity **)ptr;
8911 ptr += nr_cpu_ids * sizeof(void **);
8913 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
8914 ptr += nr_cpu_ids * sizeof(void **);
8916 #ifdef CONFIG_USER_SCHED
8917 root_task_group.se = (struct sched_entity **)ptr;
8918 ptr += nr_cpu_ids * sizeof(void **);
8920 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8921 ptr += nr_cpu_ids * sizeof(void **);
8922 #endif /* CONFIG_USER_SCHED */
8923 #endif /* CONFIG_FAIR_GROUP_SCHED */
8924 #ifdef CONFIG_RT_GROUP_SCHED
8925 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
8926 ptr += nr_cpu_ids * sizeof(void **);
8928 init_task_group.rt_rq = (struct rt_rq **)ptr;
8929 ptr += nr_cpu_ids * sizeof(void **);
8931 #ifdef CONFIG_USER_SCHED
8932 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8933 ptr += nr_cpu_ids * sizeof(void **);
8935 root_task_group.rt_rq = (struct rt_rq **)ptr;
8936 ptr += nr_cpu_ids * sizeof(void **);
8937 #endif /* CONFIG_USER_SCHED */
8938 #endif /* CONFIG_RT_GROUP_SCHED */
8939 #ifdef CONFIG_CPUMASK_OFFSTACK
8940 for_each_possible_cpu(i) {
8941 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
8942 ptr += cpumask_size();
8944 #endif /* CONFIG_CPUMASK_OFFSTACK */
8948 init_defrootdomain();
8951 init_rt_bandwidth(&def_rt_bandwidth,
8952 global_rt_period(), global_rt_runtime());
8954 #ifdef CONFIG_RT_GROUP_SCHED
8955 init_rt_bandwidth(&init_task_group.rt_bandwidth,
8956 global_rt_period(), global_rt_runtime());
8957 #ifdef CONFIG_USER_SCHED
8958 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8959 global_rt_period(), RUNTIME_INF);
8960 #endif /* CONFIG_USER_SCHED */
8961 #endif /* CONFIG_RT_GROUP_SCHED */
8963 #ifdef CONFIG_GROUP_SCHED
8964 list_add(&init_task_group.list, &task_groups);
8965 INIT_LIST_HEAD(&init_task_group.children);
8967 #ifdef CONFIG_USER_SCHED
8968 INIT_LIST_HEAD(&root_task_group.children);
8969 init_task_group.parent = &root_task_group;
8970 list_add(&init_task_group.siblings, &root_task_group.children);
8971 #endif /* CONFIG_USER_SCHED */
8972 #endif /* CONFIG_GROUP_SCHED */
8974 for_each_possible_cpu(i) {
8978 spin_lock_init(&rq->lock);
8980 init_cfs_rq(&rq->cfs, rq);
8981 init_rt_rq(&rq->rt, rq);
8982 #ifdef CONFIG_FAIR_GROUP_SCHED
8983 init_task_group.shares = init_task_group_load;
8984 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8985 #ifdef CONFIG_CGROUP_SCHED
8987 * How much cpu bandwidth does init_task_group get?
8989 * In case of task-groups formed thr' the cgroup filesystem, it
8990 * gets 100% of the cpu resources in the system. This overall
8991 * system cpu resource is divided among the tasks of
8992 * init_task_group and its child task-groups in a fair manner,
8993 * based on each entity's (task or task-group's) weight
8994 * (se->load.weight).
8996 * In other words, if init_task_group has 10 tasks of weight
8997 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8998 * then A0's share of the cpu resource is:
9000 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9002 * We achieve this by letting init_task_group's tasks sit
9003 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9005 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
9006 #elif defined CONFIG_USER_SCHED
9007 root_task_group.shares = NICE_0_LOAD;
9008 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
9010 * In case of task-groups formed thr' the user id of tasks,
9011 * init_task_group represents tasks belonging to root user.
9012 * Hence it forms a sibling of all subsequent groups formed.
9013 * In this case, init_task_group gets only a fraction of overall
9014 * system cpu resource, based on the weight assigned to root
9015 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9016 * by letting tasks of init_task_group sit in a separate cfs_rq
9017 * (init_cfs_rq) and having one entity represent this group of
9018 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9020 init_tg_cfs_entry(&init_task_group,
9021 &per_cpu(init_cfs_rq, i),
9022 &per_cpu(init_sched_entity, i), i, 1,
9023 root_task_group.se[i]);
9026 #endif /* CONFIG_FAIR_GROUP_SCHED */
9028 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9029 #ifdef CONFIG_RT_GROUP_SCHED
9030 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
9031 #ifdef CONFIG_CGROUP_SCHED
9032 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
9033 #elif defined CONFIG_USER_SCHED
9034 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
9035 init_tg_rt_entry(&init_task_group,
9036 &per_cpu(init_rt_rq, i),
9037 &per_cpu(init_sched_rt_entity, i), i, 1,
9038 root_task_group.rt_se[i]);
9042 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
9043 rq->cpu_load[j] = 0;
9047 rq->active_balance = 0;
9048 rq->next_balance = jiffies;
9052 rq->migration_thread = NULL;
9053 INIT_LIST_HEAD(&rq->migration_queue);
9054 rq_attach_root(rq, &def_root_domain);
9057 atomic_set(&rq->nr_iowait, 0);
9060 set_load_weight(&init_task);
9062 #ifdef CONFIG_PREEMPT_NOTIFIERS
9063 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
9067 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9070 #ifdef CONFIG_RT_MUTEXES
9071 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
9075 * The boot idle thread does lazy MMU switching as well:
9077 atomic_inc(&init_mm.mm_count);
9078 enter_lazy_tlb(&init_mm, current);
9081 * Make us the idle thread. Technically, schedule() should not be
9082 * called from this thread, however somewhere below it might be,
9083 * but because we are the idle thread, we just pick up running again
9084 * when this runqueue becomes "idle".
9086 init_idle(current, smp_processor_id());
9088 * During early bootup we pretend to be a normal task:
9090 current->sched_class = &fair_sched_class;
9092 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9093 alloc_bootmem_cpumask_var(&nohz_cpu_mask);
9096 alloc_bootmem_cpumask_var(&nohz.cpu_mask);
9098 alloc_bootmem_cpumask_var(&cpu_isolated_map);
9101 scheduler_running = 1;
9104 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9105 void __might_sleep(char *file, int line)
9108 static unsigned long prev_jiffy; /* ratelimiting */
9110 if ((!in_atomic() && !irqs_disabled()) ||
9111 system_state != SYSTEM_RUNNING || oops_in_progress)
9113 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9115 prev_jiffy = jiffies;
9118 "BUG: sleeping function called from invalid context at %s:%d\n",
9121 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9122 in_atomic(), irqs_disabled(),
9123 current->pid, current->comm);
9125 debug_show_held_locks(current);
9126 if (irqs_disabled())
9127 print_irqtrace_events(current);
9131 EXPORT_SYMBOL(__might_sleep);
9134 #ifdef CONFIG_MAGIC_SYSRQ
9135 static void normalize_task(struct rq *rq, struct task_struct *p)
9139 update_rq_clock(rq);
9140 on_rq = p->se.on_rq;
9142 deactivate_task(rq, p, 0);
9143 __setscheduler(rq, p, SCHED_NORMAL, 0);
9145 activate_task(rq, p, 0);
9146 resched_task(rq->curr);
9150 void normalize_rt_tasks(void)
9152 struct task_struct *g, *p;
9153 unsigned long flags;
9156 read_lock_irqsave(&tasklist_lock, flags);
9157 do_each_thread(g, p) {
9159 * Only normalize user tasks:
9164 p->se.exec_start = 0;
9165 #ifdef CONFIG_SCHEDSTATS
9166 p->se.wait_start = 0;
9167 p->se.sleep_start = 0;
9168 p->se.block_start = 0;
9173 * Renice negative nice level userspace
9176 if (TASK_NICE(p) < 0 && p->mm)
9177 set_user_nice(p, 0);
9181 spin_lock(&p->pi_lock);
9182 rq = __task_rq_lock(p);
9184 normalize_task(rq, p);
9186 __task_rq_unlock(rq);
9187 spin_unlock(&p->pi_lock);
9188 } while_each_thread(g, p);
9190 read_unlock_irqrestore(&tasklist_lock, flags);
9193 #endif /* CONFIG_MAGIC_SYSRQ */
9197 * These functions are only useful for the IA64 MCA handling.
9199 * They can only be called when the whole system has been
9200 * stopped - every CPU needs to be quiescent, and no scheduling
9201 * activity can take place. Using them for anything else would
9202 * be a serious bug, and as a result, they aren't even visible
9203 * under any other configuration.
9207 * curr_task - return the current task for a given cpu.
9208 * @cpu: the processor in question.
9210 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9212 struct task_struct *curr_task(int cpu)
9214 return cpu_curr(cpu);
9218 * set_curr_task - set the current task for a given cpu.
9219 * @cpu: the processor in question.
9220 * @p: the task pointer to set.
9222 * Description: This function must only be used when non-maskable interrupts
9223 * are serviced on a separate stack. It allows the architecture to switch the
9224 * notion of the current task on a cpu in a non-blocking manner. This function
9225 * must be called with all CPU's synchronized, and interrupts disabled, the
9226 * and caller must save the original value of the current task (see
9227 * curr_task() above) and restore that value before reenabling interrupts and
9228 * re-starting the system.
9230 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9232 void set_curr_task(int cpu, struct task_struct *p)
9239 #ifdef CONFIG_FAIR_GROUP_SCHED
9240 static void free_fair_sched_group(struct task_group *tg)
9244 for_each_possible_cpu(i) {
9246 kfree(tg->cfs_rq[i]);
9256 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9258 struct cfs_rq *cfs_rq;
9259 struct sched_entity *se;
9263 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9266 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9270 tg->shares = NICE_0_LOAD;
9272 for_each_possible_cpu(i) {
9275 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9276 GFP_KERNEL, cpu_to_node(i));
9280 se = kzalloc_node(sizeof(struct sched_entity),
9281 GFP_KERNEL, cpu_to_node(i));
9285 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
9294 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9296 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
9297 &cpu_rq(cpu)->leaf_cfs_rq_list);
9300 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9302 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
9304 #else /* !CONFG_FAIR_GROUP_SCHED */
9305 static inline void free_fair_sched_group(struct task_group *tg)
9310 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9315 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9319 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9322 #endif /* CONFIG_FAIR_GROUP_SCHED */
9324 #ifdef CONFIG_RT_GROUP_SCHED
9325 static void free_rt_sched_group(struct task_group *tg)
9329 destroy_rt_bandwidth(&tg->rt_bandwidth);
9331 for_each_possible_cpu(i) {
9333 kfree(tg->rt_rq[i]);
9335 kfree(tg->rt_se[i]);
9343 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9345 struct rt_rq *rt_rq;
9346 struct sched_rt_entity *rt_se;
9350 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
9353 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
9357 init_rt_bandwidth(&tg->rt_bandwidth,
9358 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
9360 for_each_possible_cpu(i) {
9363 rt_rq = kzalloc_node(sizeof(struct rt_rq),
9364 GFP_KERNEL, cpu_to_node(i));
9368 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
9369 GFP_KERNEL, cpu_to_node(i));
9373 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
9382 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9384 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
9385 &cpu_rq(cpu)->leaf_rt_rq_list);
9388 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9390 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
9392 #else /* !CONFIG_RT_GROUP_SCHED */
9393 static inline void free_rt_sched_group(struct task_group *tg)
9398 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9403 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9407 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9410 #endif /* CONFIG_RT_GROUP_SCHED */
9412 #ifdef CONFIG_GROUP_SCHED
9413 static void free_sched_group(struct task_group *tg)
9415 free_fair_sched_group(tg);
9416 free_rt_sched_group(tg);
9420 /* allocate runqueue etc for a new task group */
9421 struct task_group *sched_create_group(struct task_group *parent)
9423 struct task_group *tg;
9424 unsigned long flags;
9427 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
9429 return ERR_PTR(-ENOMEM);
9431 if (!alloc_fair_sched_group(tg, parent))
9434 if (!alloc_rt_sched_group(tg, parent))
9437 spin_lock_irqsave(&task_group_lock, flags);
9438 for_each_possible_cpu(i) {
9439 register_fair_sched_group(tg, i);
9440 register_rt_sched_group(tg, i);
9442 list_add_rcu(&tg->list, &task_groups);
9444 WARN_ON(!parent); /* root should already exist */
9446 tg->parent = parent;
9447 INIT_LIST_HEAD(&tg->children);
9448 list_add_rcu(&tg->siblings, &parent->children);
9449 spin_unlock_irqrestore(&task_group_lock, flags);
9454 free_sched_group(tg);
9455 return ERR_PTR(-ENOMEM);
9458 /* rcu callback to free various structures associated with a task group */
9459 static void free_sched_group_rcu(struct rcu_head *rhp)
9461 /* now it should be safe to free those cfs_rqs */
9462 free_sched_group(container_of(rhp, struct task_group, rcu));
9465 /* Destroy runqueue etc associated with a task group */
9466 void sched_destroy_group(struct task_group *tg)
9468 unsigned long flags;
9471 spin_lock_irqsave(&task_group_lock, flags);
9472 for_each_possible_cpu(i) {
9473 unregister_fair_sched_group(tg, i);
9474 unregister_rt_sched_group(tg, i);
9476 list_del_rcu(&tg->list);
9477 list_del_rcu(&tg->siblings);
9478 spin_unlock_irqrestore(&task_group_lock, flags);
9480 /* wait for possible concurrent references to cfs_rqs complete */
9481 call_rcu(&tg->rcu, free_sched_group_rcu);
9484 /* change task's runqueue when it moves between groups.
9485 * The caller of this function should have put the task in its new group
9486 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
9487 * reflect its new group.
9489 void sched_move_task(struct task_struct *tsk)
9492 unsigned long flags;
9495 rq = task_rq_lock(tsk, &flags);
9497 update_rq_clock(rq);
9499 running = task_current(rq, tsk);
9500 on_rq = tsk->se.on_rq;
9503 dequeue_task(rq, tsk, 0);
9504 if (unlikely(running))
9505 tsk->sched_class->put_prev_task(rq, tsk);
9507 set_task_rq(tsk, task_cpu(tsk));
9509 #ifdef CONFIG_FAIR_GROUP_SCHED
9510 if (tsk->sched_class->moved_group)
9511 tsk->sched_class->moved_group(tsk);
9514 if (unlikely(running))
9515 tsk->sched_class->set_curr_task(rq);
9517 enqueue_task(rq, tsk, 0);
9519 task_rq_unlock(rq, &flags);
9521 #endif /* CONFIG_GROUP_SCHED */
9523 #ifdef CONFIG_FAIR_GROUP_SCHED
9524 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
9526 struct cfs_rq *cfs_rq = se->cfs_rq;
9531 dequeue_entity(cfs_rq, se, 0);
9533 se->load.weight = shares;
9534 se->load.inv_weight = 0;
9537 enqueue_entity(cfs_rq, se, 0);
9540 static void set_se_shares(struct sched_entity *se, unsigned long shares)
9542 struct cfs_rq *cfs_rq = se->cfs_rq;
9543 struct rq *rq = cfs_rq->rq;
9544 unsigned long flags;
9546 spin_lock_irqsave(&rq->lock, flags);
9547 __set_se_shares(se, shares);
9548 spin_unlock_irqrestore(&rq->lock, flags);
9551 static DEFINE_MUTEX(shares_mutex);
9553 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9556 unsigned long flags;
9559 * We can't change the weight of the root cgroup.
9564 if (shares < MIN_SHARES)
9565 shares = MIN_SHARES;
9566 else if (shares > MAX_SHARES)
9567 shares = MAX_SHARES;
9569 mutex_lock(&shares_mutex);
9570 if (tg->shares == shares)
9573 spin_lock_irqsave(&task_group_lock, flags);
9574 for_each_possible_cpu(i)
9575 unregister_fair_sched_group(tg, i);
9576 list_del_rcu(&tg->siblings);
9577 spin_unlock_irqrestore(&task_group_lock, flags);
9579 /* wait for any ongoing reference to this group to finish */
9580 synchronize_sched();
9583 * Now we are free to modify the group's share on each cpu
9584 * w/o tripping rebalance_share or load_balance_fair.
9586 tg->shares = shares;
9587 for_each_possible_cpu(i) {
9591 cfs_rq_set_shares(tg->cfs_rq[i], 0);
9592 set_se_shares(tg->se[i], shares);
9596 * Enable load balance activity on this group, by inserting it back on
9597 * each cpu's rq->leaf_cfs_rq_list.
9599 spin_lock_irqsave(&task_group_lock, flags);
9600 for_each_possible_cpu(i)
9601 register_fair_sched_group(tg, i);
9602 list_add_rcu(&tg->siblings, &tg->parent->children);
9603 spin_unlock_irqrestore(&task_group_lock, flags);
9605 mutex_unlock(&shares_mutex);
9609 unsigned long sched_group_shares(struct task_group *tg)
9615 #ifdef CONFIG_RT_GROUP_SCHED
9617 * Ensure that the real time constraints are schedulable.
9619 static DEFINE_MUTEX(rt_constraints_mutex);
9621 static unsigned long to_ratio(u64 period, u64 runtime)
9623 if (runtime == RUNTIME_INF)
9626 return div64_u64(runtime << 20, period);
9629 /* Must be called with tasklist_lock held */
9630 static inline int tg_has_rt_tasks(struct task_group *tg)
9632 struct task_struct *g, *p;
9634 do_each_thread(g, p) {
9635 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
9637 } while_each_thread(g, p);
9642 struct rt_schedulable_data {
9643 struct task_group *tg;
9648 static int tg_schedulable(struct task_group *tg, void *data)
9650 struct rt_schedulable_data *d = data;
9651 struct task_group *child;
9652 unsigned long total, sum = 0;
9653 u64 period, runtime;
9655 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9656 runtime = tg->rt_bandwidth.rt_runtime;
9659 period = d->rt_period;
9660 runtime = d->rt_runtime;
9663 #ifdef CONFIG_USER_SCHED
9664 if (tg == &root_task_group) {
9665 period = global_rt_period();
9666 runtime = global_rt_runtime();
9671 * Cannot have more runtime than the period.
9673 if (runtime > period && runtime != RUNTIME_INF)
9677 * Ensure we don't starve existing RT tasks.
9679 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
9682 total = to_ratio(period, runtime);
9685 * Nobody can have more than the global setting allows.
9687 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
9691 * The sum of our children's runtime should not exceed our own.
9693 list_for_each_entry_rcu(child, &tg->children, siblings) {
9694 period = ktime_to_ns(child->rt_bandwidth.rt_period);
9695 runtime = child->rt_bandwidth.rt_runtime;
9697 if (child == d->tg) {
9698 period = d->rt_period;
9699 runtime = d->rt_runtime;
9702 sum += to_ratio(period, runtime);
9711 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
9713 struct rt_schedulable_data data = {
9715 .rt_period = period,
9716 .rt_runtime = runtime,
9719 return walk_tg_tree(tg_schedulable, tg_nop, &data);
9722 static int tg_set_bandwidth(struct task_group *tg,
9723 u64 rt_period, u64 rt_runtime)
9727 mutex_lock(&rt_constraints_mutex);
9728 read_lock(&tasklist_lock);
9729 err = __rt_schedulable(tg, rt_period, rt_runtime);
9733 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
9734 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
9735 tg->rt_bandwidth.rt_runtime = rt_runtime;
9737 for_each_possible_cpu(i) {
9738 struct rt_rq *rt_rq = tg->rt_rq[i];
9740 spin_lock(&rt_rq->rt_runtime_lock);
9741 rt_rq->rt_runtime = rt_runtime;
9742 spin_unlock(&rt_rq->rt_runtime_lock);
9744 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
9746 read_unlock(&tasklist_lock);
9747 mutex_unlock(&rt_constraints_mutex);
9752 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
9754 u64 rt_runtime, rt_period;
9756 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9757 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
9758 if (rt_runtime_us < 0)
9759 rt_runtime = RUNTIME_INF;
9761 return tg_set_bandwidth(tg, rt_period, rt_runtime);
9764 long sched_group_rt_runtime(struct task_group *tg)
9768 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
9771 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
9772 do_div(rt_runtime_us, NSEC_PER_USEC);
9773 return rt_runtime_us;
9776 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
9778 u64 rt_runtime, rt_period;
9780 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
9781 rt_runtime = tg->rt_bandwidth.rt_runtime;
9786 return tg_set_bandwidth(tg, rt_period, rt_runtime);
9789 long sched_group_rt_period(struct task_group *tg)
9793 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
9794 do_div(rt_period_us, NSEC_PER_USEC);
9795 return rt_period_us;
9798 static int sched_rt_global_constraints(void)
9800 u64 runtime, period;
9803 if (sysctl_sched_rt_period <= 0)
9806 runtime = global_rt_runtime();
9807 period = global_rt_period();
9810 * Sanity check on the sysctl variables.
9812 if (runtime > period && runtime != RUNTIME_INF)
9815 mutex_lock(&rt_constraints_mutex);
9816 read_lock(&tasklist_lock);
9817 ret = __rt_schedulable(NULL, 0, 0);
9818 read_unlock(&tasklist_lock);
9819 mutex_unlock(&rt_constraints_mutex);
9824 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
9826 /* Don't accept realtime tasks when there is no way for them to run */
9827 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
9833 #else /* !CONFIG_RT_GROUP_SCHED */
9834 static int sched_rt_global_constraints(void)
9836 unsigned long flags;
9839 if (sysctl_sched_rt_period <= 0)
9842 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
9843 for_each_possible_cpu(i) {
9844 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
9846 spin_lock(&rt_rq->rt_runtime_lock);
9847 rt_rq->rt_runtime = global_rt_runtime();
9848 spin_unlock(&rt_rq->rt_runtime_lock);
9850 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
9854 #endif /* CONFIG_RT_GROUP_SCHED */
9856 int sched_rt_handler(struct ctl_table *table, int write,
9857 struct file *filp, void __user *buffer, size_t *lenp,
9861 int old_period, old_runtime;
9862 static DEFINE_MUTEX(mutex);
9865 old_period = sysctl_sched_rt_period;
9866 old_runtime = sysctl_sched_rt_runtime;
9868 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
9870 if (!ret && write) {
9871 ret = sched_rt_global_constraints();
9873 sysctl_sched_rt_period = old_period;
9874 sysctl_sched_rt_runtime = old_runtime;
9876 def_rt_bandwidth.rt_runtime = global_rt_runtime();
9877 def_rt_bandwidth.rt_period =
9878 ns_to_ktime(global_rt_period());
9881 mutex_unlock(&mutex);
9886 #ifdef CONFIG_CGROUP_SCHED
9888 /* return corresponding task_group object of a cgroup */
9889 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9891 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9892 struct task_group, css);
9895 static struct cgroup_subsys_state *
9896 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9898 struct task_group *tg, *parent;
9900 if (!cgrp->parent) {
9901 /* This is early initialization for the top cgroup */
9902 return &init_task_group.css;
9905 parent = cgroup_tg(cgrp->parent);
9906 tg = sched_create_group(parent);
9908 return ERR_PTR(-ENOMEM);
9914 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9916 struct task_group *tg = cgroup_tg(cgrp);
9918 sched_destroy_group(tg);
9922 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9923 struct task_struct *tsk)
9925 #ifdef CONFIG_RT_GROUP_SCHED
9926 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
9929 /* We don't support RT-tasks being in separate groups */
9930 if (tsk->sched_class != &fair_sched_class)
9938 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9939 struct cgroup *old_cont, struct task_struct *tsk)
9941 sched_move_task(tsk);
9944 #ifdef CONFIG_FAIR_GROUP_SCHED
9945 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9948 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
9951 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9953 struct task_group *tg = cgroup_tg(cgrp);
9955 return (u64) tg->shares;
9957 #endif /* CONFIG_FAIR_GROUP_SCHED */
9959 #ifdef CONFIG_RT_GROUP_SCHED
9960 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9963 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9966 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9968 return sched_group_rt_runtime(cgroup_tg(cgrp));
9971 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9974 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9977 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9979 return sched_group_rt_period(cgroup_tg(cgrp));
9981 #endif /* CONFIG_RT_GROUP_SCHED */
9983 static struct cftype cpu_files[] = {
9984 #ifdef CONFIG_FAIR_GROUP_SCHED
9987 .read_u64 = cpu_shares_read_u64,
9988 .write_u64 = cpu_shares_write_u64,
9991 #ifdef CONFIG_RT_GROUP_SCHED
9993 .name = "rt_runtime_us",
9994 .read_s64 = cpu_rt_runtime_read,
9995 .write_s64 = cpu_rt_runtime_write,
9998 .name = "rt_period_us",
9999 .read_u64 = cpu_rt_period_read_uint,
10000 .write_u64 = cpu_rt_period_write_uint,
10005 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
10007 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
10010 struct cgroup_subsys cpu_cgroup_subsys = {
10012 .create = cpu_cgroup_create,
10013 .destroy = cpu_cgroup_destroy,
10014 .can_attach = cpu_cgroup_can_attach,
10015 .attach = cpu_cgroup_attach,
10016 .populate = cpu_cgroup_populate,
10017 .subsys_id = cpu_cgroup_subsys_id,
10021 #endif /* CONFIG_CGROUP_SCHED */
10023 #ifdef CONFIG_CGROUP_CPUACCT
10026 * CPU accounting code for task groups.
10028 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10029 * (balbir@in.ibm.com).
10032 /* track cpu usage of a group of tasks and its child groups */
10034 struct cgroup_subsys_state css;
10035 /* cpuusage holds pointer to a u64-type object on every cpu */
10037 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
10038 struct cpuacct *parent;
10041 struct cgroup_subsys cpuacct_subsys;
10043 /* return cpu accounting group corresponding to this container */
10044 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
10046 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
10047 struct cpuacct, css);
10050 /* return cpu accounting group to which this task belongs */
10051 static inline struct cpuacct *task_ca(struct task_struct *tsk)
10053 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
10054 struct cpuacct, css);
10057 /* create a new cpu accounting group */
10058 static struct cgroup_subsys_state *cpuacct_create(
10059 struct cgroup_subsys *ss, struct cgroup *cgrp)
10061 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
10067 ca->cpuusage = alloc_percpu(u64);
10071 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10072 if (percpu_counter_init(&ca->cpustat[i], 0))
10073 goto out_free_counters;
10076 ca->parent = cgroup_ca(cgrp->parent);
10082 percpu_counter_destroy(&ca->cpustat[i]);
10083 free_percpu(ca->cpuusage);
10087 return ERR_PTR(-ENOMEM);
10090 /* destroy an existing cpu accounting group */
10092 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10094 struct cpuacct *ca = cgroup_ca(cgrp);
10097 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10098 percpu_counter_destroy(&ca->cpustat[i]);
10099 free_percpu(ca->cpuusage);
10103 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
10105 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10108 #ifndef CONFIG_64BIT
10110 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10112 spin_lock_irq(&cpu_rq(cpu)->lock);
10114 spin_unlock_irq(&cpu_rq(cpu)->lock);
10122 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
10124 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10126 #ifndef CONFIG_64BIT
10128 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10130 spin_lock_irq(&cpu_rq(cpu)->lock);
10132 spin_unlock_irq(&cpu_rq(cpu)->lock);
10138 /* return total cpu usage (in nanoseconds) of a group */
10139 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
10141 struct cpuacct *ca = cgroup_ca(cgrp);
10142 u64 totalcpuusage = 0;
10145 for_each_present_cpu(i)
10146 totalcpuusage += cpuacct_cpuusage_read(ca, i);
10148 return totalcpuusage;
10151 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
10154 struct cpuacct *ca = cgroup_ca(cgrp);
10163 for_each_present_cpu(i)
10164 cpuacct_cpuusage_write(ca, i, 0);
10170 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
10171 struct seq_file *m)
10173 struct cpuacct *ca = cgroup_ca(cgroup);
10177 for_each_present_cpu(i) {
10178 percpu = cpuacct_cpuusage_read(ca, i);
10179 seq_printf(m, "%llu ", (unsigned long long) percpu);
10181 seq_printf(m, "\n");
10185 static const char *cpuacct_stat_desc[] = {
10186 [CPUACCT_STAT_USER] = "user",
10187 [CPUACCT_STAT_SYSTEM] = "system",
10190 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
10191 struct cgroup_map_cb *cb)
10193 struct cpuacct *ca = cgroup_ca(cgrp);
10196 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
10197 s64 val = percpu_counter_read(&ca->cpustat[i]);
10198 val = cputime64_to_clock_t(val);
10199 cb->fill(cb, cpuacct_stat_desc[i], val);
10204 static struct cftype files[] = {
10207 .read_u64 = cpuusage_read,
10208 .write_u64 = cpuusage_write,
10211 .name = "usage_percpu",
10212 .read_seq_string = cpuacct_percpu_seq_read,
10216 .read_map = cpuacct_stats_show,
10220 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
10222 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
10226 * charge this task's execution time to its accounting group.
10228 * called with rq->lock held.
10230 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
10232 struct cpuacct *ca;
10235 if (unlikely(!cpuacct_subsys.active))
10238 cpu = task_cpu(tsk);
10244 for (; ca; ca = ca->parent) {
10245 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10246 *cpuusage += cputime;
10253 * Charge the system/user time to the task's accounting group.
10255 static void cpuacct_update_stats(struct task_struct *tsk,
10256 enum cpuacct_stat_index idx, cputime_t val)
10258 struct cpuacct *ca;
10260 if (unlikely(!cpuacct_subsys.active))
10267 percpu_counter_add(&ca->cpustat[idx], val);
10273 struct cgroup_subsys cpuacct_subsys = {
10275 .create = cpuacct_create,
10276 .destroy = cpuacct_destroy,
10277 .populate = cpuacct_populate,
10278 .subsys_id = cpuacct_subsys_id,
10280 #endif /* CONFIG_CGROUP_CPUACCT */