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/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/reciprocal_div.h>
66 #include <linux/unistd.h>
67 #include <linux/pagemap.h>
68 #include <linux/hrtimer.h>
69 #include <linux/tick.h>
70 #include <linux/bootmem.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
75 #include <asm/irq_regs.h>
77 #include "sched_cpupri.h"
80 * Convert user-nice values [ -20 ... 0 ... 19 ]
81 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
84 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
85 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
86 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
89 * 'User priority' is the nice value converted to something we
90 * can work with better when scaling various scheduler parameters,
91 * it's a [ 0 ... 39 ] range.
93 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
94 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
95 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
98 * Helpers for converting nanosecond timing to jiffy resolution
100 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
102 #define NICE_0_LOAD SCHED_LOAD_SCALE
103 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
106 * These are the 'tuning knobs' of the scheduler:
108 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
109 * Timeslices get refilled after they expire.
111 #define DEF_TIMESLICE (100 * HZ / 1000)
114 * single value that denotes runtime == period, ie unlimited time.
116 #define RUNTIME_INF ((u64)~0ULL)
120 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
121 * Since cpu_power is a 'constant', we can use a reciprocal divide.
123 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
125 return reciprocal_divide(load, sg->reciprocal_cpu_power);
129 * Each time a sched group cpu_power is changed,
130 * we must compute its reciprocal value
132 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
134 sg->__cpu_power += val;
135 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
139 static inline int rt_policy(int policy)
141 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
146 static inline int task_has_rt_policy(struct task_struct *p)
148 return rt_policy(p->policy);
152 * This is the priority-queue data structure of the RT scheduling class:
154 struct rt_prio_array {
155 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
156 struct list_head queue[MAX_RT_PRIO];
159 struct rt_bandwidth {
160 /* nests inside the rq lock: */
161 spinlock_t rt_runtime_lock;
164 struct hrtimer rt_period_timer;
167 static struct rt_bandwidth def_rt_bandwidth;
169 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
171 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
173 struct rt_bandwidth *rt_b =
174 container_of(timer, struct rt_bandwidth, rt_period_timer);
180 now = hrtimer_cb_get_time(timer);
181 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
186 idle = do_sched_rt_period_timer(rt_b, overrun);
189 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
193 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
195 rt_b->rt_period = ns_to_ktime(period);
196 rt_b->rt_runtime = runtime;
198 spin_lock_init(&rt_b->rt_runtime_lock);
200 hrtimer_init(&rt_b->rt_period_timer,
201 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
202 rt_b->rt_period_timer.function = sched_rt_period_timer;
203 rt_b->rt_period_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
206 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
210 if (rt_b->rt_runtime == RUNTIME_INF)
213 if (hrtimer_active(&rt_b->rt_period_timer))
216 spin_lock(&rt_b->rt_runtime_lock);
218 if (hrtimer_active(&rt_b->rt_period_timer))
221 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
222 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
223 hrtimer_start(&rt_b->rt_period_timer,
224 rt_b->rt_period_timer.expires,
227 spin_unlock(&rt_b->rt_runtime_lock);
230 #ifdef CONFIG_RT_GROUP_SCHED
231 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
233 hrtimer_cancel(&rt_b->rt_period_timer);
238 * sched_domains_mutex serializes calls to arch_init_sched_domains,
239 * detach_destroy_domains and partition_sched_domains.
241 static DEFINE_MUTEX(sched_domains_mutex);
243 #ifdef CONFIG_GROUP_SCHED
245 #include <linux/cgroup.h>
249 static LIST_HEAD(task_groups);
251 /* task group related information */
253 #ifdef CONFIG_CGROUP_SCHED
254 struct cgroup_subsys_state css;
257 #ifdef CONFIG_FAIR_GROUP_SCHED
258 /* schedulable entities of this group on each cpu */
259 struct sched_entity **se;
260 /* runqueue "owned" by this group on each cpu */
261 struct cfs_rq **cfs_rq;
262 unsigned long shares;
265 #ifdef CONFIG_RT_GROUP_SCHED
266 struct sched_rt_entity **rt_se;
267 struct rt_rq **rt_rq;
269 struct rt_bandwidth rt_bandwidth;
273 struct list_head list;
275 struct task_group *parent;
276 struct list_head siblings;
277 struct list_head children;
280 #ifdef CONFIG_USER_SCHED
284 * Every UID task group (including init_task_group aka UID-0) will
285 * be a child to this group.
287 struct task_group root_task_group;
289 #ifdef CONFIG_FAIR_GROUP_SCHED
290 /* Default task group's sched entity on each cpu */
291 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
292 /* Default task group's cfs_rq on each cpu */
293 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
294 #endif /* CONFIG_FAIR_GROUP_SCHED */
296 #ifdef CONFIG_RT_GROUP_SCHED
297 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
298 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
299 #endif /* CONFIG_RT_GROUP_SCHED */
300 #else /* !CONFIG_FAIR_GROUP_SCHED */
301 #define root_task_group init_task_group
302 #endif /* CONFIG_FAIR_GROUP_SCHED */
304 /* task_group_lock serializes add/remove of task groups and also changes to
305 * a task group's cpu shares.
307 static DEFINE_SPINLOCK(task_group_lock);
309 #ifdef CONFIG_FAIR_GROUP_SCHED
310 #ifdef CONFIG_USER_SCHED
311 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
312 #else /* !CONFIG_USER_SCHED */
313 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
314 #endif /* CONFIG_USER_SCHED */
317 * A weight of 0 or 1 can cause arithmetics problems.
318 * A weight of a cfs_rq is the sum of weights of which entities
319 * are queued on this cfs_rq, so a weight of a entity should not be
320 * too large, so as the shares value of a task group.
321 * (The default weight is 1024 - so there's no practical
322 * limitation from this.)
325 #define MAX_SHARES (1UL << 18)
327 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
330 /* Default task group.
331 * Every task in system belong to this group at bootup.
333 struct task_group init_task_group;
335 /* return group to which a task belongs */
336 static inline struct task_group *task_group(struct task_struct *p)
338 struct task_group *tg;
340 #ifdef CONFIG_USER_SCHED
342 #elif defined(CONFIG_CGROUP_SCHED)
343 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
344 struct task_group, css);
346 tg = &init_task_group;
351 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
352 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
354 #ifdef CONFIG_FAIR_GROUP_SCHED
355 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
356 p->se.parent = task_group(p)->se[cpu];
359 #ifdef CONFIG_RT_GROUP_SCHED
360 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
361 p->rt.parent = task_group(p)->rt_se[cpu];
367 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
368 static inline struct task_group *task_group(struct task_struct *p)
373 #endif /* CONFIG_GROUP_SCHED */
375 /* CFS-related fields in a runqueue */
377 struct load_weight load;
378 unsigned long nr_running;
384 struct rb_root tasks_timeline;
385 struct rb_node *rb_leftmost;
387 struct list_head tasks;
388 struct list_head *balance_iterator;
391 * 'curr' points to currently running entity on this cfs_rq.
392 * It is set to NULL otherwise (i.e when none are currently running).
394 struct sched_entity *curr, *next;
396 unsigned long nr_spread_over;
398 #ifdef CONFIG_FAIR_GROUP_SCHED
399 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
402 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
403 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
404 * (like users, containers etc.)
406 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
407 * list is used during load balance.
409 struct list_head leaf_cfs_rq_list;
410 struct task_group *tg; /* group that "owns" this runqueue */
414 * the part of load.weight contributed by tasks
416 unsigned long task_weight;
419 * h_load = weight * f(tg)
421 * Where f(tg) is the recursive weight fraction assigned to
424 unsigned long h_load;
427 * this cpu's part of tg->shares
429 unsigned long shares;
432 * load.weight at the time we set shares
434 unsigned long rq_weight;
439 /* Real-Time classes' related field in a runqueue: */
441 struct rt_prio_array active;
442 unsigned long rt_nr_running;
443 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
444 int highest_prio; /* highest queued rt task prio */
447 unsigned long rt_nr_migratory;
453 /* Nests inside the rq lock: */
454 spinlock_t rt_runtime_lock;
456 #ifdef CONFIG_RT_GROUP_SCHED
457 unsigned long rt_nr_boosted;
460 struct list_head leaf_rt_rq_list;
461 struct task_group *tg;
462 struct sched_rt_entity *rt_se;
469 * We add the notion of a root-domain which will be used to define per-domain
470 * variables. Each exclusive cpuset essentially defines an island domain by
471 * fully partitioning the member cpus from any other cpuset. Whenever a new
472 * exclusive cpuset is created, we also create and attach a new root-domain
482 * The "RT overload" flag: it gets set if a CPU has more than
483 * one runnable RT task.
488 struct cpupri cpupri;
493 * By default the system creates a single root-domain with all cpus as
494 * members (mimicking the global state we have today).
496 static struct root_domain def_root_domain;
501 * This is the main, per-CPU runqueue data structure.
503 * Locking rule: those places that want to lock multiple runqueues
504 * (such as the load balancing or the thread migration code), lock
505 * acquire operations must be ordered by ascending &runqueue.
512 * nr_running and cpu_load should be in the same cacheline because
513 * remote CPUs use both these fields when doing load calculation.
515 unsigned long nr_running;
516 #define CPU_LOAD_IDX_MAX 5
517 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
518 unsigned char idle_at_tick;
520 unsigned long last_tick_seen;
521 unsigned char in_nohz_recently;
523 /* capture load from *all* tasks on this cpu: */
524 struct load_weight load;
525 unsigned long nr_load_updates;
531 #ifdef CONFIG_FAIR_GROUP_SCHED
532 /* list of leaf cfs_rq on this cpu: */
533 struct list_head leaf_cfs_rq_list;
535 #ifdef CONFIG_RT_GROUP_SCHED
536 struct list_head leaf_rt_rq_list;
540 * This is part of a global counter where only the total sum
541 * over all CPUs matters. A task can increase this counter on
542 * one CPU and if it got migrated afterwards it may decrease
543 * it on another CPU. Always updated under the runqueue lock:
545 unsigned long nr_uninterruptible;
547 struct task_struct *curr, *idle;
548 unsigned long next_balance;
549 struct mm_struct *prev_mm;
556 struct root_domain *rd;
557 struct sched_domain *sd;
559 /* For active balancing */
562 /* cpu of this runqueue: */
566 unsigned long avg_load_per_task;
568 struct task_struct *migration_thread;
569 struct list_head migration_queue;
572 #ifdef CONFIG_SCHED_HRTICK
573 unsigned long hrtick_flags;
574 ktime_t hrtick_expire;
575 struct hrtimer hrtick_timer;
578 #ifdef CONFIG_SCHEDSTATS
580 struct sched_info rq_sched_info;
582 /* sys_sched_yield() stats */
583 unsigned int yld_exp_empty;
584 unsigned int yld_act_empty;
585 unsigned int yld_both_empty;
586 unsigned int yld_count;
588 /* schedule() stats */
589 unsigned int sched_switch;
590 unsigned int sched_count;
591 unsigned int sched_goidle;
593 /* try_to_wake_up() stats */
594 unsigned int ttwu_count;
595 unsigned int ttwu_local;
598 unsigned int bkl_count;
600 struct lock_class_key rq_lock_key;
603 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
605 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
607 rq->curr->sched_class->check_preempt_curr(rq, p);
610 static inline int cpu_of(struct rq *rq)
620 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
621 * See detach_destroy_domains: synchronize_sched for details.
623 * The domain tree of any CPU may only be accessed from within
624 * preempt-disabled sections.
626 #define for_each_domain(cpu, __sd) \
627 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
629 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
630 #define this_rq() (&__get_cpu_var(runqueues))
631 #define task_rq(p) cpu_rq(task_cpu(p))
632 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
634 static inline void update_rq_clock(struct rq *rq)
636 rq->clock = sched_clock_cpu(cpu_of(rq));
640 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
642 #ifdef CONFIG_SCHED_DEBUG
643 # define const_debug __read_mostly
645 # define const_debug static const
649 * Debugging: various feature bits
652 #define SCHED_FEAT(name, enabled) \
653 __SCHED_FEAT_##name ,
656 #include "sched_features.h"
661 #define SCHED_FEAT(name, enabled) \
662 (1UL << __SCHED_FEAT_##name) * enabled |
664 const_debug unsigned int sysctl_sched_features =
665 #include "sched_features.h"
670 #ifdef CONFIG_SCHED_DEBUG
671 #define SCHED_FEAT(name, enabled) \
674 static __read_mostly char *sched_feat_names[] = {
675 #include "sched_features.h"
681 static int sched_feat_open(struct inode *inode, struct file *filp)
683 filp->private_data = inode->i_private;
688 sched_feat_read(struct file *filp, char __user *ubuf,
689 size_t cnt, loff_t *ppos)
696 for (i = 0; sched_feat_names[i]; i++) {
697 len += strlen(sched_feat_names[i]);
701 buf = kmalloc(len + 2, GFP_KERNEL);
705 for (i = 0; sched_feat_names[i]; i++) {
706 if (sysctl_sched_features & (1UL << i))
707 r += sprintf(buf + r, "%s ", sched_feat_names[i]);
709 r += sprintf(buf + r, "NO_%s ", sched_feat_names[i]);
712 r += sprintf(buf + r, "\n");
713 WARN_ON(r >= len + 2);
715 r = simple_read_from_buffer(ubuf, cnt, ppos, buf, r);
723 sched_feat_write(struct file *filp, const char __user *ubuf,
724 size_t cnt, loff_t *ppos)
734 if (copy_from_user(&buf, ubuf, cnt))
739 if (strncmp(buf, "NO_", 3) == 0) {
744 for (i = 0; sched_feat_names[i]; i++) {
745 int len = strlen(sched_feat_names[i]);
747 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
749 sysctl_sched_features &= ~(1UL << i);
751 sysctl_sched_features |= (1UL << i);
756 if (!sched_feat_names[i])
764 static struct file_operations sched_feat_fops = {
765 .open = sched_feat_open,
766 .read = sched_feat_read,
767 .write = sched_feat_write,
770 static __init int sched_init_debug(void)
772 debugfs_create_file("sched_features", 0644, NULL, NULL,
777 late_initcall(sched_init_debug);
781 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
784 * Number of tasks to iterate in a single balance run.
785 * Limited because this is done with IRQs disabled.
787 const_debug unsigned int sysctl_sched_nr_migrate = 32;
790 * ratelimit for updating the group shares.
793 const_debug unsigned int sysctl_sched_shares_ratelimit = 500000;
796 * period over which we measure -rt task cpu usage in us.
799 unsigned int sysctl_sched_rt_period = 1000000;
801 static __read_mostly int scheduler_running;
804 * part of the period that we allow rt tasks to run in us.
807 int sysctl_sched_rt_runtime = 950000;
809 static inline u64 global_rt_period(void)
811 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
814 static inline u64 global_rt_runtime(void)
816 if (sysctl_sched_rt_period < 0)
819 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
822 #ifndef prepare_arch_switch
823 # define prepare_arch_switch(next) do { } while (0)
825 #ifndef finish_arch_switch
826 # define finish_arch_switch(prev) do { } while (0)
829 static inline int task_current(struct rq *rq, struct task_struct *p)
831 return rq->curr == p;
834 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
835 static inline int task_running(struct rq *rq, struct task_struct *p)
837 return task_current(rq, p);
840 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
844 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
846 #ifdef CONFIG_DEBUG_SPINLOCK
847 /* this is a valid case when another task releases the spinlock */
848 rq->lock.owner = current;
851 * If we are tracking spinlock dependencies then we have to
852 * fix up the runqueue lock - which gets 'carried over' from
855 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
857 spin_unlock_irq(&rq->lock);
860 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
861 static inline int task_running(struct rq *rq, struct task_struct *p)
866 return task_current(rq, p);
870 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
874 * We can optimise this out completely for !SMP, because the
875 * SMP rebalancing from interrupt is the only thing that cares
880 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
881 spin_unlock_irq(&rq->lock);
883 spin_unlock(&rq->lock);
887 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
891 * After ->oncpu is cleared, the task can be moved to a different CPU.
892 * We must ensure this doesn't happen until the switch is completely
898 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
902 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
905 * __task_rq_lock - lock the runqueue a given task resides on.
906 * Must be called interrupts disabled.
908 static inline struct rq *__task_rq_lock(struct task_struct *p)
912 struct rq *rq = task_rq(p);
913 spin_lock(&rq->lock);
914 if (likely(rq == task_rq(p)))
916 spin_unlock(&rq->lock);
921 * task_rq_lock - lock the runqueue a given task resides on and disable
922 * interrupts. Note the ordering: we can safely lookup the task_rq without
923 * explicitly disabling preemption.
925 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
931 local_irq_save(*flags);
933 spin_lock(&rq->lock);
934 if (likely(rq == task_rq(p)))
936 spin_unlock_irqrestore(&rq->lock, *flags);
940 static void __task_rq_unlock(struct rq *rq)
943 spin_unlock(&rq->lock);
946 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
949 spin_unlock_irqrestore(&rq->lock, *flags);
953 * this_rq_lock - lock this runqueue and disable interrupts.
955 static struct rq *this_rq_lock(void)
962 spin_lock(&rq->lock);
967 static void __resched_task(struct task_struct *p, int tif_bit);
969 static inline void resched_task(struct task_struct *p)
971 __resched_task(p, TIF_NEED_RESCHED);
974 #ifdef CONFIG_SCHED_HRTICK
976 * Use HR-timers to deliver accurate preemption points.
978 * Its all a bit involved since we cannot program an hrt while holding the
979 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
982 * When we get rescheduled we reprogram the hrtick_timer outside of the
985 static inline void resched_hrt(struct task_struct *p)
987 __resched_task(p, TIF_HRTICK_RESCHED);
990 static inline void resched_rq(struct rq *rq)
994 spin_lock_irqsave(&rq->lock, flags);
995 resched_task(rq->curr);
996 spin_unlock_irqrestore(&rq->lock, flags);
1000 HRTICK_SET, /* re-programm hrtick_timer */
1001 HRTICK_RESET, /* not a new slice */
1002 HRTICK_BLOCK, /* stop hrtick operations */
1007 * - enabled by features
1008 * - hrtimer is actually high res
1010 static inline int hrtick_enabled(struct rq *rq)
1012 if (!sched_feat(HRTICK))
1014 if (unlikely(test_bit(HRTICK_BLOCK, &rq->hrtick_flags)))
1016 return hrtimer_is_hres_active(&rq->hrtick_timer);
1020 * Called to set the hrtick timer state.
1022 * called with rq->lock held and irqs disabled
1024 static void hrtick_start(struct rq *rq, u64 delay, int reset)
1026 assert_spin_locked(&rq->lock);
1029 * preempt at: now + delay
1032 ktime_add_ns(rq->hrtick_timer.base->get_time(), delay);
1034 * indicate we need to program the timer
1036 __set_bit(HRTICK_SET, &rq->hrtick_flags);
1038 __set_bit(HRTICK_RESET, &rq->hrtick_flags);
1041 * New slices are called from the schedule path and don't need a
1042 * forced reschedule.
1045 resched_hrt(rq->curr);
1048 static void hrtick_clear(struct rq *rq)
1050 if (hrtimer_active(&rq->hrtick_timer))
1051 hrtimer_cancel(&rq->hrtick_timer);
1055 * Update the timer from the possible pending state.
1057 static void hrtick_set(struct rq *rq)
1061 unsigned long flags;
1063 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1065 spin_lock_irqsave(&rq->lock, flags);
1066 set = __test_and_clear_bit(HRTICK_SET, &rq->hrtick_flags);
1067 reset = __test_and_clear_bit(HRTICK_RESET, &rq->hrtick_flags);
1068 time = rq->hrtick_expire;
1069 clear_thread_flag(TIF_HRTICK_RESCHED);
1070 spin_unlock_irqrestore(&rq->lock, flags);
1073 hrtimer_start(&rq->hrtick_timer, time, HRTIMER_MODE_ABS);
1074 if (reset && !hrtimer_active(&rq->hrtick_timer))
1081 * High-resolution timer tick.
1082 * Runs from hardirq context with interrupts disabled.
1084 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1086 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1088 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1090 spin_lock(&rq->lock);
1091 update_rq_clock(rq);
1092 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1093 spin_unlock(&rq->lock);
1095 return HRTIMER_NORESTART;
1099 static void hotplug_hrtick_disable(int cpu)
1101 struct rq *rq = cpu_rq(cpu);
1102 unsigned long flags;
1104 spin_lock_irqsave(&rq->lock, flags);
1105 rq->hrtick_flags = 0;
1106 __set_bit(HRTICK_BLOCK, &rq->hrtick_flags);
1107 spin_unlock_irqrestore(&rq->lock, flags);
1112 static void hotplug_hrtick_enable(int cpu)
1114 struct rq *rq = cpu_rq(cpu);
1115 unsigned long flags;
1117 spin_lock_irqsave(&rq->lock, flags);
1118 __clear_bit(HRTICK_BLOCK, &rq->hrtick_flags);
1119 spin_unlock_irqrestore(&rq->lock, flags);
1123 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1125 int cpu = (int)(long)hcpu;
1128 case CPU_UP_CANCELED:
1129 case CPU_UP_CANCELED_FROZEN:
1130 case CPU_DOWN_PREPARE:
1131 case CPU_DOWN_PREPARE_FROZEN:
1133 case CPU_DEAD_FROZEN:
1134 hotplug_hrtick_disable(cpu);
1137 case CPU_UP_PREPARE:
1138 case CPU_UP_PREPARE_FROZEN:
1139 case CPU_DOWN_FAILED:
1140 case CPU_DOWN_FAILED_FROZEN:
1142 case CPU_ONLINE_FROZEN:
1143 hotplug_hrtick_enable(cpu);
1150 static void init_hrtick(void)
1152 hotcpu_notifier(hotplug_hrtick, 0);
1154 #endif /* CONFIG_SMP */
1156 static void init_rq_hrtick(struct rq *rq)
1158 rq->hrtick_flags = 0;
1159 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1160 rq->hrtick_timer.function = hrtick;
1161 rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
1164 void hrtick_resched(void)
1167 unsigned long flags;
1169 if (!test_thread_flag(TIF_HRTICK_RESCHED))
1172 local_irq_save(flags);
1173 rq = cpu_rq(smp_processor_id());
1175 local_irq_restore(flags);
1178 static inline void hrtick_clear(struct rq *rq)
1182 static inline void hrtick_set(struct rq *rq)
1186 static inline void init_rq_hrtick(struct rq *rq)
1190 void hrtick_resched(void)
1194 static inline void init_hrtick(void)
1200 * resched_task - mark a task 'to be rescheduled now'.
1202 * On UP this means the setting of the need_resched flag, on SMP it
1203 * might also involve a cross-CPU call to trigger the scheduler on
1208 #ifndef tsk_is_polling
1209 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1212 static void __resched_task(struct task_struct *p, int tif_bit)
1216 assert_spin_locked(&task_rq(p)->lock);
1218 if (unlikely(test_tsk_thread_flag(p, tif_bit)))
1221 set_tsk_thread_flag(p, tif_bit);
1224 if (cpu == smp_processor_id())
1227 /* NEED_RESCHED must be visible before we test polling */
1229 if (!tsk_is_polling(p))
1230 smp_send_reschedule(cpu);
1233 static void resched_cpu(int cpu)
1235 struct rq *rq = cpu_rq(cpu);
1236 unsigned long flags;
1238 if (!spin_trylock_irqsave(&rq->lock, flags))
1240 resched_task(cpu_curr(cpu));
1241 spin_unlock_irqrestore(&rq->lock, flags);
1246 * When add_timer_on() enqueues a timer into the timer wheel of an
1247 * idle CPU then this timer might expire before the next timer event
1248 * which is scheduled to wake up that CPU. In case of a completely
1249 * idle system the next event might even be infinite time into the
1250 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1251 * leaves the inner idle loop so the newly added timer is taken into
1252 * account when the CPU goes back to idle and evaluates the timer
1253 * wheel for the next timer event.
1255 void wake_up_idle_cpu(int cpu)
1257 struct rq *rq = cpu_rq(cpu);
1259 if (cpu == smp_processor_id())
1263 * This is safe, as this function is called with the timer
1264 * wheel base lock of (cpu) held. When the CPU is on the way
1265 * to idle and has not yet set rq->curr to idle then it will
1266 * be serialized on the timer wheel base lock and take the new
1267 * timer into account automatically.
1269 if (rq->curr != rq->idle)
1273 * We can set TIF_RESCHED on the idle task of the other CPU
1274 * lockless. The worst case is that the other CPU runs the
1275 * idle task through an additional NOOP schedule()
1277 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1279 /* NEED_RESCHED must be visible before we test polling */
1281 if (!tsk_is_polling(rq->idle))
1282 smp_send_reschedule(cpu);
1284 #endif /* CONFIG_NO_HZ */
1286 #else /* !CONFIG_SMP */
1287 static void __resched_task(struct task_struct *p, int tif_bit)
1289 assert_spin_locked(&task_rq(p)->lock);
1290 set_tsk_thread_flag(p, tif_bit);
1292 #endif /* CONFIG_SMP */
1294 #if BITS_PER_LONG == 32
1295 # define WMULT_CONST (~0UL)
1297 # define WMULT_CONST (1UL << 32)
1300 #define WMULT_SHIFT 32
1303 * Shift right and round:
1305 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1308 * delta *= weight / lw
1310 static unsigned long
1311 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1312 struct load_weight *lw)
1316 if (!lw->inv_weight) {
1317 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1320 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1324 tmp = (u64)delta_exec * weight;
1326 * Check whether we'd overflow the 64-bit multiplication:
1328 if (unlikely(tmp > WMULT_CONST))
1329 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1332 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1334 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1337 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1343 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1350 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1351 * of tasks with abnormal "nice" values across CPUs the contribution that
1352 * each task makes to its run queue's load is weighted according to its
1353 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1354 * scaled version of the new time slice allocation that they receive on time
1358 #define WEIGHT_IDLEPRIO 2
1359 #define WMULT_IDLEPRIO (1 << 31)
1362 * Nice levels are multiplicative, with a gentle 10% change for every
1363 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1364 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1365 * that remained on nice 0.
1367 * The "10% effect" is relative and cumulative: from _any_ nice level,
1368 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1369 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1370 * If a task goes up by ~10% and another task goes down by ~10% then
1371 * the relative distance between them is ~25%.)
1373 static const int prio_to_weight[40] = {
1374 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1375 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1376 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1377 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1378 /* 0 */ 1024, 820, 655, 526, 423,
1379 /* 5 */ 335, 272, 215, 172, 137,
1380 /* 10 */ 110, 87, 70, 56, 45,
1381 /* 15 */ 36, 29, 23, 18, 15,
1385 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1387 * In cases where the weight does not change often, we can use the
1388 * precalculated inverse to speed up arithmetics by turning divisions
1389 * into multiplications:
1391 static const u32 prio_to_wmult[40] = {
1392 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1393 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1394 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1395 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1396 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1397 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1398 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1399 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1402 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1405 * runqueue iterator, to support SMP load-balancing between different
1406 * scheduling classes, without having to expose their internal data
1407 * structures to the load-balancing proper:
1409 struct rq_iterator {
1411 struct task_struct *(*start)(void *);
1412 struct task_struct *(*next)(void *);
1416 static unsigned long
1417 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1418 unsigned long max_load_move, struct sched_domain *sd,
1419 enum cpu_idle_type idle, int *all_pinned,
1420 int *this_best_prio, struct rq_iterator *iterator);
1423 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1424 struct sched_domain *sd, enum cpu_idle_type idle,
1425 struct rq_iterator *iterator);
1428 #ifdef CONFIG_CGROUP_CPUACCT
1429 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1431 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1434 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1436 update_load_add(&rq->load, load);
1439 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1441 update_load_sub(&rq->load, load);
1445 static unsigned long source_load(int cpu, int type);
1446 static unsigned long target_load(int cpu, int type);
1447 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1449 static unsigned long cpu_avg_load_per_task(int cpu)
1451 struct rq *rq = cpu_rq(cpu);
1454 rq->avg_load_per_task = rq->load.weight / rq->nr_running;
1456 return rq->avg_load_per_task;
1459 #ifdef CONFIG_FAIR_GROUP_SCHED
1461 typedef void (*tg_visitor)(struct task_group *, int, struct sched_domain *);
1464 * Iterate the full tree, calling @down when first entering a node and @up when
1465 * leaving it for the final time.
1468 walk_tg_tree(tg_visitor down, tg_visitor up, int cpu, struct sched_domain *sd)
1470 struct task_group *parent, *child;
1473 parent = &root_task_group;
1475 (*down)(parent, cpu, sd);
1476 list_for_each_entry_rcu(child, &parent->children, siblings) {
1483 (*up)(parent, cpu, sd);
1486 parent = parent->parent;
1492 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1495 * Calculate and set the cpu's group shares.
1498 __update_group_shares_cpu(struct task_group *tg, int cpu,
1499 unsigned long sd_shares, unsigned long sd_rq_weight)
1502 unsigned long shares;
1503 unsigned long rq_weight;
1508 rq_weight = tg->cfs_rq[cpu]->load.weight;
1511 * If there are currently no tasks on the cpu pretend there is one of
1512 * average load so that when a new task gets to run here it will not
1513 * get delayed by group starvation.
1517 rq_weight = NICE_0_LOAD;
1520 if (unlikely(rq_weight > sd_rq_weight))
1521 rq_weight = sd_rq_weight;
1524 * \Sum shares * rq_weight
1525 * shares = -----------------------
1529 shares = (sd_shares * rq_weight) / (sd_rq_weight + 1);
1532 * record the actual number of shares, not the boosted amount.
1534 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1535 tg->cfs_rq[cpu]->rq_weight = rq_weight;
1537 if (shares < MIN_SHARES)
1538 shares = MIN_SHARES;
1539 else if (shares > MAX_SHARES)
1540 shares = MAX_SHARES;
1542 __set_se_shares(tg->se[cpu], shares);
1546 * Re-compute the task group their per cpu shares over the given domain.
1547 * This needs to be done in a bottom-up fashion because the rq weight of a
1548 * parent group depends on the shares of its child groups.
1551 tg_shares_up(struct task_group *tg, int cpu, struct sched_domain *sd)
1553 unsigned long rq_weight = 0;
1554 unsigned long shares = 0;
1557 for_each_cpu_mask(i, sd->span) {
1558 rq_weight += tg->cfs_rq[i]->load.weight;
1559 shares += tg->cfs_rq[i]->shares;
1562 if ((!shares && rq_weight) || shares > tg->shares)
1563 shares = tg->shares;
1565 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1566 shares = tg->shares;
1569 rq_weight = cpus_weight(sd->span) * NICE_0_LOAD;
1571 for_each_cpu_mask(i, sd->span) {
1572 struct rq *rq = cpu_rq(i);
1573 unsigned long flags;
1575 spin_lock_irqsave(&rq->lock, flags);
1576 __update_group_shares_cpu(tg, i, shares, rq_weight);
1577 spin_unlock_irqrestore(&rq->lock, flags);
1582 * Compute the cpu's hierarchical load factor for each task group.
1583 * This needs to be done in a top-down fashion because the load of a child
1584 * group is a fraction of its parents load.
1587 tg_load_down(struct task_group *tg, int cpu, struct sched_domain *sd)
1592 load = cpu_rq(cpu)->load.weight;
1594 load = tg->parent->cfs_rq[cpu]->h_load;
1595 load *= tg->cfs_rq[cpu]->shares;
1596 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1599 tg->cfs_rq[cpu]->h_load = load;
1603 tg_nop(struct task_group *tg, int cpu, struct sched_domain *sd)
1607 static void update_shares(struct sched_domain *sd)
1609 u64 now = cpu_clock(raw_smp_processor_id());
1610 s64 elapsed = now - sd->last_update;
1612 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1613 sd->last_update = now;
1614 walk_tg_tree(tg_nop, tg_shares_up, 0, sd);
1618 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1620 spin_unlock(&rq->lock);
1622 spin_lock(&rq->lock);
1625 static void update_h_load(int cpu)
1627 walk_tg_tree(tg_load_down, tg_nop, cpu, NULL);
1632 static inline void update_shares(struct sched_domain *sd)
1636 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1644 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1646 #if defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)
1647 cfs_rq->shares = shares;
1651 #include "sched_stats.h"
1652 #include "sched_idletask.c"
1653 #include "sched_fair.c"
1654 #include "sched_rt.c"
1655 #ifdef CONFIG_SCHED_DEBUG
1656 # include "sched_debug.c"
1659 #define sched_class_highest (&rt_sched_class)
1660 #define for_each_class(class) \
1661 for (class = sched_class_highest; class; class = class->next)
1663 static void inc_nr_running(struct rq *rq)
1668 static void dec_nr_running(struct rq *rq)
1673 static void set_load_weight(struct task_struct *p)
1675 if (task_has_rt_policy(p)) {
1676 p->se.load.weight = prio_to_weight[0] * 2;
1677 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1682 * SCHED_IDLE tasks get minimal weight:
1684 if (p->policy == SCHED_IDLE) {
1685 p->se.load.weight = WEIGHT_IDLEPRIO;
1686 p->se.load.inv_weight = WMULT_IDLEPRIO;
1690 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1691 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1694 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1696 sched_info_queued(p);
1697 p->sched_class->enqueue_task(rq, p, wakeup);
1701 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1703 p->sched_class->dequeue_task(rq, p, sleep);
1708 * __normal_prio - return the priority that is based on the static prio
1710 static inline int __normal_prio(struct task_struct *p)
1712 return p->static_prio;
1716 * Calculate the expected normal priority: i.e. priority
1717 * without taking RT-inheritance into account. Might be
1718 * boosted by interactivity modifiers. Changes upon fork,
1719 * setprio syscalls, and whenever the interactivity
1720 * estimator recalculates.
1722 static inline int normal_prio(struct task_struct *p)
1726 if (task_has_rt_policy(p))
1727 prio = MAX_RT_PRIO-1 - p->rt_priority;
1729 prio = __normal_prio(p);
1734 * Calculate the current priority, i.e. the priority
1735 * taken into account by the scheduler. This value might
1736 * be boosted by RT tasks, or might be boosted by
1737 * interactivity modifiers. Will be RT if the task got
1738 * RT-boosted. If not then it returns p->normal_prio.
1740 static int effective_prio(struct task_struct *p)
1742 p->normal_prio = normal_prio(p);
1744 * If we are RT tasks or we were boosted to RT priority,
1745 * keep the priority unchanged. Otherwise, update priority
1746 * to the normal priority:
1748 if (!rt_prio(p->prio))
1749 return p->normal_prio;
1754 * activate_task - move a task to the runqueue.
1756 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1758 if (task_contributes_to_load(p))
1759 rq->nr_uninterruptible--;
1761 enqueue_task(rq, p, wakeup);
1766 * deactivate_task - remove a task from the runqueue.
1768 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1770 if (task_contributes_to_load(p))
1771 rq->nr_uninterruptible++;
1773 dequeue_task(rq, p, sleep);
1778 * task_curr - is this task currently executing on a CPU?
1779 * @p: the task in question.
1781 inline int task_curr(const struct task_struct *p)
1783 return cpu_curr(task_cpu(p)) == p;
1786 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1788 set_task_rq(p, cpu);
1791 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1792 * successfuly executed on another CPU. We must ensure that updates of
1793 * per-task data have been completed by this moment.
1796 task_thread_info(p)->cpu = cpu;
1800 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1801 const struct sched_class *prev_class,
1802 int oldprio, int running)
1804 if (prev_class != p->sched_class) {
1805 if (prev_class->switched_from)
1806 prev_class->switched_from(rq, p, running);
1807 p->sched_class->switched_to(rq, p, running);
1809 p->sched_class->prio_changed(rq, p, oldprio, running);
1814 /* Used instead of source_load when we know the type == 0 */
1815 static unsigned long weighted_cpuload(const int cpu)
1817 return cpu_rq(cpu)->load.weight;
1821 * Is this task likely cache-hot:
1824 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1829 * Buddy candidates are cache hot:
1831 if (sched_feat(CACHE_HOT_BUDDY) && (&p->se == cfs_rq_of(&p->se)->next))
1834 if (p->sched_class != &fair_sched_class)
1837 if (sysctl_sched_migration_cost == -1)
1839 if (sysctl_sched_migration_cost == 0)
1842 delta = now - p->se.exec_start;
1844 return delta < (s64)sysctl_sched_migration_cost;
1848 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1850 int old_cpu = task_cpu(p);
1851 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1852 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1853 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1856 clock_offset = old_rq->clock - new_rq->clock;
1858 #ifdef CONFIG_SCHEDSTATS
1859 if (p->se.wait_start)
1860 p->se.wait_start -= clock_offset;
1861 if (p->se.sleep_start)
1862 p->se.sleep_start -= clock_offset;
1863 if (p->se.block_start)
1864 p->se.block_start -= clock_offset;
1865 if (old_cpu != new_cpu) {
1866 schedstat_inc(p, se.nr_migrations);
1867 if (task_hot(p, old_rq->clock, NULL))
1868 schedstat_inc(p, se.nr_forced2_migrations);
1871 p->se.vruntime -= old_cfsrq->min_vruntime -
1872 new_cfsrq->min_vruntime;
1874 __set_task_cpu(p, new_cpu);
1877 struct migration_req {
1878 struct list_head list;
1880 struct task_struct *task;
1883 struct completion done;
1887 * The task's runqueue lock must be held.
1888 * Returns true if you have to wait for migration thread.
1891 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1893 struct rq *rq = task_rq(p);
1896 * If the task is not on a runqueue (and not running), then
1897 * it is sufficient to simply update the task's cpu field.
1899 if (!p->se.on_rq && !task_running(rq, p)) {
1900 set_task_cpu(p, dest_cpu);
1904 init_completion(&req->done);
1906 req->dest_cpu = dest_cpu;
1907 list_add(&req->list, &rq->migration_queue);
1913 * wait_task_inactive - wait for a thread to unschedule.
1915 * The caller must ensure that the task *will* unschedule sometime soon,
1916 * else this function might spin for a *long* time. This function can't
1917 * be called with interrupts off, or it may introduce deadlock with
1918 * smp_call_function() if an IPI is sent by the same process we are
1919 * waiting to become inactive.
1921 void wait_task_inactive(struct task_struct *p)
1923 unsigned long flags;
1929 * We do the initial early heuristics without holding
1930 * any task-queue locks at all. We'll only try to get
1931 * the runqueue lock when things look like they will
1937 * If the task is actively running on another CPU
1938 * still, just relax and busy-wait without holding
1941 * NOTE! Since we don't hold any locks, it's not
1942 * even sure that "rq" stays as the right runqueue!
1943 * But we don't care, since "task_running()" will
1944 * return false if the runqueue has changed and p
1945 * is actually now running somewhere else!
1947 while (task_running(rq, p))
1951 * Ok, time to look more closely! We need the rq
1952 * lock now, to be *sure*. If we're wrong, we'll
1953 * just go back and repeat.
1955 rq = task_rq_lock(p, &flags);
1956 running = task_running(rq, p);
1957 on_rq = p->se.on_rq;
1958 task_rq_unlock(rq, &flags);
1961 * Was it really running after all now that we
1962 * checked with the proper locks actually held?
1964 * Oops. Go back and try again..
1966 if (unlikely(running)) {
1972 * It's not enough that it's not actively running,
1973 * it must be off the runqueue _entirely_, and not
1976 * So if it wa still runnable (but just not actively
1977 * running right now), it's preempted, and we should
1978 * yield - it could be a while.
1980 if (unlikely(on_rq)) {
1981 schedule_timeout_uninterruptible(1);
1986 * Ahh, all good. It wasn't running, and it wasn't
1987 * runnable, which means that it will never become
1988 * running in the future either. We're all done!
1995 * kick_process - kick a running thread to enter/exit the kernel
1996 * @p: the to-be-kicked thread
1998 * Cause a process which is running on another CPU to enter
1999 * kernel-mode, without any delay. (to get signals handled.)
2001 * NOTE: this function doesnt have to take the runqueue lock,
2002 * because all it wants to ensure is that the remote task enters
2003 * the kernel. If the IPI races and the task has been migrated
2004 * to another CPU then no harm is done and the purpose has been
2007 void kick_process(struct task_struct *p)
2013 if ((cpu != smp_processor_id()) && task_curr(p))
2014 smp_send_reschedule(cpu);
2019 * Return a low guess at the load of a migration-source cpu weighted
2020 * according to the scheduling class and "nice" value.
2022 * We want to under-estimate the load of migration sources, to
2023 * balance conservatively.
2025 static unsigned long source_load(int cpu, int type)
2027 struct rq *rq = cpu_rq(cpu);
2028 unsigned long total = weighted_cpuload(cpu);
2030 if (type == 0 || !sched_feat(LB_BIAS))
2033 return min(rq->cpu_load[type-1], total);
2037 * Return a high guess at the load of a migration-target cpu weighted
2038 * according to the scheduling class and "nice" value.
2040 static unsigned long target_load(int cpu, int type)
2042 struct rq *rq = cpu_rq(cpu);
2043 unsigned long total = weighted_cpuload(cpu);
2045 if (type == 0 || !sched_feat(LB_BIAS))
2048 return max(rq->cpu_load[type-1], total);
2052 * find_idlest_group finds and returns the least busy CPU group within the
2055 static struct sched_group *
2056 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2058 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2059 unsigned long min_load = ULONG_MAX, this_load = 0;
2060 int load_idx = sd->forkexec_idx;
2061 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2064 unsigned long load, avg_load;
2068 /* Skip over this group if it has no CPUs allowed */
2069 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
2072 local_group = cpu_isset(this_cpu, group->cpumask);
2074 /* Tally up the load of all CPUs in the group */
2077 for_each_cpu_mask(i, group->cpumask) {
2078 /* Bias balancing toward cpus of our domain */
2080 load = source_load(i, load_idx);
2082 load = target_load(i, load_idx);
2087 /* Adjust by relative CPU power of the group */
2088 avg_load = sg_div_cpu_power(group,
2089 avg_load * SCHED_LOAD_SCALE);
2092 this_load = avg_load;
2094 } else if (avg_load < min_load) {
2095 min_load = avg_load;
2098 } while (group = group->next, group != sd->groups);
2100 if (!idlest || 100*this_load < imbalance*min_load)
2106 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2109 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu,
2112 unsigned long load, min_load = ULONG_MAX;
2116 /* Traverse only the allowed CPUs */
2117 cpus_and(*tmp, group->cpumask, p->cpus_allowed);
2119 for_each_cpu_mask(i, *tmp) {
2120 load = weighted_cpuload(i);
2122 if (load < min_load || (load == min_load && i == this_cpu)) {
2132 * sched_balance_self: balance the current task (running on cpu) in domains
2133 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2136 * Balance, ie. select the least loaded group.
2138 * Returns the target CPU number, or the same CPU if no balancing is needed.
2140 * preempt must be disabled.
2142 static int sched_balance_self(int cpu, int flag)
2144 struct task_struct *t = current;
2145 struct sched_domain *tmp, *sd = NULL;
2147 for_each_domain(cpu, tmp) {
2149 * If power savings logic is enabled for a domain, stop there.
2151 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2153 if (tmp->flags & flag)
2161 cpumask_t span, tmpmask;
2162 struct sched_group *group;
2163 int new_cpu, weight;
2165 if (!(sd->flags & flag)) {
2171 group = find_idlest_group(sd, t, cpu);
2177 new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask);
2178 if (new_cpu == -1 || new_cpu == cpu) {
2179 /* Now try balancing at a lower domain level of cpu */
2184 /* Now try balancing at a lower domain level of new_cpu */
2187 weight = cpus_weight(span);
2188 for_each_domain(cpu, tmp) {
2189 if (weight <= cpus_weight(tmp->span))
2191 if (tmp->flags & flag)
2194 /* while loop will break here if sd == NULL */
2200 #endif /* CONFIG_SMP */
2203 * try_to_wake_up - wake up a thread
2204 * @p: the to-be-woken-up thread
2205 * @state: the mask of task states that can be woken
2206 * @sync: do a synchronous wakeup?
2208 * Put it on the run-queue if it's not already there. The "current"
2209 * thread is always on the run-queue (except when the actual
2210 * re-schedule is in progress), and as such you're allowed to do
2211 * the simpler "current->state = TASK_RUNNING" to mark yourself
2212 * runnable without the overhead of this.
2214 * returns failure only if the task is already active.
2216 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2218 int cpu, orig_cpu, this_cpu, success = 0;
2219 unsigned long flags;
2223 if (!sched_feat(SYNC_WAKEUPS))
2227 if (sched_feat(LB_WAKEUP_UPDATE)) {
2228 struct sched_domain *sd;
2230 this_cpu = raw_smp_processor_id();
2233 for_each_domain(this_cpu, sd) {
2234 if (cpu_isset(cpu, sd->span)) {
2243 rq = task_rq_lock(p, &flags);
2244 old_state = p->state;
2245 if (!(old_state & state))
2253 this_cpu = smp_processor_id();
2256 if (unlikely(task_running(rq, p)))
2259 cpu = p->sched_class->select_task_rq(p, sync);
2260 if (cpu != orig_cpu) {
2261 set_task_cpu(p, cpu);
2262 task_rq_unlock(rq, &flags);
2263 /* might preempt at this point */
2264 rq = task_rq_lock(p, &flags);
2265 old_state = p->state;
2266 if (!(old_state & state))
2271 this_cpu = smp_processor_id();
2275 #ifdef CONFIG_SCHEDSTATS
2276 schedstat_inc(rq, ttwu_count);
2277 if (cpu == this_cpu)
2278 schedstat_inc(rq, ttwu_local);
2280 struct sched_domain *sd;
2281 for_each_domain(this_cpu, sd) {
2282 if (cpu_isset(cpu, sd->span)) {
2283 schedstat_inc(sd, ttwu_wake_remote);
2288 #endif /* CONFIG_SCHEDSTATS */
2291 #endif /* CONFIG_SMP */
2292 schedstat_inc(p, se.nr_wakeups);
2294 schedstat_inc(p, se.nr_wakeups_sync);
2295 if (orig_cpu != cpu)
2296 schedstat_inc(p, se.nr_wakeups_migrate);
2297 if (cpu == this_cpu)
2298 schedstat_inc(p, se.nr_wakeups_local);
2300 schedstat_inc(p, se.nr_wakeups_remote);
2301 update_rq_clock(rq);
2302 activate_task(rq, p, 1);
2306 check_preempt_curr(rq, p);
2308 p->state = TASK_RUNNING;
2310 if (p->sched_class->task_wake_up)
2311 p->sched_class->task_wake_up(rq, p);
2314 task_rq_unlock(rq, &flags);
2319 int wake_up_process(struct task_struct *p)
2321 return try_to_wake_up(p, TASK_ALL, 0);
2323 EXPORT_SYMBOL(wake_up_process);
2325 int wake_up_state(struct task_struct *p, unsigned int state)
2327 return try_to_wake_up(p, state, 0);
2331 * Perform scheduler related setup for a newly forked process p.
2332 * p is forked by current.
2334 * __sched_fork() is basic setup used by init_idle() too:
2336 static void __sched_fork(struct task_struct *p)
2338 p->se.exec_start = 0;
2339 p->se.sum_exec_runtime = 0;
2340 p->se.prev_sum_exec_runtime = 0;
2341 p->se.last_wakeup = 0;
2342 p->se.avg_overlap = 0;
2344 #ifdef CONFIG_SCHEDSTATS
2345 p->se.wait_start = 0;
2346 p->se.sum_sleep_runtime = 0;
2347 p->se.sleep_start = 0;
2348 p->se.block_start = 0;
2349 p->se.sleep_max = 0;
2350 p->se.block_max = 0;
2352 p->se.slice_max = 0;
2356 INIT_LIST_HEAD(&p->rt.run_list);
2358 INIT_LIST_HEAD(&p->se.group_node);
2360 #ifdef CONFIG_PREEMPT_NOTIFIERS
2361 INIT_HLIST_HEAD(&p->preempt_notifiers);
2365 * We mark the process as running here, but have not actually
2366 * inserted it onto the runqueue yet. This guarantees that
2367 * nobody will actually run it, and a signal or other external
2368 * event cannot wake it up and insert it on the runqueue either.
2370 p->state = TASK_RUNNING;
2374 * fork()/clone()-time setup:
2376 void sched_fork(struct task_struct *p, int clone_flags)
2378 int cpu = get_cpu();
2383 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2385 set_task_cpu(p, cpu);
2388 * Make sure we do not leak PI boosting priority to the child:
2390 p->prio = current->normal_prio;
2391 if (!rt_prio(p->prio))
2392 p->sched_class = &fair_sched_class;
2394 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2395 if (likely(sched_info_on()))
2396 memset(&p->sched_info, 0, sizeof(p->sched_info));
2398 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2401 #ifdef CONFIG_PREEMPT
2402 /* Want to start with kernel preemption disabled. */
2403 task_thread_info(p)->preempt_count = 1;
2409 * wake_up_new_task - wake up a newly created task for the first time.
2411 * This function will do some initial scheduler statistics housekeeping
2412 * that must be done for every newly created context, then puts the task
2413 * on the runqueue and wakes it.
2415 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2417 unsigned long flags;
2420 rq = task_rq_lock(p, &flags);
2421 BUG_ON(p->state != TASK_RUNNING);
2422 update_rq_clock(rq);
2424 p->prio = effective_prio(p);
2426 if (!p->sched_class->task_new || !current->se.on_rq) {
2427 activate_task(rq, p, 0);
2430 * Let the scheduling class do new task startup
2431 * management (if any):
2433 p->sched_class->task_new(rq, p);
2436 check_preempt_curr(rq, p);
2438 if (p->sched_class->task_wake_up)
2439 p->sched_class->task_wake_up(rq, p);
2441 task_rq_unlock(rq, &flags);
2444 #ifdef CONFIG_PREEMPT_NOTIFIERS
2447 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2448 * @notifier: notifier struct to register
2450 void preempt_notifier_register(struct preempt_notifier *notifier)
2452 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2454 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2457 * preempt_notifier_unregister - no longer interested in preemption notifications
2458 * @notifier: notifier struct to unregister
2460 * This is safe to call from within a preemption notifier.
2462 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2464 hlist_del(¬ifier->link);
2466 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2468 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2470 struct preempt_notifier *notifier;
2471 struct hlist_node *node;
2473 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2474 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2478 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2479 struct task_struct *next)
2481 struct preempt_notifier *notifier;
2482 struct hlist_node *node;
2484 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2485 notifier->ops->sched_out(notifier, next);
2488 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2490 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2495 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2496 struct task_struct *next)
2500 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2503 * prepare_task_switch - prepare to switch tasks
2504 * @rq: the runqueue preparing to switch
2505 * @prev: the current task that is being switched out
2506 * @next: the task we are going to switch to.
2508 * This is called with the rq lock held and interrupts off. It must
2509 * be paired with a subsequent finish_task_switch after the context
2512 * prepare_task_switch sets up locking and calls architecture specific
2516 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2517 struct task_struct *next)
2519 fire_sched_out_preempt_notifiers(prev, next);
2520 prepare_lock_switch(rq, next);
2521 prepare_arch_switch(next);
2525 * finish_task_switch - clean up after a task-switch
2526 * @rq: runqueue associated with task-switch
2527 * @prev: the thread we just switched away from.
2529 * finish_task_switch must be called after the context switch, paired
2530 * with a prepare_task_switch call before the context switch.
2531 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2532 * and do any other architecture-specific cleanup actions.
2534 * Note that we may have delayed dropping an mm in context_switch(). If
2535 * so, we finish that here outside of the runqueue lock. (Doing it
2536 * with the lock held can cause deadlocks; see schedule() for
2539 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2540 __releases(rq->lock)
2542 struct mm_struct *mm = rq->prev_mm;
2548 * A task struct has one reference for the use as "current".
2549 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2550 * schedule one last time. The schedule call will never return, and
2551 * the scheduled task must drop that reference.
2552 * The test for TASK_DEAD must occur while the runqueue locks are
2553 * still held, otherwise prev could be scheduled on another cpu, die
2554 * there before we look at prev->state, and then the reference would
2556 * Manfred Spraul <manfred@colorfullife.com>
2558 prev_state = prev->state;
2559 finish_arch_switch(prev);
2560 finish_lock_switch(rq, prev);
2562 if (current->sched_class->post_schedule)
2563 current->sched_class->post_schedule(rq);
2566 fire_sched_in_preempt_notifiers(current);
2569 if (unlikely(prev_state == TASK_DEAD)) {
2571 * Remove function-return probe instances associated with this
2572 * task and put them back on the free list.
2574 kprobe_flush_task(prev);
2575 put_task_struct(prev);
2580 * schedule_tail - first thing a freshly forked thread must call.
2581 * @prev: the thread we just switched away from.
2583 asmlinkage void schedule_tail(struct task_struct *prev)
2584 __releases(rq->lock)
2586 struct rq *rq = this_rq();
2588 finish_task_switch(rq, prev);
2589 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2590 /* In this case, finish_task_switch does not reenable preemption */
2593 if (current->set_child_tid)
2594 put_user(task_pid_vnr(current), current->set_child_tid);
2598 * context_switch - switch to the new MM and the new
2599 * thread's register state.
2602 context_switch(struct rq *rq, struct task_struct *prev,
2603 struct task_struct *next)
2605 struct mm_struct *mm, *oldmm;
2607 prepare_task_switch(rq, prev, next);
2609 oldmm = prev->active_mm;
2611 * For paravirt, this is coupled with an exit in switch_to to
2612 * combine the page table reload and the switch backend into
2615 arch_enter_lazy_cpu_mode();
2617 if (unlikely(!mm)) {
2618 next->active_mm = oldmm;
2619 atomic_inc(&oldmm->mm_count);
2620 enter_lazy_tlb(oldmm, next);
2622 switch_mm(oldmm, mm, next);
2624 if (unlikely(!prev->mm)) {
2625 prev->active_mm = NULL;
2626 rq->prev_mm = oldmm;
2629 * Since the runqueue lock will be released by the next
2630 * task (which is an invalid locking op but in the case
2631 * of the scheduler it's an obvious special-case), so we
2632 * do an early lockdep release here:
2634 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2635 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2638 /* Here we just switch the register state and the stack. */
2639 switch_to(prev, next, prev);
2643 * this_rq must be evaluated again because prev may have moved
2644 * CPUs since it called schedule(), thus the 'rq' on its stack
2645 * frame will be invalid.
2647 finish_task_switch(this_rq(), prev);
2651 * nr_running, nr_uninterruptible and nr_context_switches:
2653 * externally visible scheduler statistics: current number of runnable
2654 * threads, current number of uninterruptible-sleeping threads, total
2655 * number of context switches performed since bootup.
2657 unsigned long nr_running(void)
2659 unsigned long i, sum = 0;
2661 for_each_online_cpu(i)
2662 sum += cpu_rq(i)->nr_running;
2667 unsigned long nr_uninterruptible(void)
2669 unsigned long i, sum = 0;
2671 for_each_possible_cpu(i)
2672 sum += cpu_rq(i)->nr_uninterruptible;
2675 * Since we read the counters lockless, it might be slightly
2676 * inaccurate. Do not allow it to go below zero though:
2678 if (unlikely((long)sum < 0))
2684 unsigned long long nr_context_switches(void)
2687 unsigned long long sum = 0;
2689 for_each_possible_cpu(i)
2690 sum += cpu_rq(i)->nr_switches;
2695 unsigned long nr_iowait(void)
2697 unsigned long i, sum = 0;
2699 for_each_possible_cpu(i)
2700 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2705 unsigned long nr_active(void)
2707 unsigned long i, running = 0, uninterruptible = 0;
2709 for_each_online_cpu(i) {
2710 running += cpu_rq(i)->nr_running;
2711 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2714 if (unlikely((long)uninterruptible < 0))
2715 uninterruptible = 0;
2717 return running + uninterruptible;
2721 * Update rq->cpu_load[] statistics. This function is usually called every
2722 * scheduler tick (TICK_NSEC).
2724 static void update_cpu_load(struct rq *this_rq)
2726 unsigned long this_load = this_rq->load.weight;
2729 this_rq->nr_load_updates++;
2731 /* Update our load: */
2732 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2733 unsigned long old_load, new_load;
2735 /* scale is effectively 1 << i now, and >> i divides by scale */
2737 old_load = this_rq->cpu_load[i];
2738 new_load = this_load;
2740 * Round up the averaging division if load is increasing. This
2741 * prevents us from getting stuck on 9 if the load is 10, for
2744 if (new_load > old_load)
2745 new_load += scale-1;
2746 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2753 * double_rq_lock - safely lock two runqueues
2755 * Note this does not disable interrupts like task_rq_lock,
2756 * you need to do so manually before calling.
2758 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2759 __acquires(rq1->lock)
2760 __acquires(rq2->lock)
2762 BUG_ON(!irqs_disabled());
2764 spin_lock(&rq1->lock);
2765 __acquire(rq2->lock); /* Fake it out ;) */
2768 spin_lock(&rq1->lock);
2769 spin_lock(&rq2->lock);
2771 spin_lock(&rq2->lock);
2772 spin_lock(&rq1->lock);
2775 update_rq_clock(rq1);
2776 update_rq_clock(rq2);
2780 * double_rq_unlock - safely unlock two runqueues
2782 * Note this does not restore interrupts like task_rq_unlock,
2783 * you need to do so manually after calling.
2785 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2786 __releases(rq1->lock)
2787 __releases(rq2->lock)
2789 spin_unlock(&rq1->lock);
2791 spin_unlock(&rq2->lock);
2793 __release(rq2->lock);
2797 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2799 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2800 __releases(this_rq->lock)
2801 __acquires(busiest->lock)
2802 __acquires(this_rq->lock)
2806 if (unlikely(!irqs_disabled())) {
2807 /* printk() doesn't work good under rq->lock */
2808 spin_unlock(&this_rq->lock);
2811 if (unlikely(!spin_trylock(&busiest->lock))) {
2812 if (busiest < this_rq) {
2813 spin_unlock(&this_rq->lock);
2814 spin_lock(&busiest->lock);
2815 spin_lock(&this_rq->lock);
2818 spin_lock(&busiest->lock);
2824 * If dest_cpu is allowed for this process, migrate the task to it.
2825 * This is accomplished by forcing the cpu_allowed mask to only
2826 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2827 * the cpu_allowed mask is restored.
2829 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2831 struct migration_req req;
2832 unsigned long flags;
2835 rq = task_rq_lock(p, &flags);
2836 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2837 || unlikely(cpu_is_offline(dest_cpu)))
2840 /* force the process onto the specified CPU */
2841 if (migrate_task(p, dest_cpu, &req)) {
2842 /* Need to wait for migration thread (might exit: take ref). */
2843 struct task_struct *mt = rq->migration_thread;
2845 get_task_struct(mt);
2846 task_rq_unlock(rq, &flags);
2847 wake_up_process(mt);
2848 put_task_struct(mt);
2849 wait_for_completion(&req.done);
2854 task_rq_unlock(rq, &flags);
2858 * sched_exec - execve() is a valuable balancing opportunity, because at
2859 * this point the task has the smallest effective memory and cache footprint.
2861 void sched_exec(void)
2863 int new_cpu, this_cpu = get_cpu();
2864 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2866 if (new_cpu != this_cpu)
2867 sched_migrate_task(current, new_cpu);
2871 * pull_task - move a task from a remote runqueue to the local runqueue.
2872 * Both runqueues must be locked.
2874 static void pull_task(struct rq *src_rq, struct task_struct *p,
2875 struct rq *this_rq, int this_cpu)
2877 deactivate_task(src_rq, p, 0);
2878 set_task_cpu(p, this_cpu);
2879 activate_task(this_rq, p, 0);
2881 * Note that idle threads have a prio of MAX_PRIO, for this test
2882 * to be always true for them.
2884 check_preempt_curr(this_rq, p);
2888 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2891 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2892 struct sched_domain *sd, enum cpu_idle_type idle,
2896 * We do not migrate tasks that are:
2897 * 1) running (obviously), or
2898 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2899 * 3) are cache-hot on their current CPU.
2901 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2902 schedstat_inc(p, se.nr_failed_migrations_affine);
2907 if (task_running(rq, p)) {
2908 schedstat_inc(p, se.nr_failed_migrations_running);
2913 * Aggressive migration if:
2914 * 1) task is cache cold, or
2915 * 2) too many balance attempts have failed.
2918 if (!task_hot(p, rq->clock, sd) ||
2919 sd->nr_balance_failed > sd->cache_nice_tries) {
2920 #ifdef CONFIG_SCHEDSTATS
2921 if (task_hot(p, rq->clock, sd)) {
2922 schedstat_inc(sd, lb_hot_gained[idle]);
2923 schedstat_inc(p, se.nr_forced_migrations);
2929 if (task_hot(p, rq->clock, sd)) {
2930 schedstat_inc(p, se.nr_failed_migrations_hot);
2936 static unsigned long
2937 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2938 unsigned long max_load_move, struct sched_domain *sd,
2939 enum cpu_idle_type idle, int *all_pinned,
2940 int *this_best_prio, struct rq_iterator *iterator)
2942 int loops = 0, pulled = 0, pinned = 0;
2943 struct task_struct *p;
2944 long rem_load_move = max_load_move;
2946 if (max_load_move == 0)
2952 * Start the load-balancing iterator:
2954 p = iterator->start(iterator->arg);
2956 if (!p || loops++ > sysctl_sched_nr_migrate)
2959 if ((p->se.load.weight >> 1) > rem_load_move ||
2960 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2961 p = iterator->next(iterator->arg);
2965 pull_task(busiest, p, this_rq, this_cpu);
2967 rem_load_move -= p->se.load.weight;
2970 * We only want to steal up to the prescribed amount of weighted load.
2972 if (rem_load_move > 0) {
2973 if (p->prio < *this_best_prio)
2974 *this_best_prio = p->prio;
2975 p = iterator->next(iterator->arg);
2980 * Right now, this is one of only two places pull_task() is called,
2981 * so we can safely collect pull_task() stats here rather than
2982 * inside pull_task().
2984 schedstat_add(sd, lb_gained[idle], pulled);
2987 *all_pinned = pinned;
2989 return max_load_move - rem_load_move;
2993 * move_tasks tries to move up to max_load_move weighted load from busiest to
2994 * this_rq, as part of a balancing operation within domain "sd".
2995 * Returns 1 if successful and 0 otherwise.
2997 * Called with both runqueues locked.
2999 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3000 unsigned long max_load_move,
3001 struct sched_domain *sd, enum cpu_idle_type idle,
3004 const struct sched_class *class = sched_class_highest;
3005 unsigned long total_load_moved = 0;
3006 int this_best_prio = this_rq->curr->prio;
3010 class->load_balance(this_rq, this_cpu, busiest,
3011 max_load_move - total_load_moved,
3012 sd, idle, all_pinned, &this_best_prio);
3013 class = class->next;
3014 } while (class && max_load_move > total_load_moved);
3016 return total_load_moved > 0;
3020 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3021 struct sched_domain *sd, enum cpu_idle_type idle,
3022 struct rq_iterator *iterator)
3024 struct task_struct *p = iterator->start(iterator->arg);
3028 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3029 pull_task(busiest, p, this_rq, this_cpu);
3031 * Right now, this is only the second place pull_task()
3032 * is called, so we can safely collect pull_task()
3033 * stats here rather than inside pull_task().
3035 schedstat_inc(sd, lb_gained[idle]);
3039 p = iterator->next(iterator->arg);
3046 * move_one_task tries to move exactly one task from busiest to this_rq, as
3047 * part of active balancing operations within "domain".
3048 * Returns 1 if successful and 0 otherwise.
3050 * Called with both runqueues locked.
3052 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3053 struct sched_domain *sd, enum cpu_idle_type idle)
3055 const struct sched_class *class;
3057 for (class = sched_class_highest; class; class = class->next)
3058 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3065 * find_busiest_group finds and returns the busiest CPU group within the
3066 * domain. It calculates and returns the amount of weighted load which
3067 * should be moved to restore balance via the imbalance parameter.
3069 static struct sched_group *
3070 find_busiest_group(struct sched_domain *sd, int this_cpu,
3071 unsigned long *imbalance, enum cpu_idle_type idle,
3072 int *sd_idle, const cpumask_t *cpus, int *balance)
3074 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
3075 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
3076 unsigned long max_pull;
3077 unsigned long busiest_load_per_task, busiest_nr_running;
3078 unsigned long this_load_per_task, this_nr_running;
3079 int load_idx, group_imb = 0;
3080 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3081 int power_savings_balance = 1;
3082 unsigned long leader_nr_running = 0, min_load_per_task = 0;
3083 unsigned long min_nr_running = ULONG_MAX;
3084 struct sched_group *group_min = NULL, *group_leader = NULL;
3087 max_load = this_load = total_load = total_pwr = 0;
3088 busiest_load_per_task = busiest_nr_running = 0;
3089 this_load_per_task = this_nr_running = 0;
3091 if (idle == CPU_NOT_IDLE)
3092 load_idx = sd->busy_idx;
3093 else if (idle == CPU_NEWLY_IDLE)
3094 load_idx = sd->newidle_idx;
3096 load_idx = sd->idle_idx;
3099 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
3102 int __group_imb = 0;
3103 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3104 unsigned long sum_nr_running, sum_weighted_load;
3105 unsigned long sum_avg_load_per_task;
3106 unsigned long avg_load_per_task;
3108 local_group = cpu_isset(this_cpu, group->cpumask);
3111 balance_cpu = first_cpu(group->cpumask);
3113 /* Tally up the load of all CPUs in the group */
3114 sum_weighted_load = sum_nr_running = avg_load = 0;
3115 sum_avg_load_per_task = avg_load_per_task = 0;
3118 min_cpu_load = ~0UL;
3120 for_each_cpu_mask(i, group->cpumask) {
3123 if (!cpu_isset(i, *cpus))
3128 if (*sd_idle && rq->nr_running)
3131 /* Bias balancing toward cpus of our domain */
3133 if (idle_cpu(i) && !first_idle_cpu) {
3138 load = target_load(i, load_idx);
3140 load = source_load(i, load_idx);
3141 if (load > max_cpu_load)
3142 max_cpu_load = load;
3143 if (min_cpu_load > load)
3144 min_cpu_load = load;
3148 sum_nr_running += rq->nr_running;
3149 sum_weighted_load += weighted_cpuload(i);
3151 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3155 * First idle cpu or the first cpu(busiest) in this sched group
3156 * is eligible for doing load balancing at this and above
3157 * domains. In the newly idle case, we will allow all the cpu's
3158 * to do the newly idle load balance.
3160 if (idle != CPU_NEWLY_IDLE && local_group &&
3161 balance_cpu != this_cpu && balance) {
3166 total_load += avg_load;
3167 total_pwr += group->__cpu_power;
3169 /* Adjust by relative CPU power of the group */
3170 avg_load = sg_div_cpu_power(group,
3171 avg_load * SCHED_LOAD_SCALE);
3175 * Consider the group unbalanced when the imbalance is larger
3176 * than the average weight of two tasks.
3178 * APZ: with cgroup the avg task weight can vary wildly and
3179 * might not be a suitable number - should we keep a
3180 * normalized nr_running number somewhere that negates
3183 avg_load_per_task = sg_div_cpu_power(group,
3184 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3186 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3189 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3192 this_load = avg_load;
3194 this_nr_running = sum_nr_running;
3195 this_load_per_task = sum_weighted_load;
3196 } else if (avg_load > max_load &&
3197 (sum_nr_running > group_capacity || __group_imb)) {
3198 max_load = avg_load;
3200 busiest_nr_running = sum_nr_running;
3201 busiest_load_per_task = sum_weighted_load;
3202 group_imb = __group_imb;
3205 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3207 * Busy processors will not participate in power savings
3210 if (idle == CPU_NOT_IDLE ||
3211 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3215 * If the local group is idle or completely loaded
3216 * no need to do power savings balance at this domain
3218 if (local_group && (this_nr_running >= group_capacity ||
3220 power_savings_balance = 0;
3223 * If a group is already running at full capacity or idle,
3224 * don't include that group in power savings calculations
3226 if (!power_savings_balance || sum_nr_running >= group_capacity
3231 * Calculate the group which has the least non-idle load.
3232 * This is the group from where we need to pick up the load
3235 if ((sum_nr_running < min_nr_running) ||
3236 (sum_nr_running == min_nr_running &&
3237 first_cpu(group->cpumask) <
3238 first_cpu(group_min->cpumask))) {
3240 min_nr_running = sum_nr_running;
3241 min_load_per_task = sum_weighted_load /
3246 * Calculate the group which is almost near its
3247 * capacity but still has some space to pick up some load
3248 * from other group and save more power
3250 if (sum_nr_running <= group_capacity - 1) {
3251 if (sum_nr_running > leader_nr_running ||
3252 (sum_nr_running == leader_nr_running &&
3253 first_cpu(group->cpumask) >
3254 first_cpu(group_leader->cpumask))) {
3255 group_leader = group;
3256 leader_nr_running = sum_nr_running;
3261 group = group->next;
3262 } while (group != sd->groups);
3264 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3267 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3269 if (this_load >= avg_load ||
3270 100*max_load <= sd->imbalance_pct*this_load)
3273 busiest_load_per_task /= busiest_nr_running;
3275 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3278 * We're trying to get all the cpus to the average_load, so we don't
3279 * want to push ourselves above the average load, nor do we wish to
3280 * reduce the max loaded cpu below the average load, as either of these
3281 * actions would just result in more rebalancing later, and ping-pong
3282 * tasks around. Thus we look for the minimum possible imbalance.
3283 * Negative imbalances (*we* are more loaded than anyone else) will
3284 * be counted as no imbalance for these purposes -- we can't fix that
3285 * by pulling tasks to us. Be careful of negative numbers as they'll
3286 * appear as very large values with unsigned longs.
3288 if (max_load <= busiest_load_per_task)
3292 * In the presence of smp nice balancing, certain scenarios can have
3293 * max load less than avg load(as we skip the groups at or below
3294 * its cpu_power, while calculating max_load..)
3296 if (max_load < avg_load) {
3298 goto small_imbalance;
3301 /* Don't want to pull so many tasks that a group would go idle */
3302 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3304 /* How much load to actually move to equalise the imbalance */
3305 *imbalance = min(max_pull * busiest->__cpu_power,
3306 (avg_load - this_load) * this->__cpu_power)
3310 * if *imbalance is less than the average load per runnable task
3311 * there is no gaurantee that any tasks will be moved so we'll have
3312 * a think about bumping its value to force at least one task to be
3315 if (*imbalance < busiest_load_per_task) {
3316 unsigned long tmp, pwr_now, pwr_move;
3320 pwr_move = pwr_now = 0;
3322 if (this_nr_running) {
3323 this_load_per_task /= this_nr_running;
3324 if (busiest_load_per_task > this_load_per_task)
3327 this_load_per_task = cpu_avg_load_per_task(this_cpu);
3329 if (max_load - this_load + 2*busiest_load_per_task >=
3330 busiest_load_per_task * imbn) {
3331 *imbalance = busiest_load_per_task;
3336 * OK, we don't have enough imbalance to justify moving tasks,
3337 * however we may be able to increase total CPU power used by
3341 pwr_now += busiest->__cpu_power *
3342 min(busiest_load_per_task, max_load);
3343 pwr_now += this->__cpu_power *
3344 min(this_load_per_task, this_load);
3345 pwr_now /= SCHED_LOAD_SCALE;
3347 /* Amount of load we'd subtract */
3348 tmp = sg_div_cpu_power(busiest,
3349 busiest_load_per_task * SCHED_LOAD_SCALE);
3351 pwr_move += busiest->__cpu_power *
3352 min(busiest_load_per_task, max_load - tmp);
3354 /* Amount of load we'd add */
3355 if (max_load * busiest->__cpu_power <
3356 busiest_load_per_task * SCHED_LOAD_SCALE)
3357 tmp = sg_div_cpu_power(this,
3358 max_load * busiest->__cpu_power);
3360 tmp = sg_div_cpu_power(this,
3361 busiest_load_per_task * SCHED_LOAD_SCALE);
3362 pwr_move += this->__cpu_power *
3363 min(this_load_per_task, this_load + tmp);
3364 pwr_move /= SCHED_LOAD_SCALE;
3366 /* Move if we gain throughput */
3367 if (pwr_move > pwr_now)
3368 *imbalance = busiest_load_per_task;
3374 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3375 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3378 if (this == group_leader && group_leader != group_min) {
3379 *imbalance = min_load_per_task;
3389 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3392 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3393 unsigned long imbalance, const cpumask_t *cpus)
3395 struct rq *busiest = NULL, *rq;
3396 unsigned long max_load = 0;
3399 for_each_cpu_mask(i, group->cpumask) {
3402 if (!cpu_isset(i, *cpus))
3406 wl = weighted_cpuload(i);
3408 if (rq->nr_running == 1 && wl > imbalance)
3411 if (wl > max_load) {
3421 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3422 * so long as it is large enough.
3424 #define MAX_PINNED_INTERVAL 512
3427 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3428 * tasks if there is an imbalance.
3430 static int load_balance(int this_cpu, struct rq *this_rq,
3431 struct sched_domain *sd, enum cpu_idle_type idle,
3432 int *balance, cpumask_t *cpus)
3434 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3435 struct sched_group *group;
3436 unsigned long imbalance;
3438 unsigned long flags;
3443 * When power savings policy is enabled for the parent domain, idle
3444 * sibling can pick up load irrespective of busy siblings. In this case,
3445 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3446 * portraying it as CPU_NOT_IDLE.
3448 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3449 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3452 schedstat_inc(sd, lb_count[idle]);
3456 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3463 schedstat_inc(sd, lb_nobusyg[idle]);
3467 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3469 schedstat_inc(sd, lb_nobusyq[idle]);
3473 BUG_ON(busiest == this_rq);
3475 schedstat_add(sd, lb_imbalance[idle], imbalance);
3478 if (busiest->nr_running > 1) {
3480 * Attempt to move tasks. If find_busiest_group has found
3481 * an imbalance but busiest->nr_running <= 1, the group is
3482 * still unbalanced. ld_moved simply stays zero, so it is
3483 * correctly treated as an imbalance.
3485 local_irq_save(flags);
3486 double_rq_lock(this_rq, busiest);
3487 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3488 imbalance, sd, idle, &all_pinned);
3489 double_rq_unlock(this_rq, busiest);
3490 local_irq_restore(flags);
3493 * some other cpu did the load balance for us.
3495 if (ld_moved && this_cpu != smp_processor_id())
3496 resched_cpu(this_cpu);
3498 /* All tasks on this runqueue were pinned by CPU affinity */
3499 if (unlikely(all_pinned)) {
3500 cpu_clear(cpu_of(busiest), *cpus);
3501 if (!cpus_empty(*cpus))
3508 schedstat_inc(sd, lb_failed[idle]);
3509 sd->nr_balance_failed++;
3511 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3513 spin_lock_irqsave(&busiest->lock, flags);
3515 /* don't kick the migration_thread, if the curr
3516 * task on busiest cpu can't be moved to this_cpu
3518 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3519 spin_unlock_irqrestore(&busiest->lock, flags);
3521 goto out_one_pinned;
3524 if (!busiest->active_balance) {
3525 busiest->active_balance = 1;
3526 busiest->push_cpu = this_cpu;
3529 spin_unlock_irqrestore(&busiest->lock, flags);
3531 wake_up_process(busiest->migration_thread);
3534 * We've kicked active balancing, reset the failure
3537 sd->nr_balance_failed = sd->cache_nice_tries+1;
3540 sd->nr_balance_failed = 0;
3542 if (likely(!active_balance)) {
3543 /* We were unbalanced, so reset the balancing interval */
3544 sd->balance_interval = sd->min_interval;
3547 * If we've begun active balancing, start to back off. This
3548 * case may not be covered by the all_pinned logic if there
3549 * is only 1 task on the busy runqueue (because we don't call
3552 if (sd->balance_interval < sd->max_interval)
3553 sd->balance_interval *= 2;
3556 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3557 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3563 schedstat_inc(sd, lb_balanced[idle]);
3565 sd->nr_balance_failed = 0;
3568 /* tune up the balancing interval */
3569 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3570 (sd->balance_interval < sd->max_interval))
3571 sd->balance_interval *= 2;
3573 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3574 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3585 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3586 * tasks if there is an imbalance.
3588 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3589 * this_rq is locked.
3592 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3595 struct sched_group *group;
3596 struct rq *busiest = NULL;
3597 unsigned long imbalance;
3605 * When power savings policy is enabled for the parent domain, idle
3606 * sibling can pick up load irrespective of busy siblings. In this case,
3607 * let the state of idle sibling percolate up as IDLE, instead of
3608 * portraying it as CPU_NOT_IDLE.
3610 if (sd->flags & SD_SHARE_CPUPOWER &&
3611 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3614 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3616 update_shares_locked(this_rq, sd);
3617 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3618 &sd_idle, cpus, NULL);
3620 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3624 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3626 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3630 BUG_ON(busiest == this_rq);
3632 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3635 if (busiest->nr_running > 1) {
3636 /* Attempt to move tasks */
3637 double_lock_balance(this_rq, busiest);
3638 /* this_rq->clock is already updated */
3639 update_rq_clock(busiest);
3640 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3641 imbalance, sd, CPU_NEWLY_IDLE,
3643 spin_unlock(&busiest->lock);
3645 if (unlikely(all_pinned)) {
3646 cpu_clear(cpu_of(busiest), *cpus);
3647 if (!cpus_empty(*cpus))
3653 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3654 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3655 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3658 sd->nr_balance_failed = 0;
3660 update_shares_locked(this_rq, sd);
3664 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3665 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3666 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3668 sd->nr_balance_failed = 0;
3674 * idle_balance is called by schedule() if this_cpu is about to become
3675 * idle. Attempts to pull tasks from other CPUs.
3677 static void idle_balance(int this_cpu, struct rq *this_rq)
3679 struct sched_domain *sd;
3680 int pulled_task = -1;
3681 unsigned long next_balance = jiffies + HZ;
3684 for_each_domain(this_cpu, sd) {
3685 unsigned long interval;
3687 if (!(sd->flags & SD_LOAD_BALANCE))
3690 if (sd->flags & SD_BALANCE_NEWIDLE)
3691 /* If we've pulled tasks over stop searching: */
3692 pulled_task = load_balance_newidle(this_cpu, this_rq,
3695 interval = msecs_to_jiffies(sd->balance_interval);
3696 if (time_after(next_balance, sd->last_balance + interval))
3697 next_balance = sd->last_balance + interval;
3701 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3703 * We are going idle. next_balance may be set based on
3704 * a busy processor. So reset next_balance.
3706 this_rq->next_balance = next_balance;
3711 * active_load_balance is run by migration threads. It pushes running tasks
3712 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3713 * running on each physical CPU where possible, and avoids physical /
3714 * logical imbalances.
3716 * Called with busiest_rq locked.
3718 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3720 int target_cpu = busiest_rq->push_cpu;
3721 struct sched_domain *sd;
3722 struct rq *target_rq;
3724 /* Is there any task to move? */
3725 if (busiest_rq->nr_running <= 1)
3728 target_rq = cpu_rq(target_cpu);
3731 * This condition is "impossible", if it occurs
3732 * we need to fix it. Originally reported by
3733 * Bjorn Helgaas on a 128-cpu setup.
3735 BUG_ON(busiest_rq == target_rq);
3737 /* move a task from busiest_rq to target_rq */
3738 double_lock_balance(busiest_rq, target_rq);
3739 update_rq_clock(busiest_rq);
3740 update_rq_clock(target_rq);
3742 /* Search for an sd spanning us and the target CPU. */
3743 for_each_domain(target_cpu, sd) {
3744 if ((sd->flags & SD_LOAD_BALANCE) &&
3745 cpu_isset(busiest_cpu, sd->span))
3750 schedstat_inc(sd, alb_count);
3752 if (move_one_task(target_rq, target_cpu, busiest_rq,
3754 schedstat_inc(sd, alb_pushed);
3756 schedstat_inc(sd, alb_failed);
3758 spin_unlock(&target_rq->lock);
3763 atomic_t load_balancer;
3765 } nohz ____cacheline_aligned = {
3766 .load_balancer = ATOMIC_INIT(-1),
3767 .cpu_mask = CPU_MASK_NONE,
3771 * This routine will try to nominate the ilb (idle load balancing)
3772 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3773 * load balancing on behalf of all those cpus. If all the cpus in the system
3774 * go into this tickless mode, then there will be no ilb owner (as there is
3775 * no need for one) and all the cpus will sleep till the next wakeup event
3778 * For the ilb owner, tick is not stopped. And this tick will be used
3779 * for idle load balancing. ilb owner will still be part of
3782 * While stopping the tick, this cpu will become the ilb owner if there
3783 * is no other owner. And will be the owner till that cpu becomes busy
3784 * or if all cpus in the system stop their ticks at which point
3785 * there is no need for ilb owner.
3787 * When the ilb owner becomes busy, it nominates another owner, during the
3788 * next busy scheduler_tick()
3790 int select_nohz_load_balancer(int stop_tick)
3792 int cpu = smp_processor_id();
3795 cpu_set(cpu, nohz.cpu_mask);
3796 cpu_rq(cpu)->in_nohz_recently = 1;
3799 * If we are going offline and still the leader, give up!
3801 if (cpu_is_offline(cpu) &&
3802 atomic_read(&nohz.load_balancer) == cpu) {
3803 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3808 /* time for ilb owner also to sleep */
3809 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3810 if (atomic_read(&nohz.load_balancer) == cpu)
3811 atomic_set(&nohz.load_balancer, -1);
3815 if (atomic_read(&nohz.load_balancer) == -1) {
3816 /* make me the ilb owner */
3817 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3819 } else if (atomic_read(&nohz.load_balancer) == cpu)
3822 if (!cpu_isset(cpu, nohz.cpu_mask))
3825 cpu_clear(cpu, nohz.cpu_mask);
3827 if (atomic_read(&nohz.load_balancer) == cpu)
3828 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3835 static DEFINE_SPINLOCK(balancing);
3838 * It checks each scheduling domain to see if it is due to be balanced,
3839 * and initiates a balancing operation if so.
3841 * Balancing parameters are set up in arch_init_sched_domains.
3843 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3846 struct rq *rq = cpu_rq(cpu);
3847 unsigned long interval;
3848 struct sched_domain *sd;
3849 /* Earliest time when we have to do rebalance again */
3850 unsigned long next_balance = jiffies + 60*HZ;
3851 int update_next_balance = 0;
3855 for_each_domain(cpu, sd) {
3856 if (!(sd->flags & SD_LOAD_BALANCE))
3859 interval = sd->balance_interval;
3860 if (idle != CPU_IDLE)
3861 interval *= sd->busy_factor;
3863 /* scale ms to jiffies */
3864 interval = msecs_to_jiffies(interval);
3865 if (unlikely(!interval))
3867 if (interval > HZ*NR_CPUS/10)
3868 interval = HZ*NR_CPUS/10;
3870 need_serialize = sd->flags & SD_SERIALIZE;
3872 if (need_serialize) {
3873 if (!spin_trylock(&balancing))
3877 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3878 if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) {
3880 * We've pulled tasks over so either we're no
3881 * longer idle, or one of our SMT siblings is
3884 idle = CPU_NOT_IDLE;
3886 sd->last_balance = jiffies;
3889 spin_unlock(&balancing);
3891 if (time_after(next_balance, sd->last_balance + interval)) {
3892 next_balance = sd->last_balance + interval;
3893 update_next_balance = 1;
3897 * Stop the load balance at this level. There is another
3898 * CPU in our sched group which is doing load balancing more
3906 * next_balance will be updated only when there is a need.
3907 * When the cpu is attached to null domain for ex, it will not be
3910 if (likely(update_next_balance))
3911 rq->next_balance = next_balance;
3915 * run_rebalance_domains is triggered when needed from the scheduler tick.
3916 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3917 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3919 static void run_rebalance_domains(struct softirq_action *h)
3921 int this_cpu = smp_processor_id();
3922 struct rq *this_rq = cpu_rq(this_cpu);
3923 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3924 CPU_IDLE : CPU_NOT_IDLE;
3926 rebalance_domains(this_cpu, idle);
3930 * If this cpu is the owner for idle load balancing, then do the
3931 * balancing on behalf of the other idle cpus whose ticks are
3934 if (this_rq->idle_at_tick &&
3935 atomic_read(&nohz.load_balancer) == this_cpu) {
3936 cpumask_t cpus = nohz.cpu_mask;
3940 cpu_clear(this_cpu, cpus);
3941 for_each_cpu_mask(balance_cpu, cpus) {
3943 * If this cpu gets work to do, stop the load balancing
3944 * work being done for other cpus. Next load
3945 * balancing owner will pick it up.
3950 rebalance_domains(balance_cpu, CPU_IDLE);
3952 rq = cpu_rq(balance_cpu);
3953 if (time_after(this_rq->next_balance, rq->next_balance))
3954 this_rq->next_balance = rq->next_balance;
3961 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3963 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3964 * idle load balancing owner or decide to stop the periodic load balancing,
3965 * if the whole system is idle.
3967 static inline void trigger_load_balance(struct rq *rq, int cpu)
3971 * If we were in the nohz mode recently and busy at the current
3972 * scheduler tick, then check if we need to nominate new idle
3975 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3976 rq->in_nohz_recently = 0;
3978 if (atomic_read(&nohz.load_balancer) == cpu) {
3979 cpu_clear(cpu, nohz.cpu_mask);
3980 atomic_set(&nohz.load_balancer, -1);
3983 if (atomic_read(&nohz.load_balancer) == -1) {
3985 * simple selection for now: Nominate the
3986 * first cpu in the nohz list to be the next
3989 * TBD: Traverse the sched domains and nominate
3990 * the nearest cpu in the nohz.cpu_mask.
3992 int ilb = first_cpu(nohz.cpu_mask);
3994 if (ilb < nr_cpu_ids)
4000 * If this cpu is idle and doing idle load balancing for all the
4001 * cpus with ticks stopped, is it time for that to stop?
4003 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4004 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
4010 * If this cpu is idle and the idle load balancing is done by
4011 * someone else, then no need raise the SCHED_SOFTIRQ
4013 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4014 cpu_isset(cpu, nohz.cpu_mask))
4017 if (time_after_eq(jiffies, rq->next_balance))
4018 raise_softirq(SCHED_SOFTIRQ);
4021 #else /* CONFIG_SMP */
4024 * on UP we do not need to balance between CPUs:
4026 static inline void idle_balance(int cpu, struct rq *rq)
4032 DEFINE_PER_CPU(struct kernel_stat, kstat);
4034 EXPORT_PER_CPU_SYMBOL(kstat);
4037 * Return p->sum_exec_runtime plus any more ns on the sched_clock
4038 * that have not yet been banked in case the task is currently running.
4040 unsigned long long task_sched_runtime(struct task_struct *p)
4042 unsigned long flags;
4046 rq = task_rq_lock(p, &flags);
4047 ns = p->se.sum_exec_runtime;
4048 if (task_current(rq, p)) {
4049 update_rq_clock(rq);
4050 delta_exec = rq->clock - p->se.exec_start;
4051 if ((s64)delta_exec > 0)
4054 task_rq_unlock(rq, &flags);
4060 * Account user cpu time to a process.
4061 * @p: the process that the cpu time gets accounted to
4062 * @cputime: the cpu time spent in user space since the last update
4064 void account_user_time(struct task_struct *p, cputime_t cputime)
4066 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4069 p->utime = cputime_add(p->utime, cputime);
4071 /* Add user time to cpustat. */
4072 tmp = cputime_to_cputime64(cputime);
4073 if (TASK_NICE(p) > 0)
4074 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4076 cpustat->user = cputime64_add(cpustat->user, tmp);
4080 * Account guest cpu time to a process.
4081 * @p: the process that the cpu time gets accounted to
4082 * @cputime: the cpu time spent in virtual machine since the last update
4084 static void account_guest_time(struct task_struct *p, cputime_t cputime)
4087 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4089 tmp = cputime_to_cputime64(cputime);
4091 p->utime = cputime_add(p->utime, cputime);
4092 p->gtime = cputime_add(p->gtime, cputime);
4094 cpustat->user = cputime64_add(cpustat->user, tmp);
4095 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4099 * Account scaled user cpu time to a process.
4100 * @p: the process that the cpu time gets accounted to
4101 * @cputime: the cpu time spent in user space since the last update
4103 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
4105 p->utimescaled = cputime_add(p->utimescaled, cputime);
4109 * Account system cpu time to a process.
4110 * @p: the process that the cpu time gets accounted to
4111 * @hardirq_offset: the offset to subtract from hardirq_count()
4112 * @cputime: the cpu time spent in kernel space since the last update
4114 void account_system_time(struct task_struct *p, int hardirq_offset,
4117 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4118 struct rq *rq = this_rq();
4121 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4122 account_guest_time(p, cputime);
4126 p->stime = cputime_add(p->stime, cputime);
4128 /* Add system time to cpustat. */
4129 tmp = cputime_to_cputime64(cputime);
4130 if (hardirq_count() - hardirq_offset)
4131 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4132 else if (softirq_count())
4133 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4134 else if (p != rq->idle)
4135 cpustat->system = cputime64_add(cpustat->system, tmp);
4136 else if (atomic_read(&rq->nr_iowait) > 0)
4137 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4139 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4140 /* Account for system time used */
4141 acct_update_integrals(p);
4145 * Account scaled system cpu time to a process.
4146 * @p: the process that the cpu time gets accounted to
4147 * @hardirq_offset: the offset to subtract from hardirq_count()
4148 * @cputime: the cpu time spent in kernel space since the last update
4150 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
4152 p->stimescaled = cputime_add(p->stimescaled, cputime);
4156 * Account for involuntary wait time.
4157 * @p: the process from which the cpu time has been stolen
4158 * @steal: the cpu time spent in involuntary wait
4160 void account_steal_time(struct task_struct *p, cputime_t steal)
4162 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4163 cputime64_t tmp = cputime_to_cputime64(steal);
4164 struct rq *rq = this_rq();
4166 if (p == rq->idle) {
4167 p->stime = cputime_add(p->stime, steal);
4168 if (atomic_read(&rq->nr_iowait) > 0)
4169 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4171 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4173 cpustat->steal = cputime64_add(cpustat->steal, tmp);
4177 * This function gets called by the timer code, with HZ frequency.
4178 * We call it with interrupts disabled.
4180 * It also gets called by the fork code, when changing the parent's
4183 void scheduler_tick(void)
4185 int cpu = smp_processor_id();
4186 struct rq *rq = cpu_rq(cpu);
4187 struct task_struct *curr = rq->curr;
4191 spin_lock(&rq->lock);
4192 update_rq_clock(rq);
4193 update_cpu_load(rq);
4194 curr->sched_class->task_tick(rq, curr, 0);
4195 spin_unlock(&rq->lock);
4198 rq->idle_at_tick = idle_cpu(cpu);
4199 trigger_load_balance(rq, cpu);
4203 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
4205 void __kprobes add_preempt_count(int val)
4210 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4212 preempt_count() += val;
4214 * Spinlock count overflowing soon?
4216 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4219 EXPORT_SYMBOL(add_preempt_count);
4221 void __kprobes sub_preempt_count(int val)
4226 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4229 * Is the spinlock portion underflowing?
4231 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4232 !(preempt_count() & PREEMPT_MASK)))
4235 preempt_count() -= val;
4237 EXPORT_SYMBOL(sub_preempt_count);
4242 * Print scheduling while atomic bug:
4244 static noinline void __schedule_bug(struct task_struct *prev)
4246 struct pt_regs *regs = get_irq_regs();
4248 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4249 prev->comm, prev->pid, preempt_count());
4251 debug_show_held_locks(prev);
4253 if (irqs_disabled())
4254 print_irqtrace_events(prev);
4263 * Various schedule()-time debugging checks and statistics:
4265 static inline void schedule_debug(struct task_struct *prev)
4268 * Test if we are atomic. Since do_exit() needs to call into
4269 * schedule() atomically, we ignore that path for now.
4270 * Otherwise, whine if we are scheduling when we should not be.
4272 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4273 __schedule_bug(prev);
4275 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4277 schedstat_inc(this_rq(), sched_count);
4278 #ifdef CONFIG_SCHEDSTATS
4279 if (unlikely(prev->lock_depth >= 0)) {
4280 schedstat_inc(this_rq(), bkl_count);
4281 schedstat_inc(prev, sched_info.bkl_count);
4287 * Pick up the highest-prio task:
4289 static inline struct task_struct *
4290 pick_next_task(struct rq *rq, struct task_struct *prev)
4292 const struct sched_class *class;
4293 struct task_struct *p;
4296 * Optimization: we know that if all tasks are in
4297 * the fair class we can call that function directly:
4299 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4300 p = fair_sched_class.pick_next_task(rq);
4305 class = sched_class_highest;
4307 p = class->pick_next_task(rq);
4311 * Will never be NULL as the idle class always
4312 * returns a non-NULL p:
4314 class = class->next;
4319 * schedule() is the main scheduler function.
4321 asmlinkage void __sched schedule(void)
4323 struct task_struct *prev, *next;
4324 unsigned long *switch_count;
4326 int cpu, hrtick = sched_feat(HRTICK);
4330 cpu = smp_processor_id();
4334 switch_count = &prev->nivcsw;
4336 release_kernel_lock(prev);
4337 need_resched_nonpreemptible:
4339 schedule_debug(prev);
4345 * Do the rq-clock update outside the rq lock:
4347 local_irq_disable();
4348 update_rq_clock(rq);
4349 spin_lock(&rq->lock);
4350 clear_tsk_need_resched(prev);
4352 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4353 if (unlikely(signal_pending_state(prev->state, prev)))
4354 prev->state = TASK_RUNNING;
4356 deactivate_task(rq, prev, 1);
4357 switch_count = &prev->nvcsw;
4361 if (prev->sched_class->pre_schedule)
4362 prev->sched_class->pre_schedule(rq, prev);
4365 if (unlikely(!rq->nr_running))
4366 idle_balance(cpu, rq);
4368 prev->sched_class->put_prev_task(rq, prev);
4369 next = pick_next_task(rq, prev);
4371 if (likely(prev != next)) {
4372 sched_info_switch(prev, next);
4378 context_switch(rq, prev, next); /* unlocks the rq */
4380 * the context switch might have flipped the stack from under
4381 * us, hence refresh the local variables.
4383 cpu = smp_processor_id();
4386 spin_unlock_irq(&rq->lock);
4391 if (unlikely(reacquire_kernel_lock(current) < 0))
4392 goto need_resched_nonpreemptible;
4394 preempt_enable_no_resched();
4395 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4398 EXPORT_SYMBOL(schedule);
4400 #ifdef CONFIG_PREEMPT
4402 * this is the entry point to schedule() from in-kernel preemption
4403 * off of preempt_enable. Kernel preemptions off return from interrupt
4404 * occur there and call schedule directly.
4406 asmlinkage void __sched preempt_schedule(void)
4408 struct thread_info *ti = current_thread_info();
4411 * If there is a non-zero preempt_count or interrupts are disabled,
4412 * we do not want to preempt the current task. Just return..
4414 if (likely(ti->preempt_count || irqs_disabled()))
4418 add_preempt_count(PREEMPT_ACTIVE);
4420 sub_preempt_count(PREEMPT_ACTIVE);
4423 * Check again in case we missed a preemption opportunity
4424 * between schedule and now.
4427 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4429 EXPORT_SYMBOL(preempt_schedule);
4432 * this is the entry point to schedule() from kernel preemption
4433 * off of irq context.
4434 * Note, that this is called and return with irqs disabled. This will
4435 * protect us against recursive calling from irq.
4437 asmlinkage void __sched preempt_schedule_irq(void)
4439 struct thread_info *ti = current_thread_info();
4441 /* Catch callers which need to be fixed */
4442 BUG_ON(ti->preempt_count || !irqs_disabled());
4445 add_preempt_count(PREEMPT_ACTIVE);
4448 local_irq_disable();
4449 sub_preempt_count(PREEMPT_ACTIVE);
4452 * Check again in case we missed a preemption opportunity
4453 * between schedule and now.
4456 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4459 #endif /* CONFIG_PREEMPT */
4461 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4464 return try_to_wake_up(curr->private, mode, sync);
4466 EXPORT_SYMBOL(default_wake_function);
4469 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4470 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4471 * number) then we wake all the non-exclusive tasks and one exclusive task.
4473 * There are circumstances in which we can try to wake a task which has already
4474 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4475 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4477 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4478 int nr_exclusive, int sync, void *key)
4480 wait_queue_t *curr, *next;
4482 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4483 unsigned flags = curr->flags;
4485 if (curr->func(curr, mode, sync, key) &&
4486 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4492 * __wake_up - wake up threads blocked on a waitqueue.
4494 * @mode: which threads
4495 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4496 * @key: is directly passed to the wakeup function
4498 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4499 int nr_exclusive, void *key)
4501 unsigned long flags;
4503 spin_lock_irqsave(&q->lock, flags);
4504 __wake_up_common(q, mode, nr_exclusive, 0, key);
4505 spin_unlock_irqrestore(&q->lock, flags);
4507 EXPORT_SYMBOL(__wake_up);
4510 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4512 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4514 __wake_up_common(q, mode, 1, 0, NULL);
4518 * __wake_up_sync - wake up threads blocked on a waitqueue.
4520 * @mode: which threads
4521 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4523 * The sync wakeup differs that the waker knows that it will schedule
4524 * away soon, so while the target thread will be woken up, it will not
4525 * be migrated to another CPU - ie. the two threads are 'synchronized'
4526 * with each other. This can prevent needless bouncing between CPUs.
4528 * On UP it can prevent extra preemption.
4531 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4533 unsigned long flags;
4539 if (unlikely(!nr_exclusive))
4542 spin_lock_irqsave(&q->lock, flags);
4543 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4544 spin_unlock_irqrestore(&q->lock, flags);
4546 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4548 void complete(struct completion *x)
4550 unsigned long flags;
4552 spin_lock_irqsave(&x->wait.lock, flags);
4554 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4555 spin_unlock_irqrestore(&x->wait.lock, flags);
4557 EXPORT_SYMBOL(complete);
4559 void complete_all(struct completion *x)
4561 unsigned long flags;
4563 spin_lock_irqsave(&x->wait.lock, flags);
4564 x->done += UINT_MAX/2;
4565 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4566 spin_unlock_irqrestore(&x->wait.lock, flags);
4568 EXPORT_SYMBOL(complete_all);
4570 static inline long __sched
4571 do_wait_for_common(struct completion *x, long timeout, int state)
4574 DECLARE_WAITQUEUE(wait, current);
4576 wait.flags |= WQ_FLAG_EXCLUSIVE;
4577 __add_wait_queue_tail(&x->wait, &wait);
4579 if ((state == TASK_INTERRUPTIBLE &&
4580 signal_pending(current)) ||
4581 (state == TASK_KILLABLE &&
4582 fatal_signal_pending(current))) {
4583 timeout = -ERESTARTSYS;
4586 __set_current_state(state);
4587 spin_unlock_irq(&x->wait.lock);
4588 timeout = schedule_timeout(timeout);
4589 spin_lock_irq(&x->wait.lock);
4590 } while (!x->done && timeout);
4591 __remove_wait_queue(&x->wait, &wait);
4596 return timeout ?: 1;
4600 wait_for_common(struct completion *x, long timeout, int state)
4604 spin_lock_irq(&x->wait.lock);
4605 timeout = do_wait_for_common(x, timeout, state);
4606 spin_unlock_irq(&x->wait.lock);
4610 void __sched wait_for_completion(struct completion *x)
4612 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4614 EXPORT_SYMBOL(wait_for_completion);
4616 unsigned long __sched
4617 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4619 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4621 EXPORT_SYMBOL(wait_for_completion_timeout);
4623 int __sched wait_for_completion_interruptible(struct completion *x)
4625 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4626 if (t == -ERESTARTSYS)
4630 EXPORT_SYMBOL(wait_for_completion_interruptible);
4632 unsigned long __sched
4633 wait_for_completion_interruptible_timeout(struct completion *x,
4634 unsigned long timeout)
4636 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4638 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4640 int __sched wait_for_completion_killable(struct completion *x)
4642 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4643 if (t == -ERESTARTSYS)
4647 EXPORT_SYMBOL(wait_for_completion_killable);
4650 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4652 unsigned long flags;
4655 init_waitqueue_entry(&wait, current);
4657 __set_current_state(state);
4659 spin_lock_irqsave(&q->lock, flags);
4660 __add_wait_queue(q, &wait);
4661 spin_unlock(&q->lock);
4662 timeout = schedule_timeout(timeout);
4663 spin_lock_irq(&q->lock);
4664 __remove_wait_queue(q, &wait);
4665 spin_unlock_irqrestore(&q->lock, flags);
4670 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4672 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4674 EXPORT_SYMBOL(interruptible_sleep_on);
4677 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4679 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4681 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4683 void __sched sleep_on(wait_queue_head_t *q)
4685 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4687 EXPORT_SYMBOL(sleep_on);
4689 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4691 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4693 EXPORT_SYMBOL(sleep_on_timeout);
4695 #ifdef CONFIG_RT_MUTEXES
4698 * rt_mutex_setprio - set the current priority of a task
4700 * @prio: prio value (kernel-internal form)
4702 * This function changes the 'effective' priority of a task. It does
4703 * not touch ->normal_prio like __setscheduler().
4705 * Used by the rt_mutex code to implement priority inheritance logic.
4707 void rt_mutex_setprio(struct task_struct *p, int prio)
4709 unsigned long flags;
4710 int oldprio, on_rq, running;
4712 const struct sched_class *prev_class = p->sched_class;
4714 BUG_ON(prio < 0 || prio > MAX_PRIO);
4716 rq = task_rq_lock(p, &flags);
4717 update_rq_clock(rq);
4720 on_rq = p->se.on_rq;
4721 running = task_current(rq, p);
4723 dequeue_task(rq, p, 0);
4725 p->sched_class->put_prev_task(rq, p);
4728 p->sched_class = &rt_sched_class;
4730 p->sched_class = &fair_sched_class;
4735 p->sched_class->set_curr_task(rq);
4737 enqueue_task(rq, p, 0);
4739 check_class_changed(rq, p, prev_class, oldprio, running);
4741 task_rq_unlock(rq, &flags);
4746 void set_user_nice(struct task_struct *p, long nice)
4748 int old_prio, delta, on_rq;
4749 unsigned long flags;
4752 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4755 * We have to be careful, if called from sys_setpriority(),
4756 * the task might be in the middle of scheduling on another CPU.
4758 rq = task_rq_lock(p, &flags);
4759 update_rq_clock(rq);
4761 * The RT priorities are set via sched_setscheduler(), but we still
4762 * allow the 'normal' nice value to be set - but as expected
4763 * it wont have any effect on scheduling until the task is
4764 * SCHED_FIFO/SCHED_RR:
4766 if (task_has_rt_policy(p)) {
4767 p->static_prio = NICE_TO_PRIO(nice);
4770 on_rq = p->se.on_rq;
4772 dequeue_task(rq, p, 0);
4774 p->static_prio = NICE_TO_PRIO(nice);
4777 p->prio = effective_prio(p);
4778 delta = p->prio - old_prio;
4781 enqueue_task(rq, p, 0);
4783 * If the task increased its priority or is running and
4784 * lowered its priority, then reschedule its CPU:
4786 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4787 resched_task(rq->curr);
4790 task_rq_unlock(rq, &flags);
4792 EXPORT_SYMBOL(set_user_nice);
4795 * can_nice - check if a task can reduce its nice value
4799 int can_nice(const struct task_struct *p, const int nice)
4801 /* convert nice value [19,-20] to rlimit style value [1,40] */
4802 int nice_rlim = 20 - nice;
4804 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4805 capable(CAP_SYS_NICE));
4808 #ifdef __ARCH_WANT_SYS_NICE
4811 * sys_nice - change the priority of the current process.
4812 * @increment: priority increment
4814 * sys_setpriority is a more generic, but much slower function that
4815 * does similar things.
4817 asmlinkage long sys_nice(int increment)
4822 * Setpriority might change our priority at the same moment.
4823 * We don't have to worry. Conceptually one call occurs first
4824 * and we have a single winner.
4826 if (increment < -40)
4831 nice = PRIO_TO_NICE(current->static_prio) + increment;
4837 if (increment < 0 && !can_nice(current, nice))
4840 retval = security_task_setnice(current, nice);
4844 set_user_nice(current, nice);
4851 * task_prio - return the priority value of a given task.
4852 * @p: the task in question.
4854 * This is the priority value as seen by users in /proc.
4855 * RT tasks are offset by -200. Normal tasks are centered
4856 * around 0, value goes from -16 to +15.
4858 int task_prio(const struct task_struct *p)
4860 return p->prio - MAX_RT_PRIO;
4864 * task_nice - return the nice value of a given task.
4865 * @p: the task in question.
4867 int task_nice(const struct task_struct *p)
4869 return TASK_NICE(p);
4871 EXPORT_SYMBOL(task_nice);
4874 * idle_cpu - is a given cpu idle currently?
4875 * @cpu: the processor in question.
4877 int idle_cpu(int cpu)
4879 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4883 * idle_task - return the idle task for a given cpu.
4884 * @cpu: the processor in question.
4886 struct task_struct *idle_task(int cpu)
4888 return cpu_rq(cpu)->idle;
4892 * find_process_by_pid - find a process with a matching PID value.
4893 * @pid: the pid in question.
4895 static struct task_struct *find_process_by_pid(pid_t pid)
4897 return pid ? find_task_by_vpid(pid) : current;
4900 /* Actually do priority change: must hold rq lock. */
4902 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4904 BUG_ON(p->se.on_rq);
4907 switch (p->policy) {
4911 p->sched_class = &fair_sched_class;
4915 p->sched_class = &rt_sched_class;
4919 p->rt_priority = prio;
4920 p->normal_prio = normal_prio(p);
4921 /* we are holding p->pi_lock already */
4922 p->prio = rt_mutex_getprio(p);
4927 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4928 * @p: the task in question.
4929 * @policy: new policy.
4930 * @param: structure containing the new RT priority.
4932 * NOTE that the task may be already dead.
4934 int sched_setscheduler(struct task_struct *p, int policy,
4935 struct sched_param *param)
4937 int retval, oldprio, oldpolicy = -1, on_rq, running;
4938 unsigned long flags;
4939 const struct sched_class *prev_class = p->sched_class;
4942 /* may grab non-irq protected spin_locks */
4943 BUG_ON(in_interrupt());
4945 /* double check policy once rq lock held */
4947 policy = oldpolicy = p->policy;
4948 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4949 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4950 policy != SCHED_IDLE)
4953 * Valid priorities for SCHED_FIFO and SCHED_RR are
4954 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4955 * SCHED_BATCH and SCHED_IDLE is 0.
4957 if (param->sched_priority < 0 ||
4958 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4959 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4961 if (rt_policy(policy) != (param->sched_priority != 0))
4965 * Allow unprivileged RT tasks to decrease priority:
4967 if (!capable(CAP_SYS_NICE)) {
4968 if (rt_policy(policy)) {
4969 unsigned long rlim_rtprio;
4971 if (!lock_task_sighand(p, &flags))
4973 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4974 unlock_task_sighand(p, &flags);
4976 /* can't set/change the rt policy */
4977 if (policy != p->policy && !rlim_rtprio)
4980 /* can't increase priority */
4981 if (param->sched_priority > p->rt_priority &&
4982 param->sched_priority > rlim_rtprio)
4986 * Like positive nice levels, dont allow tasks to
4987 * move out of SCHED_IDLE either:
4989 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4992 /* can't change other user's priorities */
4993 if ((current->euid != p->euid) &&
4994 (current->euid != p->uid))
4998 #ifdef CONFIG_RT_GROUP_SCHED
5000 * Do not allow realtime tasks into groups that have no runtime
5003 if (rt_policy(policy) && task_group(p)->rt_bandwidth.rt_runtime == 0)
5007 retval = security_task_setscheduler(p, policy, param);
5011 * make sure no PI-waiters arrive (or leave) while we are
5012 * changing the priority of the task:
5014 spin_lock_irqsave(&p->pi_lock, flags);
5016 * To be able to change p->policy safely, the apropriate
5017 * runqueue lock must be held.
5019 rq = __task_rq_lock(p);
5020 /* recheck policy now with rq lock held */
5021 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5022 policy = oldpolicy = -1;
5023 __task_rq_unlock(rq);
5024 spin_unlock_irqrestore(&p->pi_lock, flags);
5027 update_rq_clock(rq);
5028 on_rq = p->se.on_rq;
5029 running = task_current(rq, p);
5031 deactivate_task(rq, p, 0);
5033 p->sched_class->put_prev_task(rq, p);
5036 __setscheduler(rq, p, policy, param->sched_priority);
5039 p->sched_class->set_curr_task(rq);
5041 activate_task(rq, p, 0);
5043 check_class_changed(rq, p, prev_class, oldprio, running);
5045 __task_rq_unlock(rq);
5046 spin_unlock_irqrestore(&p->pi_lock, flags);
5048 rt_mutex_adjust_pi(p);
5052 EXPORT_SYMBOL_GPL(sched_setscheduler);
5055 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5057 struct sched_param lparam;
5058 struct task_struct *p;
5061 if (!param || pid < 0)
5063 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5068 p = find_process_by_pid(pid);
5070 retval = sched_setscheduler(p, policy, &lparam);
5077 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5078 * @pid: the pid in question.
5079 * @policy: new policy.
5080 * @param: structure containing the new RT priority.
5083 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5085 /* negative values for policy are not valid */
5089 return do_sched_setscheduler(pid, policy, param);
5093 * sys_sched_setparam - set/change the RT priority of a thread
5094 * @pid: the pid in question.
5095 * @param: structure containing the new RT priority.
5097 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
5099 return do_sched_setscheduler(pid, -1, param);
5103 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5104 * @pid: the pid in question.
5106 asmlinkage long sys_sched_getscheduler(pid_t pid)
5108 struct task_struct *p;
5115 read_lock(&tasklist_lock);
5116 p = find_process_by_pid(pid);
5118 retval = security_task_getscheduler(p);
5122 read_unlock(&tasklist_lock);
5127 * sys_sched_getscheduler - get the RT priority of a thread
5128 * @pid: the pid in question.
5129 * @param: structure containing the RT priority.
5131 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
5133 struct sched_param lp;
5134 struct task_struct *p;
5137 if (!param || pid < 0)
5140 read_lock(&tasklist_lock);
5141 p = find_process_by_pid(pid);
5146 retval = security_task_getscheduler(p);
5150 lp.sched_priority = p->rt_priority;
5151 read_unlock(&tasklist_lock);
5154 * This one might sleep, we cannot do it with a spinlock held ...
5156 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5161 read_unlock(&tasklist_lock);
5165 long sched_setaffinity(pid_t pid, const cpumask_t *in_mask)
5167 cpumask_t cpus_allowed;
5168 cpumask_t new_mask = *in_mask;
5169 struct task_struct *p;
5173 read_lock(&tasklist_lock);
5175 p = find_process_by_pid(pid);
5177 read_unlock(&tasklist_lock);
5183 * It is not safe to call set_cpus_allowed with the
5184 * tasklist_lock held. We will bump the task_struct's
5185 * usage count and then drop tasklist_lock.
5188 read_unlock(&tasklist_lock);
5191 if ((current->euid != p->euid) && (current->euid != p->uid) &&
5192 !capable(CAP_SYS_NICE))
5195 retval = security_task_setscheduler(p, 0, NULL);
5199 cpuset_cpus_allowed(p, &cpus_allowed);
5200 cpus_and(new_mask, new_mask, cpus_allowed);
5202 retval = set_cpus_allowed_ptr(p, &new_mask);
5205 cpuset_cpus_allowed(p, &cpus_allowed);
5206 if (!cpus_subset(new_mask, cpus_allowed)) {
5208 * We must have raced with a concurrent cpuset
5209 * update. Just reset the cpus_allowed to the
5210 * cpuset's cpus_allowed
5212 new_mask = cpus_allowed;
5222 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5223 cpumask_t *new_mask)
5225 if (len < sizeof(cpumask_t)) {
5226 memset(new_mask, 0, sizeof(cpumask_t));
5227 } else if (len > sizeof(cpumask_t)) {
5228 len = sizeof(cpumask_t);
5230 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5234 * sys_sched_setaffinity - set the cpu affinity of a process
5235 * @pid: pid of the process
5236 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5237 * @user_mask_ptr: user-space pointer to the new cpu mask
5239 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
5240 unsigned long __user *user_mask_ptr)
5245 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
5249 return sched_setaffinity(pid, &new_mask);
5252 long sched_getaffinity(pid_t pid, cpumask_t *mask)
5254 struct task_struct *p;
5258 read_lock(&tasklist_lock);
5261 p = find_process_by_pid(pid);
5265 retval = security_task_getscheduler(p);
5269 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
5272 read_unlock(&tasklist_lock);
5279 * sys_sched_getaffinity - get the cpu affinity of a process
5280 * @pid: pid of the process
5281 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5282 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5284 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5285 unsigned long __user *user_mask_ptr)
5290 if (len < sizeof(cpumask_t))
5293 ret = sched_getaffinity(pid, &mask);
5297 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5300 return sizeof(cpumask_t);
5304 * sys_sched_yield - yield the current processor to other threads.
5306 * This function yields the current CPU to other tasks. If there are no
5307 * other threads running on this CPU then this function will return.
5309 asmlinkage long sys_sched_yield(void)
5311 struct rq *rq = this_rq_lock();
5313 schedstat_inc(rq, yld_count);
5314 current->sched_class->yield_task(rq);
5317 * Since we are going to call schedule() anyway, there's
5318 * no need to preempt or enable interrupts:
5320 __release(rq->lock);
5321 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5322 _raw_spin_unlock(&rq->lock);
5323 preempt_enable_no_resched();
5330 static void __cond_resched(void)
5332 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5333 __might_sleep(__FILE__, __LINE__);
5336 * The BKS might be reacquired before we have dropped
5337 * PREEMPT_ACTIVE, which could trigger a second
5338 * cond_resched() call.
5341 add_preempt_count(PREEMPT_ACTIVE);
5343 sub_preempt_count(PREEMPT_ACTIVE);
5344 } while (need_resched());
5347 int __sched _cond_resched(void)
5349 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5350 system_state == SYSTEM_RUNNING) {
5356 EXPORT_SYMBOL(_cond_resched);
5359 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5360 * call schedule, and on return reacquire the lock.
5362 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5363 * operations here to prevent schedule() from being called twice (once via
5364 * spin_unlock(), once by hand).
5366 int cond_resched_lock(spinlock_t *lock)
5368 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5371 if (spin_needbreak(lock) || resched) {
5373 if (resched && need_resched())
5382 EXPORT_SYMBOL(cond_resched_lock);
5384 int __sched cond_resched_softirq(void)
5386 BUG_ON(!in_softirq());
5388 if (need_resched() && system_state == SYSTEM_RUNNING) {
5396 EXPORT_SYMBOL(cond_resched_softirq);
5399 * yield - yield the current processor to other threads.
5401 * This is a shortcut for kernel-space yielding - it marks the
5402 * thread runnable and calls sys_sched_yield().
5404 void __sched yield(void)
5406 set_current_state(TASK_RUNNING);
5409 EXPORT_SYMBOL(yield);
5412 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5413 * that process accounting knows that this is a task in IO wait state.
5415 * But don't do that if it is a deliberate, throttling IO wait (this task
5416 * has set its backing_dev_info: the queue against which it should throttle)
5418 void __sched io_schedule(void)
5420 struct rq *rq = &__raw_get_cpu_var(runqueues);
5422 delayacct_blkio_start();
5423 atomic_inc(&rq->nr_iowait);
5425 atomic_dec(&rq->nr_iowait);
5426 delayacct_blkio_end();
5428 EXPORT_SYMBOL(io_schedule);
5430 long __sched io_schedule_timeout(long timeout)
5432 struct rq *rq = &__raw_get_cpu_var(runqueues);
5435 delayacct_blkio_start();
5436 atomic_inc(&rq->nr_iowait);
5437 ret = schedule_timeout(timeout);
5438 atomic_dec(&rq->nr_iowait);
5439 delayacct_blkio_end();
5444 * sys_sched_get_priority_max - return maximum RT priority.
5445 * @policy: scheduling class.
5447 * this syscall returns the maximum rt_priority that can be used
5448 * by a given scheduling class.
5450 asmlinkage long sys_sched_get_priority_max(int policy)
5457 ret = MAX_USER_RT_PRIO-1;
5469 * sys_sched_get_priority_min - return minimum RT priority.
5470 * @policy: scheduling class.
5472 * this syscall returns the minimum rt_priority that can be used
5473 * by a given scheduling class.
5475 asmlinkage long sys_sched_get_priority_min(int policy)
5493 * sys_sched_rr_get_interval - return the default timeslice of a process.
5494 * @pid: pid of the process.
5495 * @interval: userspace pointer to the timeslice value.
5497 * this syscall writes the default timeslice value of a given process
5498 * into the user-space timespec buffer. A value of '0' means infinity.
5501 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5503 struct task_struct *p;
5504 unsigned int time_slice;
5512 read_lock(&tasklist_lock);
5513 p = find_process_by_pid(pid);
5517 retval = security_task_getscheduler(p);
5522 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5523 * tasks that are on an otherwise idle runqueue:
5526 if (p->policy == SCHED_RR) {
5527 time_slice = DEF_TIMESLICE;
5528 } else if (p->policy != SCHED_FIFO) {
5529 struct sched_entity *se = &p->se;
5530 unsigned long flags;
5533 rq = task_rq_lock(p, &flags);
5534 if (rq->cfs.load.weight)
5535 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5536 task_rq_unlock(rq, &flags);
5538 read_unlock(&tasklist_lock);
5539 jiffies_to_timespec(time_slice, &t);
5540 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5544 read_unlock(&tasklist_lock);
5548 static const char stat_nam[] = "RSDTtZX";
5550 void sched_show_task(struct task_struct *p)
5552 unsigned long free = 0;
5555 state = p->state ? __ffs(p->state) + 1 : 0;
5556 printk(KERN_INFO "%-13.13s %c", p->comm,
5557 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5558 #if BITS_PER_LONG == 32
5559 if (state == TASK_RUNNING)
5560 printk(KERN_CONT " running ");
5562 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5564 if (state == TASK_RUNNING)
5565 printk(KERN_CONT " running task ");
5567 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5569 #ifdef CONFIG_DEBUG_STACK_USAGE
5571 unsigned long *n = end_of_stack(p);
5574 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5577 printk(KERN_CONT "%5lu %5d %6d\n", free,
5578 task_pid_nr(p), task_pid_nr(p->real_parent));
5580 show_stack(p, NULL);
5583 void show_state_filter(unsigned long state_filter)
5585 struct task_struct *g, *p;
5587 #if BITS_PER_LONG == 32
5589 " task PC stack pid father\n");
5592 " task PC stack pid father\n");
5594 read_lock(&tasklist_lock);
5595 do_each_thread(g, p) {
5597 * reset the NMI-timeout, listing all files on a slow
5598 * console might take alot of time:
5600 touch_nmi_watchdog();
5601 if (!state_filter || (p->state & state_filter))
5603 } while_each_thread(g, p);
5605 touch_all_softlockup_watchdogs();
5607 #ifdef CONFIG_SCHED_DEBUG
5608 sysrq_sched_debug_show();
5610 read_unlock(&tasklist_lock);
5612 * Only show locks if all tasks are dumped:
5614 if (state_filter == -1)
5615 debug_show_all_locks();
5618 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5620 idle->sched_class = &idle_sched_class;
5624 * init_idle - set up an idle thread for a given CPU
5625 * @idle: task in question
5626 * @cpu: cpu the idle task belongs to
5628 * NOTE: this function does not set the idle thread's NEED_RESCHED
5629 * flag, to make booting more robust.
5631 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5633 struct rq *rq = cpu_rq(cpu);
5634 unsigned long flags;
5637 idle->se.exec_start = sched_clock();
5639 idle->prio = idle->normal_prio = MAX_PRIO;
5640 idle->cpus_allowed = cpumask_of_cpu(cpu);
5641 __set_task_cpu(idle, cpu);
5643 spin_lock_irqsave(&rq->lock, flags);
5644 rq->curr = rq->idle = idle;
5645 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5648 spin_unlock_irqrestore(&rq->lock, flags);
5650 /* Set the preempt count _outside_ the spinlocks! */
5651 #if defined(CONFIG_PREEMPT)
5652 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5654 task_thread_info(idle)->preempt_count = 0;
5657 * The idle tasks have their own, simple scheduling class:
5659 idle->sched_class = &idle_sched_class;
5663 * In a system that switches off the HZ timer nohz_cpu_mask
5664 * indicates which cpus entered this state. This is used
5665 * in the rcu update to wait only for active cpus. For system
5666 * which do not switch off the HZ timer nohz_cpu_mask should
5667 * always be CPU_MASK_NONE.
5669 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5672 * Increase the granularity value when there are more CPUs,
5673 * because with more CPUs the 'effective latency' as visible
5674 * to users decreases. But the relationship is not linear,
5675 * so pick a second-best guess by going with the log2 of the
5678 * This idea comes from the SD scheduler of Con Kolivas:
5680 static inline void sched_init_granularity(void)
5682 unsigned int factor = 1 + ilog2(num_online_cpus());
5683 const unsigned long limit = 200000000;
5685 sysctl_sched_min_granularity *= factor;
5686 if (sysctl_sched_min_granularity > limit)
5687 sysctl_sched_min_granularity = limit;
5689 sysctl_sched_latency *= factor;
5690 if (sysctl_sched_latency > limit)
5691 sysctl_sched_latency = limit;
5693 sysctl_sched_wakeup_granularity *= factor;
5698 * This is how migration works:
5700 * 1) we queue a struct migration_req structure in the source CPU's
5701 * runqueue and wake up that CPU's migration thread.
5702 * 2) we down() the locked semaphore => thread blocks.
5703 * 3) migration thread wakes up (implicitly it forces the migrated
5704 * thread off the CPU)
5705 * 4) it gets the migration request and checks whether the migrated
5706 * task is still in the wrong runqueue.
5707 * 5) if it's in the wrong runqueue then the migration thread removes
5708 * it and puts it into the right queue.
5709 * 6) migration thread up()s the semaphore.
5710 * 7) we wake up and the migration is done.
5714 * Change a given task's CPU affinity. Migrate the thread to a
5715 * proper CPU and schedule it away if the CPU it's executing on
5716 * is removed from the allowed bitmask.
5718 * NOTE: the caller must have a valid reference to the task, the
5719 * task must not exit() & deallocate itself prematurely. The
5720 * call is not atomic; no spinlocks may be held.
5722 int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask)
5724 struct migration_req req;
5725 unsigned long flags;
5729 rq = task_rq_lock(p, &flags);
5730 if (!cpus_intersects(*new_mask, cpu_online_map)) {
5735 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5736 !cpus_equal(p->cpus_allowed, *new_mask))) {
5741 if (p->sched_class->set_cpus_allowed)
5742 p->sched_class->set_cpus_allowed(p, new_mask);
5744 p->cpus_allowed = *new_mask;
5745 p->rt.nr_cpus_allowed = cpus_weight(*new_mask);
5748 /* Can the task run on the task's current CPU? If so, we're done */
5749 if (cpu_isset(task_cpu(p), *new_mask))
5752 if (migrate_task(p, any_online_cpu(*new_mask), &req)) {
5753 /* Need help from migration thread: drop lock and wait. */
5754 task_rq_unlock(rq, &flags);
5755 wake_up_process(rq->migration_thread);
5756 wait_for_completion(&req.done);
5757 tlb_migrate_finish(p->mm);
5761 task_rq_unlock(rq, &flags);
5765 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5768 * Move (not current) task off this cpu, onto dest cpu. We're doing
5769 * this because either it can't run here any more (set_cpus_allowed()
5770 * away from this CPU, or CPU going down), or because we're
5771 * attempting to rebalance this task on exec (sched_exec).
5773 * So we race with normal scheduler movements, but that's OK, as long
5774 * as the task is no longer on this CPU.
5776 * Returns non-zero if task was successfully migrated.
5778 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5780 struct rq *rq_dest, *rq_src;
5783 if (unlikely(cpu_is_offline(dest_cpu)))
5786 rq_src = cpu_rq(src_cpu);
5787 rq_dest = cpu_rq(dest_cpu);
5789 double_rq_lock(rq_src, rq_dest);
5790 /* Already moved. */
5791 if (task_cpu(p) != src_cpu)
5793 /* Affinity changed (again). */
5794 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5797 on_rq = p->se.on_rq;
5799 deactivate_task(rq_src, p, 0);
5801 set_task_cpu(p, dest_cpu);
5803 activate_task(rq_dest, p, 0);
5804 check_preempt_curr(rq_dest, p);
5808 double_rq_unlock(rq_src, rq_dest);
5813 * migration_thread - this is a highprio system thread that performs
5814 * thread migration by bumping thread off CPU then 'pushing' onto
5817 static int migration_thread(void *data)
5819 int cpu = (long)data;
5823 BUG_ON(rq->migration_thread != current);
5825 set_current_state(TASK_INTERRUPTIBLE);
5826 while (!kthread_should_stop()) {
5827 struct migration_req *req;
5828 struct list_head *head;
5830 spin_lock_irq(&rq->lock);
5832 if (cpu_is_offline(cpu)) {
5833 spin_unlock_irq(&rq->lock);
5837 if (rq->active_balance) {
5838 active_load_balance(rq, cpu);
5839 rq->active_balance = 0;
5842 head = &rq->migration_queue;
5844 if (list_empty(head)) {
5845 spin_unlock_irq(&rq->lock);
5847 set_current_state(TASK_INTERRUPTIBLE);
5850 req = list_entry(head->next, struct migration_req, list);
5851 list_del_init(head->next);
5853 spin_unlock(&rq->lock);
5854 __migrate_task(req->task, cpu, req->dest_cpu);
5857 complete(&req->done);
5859 __set_current_state(TASK_RUNNING);
5863 /* Wait for kthread_stop */
5864 set_current_state(TASK_INTERRUPTIBLE);
5865 while (!kthread_should_stop()) {
5867 set_current_state(TASK_INTERRUPTIBLE);
5869 __set_current_state(TASK_RUNNING);
5873 #ifdef CONFIG_HOTPLUG_CPU
5875 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5879 local_irq_disable();
5880 ret = __migrate_task(p, src_cpu, dest_cpu);
5886 * Figure out where task on dead CPU should go, use force if necessary.
5887 * NOTE: interrupts should be disabled by the caller
5889 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5891 unsigned long flags;
5898 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5899 cpus_and(mask, mask, p->cpus_allowed);
5900 dest_cpu = any_online_cpu(mask);
5902 /* On any allowed CPU? */
5903 if (dest_cpu >= nr_cpu_ids)
5904 dest_cpu = any_online_cpu(p->cpus_allowed);
5906 /* No more Mr. Nice Guy. */
5907 if (dest_cpu >= nr_cpu_ids) {
5908 cpumask_t cpus_allowed;
5910 cpuset_cpus_allowed_locked(p, &cpus_allowed);
5912 * Try to stay on the same cpuset, where the
5913 * current cpuset may be a subset of all cpus.
5914 * The cpuset_cpus_allowed_locked() variant of
5915 * cpuset_cpus_allowed() will not block. It must be
5916 * called within calls to cpuset_lock/cpuset_unlock.
5918 rq = task_rq_lock(p, &flags);
5919 p->cpus_allowed = cpus_allowed;
5920 dest_cpu = any_online_cpu(p->cpus_allowed);
5921 task_rq_unlock(rq, &flags);
5924 * Don't tell them about moving exiting tasks or
5925 * kernel threads (both mm NULL), since they never
5928 if (p->mm && printk_ratelimit()) {
5929 printk(KERN_INFO "process %d (%s) no "
5930 "longer affine to cpu%d\n",
5931 task_pid_nr(p), p->comm, dead_cpu);
5934 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
5938 * While a dead CPU has no uninterruptible tasks queued at this point,
5939 * it might still have a nonzero ->nr_uninterruptible counter, because
5940 * for performance reasons the counter is not stricly tracking tasks to
5941 * their home CPUs. So we just add the counter to another CPU's counter,
5942 * to keep the global sum constant after CPU-down:
5944 static void migrate_nr_uninterruptible(struct rq *rq_src)
5946 struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR));
5947 unsigned long flags;
5949 local_irq_save(flags);
5950 double_rq_lock(rq_src, rq_dest);
5951 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5952 rq_src->nr_uninterruptible = 0;
5953 double_rq_unlock(rq_src, rq_dest);
5954 local_irq_restore(flags);
5957 /* Run through task list and migrate tasks from the dead cpu. */
5958 static void migrate_live_tasks(int src_cpu)
5960 struct task_struct *p, *t;
5962 read_lock(&tasklist_lock);
5964 do_each_thread(t, p) {
5968 if (task_cpu(p) == src_cpu)
5969 move_task_off_dead_cpu(src_cpu, p);
5970 } while_each_thread(t, p);
5972 read_unlock(&tasklist_lock);
5976 * Schedules idle task to be the next runnable task on current CPU.
5977 * It does so by boosting its priority to highest possible.
5978 * Used by CPU offline code.
5980 void sched_idle_next(void)
5982 int this_cpu = smp_processor_id();
5983 struct rq *rq = cpu_rq(this_cpu);
5984 struct task_struct *p = rq->idle;
5985 unsigned long flags;
5987 /* cpu has to be offline */
5988 BUG_ON(cpu_online(this_cpu));
5991 * Strictly not necessary since rest of the CPUs are stopped by now
5992 * and interrupts disabled on the current cpu.
5994 spin_lock_irqsave(&rq->lock, flags);
5996 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5998 update_rq_clock(rq);
5999 activate_task(rq, p, 0);
6001 spin_unlock_irqrestore(&rq->lock, flags);
6005 * Ensures that the idle task is using init_mm right before its cpu goes
6008 void idle_task_exit(void)
6010 struct mm_struct *mm = current->active_mm;
6012 BUG_ON(cpu_online(smp_processor_id()));
6015 switch_mm(mm, &init_mm, current);
6019 /* called under rq->lock with disabled interrupts */
6020 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
6022 struct rq *rq = cpu_rq(dead_cpu);
6024 /* Must be exiting, otherwise would be on tasklist. */
6025 BUG_ON(!p->exit_state);
6027 /* Cannot have done final schedule yet: would have vanished. */
6028 BUG_ON(p->state == TASK_DEAD);
6033 * Drop lock around migration; if someone else moves it,
6034 * that's OK. No task can be added to this CPU, so iteration is
6037 spin_unlock_irq(&rq->lock);
6038 move_task_off_dead_cpu(dead_cpu, p);
6039 spin_lock_irq(&rq->lock);
6044 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6045 static void migrate_dead_tasks(unsigned int dead_cpu)
6047 struct rq *rq = cpu_rq(dead_cpu);
6048 struct task_struct *next;
6051 if (!rq->nr_running)
6053 update_rq_clock(rq);
6054 next = pick_next_task(rq, rq->curr);
6057 migrate_dead(dead_cpu, next);
6061 #endif /* CONFIG_HOTPLUG_CPU */
6063 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6065 static struct ctl_table sd_ctl_dir[] = {
6067 .procname = "sched_domain",
6073 static struct ctl_table sd_ctl_root[] = {
6075 .ctl_name = CTL_KERN,
6076 .procname = "kernel",
6078 .child = sd_ctl_dir,
6083 static struct ctl_table *sd_alloc_ctl_entry(int n)
6085 struct ctl_table *entry =
6086 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6091 static void sd_free_ctl_entry(struct ctl_table **tablep)
6093 struct ctl_table *entry;
6096 * In the intermediate directories, both the child directory and
6097 * procname are dynamically allocated and could fail but the mode
6098 * will always be set. In the lowest directory the names are
6099 * static strings and all have proc handlers.
6101 for (entry = *tablep; entry->mode; entry++) {
6103 sd_free_ctl_entry(&entry->child);
6104 if (entry->proc_handler == NULL)
6105 kfree(entry->procname);
6113 set_table_entry(struct ctl_table *entry,
6114 const char *procname, void *data, int maxlen,
6115 mode_t mode, proc_handler *proc_handler)
6117 entry->procname = procname;
6119 entry->maxlen = maxlen;
6121 entry->proc_handler = proc_handler;
6124 static struct ctl_table *
6125 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6127 struct ctl_table *table = sd_alloc_ctl_entry(12);
6132 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6133 sizeof(long), 0644, proc_doulongvec_minmax);
6134 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6135 sizeof(long), 0644, proc_doulongvec_minmax);
6136 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6137 sizeof(int), 0644, proc_dointvec_minmax);
6138 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6139 sizeof(int), 0644, proc_dointvec_minmax);
6140 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6141 sizeof(int), 0644, proc_dointvec_minmax);
6142 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6143 sizeof(int), 0644, proc_dointvec_minmax);
6144 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6145 sizeof(int), 0644, proc_dointvec_minmax);
6146 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6147 sizeof(int), 0644, proc_dointvec_minmax);
6148 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6149 sizeof(int), 0644, proc_dointvec_minmax);
6150 set_table_entry(&table[9], "cache_nice_tries",
6151 &sd->cache_nice_tries,
6152 sizeof(int), 0644, proc_dointvec_minmax);
6153 set_table_entry(&table[10], "flags", &sd->flags,
6154 sizeof(int), 0644, proc_dointvec_minmax);
6155 /* &table[11] is terminator */
6160 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6162 struct ctl_table *entry, *table;
6163 struct sched_domain *sd;
6164 int domain_num = 0, i;
6167 for_each_domain(cpu, sd)
6169 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6174 for_each_domain(cpu, sd) {
6175 snprintf(buf, 32, "domain%d", i);
6176 entry->procname = kstrdup(buf, GFP_KERNEL);
6178 entry->child = sd_alloc_ctl_domain_table(sd);
6185 static struct ctl_table_header *sd_sysctl_header;
6186 static void register_sched_domain_sysctl(void)
6188 int i, cpu_num = num_online_cpus();
6189 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6192 WARN_ON(sd_ctl_dir[0].child);
6193 sd_ctl_dir[0].child = entry;
6198 for_each_online_cpu(i) {
6199 snprintf(buf, 32, "cpu%d", i);
6200 entry->procname = kstrdup(buf, GFP_KERNEL);
6202 entry->child = sd_alloc_ctl_cpu_table(i);
6206 WARN_ON(sd_sysctl_header);
6207 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6210 /* may be called multiple times per register */
6211 static void unregister_sched_domain_sysctl(void)
6213 if (sd_sysctl_header)
6214 unregister_sysctl_table(sd_sysctl_header);
6215 sd_sysctl_header = NULL;
6216 if (sd_ctl_dir[0].child)
6217 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6220 static void register_sched_domain_sysctl(void)
6223 static void unregister_sched_domain_sysctl(void)
6228 static void set_rq_online(struct rq *rq)
6231 const struct sched_class *class;
6233 cpu_set(rq->cpu, rq->rd->online);
6236 for_each_class(class) {
6237 if (class->rq_online)
6238 class->rq_online(rq);
6243 static void set_rq_offline(struct rq *rq)
6246 const struct sched_class *class;
6248 for_each_class(class) {
6249 if (class->rq_offline)
6250 class->rq_offline(rq);
6253 cpu_clear(rq->cpu, rq->rd->online);
6259 * migration_call - callback that gets triggered when a CPU is added.
6260 * Here we can start up the necessary migration thread for the new CPU.
6262 static int __cpuinit
6263 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6265 struct task_struct *p;
6266 int cpu = (long)hcpu;
6267 unsigned long flags;
6272 case CPU_UP_PREPARE:
6273 case CPU_UP_PREPARE_FROZEN:
6274 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6277 kthread_bind(p, cpu);
6278 /* Must be high prio: stop_machine expects to yield to it. */
6279 rq = task_rq_lock(p, &flags);
6280 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6281 task_rq_unlock(rq, &flags);
6282 cpu_rq(cpu)->migration_thread = p;
6286 case CPU_ONLINE_FROZEN:
6287 /* Strictly unnecessary, as first user will wake it. */
6288 wake_up_process(cpu_rq(cpu)->migration_thread);
6290 /* Update our root-domain */
6292 spin_lock_irqsave(&rq->lock, flags);
6294 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6298 spin_unlock_irqrestore(&rq->lock, flags);
6301 #ifdef CONFIG_HOTPLUG_CPU
6302 case CPU_UP_CANCELED:
6303 case CPU_UP_CANCELED_FROZEN:
6304 if (!cpu_rq(cpu)->migration_thread)
6306 /* Unbind it from offline cpu so it can run. Fall thru. */
6307 kthread_bind(cpu_rq(cpu)->migration_thread,
6308 any_online_cpu(cpu_online_map));
6309 kthread_stop(cpu_rq(cpu)->migration_thread);
6310 cpu_rq(cpu)->migration_thread = NULL;
6314 case CPU_DEAD_FROZEN:
6315 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6316 migrate_live_tasks(cpu);
6318 kthread_stop(rq->migration_thread);
6319 rq->migration_thread = NULL;
6320 /* Idle task back to normal (off runqueue, low prio) */
6321 spin_lock_irq(&rq->lock);
6322 update_rq_clock(rq);
6323 deactivate_task(rq, rq->idle, 0);
6324 rq->idle->static_prio = MAX_PRIO;
6325 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6326 rq->idle->sched_class = &idle_sched_class;
6327 migrate_dead_tasks(cpu);
6328 spin_unlock_irq(&rq->lock);
6330 migrate_nr_uninterruptible(rq);
6331 BUG_ON(rq->nr_running != 0);
6334 * No need to migrate the tasks: it was best-effort if
6335 * they didn't take sched_hotcpu_mutex. Just wake up
6338 spin_lock_irq(&rq->lock);
6339 while (!list_empty(&rq->migration_queue)) {
6340 struct migration_req *req;
6342 req = list_entry(rq->migration_queue.next,
6343 struct migration_req, list);
6344 list_del_init(&req->list);
6345 complete(&req->done);
6347 spin_unlock_irq(&rq->lock);
6351 case CPU_DYING_FROZEN:
6352 /* Update our root-domain */
6354 spin_lock_irqsave(&rq->lock, flags);
6356 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6359 spin_unlock_irqrestore(&rq->lock, flags);
6366 /* Register at highest priority so that task migration (migrate_all_tasks)
6367 * happens before everything else.
6369 static struct notifier_block __cpuinitdata migration_notifier = {
6370 .notifier_call = migration_call,
6374 void __init migration_init(void)
6376 void *cpu = (void *)(long)smp_processor_id();
6379 /* Start one for the boot CPU: */
6380 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6381 BUG_ON(err == NOTIFY_BAD);
6382 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6383 register_cpu_notifier(&migration_notifier);
6389 #ifdef CONFIG_SCHED_DEBUG
6391 static inline const char *sd_level_to_string(enum sched_domain_level lvl)
6404 case SD_LV_ALLNODES:
6413 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6414 cpumask_t *groupmask)
6416 struct sched_group *group = sd->groups;
6419 cpulist_scnprintf(str, sizeof(str), sd->span);
6420 cpus_clear(*groupmask);
6422 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6424 if (!(sd->flags & SD_LOAD_BALANCE)) {
6425 printk("does not load-balance\n");
6427 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6432 printk(KERN_CONT "span %s level %s\n",
6433 str, sd_level_to_string(sd->level));
6435 if (!cpu_isset(cpu, sd->span)) {
6436 printk(KERN_ERR "ERROR: domain->span does not contain "
6439 if (!cpu_isset(cpu, group->cpumask)) {
6440 printk(KERN_ERR "ERROR: domain->groups does not contain"
6444 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6448 printk(KERN_ERR "ERROR: group is NULL\n");
6452 if (!group->__cpu_power) {
6453 printk(KERN_CONT "\n");
6454 printk(KERN_ERR "ERROR: domain->cpu_power not "
6459 if (!cpus_weight(group->cpumask)) {
6460 printk(KERN_CONT "\n");
6461 printk(KERN_ERR "ERROR: empty group\n");
6465 if (cpus_intersects(*groupmask, group->cpumask)) {
6466 printk(KERN_CONT "\n");
6467 printk(KERN_ERR "ERROR: repeated CPUs\n");
6471 cpus_or(*groupmask, *groupmask, group->cpumask);
6473 cpulist_scnprintf(str, sizeof(str), group->cpumask);
6474 printk(KERN_CONT " %s", str);
6476 group = group->next;
6477 } while (group != sd->groups);
6478 printk(KERN_CONT "\n");
6480 if (!cpus_equal(sd->span, *groupmask))
6481 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6483 if (sd->parent && !cpus_subset(*groupmask, sd->parent->span))
6484 printk(KERN_ERR "ERROR: parent span is not a superset "
6485 "of domain->span\n");
6489 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6491 cpumask_t *groupmask;
6495 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6499 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6501 groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6503 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6508 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6517 #else /* !CONFIG_SCHED_DEBUG */
6518 # define sched_domain_debug(sd, cpu) do { } while (0)
6519 #endif /* CONFIG_SCHED_DEBUG */
6521 static int sd_degenerate(struct sched_domain *sd)
6523 if (cpus_weight(sd->span) == 1)
6526 /* Following flags need at least 2 groups */
6527 if (sd->flags & (SD_LOAD_BALANCE |
6528 SD_BALANCE_NEWIDLE |
6532 SD_SHARE_PKG_RESOURCES)) {
6533 if (sd->groups != sd->groups->next)
6537 /* Following flags don't use groups */
6538 if (sd->flags & (SD_WAKE_IDLE |
6547 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6549 unsigned long cflags = sd->flags, pflags = parent->flags;
6551 if (sd_degenerate(parent))
6554 if (!cpus_equal(sd->span, parent->span))
6557 /* Does parent contain flags not in child? */
6558 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6559 if (cflags & SD_WAKE_AFFINE)
6560 pflags &= ~SD_WAKE_BALANCE;
6561 /* Flags needing groups don't count if only 1 group in parent */
6562 if (parent->groups == parent->groups->next) {
6563 pflags &= ~(SD_LOAD_BALANCE |
6564 SD_BALANCE_NEWIDLE |
6568 SD_SHARE_PKG_RESOURCES);
6570 if (~cflags & pflags)
6576 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6578 unsigned long flags;
6580 spin_lock_irqsave(&rq->lock, flags);
6583 struct root_domain *old_rd = rq->rd;
6585 if (cpu_isset(rq->cpu, old_rd->online))
6588 cpu_clear(rq->cpu, old_rd->span);
6590 if (atomic_dec_and_test(&old_rd->refcount))
6594 atomic_inc(&rd->refcount);
6597 cpu_set(rq->cpu, rd->span);
6598 if (cpu_isset(rq->cpu, cpu_online_map))
6601 spin_unlock_irqrestore(&rq->lock, flags);
6604 static void init_rootdomain(struct root_domain *rd)
6606 memset(rd, 0, sizeof(*rd));
6608 cpus_clear(rd->span);
6609 cpus_clear(rd->online);
6611 cpupri_init(&rd->cpupri);
6614 static void init_defrootdomain(void)
6616 init_rootdomain(&def_root_domain);
6617 atomic_set(&def_root_domain.refcount, 1);
6620 static struct root_domain *alloc_rootdomain(void)
6622 struct root_domain *rd;
6624 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6628 init_rootdomain(rd);
6634 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6635 * hold the hotplug lock.
6638 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6640 struct rq *rq = cpu_rq(cpu);
6641 struct sched_domain *tmp;
6643 /* Remove the sched domains which do not contribute to scheduling. */
6644 for (tmp = sd; tmp; tmp = tmp->parent) {
6645 struct sched_domain *parent = tmp->parent;
6648 if (sd_parent_degenerate(tmp, parent)) {
6649 tmp->parent = parent->parent;
6651 parent->parent->child = tmp;
6655 if (sd && sd_degenerate(sd)) {
6661 sched_domain_debug(sd, cpu);
6663 rq_attach_root(rq, rd);
6664 rcu_assign_pointer(rq->sd, sd);
6667 /* cpus with isolated domains */
6668 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
6670 /* Setup the mask of cpus configured for isolated domains */
6671 static int __init isolated_cpu_setup(char *str)
6673 int ints[NR_CPUS], i;
6675 str = get_options(str, ARRAY_SIZE(ints), ints);
6676 cpus_clear(cpu_isolated_map);
6677 for (i = 1; i <= ints[0]; i++)
6678 if (ints[i] < NR_CPUS)
6679 cpu_set(ints[i], cpu_isolated_map);
6683 __setup("isolcpus=", isolated_cpu_setup);
6686 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6687 * to a function which identifies what group(along with sched group) a CPU
6688 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6689 * (due to the fact that we keep track of groups covered with a cpumask_t).
6691 * init_sched_build_groups will build a circular linked list of the groups
6692 * covered by the given span, and will set each group's ->cpumask correctly,
6693 * and ->cpu_power to 0.
6696 init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map,
6697 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6698 struct sched_group **sg,
6699 cpumask_t *tmpmask),
6700 cpumask_t *covered, cpumask_t *tmpmask)
6702 struct sched_group *first = NULL, *last = NULL;
6705 cpus_clear(*covered);
6707 for_each_cpu_mask(i, *span) {
6708 struct sched_group *sg;
6709 int group = group_fn(i, cpu_map, &sg, tmpmask);
6712 if (cpu_isset(i, *covered))
6715 cpus_clear(sg->cpumask);
6716 sg->__cpu_power = 0;
6718 for_each_cpu_mask(j, *span) {
6719 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6722 cpu_set(j, *covered);
6723 cpu_set(j, sg->cpumask);
6734 #define SD_NODES_PER_DOMAIN 16
6739 * find_next_best_node - find the next node to include in a sched_domain
6740 * @node: node whose sched_domain we're building
6741 * @used_nodes: nodes already in the sched_domain
6743 * Find the next node to include in a given scheduling domain. Simply
6744 * finds the closest node not already in the @used_nodes map.
6746 * Should use nodemask_t.
6748 static int find_next_best_node(int node, nodemask_t *used_nodes)
6750 int i, n, val, min_val, best_node = 0;
6754 for (i = 0; i < MAX_NUMNODES; i++) {
6755 /* Start at @node */
6756 n = (node + i) % MAX_NUMNODES;
6758 if (!nr_cpus_node(n))
6761 /* Skip already used nodes */
6762 if (node_isset(n, *used_nodes))
6765 /* Simple min distance search */
6766 val = node_distance(node, n);
6768 if (val < min_val) {
6774 node_set(best_node, *used_nodes);
6779 * sched_domain_node_span - get a cpumask for a node's sched_domain
6780 * @node: node whose cpumask we're constructing
6781 * @span: resulting cpumask
6783 * Given a node, construct a good cpumask for its sched_domain to span. It
6784 * should be one that prevents unnecessary balancing, but also spreads tasks
6787 static void sched_domain_node_span(int node, cpumask_t *span)
6789 nodemask_t used_nodes;
6790 node_to_cpumask_ptr(nodemask, node);
6794 nodes_clear(used_nodes);
6796 cpus_or(*span, *span, *nodemask);
6797 node_set(node, used_nodes);
6799 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6800 int next_node = find_next_best_node(node, &used_nodes);
6802 node_to_cpumask_ptr_next(nodemask, next_node);
6803 cpus_or(*span, *span, *nodemask);
6806 #endif /* CONFIG_NUMA */
6808 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6811 * SMT sched-domains:
6813 #ifdef CONFIG_SCHED_SMT
6814 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6815 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6818 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6822 *sg = &per_cpu(sched_group_cpus, cpu);
6825 #endif /* CONFIG_SCHED_SMT */
6828 * multi-core sched-domains:
6830 #ifdef CONFIG_SCHED_MC
6831 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6832 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6833 #endif /* CONFIG_SCHED_MC */
6835 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6837 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6842 *mask = per_cpu(cpu_sibling_map, cpu);
6843 cpus_and(*mask, *mask, *cpu_map);
6844 group = first_cpu(*mask);
6846 *sg = &per_cpu(sched_group_core, group);
6849 #elif defined(CONFIG_SCHED_MC)
6851 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6855 *sg = &per_cpu(sched_group_core, cpu);
6860 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6861 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6864 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6868 #ifdef CONFIG_SCHED_MC
6869 *mask = cpu_coregroup_map(cpu);
6870 cpus_and(*mask, *mask, *cpu_map);
6871 group = first_cpu(*mask);
6872 #elif defined(CONFIG_SCHED_SMT)
6873 *mask = per_cpu(cpu_sibling_map, cpu);
6874 cpus_and(*mask, *mask, *cpu_map);
6875 group = first_cpu(*mask);
6880 *sg = &per_cpu(sched_group_phys, group);
6886 * The init_sched_build_groups can't handle what we want to do with node
6887 * groups, so roll our own. Now each node has its own list of groups which
6888 * gets dynamically allocated.
6890 static DEFINE_PER_CPU(struct sched_domain, node_domains);
6891 static struct sched_group ***sched_group_nodes_bycpu;
6893 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6894 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
6896 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6897 struct sched_group **sg, cpumask_t *nodemask)
6901 *nodemask = node_to_cpumask(cpu_to_node(cpu));
6902 cpus_and(*nodemask, *nodemask, *cpu_map);
6903 group = first_cpu(*nodemask);
6906 *sg = &per_cpu(sched_group_allnodes, group);
6910 static void init_numa_sched_groups_power(struct sched_group *group_head)
6912 struct sched_group *sg = group_head;
6918 for_each_cpu_mask(j, sg->cpumask) {
6919 struct sched_domain *sd;
6921 sd = &per_cpu(phys_domains, j);
6922 if (j != first_cpu(sd->groups->cpumask)) {
6924 * Only add "power" once for each
6930 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
6933 } while (sg != group_head);
6935 #endif /* CONFIG_NUMA */
6938 /* Free memory allocated for various sched_group structures */
6939 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
6943 for_each_cpu_mask(cpu, *cpu_map) {
6944 struct sched_group **sched_group_nodes
6945 = sched_group_nodes_bycpu[cpu];
6947 if (!sched_group_nodes)
6950 for (i = 0; i < MAX_NUMNODES; i++) {
6951 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6953 *nodemask = node_to_cpumask(i);
6954 cpus_and(*nodemask, *nodemask, *cpu_map);
6955 if (cpus_empty(*nodemask))
6965 if (oldsg != sched_group_nodes[i])
6968 kfree(sched_group_nodes);
6969 sched_group_nodes_bycpu[cpu] = NULL;
6972 #else /* !CONFIG_NUMA */
6973 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
6976 #endif /* CONFIG_NUMA */
6979 * Initialize sched groups cpu_power.
6981 * cpu_power indicates the capacity of sched group, which is used while
6982 * distributing the load between different sched groups in a sched domain.
6983 * Typically cpu_power for all the groups in a sched domain will be same unless
6984 * there are asymmetries in the topology. If there are asymmetries, group
6985 * having more cpu_power will pickup more load compared to the group having
6988 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6989 * the maximum number of tasks a group can handle in the presence of other idle
6990 * or lightly loaded groups in the same sched domain.
6992 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6994 struct sched_domain *child;
6995 struct sched_group *group;
6997 WARN_ON(!sd || !sd->groups);
6999 if (cpu != first_cpu(sd->groups->cpumask))
7004 sd->groups->__cpu_power = 0;
7007 * For perf policy, if the groups in child domain share resources
7008 * (for example cores sharing some portions of the cache hierarchy
7009 * or SMT), then set this domain groups cpu_power such that each group
7010 * can handle only one task, when there are other idle groups in the
7011 * same sched domain.
7013 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
7015 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
7016 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
7021 * add cpu_power of each child group to this groups cpu_power
7023 group = child->groups;
7025 sg_inc_cpu_power(sd->groups, group->__cpu_power);
7026 group = group->next;
7027 } while (group != child->groups);
7031 * Initializers for schedule domains
7032 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7035 #define SD_INIT(sd, type) sd_init_##type(sd)
7036 #define SD_INIT_FUNC(type) \
7037 static noinline void sd_init_##type(struct sched_domain *sd) \
7039 memset(sd, 0, sizeof(*sd)); \
7040 *sd = SD_##type##_INIT; \
7041 sd->level = SD_LV_##type; \
7046 SD_INIT_FUNC(ALLNODES)
7049 #ifdef CONFIG_SCHED_SMT
7050 SD_INIT_FUNC(SIBLING)
7052 #ifdef CONFIG_SCHED_MC
7057 * To minimize stack usage kmalloc room for cpumasks and share the
7058 * space as the usage in build_sched_domains() dictates. Used only
7059 * if the amount of space is significant.
7062 cpumask_t tmpmask; /* make this one first */
7065 cpumask_t this_sibling_map;
7066 cpumask_t this_core_map;
7068 cpumask_t send_covered;
7071 cpumask_t domainspan;
7073 cpumask_t notcovered;
7078 #define SCHED_CPUMASK_ALLOC 1
7079 #define SCHED_CPUMASK_FREE(v) kfree(v)
7080 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
7082 #define SCHED_CPUMASK_ALLOC 0
7083 #define SCHED_CPUMASK_FREE(v)
7084 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
7087 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
7088 ((unsigned long)(a) + offsetof(struct allmasks, v))
7090 static int default_relax_domain_level = -1;
7092 static int __init setup_relax_domain_level(char *str)
7096 val = simple_strtoul(str, NULL, 0);
7097 if (val < SD_LV_MAX)
7098 default_relax_domain_level = val;
7102 __setup("relax_domain_level=", setup_relax_domain_level);
7104 static void set_domain_attribute(struct sched_domain *sd,
7105 struct sched_domain_attr *attr)
7109 if (!attr || attr->relax_domain_level < 0) {
7110 if (default_relax_domain_level < 0)
7113 request = default_relax_domain_level;
7115 request = attr->relax_domain_level;
7116 if (request < sd->level) {
7117 /* turn off idle balance on this domain */
7118 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
7120 /* turn on idle balance on this domain */
7121 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
7126 * Build sched domains for a given set of cpus and attach the sched domains
7127 * to the individual cpus
7129 static int __build_sched_domains(const cpumask_t *cpu_map,
7130 struct sched_domain_attr *attr)
7133 struct root_domain *rd;
7134 SCHED_CPUMASK_DECLARE(allmasks);
7137 struct sched_group **sched_group_nodes = NULL;
7138 int sd_allnodes = 0;
7141 * Allocate the per-node list of sched groups
7143 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
7145 if (!sched_group_nodes) {
7146 printk(KERN_WARNING "Can not alloc sched group node list\n");
7151 rd = alloc_rootdomain();
7153 printk(KERN_WARNING "Cannot alloc root domain\n");
7155 kfree(sched_group_nodes);
7160 #if SCHED_CPUMASK_ALLOC
7161 /* get space for all scratch cpumask variables */
7162 allmasks = kmalloc(sizeof(*allmasks), GFP_KERNEL);
7164 printk(KERN_WARNING "Cannot alloc cpumask array\n");
7167 kfree(sched_group_nodes);
7172 tmpmask = (cpumask_t *)allmasks;
7176 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
7180 * Set up domains for cpus specified by the cpu_map.
7182 for_each_cpu_mask(i, *cpu_map) {
7183 struct sched_domain *sd = NULL, *p;
7184 SCHED_CPUMASK_VAR(nodemask, allmasks);
7186 *nodemask = node_to_cpumask(cpu_to_node(i));
7187 cpus_and(*nodemask, *nodemask, *cpu_map);
7190 if (cpus_weight(*cpu_map) >
7191 SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) {
7192 sd = &per_cpu(allnodes_domains, i);
7193 SD_INIT(sd, ALLNODES);
7194 set_domain_attribute(sd, attr);
7195 sd->span = *cpu_map;
7196 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
7202 sd = &per_cpu(node_domains, i);
7204 set_domain_attribute(sd, attr);
7205 sched_domain_node_span(cpu_to_node(i), &sd->span);
7209 cpus_and(sd->span, sd->span, *cpu_map);
7213 sd = &per_cpu(phys_domains, i);
7215 set_domain_attribute(sd, attr);
7216 sd->span = *nodemask;
7220 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
7222 #ifdef CONFIG_SCHED_MC
7224 sd = &per_cpu(core_domains, i);
7226 set_domain_attribute(sd, attr);
7227 sd->span = cpu_coregroup_map(i);
7228 cpus_and(sd->span, sd->span, *cpu_map);
7231 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
7234 #ifdef CONFIG_SCHED_SMT
7236 sd = &per_cpu(cpu_domains, i);
7237 SD_INIT(sd, SIBLING);
7238 set_domain_attribute(sd, attr);
7239 sd->span = per_cpu(cpu_sibling_map, i);
7240 cpus_and(sd->span, sd->span, *cpu_map);
7243 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
7247 #ifdef CONFIG_SCHED_SMT
7248 /* Set up CPU (sibling) groups */
7249 for_each_cpu_mask(i, *cpu_map) {
7250 SCHED_CPUMASK_VAR(this_sibling_map, allmasks);
7251 SCHED_CPUMASK_VAR(send_covered, allmasks);
7253 *this_sibling_map = per_cpu(cpu_sibling_map, i);
7254 cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map);
7255 if (i != first_cpu(*this_sibling_map))
7258 init_sched_build_groups(this_sibling_map, cpu_map,
7260 send_covered, tmpmask);
7264 #ifdef CONFIG_SCHED_MC
7265 /* Set up multi-core groups */
7266 for_each_cpu_mask(i, *cpu_map) {
7267 SCHED_CPUMASK_VAR(this_core_map, allmasks);
7268 SCHED_CPUMASK_VAR(send_covered, allmasks);
7270 *this_core_map = cpu_coregroup_map(i);
7271 cpus_and(*this_core_map, *this_core_map, *cpu_map);
7272 if (i != first_cpu(*this_core_map))
7275 init_sched_build_groups(this_core_map, cpu_map,
7277 send_covered, tmpmask);
7281 /* Set up physical groups */
7282 for (i = 0; i < MAX_NUMNODES; i++) {
7283 SCHED_CPUMASK_VAR(nodemask, allmasks);
7284 SCHED_CPUMASK_VAR(send_covered, allmasks);
7286 *nodemask = node_to_cpumask(i);
7287 cpus_and(*nodemask, *nodemask, *cpu_map);
7288 if (cpus_empty(*nodemask))
7291 init_sched_build_groups(nodemask, cpu_map,
7293 send_covered, tmpmask);
7297 /* Set up node groups */
7299 SCHED_CPUMASK_VAR(send_covered, allmasks);
7301 init_sched_build_groups(cpu_map, cpu_map,
7302 &cpu_to_allnodes_group,
7303 send_covered, tmpmask);
7306 for (i = 0; i < MAX_NUMNODES; i++) {
7307 /* Set up node groups */
7308 struct sched_group *sg, *prev;
7309 SCHED_CPUMASK_VAR(nodemask, allmasks);
7310 SCHED_CPUMASK_VAR(domainspan, allmasks);
7311 SCHED_CPUMASK_VAR(covered, allmasks);
7314 *nodemask = node_to_cpumask(i);
7315 cpus_clear(*covered);
7317 cpus_and(*nodemask, *nodemask, *cpu_map);
7318 if (cpus_empty(*nodemask)) {
7319 sched_group_nodes[i] = NULL;
7323 sched_domain_node_span(i, domainspan);
7324 cpus_and(*domainspan, *domainspan, *cpu_map);
7326 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
7328 printk(KERN_WARNING "Can not alloc domain group for "
7332 sched_group_nodes[i] = sg;
7333 for_each_cpu_mask(j, *nodemask) {
7334 struct sched_domain *sd;
7336 sd = &per_cpu(node_domains, j);
7339 sg->__cpu_power = 0;
7340 sg->cpumask = *nodemask;
7342 cpus_or(*covered, *covered, *nodemask);
7345 for (j = 0; j < MAX_NUMNODES; j++) {
7346 SCHED_CPUMASK_VAR(notcovered, allmasks);
7347 int n = (i + j) % MAX_NUMNODES;
7348 node_to_cpumask_ptr(pnodemask, n);
7350 cpus_complement(*notcovered, *covered);
7351 cpus_and(*tmpmask, *notcovered, *cpu_map);
7352 cpus_and(*tmpmask, *tmpmask, *domainspan);
7353 if (cpus_empty(*tmpmask))
7356 cpus_and(*tmpmask, *tmpmask, *pnodemask);
7357 if (cpus_empty(*tmpmask))
7360 sg = kmalloc_node(sizeof(struct sched_group),
7364 "Can not alloc domain group for node %d\n", j);
7367 sg->__cpu_power = 0;
7368 sg->cpumask = *tmpmask;
7369 sg->next = prev->next;
7370 cpus_or(*covered, *covered, *tmpmask);
7377 /* Calculate CPU power for physical packages and nodes */
7378 #ifdef CONFIG_SCHED_SMT
7379 for_each_cpu_mask(i, *cpu_map) {
7380 struct sched_domain *sd = &per_cpu(cpu_domains, i);
7382 init_sched_groups_power(i, sd);
7385 #ifdef CONFIG_SCHED_MC
7386 for_each_cpu_mask(i, *cpu_map) {
7387 struct sched_domain *sd = &per_cpu(core_domains, i);
7389 init_sched_groups_power(i, sd);
7393 for_each_cpu_mask(i, *cpu_map) {
7394 struct sched_domain *sd = &per_cpu(phys_domains, i);
7396 init_sched_groups_power(i, sd);
7400 for (i = 0; i < MAX_NUMNODES; i++)
7401 init_numa_sched_groups_power(sched_group_nodes[i]);
7404 struct sched_group *sg;
7406 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg,
7408 init_numa_sched_groups_power(sg);
7412 /* Attach the domains */
7413 for_each_cpu_mask(i, *cpu_map) {
7414 struct sched_domain *sd;
7415 #ifdef CONFIG_SCHED_SMT
7416 sd = &per_cpu(cpu_domains, i);
7417 #elif defined(CONFIG_SCHED_MC)
7418 sd = &per_cpu(core_domains, i);
7420 sd = &per_cpu(phys_domains, i);
7422 cpu_attach_domain(sd, rd, i);
7425 SCHED_CPUMASK_FREE((void *)allmasks);
7430 free_sched_groups(cpu_map, tmpmask);
7431 SCHED_CPUMASK_FREE((void *)allmasks);
7436 static int build_sched_domains(const cpumask_t *cpu_map)
7438 return __build_sched_domains(cpu_map, NULL);
7441 static cpumask_t *doms_cur; /* current sched domains */
7442 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7443 static struct sched_domain_attr *dattr_cur;
7444 /* attribues of custom domains in 'doms_cur' */
7447 * Special case: If a kmalloc of a doms_cur partition (array of
7448 * cpumask_t) fails, then fallback to a single sched domain,
7449 * as determined by the single cpumask_t fallback_doms.
7451 static cpumask_t fallback_doms;
7453 void __attribute__((weak)) arch_update_cpu_topology(void)
7458 * Free current domain masks.
7459 * Called after all cpus are attached to NULL domain.
7461 static void free_sched_domains(void)
7464 if (doms_cur != &fallback_doms)
7466 doms_cur = &fallback_doms;
7470 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7471 * For now this just excludes isolated cpus, but could be used to
7472 * exclude other special cases in the future.
7474 static int arch_init_sched_domains(const cpumask_t *cpu_map)
7478 arch_update_cpu_topology();
7480 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
7482 doms_cur = &fallback_doms;
7483 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
7485 err = build_sched_domains(doms_cur);
7486 register_sched_domain_sysctl();
7491 static void arch_destroy_sched_domains(const cpumask_t *cpu_map,
7494 free_sched_groups(cpu_map, tmpmask);
7498 * Detach sched domains from a group of cpus specified in cpu_map
7499 * These cpus will now be attached to the NULL domain
7501 static void detach_destroy_domains(const cpumask_t *cpu_map)
7506 unregister_sched_domain_sysctl();
7508 for_each_cpu_mask(i, *cpu_map)
7509 cpu_attach_domain(NULL, &def_root_domain, i);
7510 synchronize_sched();
7511 arch_destroy_sched_domains(cpu_map, &tmpmask);
7514 /* handle null as "default" */
7515 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7516 struct sched_domain_attr *new, int idx_new)
7518 struct sched_domain_attr tmp;
7525 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7526 new ? (new + idx_new) : &tmp,
7527 sizeof(struct sched_domain_attr));
7531 * Partition sched domains as specified by the 'ndoms_new'
7532 * cpumasks in the array doms_new[] of cpumasks. This compares
7533 * doms_new[] to the current sched domain partitioning, doms_cur[].
7534 * It destroys each deleted domain and builds each new domain.
7536 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7537 * The masks don't intersect (don't overlap.) We should setup one
7538 * sched domain for each mask. CPUs not in any of the cpumasks will
7539 * not be load balanced. If the same cpumask appears both in the
7540 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7543 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7544 * ownership of it and will kfree it when done with it. If the caller
7545 * failed the kmalloc call, then it can pass in doms_new == NULL,
7546 * and partition_sched_domains() will fallback to the single partition
7549 * Call with hotplug lock held
7551 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new,
7552 struct sched_domain_attr *dattr_new)
7556 mutex_lock(&sched_domains_mutex);
7558 /* always unregister in case we don't destroy any domains */
7559 unregister_sched_domain_sysctl();
7561 if (doms_new == NULL) {
7563 doms_new = &fallback_doms;
7564 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
7568 /* Destroy deleted domains */
7569 for (i = 0; i < ndoms_cur; i++) {
7570 for (j = 0; j < ndoms_new; j++) {
7571 if (cpus_equal(doms_cur[i], doms_new[j])
7572 && dattrs_equal(dattr_cur, i, dattr_new, j))
7575 /* no match - a current sched domain not in new doms_new[] */
7576 detach_destroy_domains(doms_cur + i);
7581 /* Build new domains */
7582 for (i = 0; i < ndoms_new; i++) {
7583 for (j = 0; j < ndoms_cur; j++) {
7584 if (cpus_equal(doms_new[i], doms_cur[j])
7585 && dattrs_equal(dattr_new, i, dattr_cur, j))
7588 /* no match - add a new doms_new */
7589 __build_sched_domains(doms_new + i,
7590 dattr_new ? dattr_new + i : NULL);
7595 /* Remember the new sched domains */
7596 if (doms_cur != &fallback_doms)
7598 kfree(dattr_cur); /* kfree(NULL) is safe */
7599 doms_cur = doms_new;
7600 dattr_cur = dattr_new;
7601 ndoms_cur = ndoms_new;
7603 register_sched_domain_sysctl();
7605 mutex_unlock(&sched_domains_mutex);
7608 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7609 int arch_reinit_sched_domains(void)
7614 mutex_lock(&sched_domains_mutex);
7615 detach_destroy_domains(&cpu_online_map);
7616 free_sched_domains();
7617 err = arch_init_sched_domains(&cpu_online_map);
7618 mutex_unlock(&sched_domains_mutex);
7624 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7628 if (buf[0] != '0' && buf[0] != '1')
7632 sched_smt_power_savings = (buf[0] == '1');
7634 sched_mc_power_savings = (buf[0] == '1');
7636 ret = arch_reinit_sched_domains();
7638 return ret ? ret : count;
7641 #ifdef CONFIG_SCHED_MC
7642 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
7644 return sprintf(page, "%u\n", sched_mc_power_savings);
7646 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
7647 const char *buf, size_t count)
7649 return sched_power_savings_store(buf, count, 0);
7651 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
7652 sched_mc_power_savings_store);
7655 #ifdef CONFIG_SCHED_SMT
7656 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
7658 return sprintf(page, "%u\n", sched_smt_power_savings);
7660 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
7661 const char *buf, size_t count)
7663 return sched_power_savings_store(buf, count, 1);
7665 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
7666 sched_smt_power_savings_store);
7669 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7673 #ifdef CONFIG_SCHED_SMT
7675 err = sysfs_create_file(&cls->kset.kobj,
7676 &attr_sched_smt_power_savings.attr);
7678 #ifdef CONFIG_SCHED_MC
7679 if (!err && mc_capable())
7680 err = sysfs_create_file(&cls->kset.kobj,
7681 &attr_sched_mc_power_savings.attr);
7685 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7688 * Force a reinitialization of the sched domains hierarchy. The domains
7689 * and groups cannot be updated in place without racing with the balancing
7690 * code, so we temporarily attach all running cpus to the NULL domain
7691 * which will prevent rebalancing while the sched domains are recalculated.
7693 static int update_sched_domains(struct notifier_block *nfb,
7694 unsigned long action, void *hcpu)
7696 int cpu = (int)(long)hcpu;
7699 case CPU_DOWN_PREPARE:
7700 case CPU_DOWN_PREPARE_FROZEN:
7701 disable_runtime(cpu_rq(cpu));
7703 case CPU_UP_PREPARE:
7704 case CPU_UP_PREPARE_FROZEN:
7705 detach_destroy_domains(&cpu_online_map);
7706 free_sched_domains();
7710 case CPU_DOWN_FAILED:
7711 case CPU_DOWN_FAILED_FROZEN:
7713 case CPU_ONLINE_FROZEN:
7714 enable_runtime(cpu_rq(cpu));
7716 case CPU_UP_CANCELED:
7717 case CPU_UP_CANCELED_FROZEN:
7719 case CPU_DEAD_FROZEN:
7721 * Fall through and re-initialise the domains.
7728 #ifndef CONFIG_CPUSETS
7730 * Create default domain partitioning if cpusets are disabled.
7731 * Otherwise we let cpusets rebuild the domains based on the
7735 /* The hotplug lock is already held by cpu_up/cpu_down */
7736 arch_init_sched_domains(&cpu_online_map);
7742 void __init sched_init_smp(void)
7744 cpumask_t non_isolated_cpus;
7746 #if defined(CONFIG_NUMA)
7747 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7749 BUG_ON(sched_group_nodes_bycpu == NULL);
7752 mutex_lock(&sched_domains_mutex);
7753 arch_init_sched_domains(&cpu_online_map);
7754 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
7755 if (cpus_empty(non_isolated_cpus))
7756 cpu_set(smp_processor_id(), non_isolated_cpus);
7757 mutex_unlock(&sched_domains_mutex);
7759 /* XXX: Theoretical race here - CPU may be hotplugged now */
7760 hotcpu_notifier(update_sched_domains, 0);
7763 /* Move init over to a non-isolated CPU */
7764 if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0)
7766 sched_init_granularity();
7769 void __init sched_init_smp(void)
7771 sched_init_granularity();
7773 #endif /* CONFIG_SMP */
7775 int in_sched_functions(unsigned long addr)
7777 return in_lock_functions(addr) ||
7778 (addr >= (unsigned long)__sched_text_start
7779 && addr < (unsigned long)__sched_text_end);
7782 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7784 cfs_rq->tasks_timeline = RB_ROOT;
7785 INIT_LIST_HEAD(&cfs_rq->tasks);
7786 #ifdef CONFIG_FAIR_GROUP_SCHED
7789 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7792 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7794 struct rt_prio_array *array;
7797 array = &rt_rq->active;
7798 for (i = 0; i < MAX_RT_PRIO; i++) {
7799 INIT_LIST_HEAD(array->queue + i);
7800 __clear_bit(i, array->bitmap);
7802 /* delimiter for bitsearch: */
7803 __set_bit(MAX_RT_PRIO, array->bitmap);
7805 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7806 rt_rq->highest_prio = MAX_RT_PRIO;
7809 rt_rq->rt_nr_migratory = 0;
7810 rt_rq->overloaded = 0;
7814 rt_rq->rt_throttled = 0;
7815 rt_rq->rt_runtime = 0;
7816 spin_lock_init(&rt_rq->rt_runtime_lock);
7818 #ifdef CONFIG_RT_GROUP_SCHED
7819 rt_rq->rt_nr_boosted = 0;
7824 #ifdef CONFIG_FAIR_GROUP_SCHED
7825 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7826 struct sched_entity *se, int cpu, int add,
7827 struct sched_entity *parent)
7829 struct rq *rq = cpu_rq(cpu);
7830 tg->cfs_rq[cpu] = cfs_rq;
7831 init_cfs_rq(cfs_rq, rq);
7834 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7837 /* se could be NULL for init_task_group */
7842 se->cfs_rq = &rq->cfs;
7844 se->cfs_rq = parent->my_q;
7847 se->load.weight = tg->shares;
7848 se->load.inv_weight = 0;
7849 se->parent = parent;
7853 #ifdef CONFIG_RT_GROUP_SCHED
7854 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7855 struct sched_rt_entity *rt_se, int cpu, int add,
7856 struct sched_rt_entity *parent)
7858 struct rq *rq = cpu_rq(cpu);
7860 tg->rt_rq[cpu] = rt_rq;
7861 init_rt_rq(rt_rq, rq);
7863 rt_rq->rt_se = rt_se;
7864 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7866 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7868 tg->rt_se[cpu] = rt_se;
7873 rt_se->rt_rq = &rq->rt;
7875 rt_se->rt_rq = parent->my_q;
7877 rt_se->my_q = rt_rq;
7878 rt_se->parent = parent;
7879 INIT_LIST_HEAD(&rt_se->run_list);
7883 void __init sched_init(void)
7886 unsigned long alloc_size = 0, ptr;
7888 #ifdef CONFIG_FAIR_GROUP_SCHED
7889 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7891 #ifdef CONFIG_RT_GROUP_SCHED
7892 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7894 #ifdef CONFIG_USER_SCHED
7898 * As sched_init() is called before page_alloc is setup,
7899 * we use alloc_bootmem().
7902 ptr = (unsigned long)alloc_bootmem(alloc_size);
7904 #ifdef CONFIG_FAIR_GROUP_SCHED
7905 init_task_group.se = (struct sched_entity **)ptr;
7906 ptr += nr_cpu_ids * sizeof(void **);
7908 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7909 ptr += nr_cpu_ids * sizeof(void **);
7911 #ifdef CONFIG_USER_SCHED
7912 root_task_group.se = (struct sched_entity **)ptr;
7913 ptr += nr_cpu_ids * sizeof(void **);
7915 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7916 ptr += nr_cpu_ids * sizeof(void **);
7917 #endif /* CONFIG_USER_SCHED */
7918 #endif /* CONFIG_FAIR_GROUP_SCHED */
7919 #ifdef CONFIG_RT_GROUP_SCHED
7920 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7921 ptr += nr_cpu_ids * sizeof(void **);
7923 init_task_group.rt_rq = (struct rt_rq **)ptr;
7924 ptr += nr_cpu_ids * sizeof(void **);
7926 #ifdef CONFIG_USER_SCHED
7927 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7928 ptr += nr_cpu_ids * sizeof(void **);
7930 root_task_group.rt_rq = (struct rt_rq **)ptr;
7931 ptr += nr_cpu_ids * sizeof(void **);
7932 #endif /* CONFIG_USER_SCHED */
7933 #endif /* CONFIG_RT_GROUP_SCHED */
7937 init_defrootdomain();
7940 init_rt_bandwidth(&def_rt_bandwidth,
7941 global_rt_period(), global_rt_runtime());
7943 #ifdef CONFIG_RT_GROUP_SCHED
7944 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7945 global_rt_period(), global_rt_runtime());
7946 #ifdef CONFIG_USER_SCHED
7947 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7948 global_rt_period(), RUNTIME_INF);
7949 #endif /* CONFIG_USER_SCHED */
7950 #endif /* CONFIG_RT_GROUP_SCHED */
7952 #ifdef CONFIG_GROUP_SCHED
7953 list_add(&init_task_group.list, &task_groups);
7954 INIT_LIST_HEAD(&init_task_group.children);
7956 #ifdef CONFIG_USER_SCHED
7957 INIT_LIST_HEAD(&root_task_group.children);
7958 init_task_group.parent = &root_task_group;
7959 list_add(&init_task_group.siblings, &root_task_group.children);
7960 #endif /* CONFIG_USER_SCHED */
7961 #endif /* CONFIG_GROUP_SCHED */
7963 for_each_possible_cpu(i) {
7967 spin_lock_init(&rq->lock);
7968 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
7970 init_cfs_rq(&rq->cfs, rq);
7971 init_rt_rq(&rq->rt, rq);
7972 #ifdef CONFIG_FAIR_GROUP_SCHED
7973 init_task_group.shares = init_task_group_load;
7974 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7975 #ifdef CONFIG_CGROUP_SCHED
7977 * How much cpu bandwidth does init_task_group get?
7979 * In case of task-groups formed thr' the cgroup filesystem, it
7980 * gets 100% of the cpu resources in the system. This overall
7981 * system cpu resource is divided among the tasks of
7982 * init_task_group and its child task-groups in a fair manner,
7983 * based on each entity's (task or task-group's) weight
7984 * (se->load.weight).
7986 * In other words, if init_task_group has 10 tasks of weight
7987 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7988 * then A0's share of the cpu resource is:
7990 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7992 * We achieve this by letting init_task_group's tasks sit
7993 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7995 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
7996 #elif defined CONFIG_USER_SCHED
7997 root_task_group.shares = NICE_0_LOAD;
7998 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
8000 * In case of task-groups formed thr' the user id of tasks,
8001 * init_task_group represents tasks belonging to root user.
8002 * Hence it forms a sibling of all subsequent groups formed.
8003 * In this case, init_task_group gets only a fraction of overall
8004 * system cpu resource, based on the weight assigned to root
8005 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8006 * by letting tasks of init_task_group sit in a separate cfs_rq
8007 * (init_cfs_rq) and having one entity represent this group of
8008 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8010 init_tg_cfs_entry(&init_task_group,
8011 &per_cpu(init_cfs_rq, i),
8012 &per_cpu(init_sched_entity, i), i, 1,
8013 root_task_group.se[i]);
8016 #endif /* CONFIG_FAIR_GROUP_SCHED */
8018 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8019 #ifdef CONFIG_RT_GROUP_SCHED
8020 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8021 #ifdef CONFIG_CGROUP_SCHED
8022 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
8023 #elif defined CONFIG_USER_SCHED
8024 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
8025 init_tg_rt_entry(&init_task_group,
8026 &per_cpu(init_rt_rq, i),
8027 &per_cpu(init_sched_rt_entity, i), i, 1,
8028 root_task_group.rt_se[i]);
8032 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8033 rq->cpu_load[j] = 0;
8037 rq->active_balance = 0;
8038 rq->next_balance = jiffies;
8042 rq->migration_thread = NULL;
8043 INIT_LIST_HEAD(&rq->migration_queue);
8044 rq_attach_root(rq, &def_root_domain);
8047 atomic_set(&rq->nr_iowait, 0);
8050 set_load_weight(&init_task);
8052 #ifdef CONFIG_PREEMPT_NOTIFIERS
8053 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8057 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
8060 #ifdef CONFIG_RT_MUTEXES
8061 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
8065 * The boot idle thread does lazy MMU switching as well:
8067 atomic_inc(&init_mm.mm_count);
8068 enter_lazy_tlb(&init_mm, current);
8071 * Make us the idle thread. Technically, schedule() should not be
8072 * called from this thread, however somewhere below it might be,
8073 * but because we are the idle thread, we just pick up running again
8074 * when this runqueue becomes "idle".
8076 init_idle(current, smp_processor_id());
8078 * During early bootup we pretend to be a normal task:
8080 current->sched_class = &fair_sched_class;
8082 scheduler_running = 1;
8085 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8086 void __might_sleep(char *file, int line)
8089 static unsigned long prev_jiffy; /* ratelimiting */
8091 if ((in_atomic() || irqs_disabled()) &&
8092 system_state == SYSTEM_RUNNING && !oops_in_progress) {
8093 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8095 prev_jiffy = jiffies;
8096 printk(KERN_ERR "BUG: sleeping function called from invalid"
8097 " context at %s:%d\n", file, line);
8098 printk("in_atomic():%d, irqs_disabled():%d\n",
8099 in_atomic(), irqs_disabled());
8100 debug_show_held_locks(current);
8101 if (irqs_disabled())
8102 print_irqtrace_events(current);
8107 EXPORT_SYMBOL(__might_sleep);
8110 #ifdef CONFIG_MAGIC_SYSRQ
8111 static void normalize_task(struct rq *rq, struct task_struct *p)
8115 update_rq_clock(rq);
8116 on_rq = p->se.on_rq;
8118 deactivate_task(rq, p, 0);
8119 __setscheduler(rq, p, SCHED_NORMAL, 0);
8121 activate_task(rq, p, 0);
8122 resched_task(rq->curr);
8126 void normalize_rt_tasks(void)
8128 struct task_struct *g, *p;
8129 unsigned long flags;
8132 read_lock_irqsave(&tasklist_lock, flags);
8133 do_each_thread(g, p) {
8135 * Only normalize user tasks:
8140 p->se.exec_start = 0;
8141 #ifdef CONFIG_SCHEDSTATS
8142 p->se.wait_start = 0;
8143 p->se.sleep_start = 0;
8144 p->se.block_start = 0;
8149 * Renice negative nice level userspace
8152 if (TASK_NICE(p) < 0 && p->mm)
8153 set_user_nice(p, 0);
8157 spin_lock(&p->pi_lock);
8158 rq = __task_rq_lock(p);
8160 normalize_task(rq, p);
8162 __task_rq_unlock(rq);
8163 spin_unlock(&p->pi_lock);
8164 } while_each_thread(g, p);
8166 read_unlock_irqrestore(&tasklist_lock, flags);
8169 #endif /* CONFIG_MAGIC_SYSRQ */
8173 * These functions are only useful for the IA64 MCA handling.
8175 * They can only be called when the whole system has been
8176 * stopped - every CPU needs to be quiescent, and no scheduling
8177 * activity can take place. Using them for anything else would
8178 * be a serious bug, and as a result, they aren't even visible
8179 * under any other configuration.
8183 * curr_task - return the current task for a given cpu.
8184 * @cpu: the processor in question.
8186 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8188 struct task_struct *curr_task(int cpu)
8190 return cpu_curr(cpu);
8194 * set_curr_task - set the current task for a given cpu.
8195 * @cpu: the processor in question.
8196 * @p: the task pointer to set.
8198 * Description: This function must only be used when non-maskable interrupts
8199 * are serviced on a separate stack. It allows the architecture to switch the
8200 * notion of the current task on a cpu in a non-blocking manner. This function
8201 * must be called with all CPU's synchronized, and interrupts disabled, the
8202 * and caller must save the original value of the current task (see
8203 * curr_task() above) and restore that value before reenabling interrupts and
8204 * re-starting the system.
8206 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8208 void set_curr_task(int cpu, struct task_struct *p)
8215 #ifdef CONFIG_FAIR_GROUP_SCHED
8216 static void free_fair_sched_group(struct task_group *tg)
8220 for_each_possible_cpu(i) {
8222 kfree(tg->cfs_rq[i]);
8232 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8234 struct cfs_rq *cfs_rq;
8235 struct sched_entity *se, *parent_se;
8239 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8242 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8246 tg->shares = NICE_0_LOAD;
8248 for_each_possible_cpu(i) {
8251 cfs_rq = kmalloc_node(sizeof(struct cfs_rq),
8252 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8256 se = kmalloc_node(sizeof(struct sched_entity),
8257 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8261 parent_se = parent ? parent->se[i] : NULL;
8262 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent_se);
8271 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8273 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8274 &cpu_rq(cpu)->leaf_cfs_rq_list);
8277 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8279 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8281 #else /* !CONFG_FAIR_GROUP_SCHED */
8282 static inline void free_fair_sched_group(struct task_group *tg)
8287 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8292 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8296 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8299 #endif /* CONFIG_FAIR_GROUP_SCHED */
8301 #ifdef CONFIG_RT_GROUP_SCHED
8302 static void free_rt_sched_group(struct task_group *tg)
8306 destroy_rt_bandwidth(&tg->rt_bandwidth);
8308 for_each_possible_cpu(i) {
8310 kfree(tg->rt_rq[i]);
8312 kfree(tg->rt_se[i]);
8320 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8322 struct rt_rq *rt_rq;
8323 struct sched_rt_entity *rt_se, *parent_se;
8327 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8330 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8334 init_rt_bandwidth(&tg->rt_bandwidth,
8335 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8337 for_each_possible_cpu(i) {
8340 rt_rq = kmalloc_node(sizeof(struct rt_rq),
8341 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8345 rt_se = kmalloc_node(sizeof(struct sched_rt_entity),
8346 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8350 parent_se = parent ? parent->rt_se[i] : NULL;
8351 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent_se);
8360 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8362 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8363 &cpu_rq(cpu)->leaf_rt_rq_list);
8366 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8368 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8370 #else /* !CONFIG_RT_GROUP_SCHED */
8371 static inline void free_rt_sched_group(struct task_group *tg)
8376 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8381 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8385 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8388 #endif /* CONFIG_RT_GROUP_SCHED */
8390 #ifdef CONFIG_GROUP_SCHED
8391 static void free_sched_group(struct task_group *tg)
8393 free_fair_sched_group(tg);
8394 free_rt_sched_group(tg);
8398 /* allocate runqueue etc for a new task group */
8399 struct task_group *sched_create_group(struct task_group *parent)
8401 struct task_group *tg;
8402 unsigned long flags;
8405 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8407 return ERR_PTR(-ENOMEM);
8409 if (!alloc_fair_sched_group(tg, parent))
8412 if (!alloc_rt_sched_group(tg, parent))
8415 spin_lock_irqsave(&task_group_lock, flags);
8416 for_each_possible_cpu(i) {
8417 register_fair_sched_group(tg, i);
8418 register_rt_sched_group(tg, i);
8420 list_add_rcu(&tg->list, &task_groups);
8422 WARN_ON(!parent); /* root should already exist */
8424 tg->parent = parent;
8425 list_add_rcu(&tg->siblings, &parent->children);
8426 INIT_LIST_HEAD(&tg->children);
8427 spin_unlock_irqrestore(&task_group_lock, flags);
8432 free_sched_group(tg);
8433 return ERR_PTR(-ENOMEM);
8436 /* rcu callback to free various structures associated with a task group */
8437 static void free_sched_group_rcu(struct rcu_head *rhp)
8439 /* now it should be safe to free those cfs_rqs */
8440 free_sched_group(container_of(rhp, struct task_group, rcu));
8443 /* Destroy runqueue etc associated with a task group */
8444 void sched_destroy_group(struct task_group *tg)
8446 unsigned long flags;
8449 spin_lock_irqsave(&task_group_lock, flags);
8450 for_each_possible_cpu(i) {
8451 unregister_fair_sched_group(tg, i);
8452 unregister_rt_sched_group(tg, i);
8454 list_del_rcu(&tg->list);
8455 list_del_rcu(&tg->siblings);
8456 spin_unlock_irqrestore(&task_group_lock, flags);
8458 /* wait for possible concurrent references to cfs_rqs complete */
8459 call_rcu(&tg->rcu, free_sched_group_rcu);
8462 /* change task's runqueue when it moves between groups.
8463 * The caller of this function should have put the task in its new group
8464 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8465 * reflect its new group.
8467 void sched_move_task(struct task_struct *tsk)
8470 unsigned long flags;
8473 rq = task_rq_lock(tsk, &flags);
8475 update_rq_clock(rq);
8477 running = task_current(rq, tsk);
8478 on_rq = tsk->se.on_rq;
8481 dequeue_task(rq, tsk, 0);
8482 if (unlikely(running))
8483 tsk->sched_class->put_prev_task(rq, tsk);
8485 set_task_rq(tsk, task_cpu(tsk));
8487 #ifdef CONFIG_FAIR_GROUP_SCHED
8488 if (tsk->sched_class->moved_group)
8489 tsk->sched_class->moved_group(tsk);
8492 if (unlikely(running))
8493 tsk->sched_class->set_curr_task(rq);
8495 enqueue_task(rq, tsk, 0);
8497 task_rq_unlock(rq, &flags);
8499 #endif /* CONFIG_GROUP_SCHED */
8501 #ifdef CONFIG_FAIR_GROUP_SCHED
8502 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8504 struct cfs_rq *cfs_rq = se->cfs_rq;
8509 dequeue_entity(cfs_rq, se, 0);
8511 se->load.weight = shares;
8512 se->load.inv_weight = 0;
8515 enqueue_entity(cfs_rq, se, 0);
8518 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8520 struct cfs_rq *cfs_rq = se->cfs_rq;
8521 struct rq *rq = cfs_rq->rq;
8522 unsigned long flags;
8524 spin_lock_irqsave(&rq->lock, flags);
8525 __set_se_shares(se, shares);
8526 spin_unlock_irqrestore(&rq->lock, flags);
8529 static DEFINE_MUTEX(shares_mutex);
8531 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8534 unsigned long flags;
8537 * We can't change the weight of the root cgroup.
8542 if (shares < MIN_SHARES)
8543 shares = MIN_SHARES;
8544 else if (shares > MAX_SHARES)
8545 shares = MAX_SHARES;
8547 mutex_lock(&shares_mutex);
8548 if (tg->shares == shares)
8551 spin_lock_irqsave(&task_group_lock, flags);
8552 for_each_possible_cpu(i)
8553 unregister_fair_sched_group(tg, i);
8554 list_del_rcu(&tg->siblings);
8555 spin_unlock_irqrestore(&task_group_lock, flags);
8557 /* wait for any ongoing reference to this group to finish */
8558 synchronize_sched();
8561 * Now we are free to modify the group's share on each cpu
8562 * w/o tripping rebalance_share or load_balance_fair.
8564 tg->shares = shares;
8565 for_each_possible_cpu(i) {
8569 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8570 set_se_shares(tg->se[i], shares);
8574 * Enable load balance activity on this group, by inserting it back on
8575 * each cpu's rq->leaf_cfs_rq_list.
8577 spin_lock_irqsave(&task_group_lock, flags);
8578 for_each_possible_cpu(i)
8579 register_fair_sched_group(tg, i);
8580 list_add_rcu(&tg->siblings, &tg->parent->children);
8581 spin_unlock_irqrestore(&task_group_lock, flags);
8583 mutex_unlock(&shares_mutex);
8587 unsigned long sched_group_shares(struct task_group *tg)
8593 #ifdef CONFIG_RT_GROUP_SCHED
8595 * Ensure that the real time constraints are schedulable.
8597 static DEFINE_MUTEX(rt_constraints_mutex);
8599 static unsigned long to_ratio(u64 period, u64 runtime)
8601 if (runtime == RUNTIME_INF)
8604 return div64_u64(runtime << 16, period);
8607 #ifdef CONFIG_CGROUP_SCHED
8608 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8610 struct task_group *tgi, *parent = tg->parent;
8611 unsigned long total = 0;
8614 if (global_rt_period() < period)
8617 return to_ratio(period, runtime) <
8618 to_ratio(global_rt_period(), global_rt_runtime());
8621 if (ktime_to_ns(parent->rt_bandwidth.rt_period) < period)
8625 list_for_each_entry_rcu(tgi, &parent->children, siblings) {
8629 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8630 tgi->rt_bandwidth.rt_runtime);
8634 return total + to_ratio(period, runtime) <=
8635 to_ratio(ktime_to_ns(parent->rt_bandwidth.rt_period),
8636 parent->rt_bandwidth.rt_runtime);
8638 #elif defined CONFIG_USER_SCHED
8639 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8641 struct task_group *tgi;
8642 unsigned long total = 0;
8643 unsigned long global_ratio =
8644 to_ratio(global_rt_period(), global_rt_runtime());
8647 list_for_each_entry_rcu(tgi, &task_groups, list) {
8651 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8652 tgi->rt_bandwidth.rt_runtime);
8656 return total + to_ratio(period, runtime) < global_ratio;
8660 /* Must be called with tasklist_lock held */
8661 static inline int tg_has_rt_tasks(struct task_group *tg)
8663 struct task_struct *g, *p;
8664 do_each_thread(g, p) {
8665 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8667 } while_each_thread(g, p);
8671 static int tg_set_bandwidth(struct task_group *tg,
8672 u64 rt_period, u64 rt_runtime)
8676 mutex_lock(&rt_constraints_mutex);
8677 read_lock(&tasklist_lock);
8678 if (rt_runtime == 0 && tg_has_rt_tasks(tg)) {
8682 if (!__rt_schedulable(tg, rt_period, rt_runtime)) {
8687 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8688 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8689 tg->rt_bandwidth.rt_runtime = rt_runtime;
8691 for_each_possible_cpu(i) {
8692 struct rt_rq *rt_rq = tg->rt_rq[i];
8694 spin_lock(&rt_rq->rt_runtime_lock);
8695 rt_rq->rt_runtime = rt_runtime;
8696 spin_unlock(&rt_rq->rt_runtime_lock);
8698 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8700 read_unlock(&tasklist_lock);
8701 mutex_unlock(&rt_constraints_mutex);
8706 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8708 u64 rt_runtime, rt_period;
8710 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8711 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8712 if (rt_runtime_us < 0)
8713 rt_runtime = RUNTIME_INF;
8715 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8718 long sched_group_rt_runtime(struct task_group *tg)
8722 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8725 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8726 do_div(rt_runtime_us, NSEC_PER_USEC);
8727 return rt_runtime_us;
8730 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8732 u64 rt_runtime, rt_period;
8734 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8735 rt_runtime = tg->rt_bandwidth.rt_runtime;
8737 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8740 long sched_group_rt_period(struct task_group *tg)
8744 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8745 do_div(rt_period_us, NSEC_PER_USEC);
8746 return rt_period_us;
8749 static int sched_rt_global_constraints(void)
8751 struct task_group *tg = &root_task_group;
8752 u64 rt_runtime, rt_period;
8755 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8756 rt_runtime = tg->rt_bandwidth.rt_runtime;
8758 mutex_lock(&rt_constraints_mutex);
8759 if (!__rt_schedulable(tg, rt_period, rt_runtime))
8761 mutex_unlock(&rt_constraints_mutex);
8765 #else /* !CONFIG_RT_GROUP_SCHED */
8766 static int sched_rt_global_constraints(void)
8768 unsigned long flags;
8771 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8772 for_each_possible_cpu(i) {
8773 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8775 spin_lock(&rt_rq->rt_runtime_lock);
8776 rt_rq->rt_runtime = global_rt_runtime();
8777 spin_unlock(&rt_rq->rt_runtime_lock);
8779 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8783 #endif /* CONFIG_RT_GROUP_SCHED */
8785 int sched_rt_handler(struct ctl_table *table, int write,
8786 struct file *filp, void __user *buffer, size_t *lenp,
8790 int old_period, old_runtime;
8791 static DEFINE_MUTEX(mutex);
8794 old_period = sysctl_sched_rt_period;
8795 old_runtime = sysctl_sched_rt_runtime;
8797 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
8799 if (!ret && write) {
8800 ret = sched_rt_global_constraints();
8802 sysctl_sched_rt_period = old_period;
8803 sysctl_sched_rt_runtime = old_runtime;
8805 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8806 def_rt_bandwidth.rt_period =
8807 ns_to_ktime(global_rt_period());
8810 mutex_unlock(&mutex);
8815 #ifdef CONFIG_CGROUP_SCHED
8817 /* return corresponding task_group object of a cgroup */
8818 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8820 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8821 struct task_group, css);
8824 static struct cgroup_subsys_state *
8825 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8827 struct task_group *tg, *parent;
8829 if (!cgrp->parent) {
8830 /* This is early initialization for the top cgroup */
8831 init_task_group.css.cgroup = cgrp;
8832 return &init_task_group.css;
8835 parent = cgroup_tg(cgrp->parent);
8836 tg = sched_create_group(parent);
8838 return ERR_PTR(-ENOMEM);
8840 /* Bind the cgroup to task_group object we just created */
8841 tg->css.cgroup = cgrp;
8847 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8849 struct task_group *tg = cgroup_tg(cgrp);
8851 sched_destroy_group(tg);
8855 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8856 struct task_struct *tsk)
8858 #ifdef CONFIG_RT_GROUP_SCHED
8859 /* Don't accept realtime tasks when there is no way for them to run */
8860 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
8863 /* We don't support RT-tasks being in separate groups */
8864 if (tsk->sched_class != &fair_sched_class)
8872 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8873 struct cgroup *old_cont, struct task_struct *tsk)
8875 sched_move_task(tsk);
8878 #ifdef CONFIG_FAIR_GROUP_SCHED
8879 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8882 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8885 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8887 struct task_group *tg = cgroup_tg(cgrp);
8889 return (u64) tg->shares;
8891 #endif /* CONFIG_FAIR_GROUP_SCHED */
8893 #ifdef CONFIG_RT_GROUP_SCHED
8894 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8897 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8900 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8902 return sched_group_rt_runtime(cgroup_tg(cgrp));
8905 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8908 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8911 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8913 return sched_group_rt_period(cgroup_tg(cgrp));
8915 #endif /* CONFIG_RT_GROUP_SCHED */
8917 static struct cftype cpu_files[] = {
8918 #ifdef CONFIG_FAIR_GROUP_SCHED
8921 .read_u64 = cpu_shares_read_u64,
8922 .write_u64 = cpu_shares_write_u64,
8925 #ifdef CONFIG_RT_GROUP_SCHED
8927 .name = "rt_runtime_us",
8928 .read_s64 = cpu_rt_runtime_read,
8929 .write_s64 = cpu_rt_runtime_write,
8932 .name = "rt_period_us",
8933 .read_u64 = cpu_rt_period_read_uint,
8934 .write_u64 = cpu_rt_period_write_uint,
8939 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8941 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8944 struct cgroup_subsys cpu_cgroup_subsys = {
8946 .create = cpu_cgroup_create,
8947 .destroy = cpu_cgroup_destroy,
8948 .can_attach = cpu_cgroup_can_attach,
8949 .attach = cpu_cgroup_attach,
8950 .populate = cpu_cgroup_populate,
8951 .subsys_id = cpu_cgroup_subsys_id,
8955 #endif /* CONFIG_CGROUP_SCHED */
8957 #ifdef CONFIG_CGROUP_CPUACCT
8960 * CPU accounting code for task groups.
8962 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8963 * (balbir@in.ibm.com).
8966 /* track cpu usage of a group of tasks */
8968 struct cgroup_subsys_state css;
8969 /* cpuusage holds pointer to a u64-type object on every cpu */
8973 struct cgroup_subsys cpuacct_subsys;
8975 /* return cpu accounting group corresponding to this container */
8976 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8978 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8979 struct cpuacct, css);
8982 /* return cpu accounting group to which this task belongs */
8983 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8985 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8986 struct cpuacct, css);
8989 /* create a new cpu accounting group */
8990 static struct cgroup_subsys_state *cpuacct_create(
8991 struct cgroup_subsys *ss, struct cgroup *cgrp)
8993 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8996 return ERR_PTR(-ENOMEM);
8998 ca->cpuusage = alloc_percpu(u64);
8999 if (!ca->cpuusage) {
9001 return ERR_PTR(-ENOMEM);
9007 /* destroy an existing cpu accounting group */
9009 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9011 struct cpuacct *ca = cgroup_ca(cgrp);
9013 free_percpu(ca->cpuusage);
9017 /* return total cpu usage (in nanoseconds) of a group */
9018 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9020 struct cpuacct *ca = cgroup_ca(cgrp);
9021 u64 totalcpuusage = 0;
9024 for_each_possible_cpu(i) {
9025 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9028 * Take rq->lock to make 64-bit addition safe on 32-bit
9031 spin_lock_irq(&cpu_rq(i)->lock);
9032 totalcpuusage += *cpuusage;
9033 spin_unlock_irq(&cpu_rq(i)->lock);
9036 return totalcpuusage;
9039 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9042 struct cpuacct *ca = cgroup_ca(cgrp);
9051 for_each_possible_cpu(i) {
9052 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9054 spin_lock_irq(&cpu_rq(i)->lock);
9056 spin_unlock_irq(&cpu_rq(i)->lock);
9062 static struct cftype files[] = {
9065 .read_u64 = cpuusage_read,
9066 .write_u64 = cpuusage_write,
9070 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9072 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9076 * charge this task's execution time to its accounting group.
9078 * called with rq->lock held.
9080 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9084 if (!cpuacct_subsys.active)
9089 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
9091 *cpuusage += cputime;
9095 struct cgroup_subsys cpuacct_subsys = {
9097 .create = cpuacct_create,
9098 .destroy = cpuacct_destroy,
9099 .populate = cpuacct_populate,
9100 .subsys_id = cpuacct_subsys_id,
9102 #endif /* CONFIG_CGROUP_CPUACCT */