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 #ifdef CONFIG_FAIR_GROUP_SCHED
1645 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1648 cfs_rq->shares = shares;
1653 #include "sched_stats.h"
1654 #include "sched_idletask.c"
1655 #include "sched_fair.c"
1656 #include "sched_rt.c"
1657 #ifdef CONFIG_SCHED_DEBUG
1658 # include "sched_debug.c"
1661 #define sched_class_highest (&rt_sched_class)
1662 #define for_each_class(class) \
1663 for (class = sched_class_highest; class; class = class->next)
1665 static void inc_nr_running(struct rq *rq)
1670 static void dec_nr_running(struct rq *rq)
1675 static void set_load_weight(struct task_struct *p)
1677 if (task_has_rt_policy(p)) {
1678 p->se.load.weight = prio_to_weight[0] * 2;
1679 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1684 * SCHED_IDLE tasks get minimal weight:
1686 if (p->policy == SCHED_IDLE) {
1687 p->se.load.weight = WEIGHT_IDLEPRIO;
1688 p->se.load.inv_weight = WMULT_IDLEPRIO;
1692 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1693 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1696 static void update_avg(u64 *avg, u64 sample)
1698 s64 diff = sample - *avg;
1702 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1704 sched_info_queued(p);
1705 p->sched_class->enqueue_task(rq, p, wakeup);
1709 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1711 if (sleep && p->se.last_wakeup) {
1712 update_avg(&p->se.avg_overlap,
1713 p->se.sum_exec_runtime - p->se.last_wakeup);
1714 p->se.last_wakeup = 0;
1717 p->sched_class->dequeue_task(rq, p, sleep);
1722 * __normal_prio - return the priority that is based on the static prio
1724 static inline int __normal_prio(struct task_struct *p)
1726 return p->static_prio;
1730 * Calculate the expected normal priority: i.e. priority
1731 * without taking RT-inheritance into account. Might be
1732 * boosted by interactivity modifiers. Changes upon fork,
1733 * setprio syscalls, and whenever the interactivity
1734 * estimator recalculates.
1736 static inline int normal_prio(struct task_struct *p)
1740 if (task_has_rt_policy(p))
1741 prio = MAX_RT_PRIO-1 - p->rt_priority;
1743 prio = __normal_prio(p);
1748 * Calculate the current priority, i.e. the priority
1749 * taken into account by the scheduler. This value might
1750 * be boosted by RT tasks, or might be boosted by
1751 * interactivity modifiers. Will be RT if the task got
1752 * RT-boosted. If not then it returns p->normal_prio.
1754 static int effective_prio(struct task_struct *p)
1756 p->normal_prio = normal_prio(p);
1758 * If we are RT tasks or we were boosted to RT priority,
1759 * keep the priority unchanged. Otherwise, update priority
1760 * to the normal priority:
1762 if (!rt_prio(p->prio))
1763 return p->normal_prio;
1768 * activate_task - move a task to the runqueue.
1770 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1772 if (task_contributes_to_load(p))
1773 rq->nr_uninterruptible--;
1775 enqueue_task(rq, p, wakeup);
1780 * deactivate_task - remove a task from the runqueue.
1782 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1784 if (task_contributes_to_load(p))
1785 rq->nr_uninterruptible++;
1787 dequeue_task(rq, p, sleep);
1792 * task_curr - is this task currently executing on a CPU?
1793 * @p: the task in question.
1795 inline int task_curr(const struct task_struct *p)
1797 return cpu_curr(task_cpu(p)) == p;
1800 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1802 set_task_rq(p, cpu);
1805 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1806 * successfuly executed on another CPU. We must ensure that updates of
1807 * per-task data have been completed by this moment.
1810 task_thread_info(p)->cpu = cpu;
1814 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1815 const struct sched_class *prev_class,
1816 int oldprio, int running)
1818 if (prev_class != p->sched_class) {
1819 if (prev_class->switched_from)
1820 prev_class->switched_from(rq, p, running);
1821 p->sched_class->switched_to(rq, p, running);
1823 p->sched_class->prio_changed(rq, p, oldprio, running);
1828 /* Used instead of source_load when we know the type == 0 */
1829 static unsigned long weighted_cpuload(const int cpu)
1831 return cpu_rq(cpu)->load.weight;
1835 * Is this task likely cache-hot:
1838 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1843 * Buddy candidates are cache hot:
1845 if (sched_feat(CACHE_HOT_BUDDY) && (&p->se == cfs_rq_of(&p->se)->next))
1848 if (p->sched_class != &fair_sched_class)
1851 if (sysctl_sched_migration_cost == -1)
1853 if (sysctl_sched_migration_cost == 0)
1856 delta = now - p->se.exec_start;
1858 return delta < (s64)sysctl_sched_migration_cost;
1862 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1864 int old_cpu = task_cpu(p);
1865 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1866 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1867 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1870 clock_offset = old_rq->clock - new_rq->clock;
1872 #ifdef CONFIG_SCHEDSTATS
1873 if (p->se.wait_start)
1874 p->se.wait_start -= clock_offset;
1875 if (p->se.sleep_start)
1876 p->se.sleep_start -= clock_offset;
1877 if (p->se.block_start)
1878 p->se.block_start -= clock_offset;
1879 if (old_cpu != new_cpu) {
1880 schedstat_inc(p, se.nr_migrations);
1881 if (task_hot(p, old_rq->clock, NULL))
1882 schedstat_inc(p, se.nr_forced2_migrations);
1885 p->se.vruntime -= old_cfsrq->min_vruntime -
1886 new_cfsrq->min_vruntime;
1888 __set_task_cpu(p, new_cpu);
1891 struct migration_req {
1892 struct list_head list;
1894 struct task_struct *task;
1897 struct completion done;
1901 * The task's runqueue lock must be held.
1902 * Returns true if you have to wait for migration thread.
1905 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1907 struct rq *rq = task_rq(p);
1910 * If the task is not on a runqueue (and not running), then
1911 * it is sufficient to simply update the task's cpu field.
1913 if (!p->se.on_rq && !task_running(rq, p)) {
1914 set_task_cpu(p, dest_cpu);
1918 init_completion(&req->done);
1920 req->dest_cpu = dest_cpu;
1921 list_add(&req->list, &rq->migration_queue);
1927 * wait_task_inactive - wait for a thread to unschedule.
1929 * The caller must ensure that the task *will* unschedule sometime soon,
1930 * else this function might spin for a *long* time. This function can't
1931 * be called with interrupts off, or it may introduce deadlock with
1932 * smp_call_function() if an IPI is sent by the same process we are
1933 * waiting to become inactive.
1935 void wait_task_inactive(struct task_struct *p)
1937 unsigned long flags;
1943 * We do the initial early heuristics without holding
1944 * any task-queue locks at all. We'll only try to get
1945 * the runqueue lock when things look like they will
1951 * If the task is actively running on another CPU
1952 * still, just relax and busy-wait without holding
1955 * NOTE! Since we don't hold any locks, it's not
1956 * even sure that "rq" stays as the right runqueue!
1957 * But we don't care, since "task_running()" will
1958 * return false if the runqueue has changed and p
1959 * is actually now running somewhere else!
1961 while (task_running(rq, p))
1965 * Ok, time to look more closely! We need the rq
1966 * lock now, to be *sure*. If we're wrong, we'll
1967 * just go back and repeat.
1969 rq = task_rq_lock(p, &flags);
1970 running = task_running(rq, p);
1971 on_rq = p->se.on_rq;
1972 task_rq_unlock(rq, &flags);
1975 * Was it really running after all now that we
1976 * checked with the proper locks actually held?
1978 * Oops. Go back and try again..
1980 if (unlikely(running)) {
1986 * It's not enough that it's not actively running,
1987 * it must be off the runqueue _entirely_, and not
1990 * So if it wa still runnable (but just not actively
1991 * running right now), it's preempted, and we should
1992 * yield - it could be a while.
1994 if (unlikely(on_rq)) {
1995 schedule_timeout_uninterruptible(1);
2000 * Ahh, all good. It wasn't running, and it wasn't
2001 * runnable, which means that it will never become
2002 * running in the future either. We're all done!
2009 * kick_process - kick a running thread to enter/exit the kernel
2010 * @p: the to-be-kicked thread
2012 * Cause a process which is running on another CPU to enter
2013 * kernel-mode, without any delay. (to get signals handled.)
2015 * NOTE: this function doesnt have to take the runqueue lock,
2016 * because all it wants to ensure is that the remote task enters
2017 * the kernel. If the IPI races and the task has been migrated
2018 * to another CPU then no harm is done and the purpose has been
2021 void kick_process(struct task_struct *p)
2027 if ((cpu != smp_processor_id()) && task_curr(p))
2028 smp_send_reschedule(cpu);
2033 * Return a low guess at the load of a migration-source cpu weighted
2034 * according to the scheduling class and "nice" value.
2036 * We want to under-estimate the load of migration sources, to
2037 * balance conservatively.
2039 static unsigned long source_load(int cpu, int type)
2041 struct rq *rq = cpu_rq(cpu);
2042 unsigned long total = weighted_cpuload(cpu);
2044 if (type == 0 || !sched_feat(LB_BIAS))
2047 return min(rq->cpu_load[type-1], total);
2051 * Return a high guess at the load of a migration-target cpu weighted
2052 * according to the scheduling class and "nice" value.
2054 static unsigned long target_load(int cpu, int type)
2056 struct rq *rq = cpu_rq(cpu);
2057 unsigned long total = weighted_cpuload(cpu);
2059 if (type == 0 || !sched_feat(LB_BIAS))
2062 return max(rq->cpu_load[type-1], total);
2066 * find_idlest_group finds and returns the least busy CPU group within the
2069 static struct sched_group *
2070 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2072 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2073 unsigned long min_load = ULONG_MAX, this_load = 0;
2074 int load_idx = sd->forkexec_idx;
2075 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2078 unsigned long load, avg_load;
2082 /* Skip over this group if it has no CPUs allowed */
2083 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
2086 local_group = cpu_isset(this_cpu, group->cpumask);
2088 /* Tally up the load of all CPUs in the group */
2091 for_each_cpu_mask(i, group->cpumask) {
2092 /* Bias balancing toward cpus of our domain */
2094 load = source_load(i, load_idx);
2096 load = target_load(i, load_idx);
2101 /* Adjust by relative CPU power of the group */
2102 avg_load = sg_div_cpu_power(group,
2103 avg_load * SCHED_LOAD_SCALE);
2106 this_load = avg_load;
2108 } else if (avg_load < min_load) {
2109 min_load = avg_load;
2112 } while (group = group->next, group != sd->groups);
2114 if (!idlest || 100*this_load < imbalance*min_load)
2120 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2123 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu,
2126 unsigned long load, min_load = ULONG_MAX;
2130 /* Traverse only the allowed CPUs */
2131 cpus_and(*tmp, group->cpumask, p->cpus_allowed);
2133 for_each_cpu_mask(i, *tmp) {
2134 load = weighted_cpuload(i);
2136 if (load < min_load || (load == min_load && i == this_cpu)) {
2146 * sched_balance_self: balance the current task (running on cpu) in domains
2147 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2150 * Balance, ie. select the least loaded group.
2152 * Returns the target CPU number, or the same CPU if no balancing is needed.
2154 * preempt must be disabled.
2156 static int sched_balance_self(int cpu, int flag)
2158 struct task_struct *t = current;
2159 struct sched_domain *tmp, *sd = NULL;
2161 for_each_domain(cpu, tmp) {
2163 * If power savings logic is enabled for a domain, stop there.
2165 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2167 if (tmp->flags & flag)
2175 cpumask_t span, tmpmask;
2176 struct sched_group *group;
2177 int new_cpu, weight;
2179 if (!(sd->flags & flag)) {
2185 group = find_idlest_group(sd, t, cpu);
2191 new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask);
2192 if (new_cpu == -1 || new_cpu == cpu) {
2193 /* Now try balancing at a lower domain level of cpu */
2198 /* Now try balancing at a lower domain level of new_cpu */
2201 weight = cpus_weight(span);
2202 for_each_domain(cpu, tmp) {
2203 if (weight <= cpus_weight(tmp->span))
2205 if (tmp->flags & flag)
2208 /* while loop will break here if sd == NULL */
2214 #endif /* CONFIG_SMP */
2217 * try_to_wake_up - wake up a thread
2218 * @p: the to-be-woken-up thread
2219 * @state: the mask of task states that can be woken
2220 * @sync: do a synchronous wakeup?
2222 * Put it on the run-queue if it's not already there. The "current"
2223 * thread is always on the run-queue (except when the actual
2224 * re-schedule is in progress), and as such you're allowed to do
2225 * the simpler "current->state = TASK_RUNNING" to mark yourself
2226 * runnable without the overhead of this.
2228 * returns failure only if the task is already active.
2230 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2232 int cpu, orig_cpu, this_cpu, success = 0;
2233 unsigned long flags;
2237 if (!sched_feat(SYNC_WAKEUPS))
2241 if (sched_feat(LB_WAKEUP_UPDATE)) {
2242 struct sched_domain *sd;
2244 this_cpu = raw_smp_processor_id();
2247 for_each_domain(this_cpu, sd) {
2248 if (cpu_isset(cpu, sd->span)) {
2257 rq = task_rq_lock(p, &flags);
2258 old_state = p->state;
2259 if (!(old_state & state))
2267 this_cpu = smp_processor_id();
2270 if (unlikely(task_running(rq, p)))
2273 cpu = p->sched_class->select_task_rq(p, sync);
2274 if (cpu != orig_cpu) {
2275 set_task_cpu(p, cpu);
2276 task_rq_unlock(rq, &flags);
2277 /* might preempt at this point */
2278 rq = task_rq_lock(p, &flags);
2279 old_state = p->state;
2280 if (!(old_state & state))
2285 this_cpu = smp_processor_id();
2289 #ifdef CONFIG_SCHEDSTATS
2290 schedstat_inc(rq, ttwu_count);
2291 if (cpu == this_cpu)
2292 schedstat_inc(rq, ttwu_local);
2294 struct sched_domain *sd;
2295 for_each_domain(this_cpu, sd) {
2296 if (cpu_isset(cpu, sd->span)) {
2297 schedstat_inc(sd, ttwu_wake_remote);
2302 #endif /* CONFIG_SCHEDSTATS */
2305 #endif /* CONFIG_SMP */
2306 schedstat_inc(p, se.nr_wakeups);
2308 schedstat_inc(p, se.nr_wakeups_sync);
2309 if (orig_cpu != cpu)
2310 schedstat_inc(p, se.nr_wakeups_migrate);
2311 if (cpu == this_cpu)
2312 schedstat_inc(p, se.nr_wakeups_local);
2314 schedstat_inc(p, se.nr_wakeups_remote);
2315 update_rq_clock(rq);
2316 activate_task(rq, p, 1);
2320 check_preempt_curr(rq, p);
2322 p->state = TASK_RUNNING;
2324 if (p->sched_class->task_wake_up)
2325 p->sched_class->task_wake_up(rq, p);
2328 current->se.last_wakeup = current->se.sum_exec_runtime;
2330 task_rq_unlock(rq, &flags);
2335 int wake_up_process(struct task_struct *p)
2337 return try_to_wake_up(p, TASK_ALL, 0);
2339 EXPORT_SYMBOL(wake_up_process);
2341 int wake_up_state(struct task_struct *p, unsigned int state)
2343 return try_to_wake_up(p, state, 0);
2347 * Perform scheduler related setup for a newly forked process p.
2348 * p is forked by current.
2350 * __sched_fork() is basic setup used by init_idle() too:
2352 static void __sched_fork(struct task_struct *p)
2354 p->se.exec_start = 0;
2355 p->se.sum_exec_runtime = 0;
2356 p->se.prev_sum_exec_runtime = 0;
2357 p->se.last_wakeup = 0;
2358 p->se.avg_overlap = 0;
2360 #ifdef CONFIG_SCHEDSTATS
2361 p->se.wait_start = 0;
2362 p->se.sum_sleep_runtime = 0;
2363 p->se.sleep_start = 0;
2364 p->se.block_start = 0;
2365 p->se.sleep_max = 0;
2366 p->se.block_max = 0;
2368 p->se.slice_max = 0;
2372 INIT_LIST_HEAD(&p->rt.run_list);
2374 INIT_LIST_HEAD(&p->se.group_node);
2376 #ifdef CONFIG_PREEMPT_NOTIFIERS
2377 INIT_HLIST_HEAD(&p->preempt_notifiers);
2381 * We mark the process as running here, but have not actually
2382 * inserted it onto the runqueue yet. This guarantees that
2383 * nobody will actually run it, and a signal or other external
2384 * event cannot wake it up and insert it on the runqueue either.
2386 p->state = TASK_RUNNING;
2390 * fork()/clone()-time setup:
2392 void sched_fork(struct task_struct *p, int clone_flags)
2394 int cpu = get_cpu();
2399 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2401 set_task_cpu(p, cpu);
2404 * Make sure we do not leak PI boosting priority to the child:
2406 p->prio = current->normal_prio;
2407 if (!rt_prio(p->prio))
2408 p->sched_class = &fair_sched_class;
2410 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2411 if (likely(sched_info_on()))
2412 memset(&p->sched_info, 0, sizeof(p->sched_info));
2414 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2417 #ifdef CONFIG_PREEMPT
2418 /* Want to start with kernel preemption disabled. */
2419 task_thread_info(p)->preempt_count = 1;
2425 * wake_up_new_task - wake up a newly created task for the first time.
2427 * This function will do some initial scheduler statistics housekeeping
2428 * that must be done for every newly created context, then puts the task
2429 * on the runqueue and wakes it.
2431 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2433 unsigned long flags;
2436 rq = task_rq_lock(p, &flags);
2437 BUG_ON(p->state != TASK_RUNNING);
2438 update_rq_clock(rq);
2440 p->prio = effective_prio(p);
2442 if (!p->sched_class->task_new || !current->se.on_rq) {
2443 activate_task(rq, p, 0);
2446 * Let the scheduling class do new task startup
2447 * management (if any):
2449 p->sched_class->task_new(rq, p);
2452 check_preempt_curr(rq, p);
2454 if (p->sched_class->task_wake_up)
2455 p->sched_class->task_wake_up(rq, p);
2457 task_rq_unlock(rq, &flags);
2460 #ifdef CONFIG_PREEMPT_NOTIFIERS
2463 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2464 * @notifier: notifier struct to register
2466 void preempt_notifier_register(struct preempt_notifier *notifier)
2468 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2470 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2473 * preempt_notifier_unregister - no longer interested in preemption notifications
2474 * @notifier: notifier struct to unregister
2476 * This is safe to call from within a preemption notifier.
2478 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2480 hlist_del(¬ifier->link);
2482 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2484 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2486 struct preempt_notifier *notifier;
2487 struct hlist_node *node;
2489 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2490 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2494 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2495 struct task_struct *next)
2497 struct preempt_notifier *notifier;
2498 struct hlist_node *node;
2500 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2501 notifier->ops->sched_out(notifier, next);
2504 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2506 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2511 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2512 struct task_struct *next)
2516 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2519 * prepare_task_switch - prepare to switch tasks
2520 * @rq: the runqueue preparing to switch
2521 * @prev: the current task that is being switched out
2522 * @next: the task we are going to switch to.
2524 * This is called with the rq lock held and interrupts off. It must
2525 * be paired with a subsequent finish_task_switch after the context
2528 * prepare_task_switch sets up locking and calls architecture specific
2532 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2533 struct task_struct *next)
2535 fire_sched_out_preempt_notifiers(prev, next);
2536 prepare_lock_switch(rq, next);
2537 prepare_arch_switch(next);
2541 * finish_task_switch - clean up after a task-switch
2542 * @rq: runqueue associated with task-switch
2543 * @prev: the thread we just switched away from.
2545 * finish_task_switch must be called after the context switch, paired
2546 * with a prepare_task_switch call before the context switch.
2547 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2548 * and do any other architecture-specific cleanup actions.
2550 * Note that we may have delayed dropping an mm in context_switch(). If
2551 * so, we finish that here outside of the runqueue lock. (Doing it
2552 * with the lock held can cause deadlocks; see schedule() for
2555 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2556 __releases(rq->lock)
2558 struct mm_struct *mm = rq->prev_mm;
2564 * A task struct has one reference for the use as "current".
2565 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2566 * schedule one last time. The schedule call will never return, and
2567 * the scheduled task must drop that reference.
2568 * The test for TASK_DEAD must occur while the runqueue locks are
2569 * still held, otherwise prev could be scheduled on another cpu, die
2570 * there before we look at prev->state, and then the reference would
2572 * Manfred Spraul <manfred@colorfullife.com>
2574 prev_state = prev->state;
2575 finish_arch_switch(prev);
2576 finish_lock_switch(rq, prev);
2578 if (current->sched_class->post_schedule)
2579 current->sched_class->post_schedule(rq);
2582 fire_sched_in_preempt_notifiers(current);
2585 if (unlikely(prev_state == TASK_DEAD)) {
2587 * Remove function-return probe instances associated with this
2588 * task and put them back on the free list.
2590 kprobe_flush_task(prev);
2591 put_task_struct(prev);
2596 * schedule_tail - first thing a freshly forked thread must call.
2597 * @prev: the thread we just switched away from.
2599 asmlinkage void schedule_tail(struct task_struct *prev)
2600 __releases(rq->lock)
2602 struct rq *rq = this_rq();
2604 finish_task_switch(rq, prev);
2605 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2606 /* In this case, finish_task_switch does not reenable preemption */
2609 if (current->set_child_tid)
2610 put_user(task_pid_vnr(current), current->set_child_tid);
2614 * context_switch - switch to the new MM and the new
2615 * thread's register state.
2618 context_switch(struct rq *rq, struct task_struct *prev,
2619 struct task_struct *next)
2621 struct mm_struct *mm, *oldmm;
2623 prepare_task_switch(rq, prev, next);
2625 oldmm = prev->active_mm;
2627 * For paravirt, this is coupled with an exit in switch_to to
2628 * combine the page table reload and the switch backend into
2631 arch_enter_lazy_cpu_mode();
2633 if (unlikely(!mm)) {
2634 next->active_mm = oldmm;
2635 atomic_inc(&oldmm->mm_count);
2636 enter_lazy_tlb(oldmm, next);
2638 switch_mm(oldmm, mm, next);
2640 if (unlikely(!prev->mm)) {
2641 prev->active_mm = NULL;
2642 rq->prev_mm = oldmm;
2645 * Since the runqueue lock will be released by the next
2646 * task (which is an invalid locking op but in the case
2647 * of the scheduler it's an obvious special-case), so we
2648 * do an early lockdep release here:
2650 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2651 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2654 /* Here we just switch the register state and the stack. */
2655 switch_to(prev, next, prev);
2659 * this_rq must be evaluated again because prev may have moved
2660 * CPUs since it called schedule(), thus the 'rq' on its stack
2661 * frame will be invalid.
2663 finish_task_switch(this_rq(), prev);
2667 * nr_running, nr_uninterruptible and nr_context_switches:
2669 * externally visible scheduler statistics: current number of runnable
2670 * threads, current number of uninterruptible-sleeping threads, total
2671 * number of context switches performed since bootup.
2673 unsigned long nr_running(void)
2675 unsigned long i, sum = 0;
2677 for_each_online_cpu(i)
2678 sum += cpu_rq(i)->nr_running;
2683 unsigned long nr_uninterruptible(void)
2685 unsigned long i, sum = 0;
2687 for_each_possible_cpu(i)
2688 sum += cpu_rq(i)->nr_uninterruptible;
2691 * Since we read the counters lockless, it might be slightly
2692 * inaccurate. Do not allow it to go below zero though:
2694 if (unlikely((long)sum < 0))
2700 unsigned long long nr_context_switches(void)
2703 unsigned long long sum = 0;
2705 for_each_possible_cpu(i)
2706 sum += cpu_rq(i)->nr_switches;
2711 unsigned long nr_iowait(void)
2713 unsigned long i, sum = 0;
2715 for_each_possible_cpu(i)
2716 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2721 unsigned long nr_active(void)
2723 unsigned long i, running = 0, uninterruptible = 0;
2725 for_each_online_cpu(i) {
2726 running += cpu_rq(i)->nr_running;
2727 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2730 if (unlikely((long)uninterruptible < 0))
2731 uninterruptible = 0;
2733 return running + uninterruptible;
2737 * Update rq->cpu_load[] statistics. This function is usually called every
2738 * scheduler tick (TICK_NSEC).
2740 static void update_cpu_load(struct rq *this_rq)
2742 unsigned long this_load = this_rq->load.weight;
2745 this_rq->nr_load_updates++;
2747 /* Update our load: */
2748 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2749 unsigned long old_load, new_load;
2751 /* scale is effectively 1 << i now, and >> i divides by scale */
2753 old_load = this_rq->cpu_load[i];
2754 new_load = this_load;
2756 * Round up the averaging division if load is increasing. This
2757 * prevents us from getting stuck on 9 if the load is 10, for
2760 if (new_load > old_load)
2761 new_load += scale-1;
2762 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2769 * double_rq_lock - safely lock two runqueues
2771 * Note this does not disable interrupts like task_rq_lock,
2772 * you need to do so manually before calling.
2774 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2775 __acquires(rq1->lock)
2776 __acquires(rq2->lock)
2778 BUG_ON(!irqs_disabled());
2780 spin_lock(&rq1->lock);
2781 __acquire(rq2->lock); /* Fake it out ;) */
2784 spin_lock(&rq1->lock);
2785 spin_lock(&rq2->lock);
2787 spin_lock(&rq2->lock);
2788 spin_lock(&rq1->lock);
2791 update_rq_clock(rq1);
2792 update_rq_clock(rq2);
2796 * double_rq_unlock - safely unlock two runqueues
2798 * Note this does not restore interrupts like task_rq_unlock,
2799 * you need to do so manually after calling.
2801 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2802 __releases(rq1->lock)
2803 __releases(rq2->lock)
2805 spin_unlock(&rq1->lock);
2807 spin_unlock(&rq2->lock);
2809 __release(rq2->lock);
2813 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2815 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2816 __releases(this_rq->lock)
2817 __acquires(busiest->lock)
2818 __acquires(this_rq->lock)
2822 if (unlikely(!irqs_disabled())) {
2823 /* printk() doesn't work good under rq->lock */
2824 spin_unlock(&this_rq->lock);
2827 if (unlikely(!spin_trylock(&busiest->lock))) {
2828 if (busiest < this_rq) {
2829 spin_unlock(&this_rq->lock);
2830 spin_lock(&busiest->lock);
2831 spin_lock(&this_rq->lock);
2834 spin_lock(&busiest->lock);
2840 * If dest_cpu is allowed for this process, migrate the task to it.
2841 * This is accomplished by forcing the cpu_allowed mask to only
2842 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2843 * the cpu_allowed mask is restored.
2845 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2847 struct migration_req req;
2848 unsigned long flags;
2851 rq = task_rq_lock(p, &flags);
2852 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2853 || unlikely(cpu_is_offline(dest_cpu)))
2856 /* force the process onto the specified CPU */
2857 if (migrate_task(p, dest_cpu, &req)) {
2858 /* Need to wait for migration thread (might exit: take ref). */
2859 struct task_struct *mt = rq->migration_thread;
2861 get_task_struct(mt);
2862 task_rq_unlock(rq, &flags);
2863 wake_up_process(mt);
2864 put_task_struct(mt);
2865 wait_for_completion(&req.done);
2870 task_rq_unlock(rq, &flags);
2874 * sched_exec - execve() is a valuable balancing opportunity, because at
2875 * this point the task has the smallest effective memory and cache footprint.
2877 void sched_exec(void)
2879 int new_cpu, this_cpu = get_cpu();
2880 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2882 if (new_cpu != this_cpu)
2883 sched_migrate_task(current, new_cpu);
2887 * pull_task - move a task from a remote runqueue to the local runqueue.
2888 * Both runqueues must be locked.
2890 static void pull_task(struct rq *src_rq, struct task_struct *p,
2891 struct rq *this_rq, int this_cpu)
2893 deactivate_task(src_rq, p, 0);
2894 set_task_cpu(p, this_cpu);
2895 activate_task(this_rq, p, 0);
2897 * Note that idle threads have a prio of MAX_PRIO, for this test
2898 * to be always true for them.
2900 check_preempt_curr(this_rq, p);
2904 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2907 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2908 struct sched_domain *sd, enum cpu_idle_type idle,
2912 * We do not migrate tasks that are:
2913 * 1) running (obviously), or
2914 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2915 * 3) are cache-hot on their current CPU.
2917 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2918 schedstat_inc(p, se.nr_failed_migrations_affine);
2923 if (task_running(rq, p)) {
2924 schedstat_inc(p, se.nr_failed_migrations_running);
2929 * Aggressive migration if:
2930 * 1) task is cache cold, or
2931 * 2) too many balance attempts have failed.
2934 if (!task_hot(p, rq->clock, sd) ||
2935 sd->nr_balance_failed > sd->cache_nice_tries) {
2936 #ifdef CONFIG_SCHEDSTATS
2937 if (task_hot(p, rq->clock, sd)) {
2938 schedstat_inc(sd, lb_hot_gained[idle]);
2939 schedstat_inc(p, se.nr_forced_migrations);
2945 if (task_hot(p, rq->clock, sd)) {
2946 schedstat_inc(p, se.nr_failed_migrations_hot);
2952 static unsigned long
2953 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2954 unsigned long max_load_move, struct sched_domain *sd,
2955 enum cpu_idle_type idle, int *all_pinned,
2956 int *this_best_prio, struct rq_iterator *iterator)
2958 int loops = 0, pulled = 0, pinned = 0;
2959 struct task_struct *p;
2960 long rem_load_move = max_load_move;
2962 if (max_load_move == 0)
2968 * Start the load-balancing iterator:
2970 p = iterator->start(iterator->arg);
2972 if (!p || loops++ > sysctl_sched_nr_migrate)
2975 if ((p->se.load.weight >> 1) > rem_load_move ||
2976 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2977 p = iterator->next(iterator->arg);
2981 pull_task(busiest, p, this_rq, this_cpu);
2983 rem_load_move -= p->se.load.weight;
2986 * We only want to steal up to the prescribed amount of weighted load.
2988 if (rem_load_move > 0) {
2989 if (p->prio < *this_best_prio)
2990 *this_best_prio = p->prio;
2991 p = iterator->next(iterator->arg);
2996 * Right now, this is one of only two places pull_task() is called,
2997 * so we can safely collect pull_task() stats here rather than
2998 * inside pull_task().
3000 schedstat_add(sd, lb_gained[idle], pulled);
3003 *all_pinned = pinned;
3005 return max_load_move - rem_load_move;
3009 * move_tasks tries to move up to max_load_move weighted load from busiest to
3010 * this_rq, as part of a balancing operation within domain "sd".
3011 * Returns 1 if successful and 0 otherwise.
3013 * Called with both runqueues locked.
3015 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3016 unsigned long max_load_move,
3017 struct sched_domain *sd, enum cpu_idle_type idle,
3020 const struct sched_class *class = sched_class_highest;
3021 unsigned long total_load_moved = 0;
3022 int this_best_prio = this_rq->curr->prio;
3026 class->load_balance(this_rq, this_cpu, busiest,
3027 max_load_move - total_load_moved,
3028 sd, idle, all_pinned, &this_best_prio);
3029 class = class->next;
3031 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3034 } while (class && max_load_move > total_load_moved);
3036 return total_load_moved > 0;
3040 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3041 struct sched_domain *sd, enum cpu_idle_type idle,
3042 struct rq_iterator *iterator)
3044 struct task_struct *p = iterator->start(iterator->arg);
3048 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3049 pull_task(busiest, p, this_rq, this_cpu);
3051 * Right now, this is only the second place pull_task()
3052 * is called, so we can safely collect pull_task()
3053 * stats here rather than inside pull_task().
3055 schedstat_inc(sd, lb_gained[idle]);
3059 p = iterator->next(iterator->arg);
3066 * move_one_task tries to move exactly one task from busiest to this_rq, as
3067 * part of active balancing operations within "domain".
3068 * Returns 1 if successful and 0 otherwise.
3070 * Called with both runqueues locked.
3072 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3073 struct sched_domain *sd, enum cpu_idle_type idle)
3075 const struct sched_class *class;
3077 for (class = sched_class_highest; class; class = class->next)
3078 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3085 * find_busiest_group finds and returns the busiest CPU group within the
3086 * domain. It calculates and returns the amount of weighted load which
3087 * should be moved to restore balance via the imbalance parameter.
3089 static struct sched_group *
3090 find_busiest_group(struct sched_domain *sd, int this_cpu,
3091 unsigned long *imbalance, enum cpu_idle_type idle,
3092 int *sd_idle, const cpumask_t *cpus, int *balance)
3094 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
3095 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
3096 unsigned long max_pull;
3097 unsigned long busiest_load_per_task, busiest_nr_running;
3098 unsigned long this_load_per_task, this_nr_running;
3099 int load_idx, group_imb = 0;
3100 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3101 int power_savings_balance = 1;
3102 unsigned long leader_nr_running = 0, min_load_per_task = 0;
3103 unsigned long min_nr_running = ULONG_MAX;
3104 struct sched_group *group_min = NULL, *group_leader = NULL;
3107 max_load = this_load = total_load = total_pwr = 0;
3108 busiest_load_per_task = busiest_nr_running = 0;
3109 this_load_per_task = this_nr_running = 0;
3111 if (idle == CPU_NOT_IDLE)
3112 load_idx = sd->busy_idx;
3113 else if (idle == CPU_NEWLY_IDLE)
3114 load_idx = sd->newidle_idx;
3116 load_idx = sd->idle_idx;
3119 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
3122 int __group_imb = 0;
3123 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3124 unsigned long sum_nr_running, sum_weighted_load;
3125 unsigned long sum_avg_load_per_task;
3126 unsigned long avg_load_per_task;
3128 local_group = cpu_isset(this_cpu, group->cpumask);
3131 balance_cpu = first_cpu(group->cpumask);
3133 /* Tally up the load of all CPUs in the group */
3134 sum_weighted_load = sum_nr_running = avg_load = 0;
3135 sum_avg_load_per_task = avg_load_per_task = 0;
3138 min_cpu_load = ~0UL;
3140 for_each_cpu_mask(i, group->cpumask) {
3143 if (!cpu_isset(i, *cpus))
3148 if (*sd_idle && rq->nr_running)
3151 /* Bias balancing toward cpus of our domain */
3153 if (idle_cpu(i) && !first_idle_cpu) {
3158 load = target_load(i, load_idx);
3160 load = source_load(i, load_idx);
3161 if (load > max_cpu_load)
3162 max_cpu_load = load;
3163 if (min_cpu_load > load)
3164 min_cpu_load = load;
3168 sum_nr_running += rq->nr_running;
3169 sum_weighted_load += weighted_cpuload(i);
3171 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3175 * First idle cpu or the first cpu(busiest) in this sched group
3176 * is eligible for doing load balancing at this and above
3177 * domains. In the newly idle case, we will allow all the cpu's
3178 * to do the newly idle load balance.
3180 if (idle != CPU_NEWLY_IDLE && local_group &&
3181 balance_cpu != this_cpu && balance) {
3186 total_load += avg_load;
3187 total_pwr += group->__cpu_power;
3189 /* Adjust by relative CPU power of the group */
3190 avg_load = sg_div_cpu_power(group,
3191 avg_load * SCHED_LOAD_SCALE);
3195 * Consider the group unbalanced when the imbalance is larger
3196 * than the average weight of two tasks.
3198 * APZ: with cgroup the avg task weight can vary wildly and
3199 * might not be a suitable number - should we keep a
3200 * normalized nr_running number somewhere that negates
3203 avg_load_per_task = sg_div_cpu_power(group,
3204 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3206 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3209 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3212 this_load = avg_load;
3214 this_nr_running = sum_nr_running;
3215 this_load_per_task = sum_weighted_load;
3216 } else if (avg_load > max_load &&
3217 (sum_nr_running > group_capacity || __group_imb)) {
3218 max_load = avg_load;
3220 busiest_nr_running = sum_nr_running;
3221 busiest_load_per_task = sum_weighted_load;
3222 group_imb = __group_imb;
3225 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3227 * Busy processors will not participate in power savings
3230 if (idle == CPU_NOT_IDLE ||
3231 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3235 * If the local group is idle or completely loaded
3236 * no need to do power savings balance at this domain
3238 if (local_group && (this_nr_running >= group_capacity ||
3240 power_savings_balance = 0;
3243 * If a group is already running at full capacity or idle,
3244 * don't include that group in power savings calculations
3246 if (!power_savings_balance || sum_nr_running >= group_capacity
3251 * Calculate the group which has the least non-idle load.
3252 * This is the group from where we need to pick up the load
3255 if ((sum_nr_running < min_nr_running) ||
3256 (sum_nr_running == min_nr_running &&
3257 first_cpu(group->cpumask) <
3258 first_cpu(group_min->cpumask))) {
3260 min_nr_running = sum_nr_running;
3261 min_load_per_task = sum_weighted_load /
3266 * Calculate the group which is almost near its
3267 * capacity but still has some space to pick up some load
3268 * from other group and save more power
3270 if (sum_nr_running <= group_capacity - 1) {
3271 if (sum_nr_running > leader_nr_running ||
3272 (sum_nr_running == leader_nr_running &&
3273 first_cpu(group->cpumask) >
3274 first_cpu(group_leader->cpumask))) {
3275 group_leader = group;
3276 leader_nr_running = sum_nr_running;
3281 group = group->next;
3282 } while (group != sd->groups);
3284 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3287 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3289 if (this_load >= avg_load ||
3290 100*max_load <= sd->imbalance_pct*this_load)
3293 busiest_load_per_task /= busiest_nr_running;
3295 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3298 * We're trying to get all the cpus to the average_load, so we don't
3299 * want to push ourselves above the average load, nor do we wish to
3300 * reduce the max loaded cpu below the average load, as either of these
3301 * actions would just result in more rebalancing later, and ping-pong
3302 * tasks around. Thus we look for the minimum possible imbalance.
3303 * Negative imbalances (*we* are more loaded than anyone else) will
3304 * be counted as no imbalance for these purposes -- we can't fix that
3305 * by pulling tasks to us. Be careful of negative numbers as they'll
3306 * appear as very large values with unsigned longs.
3308 if (max_load <= busiest_load_per_task)
3312 * In the presence of smp nice balancing, certain scenarios can have
3313 * max load less than avg load(as we skip the groups at or below
3314 * its cpu_power, while calculating max_load..)
3316 if (max_load < avg_load) {
3318 goto small_imbalance;
3321 /* Don't want to pull so many tasks that a group would go idle */
3322 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3324 /* How much load to actually move to equalise the imbalance */
3325 *imbalance = min(max_pull * busiest->__cpu_power,
3326 (avg_load - this_load) * this->__cpu_power)
3330 * if *imbalance is less than the average load per runnable task
3331 * there is no gaurantee that any tasks will be moved so we'll have
3332 * a think about bumping its value to force at least one task to be
3335 if (*imbalance < busiest_load_per_task) {
3336 unsigned long tmp, pwr_now, pwr_move;
3340 pwr_move = pwr_now = 0;
3342 if (this_nr_running) {
3343 this_load_per_task /= this_nr_running;
3344 if (busiest_load_per_task > this_load_per_task)
3347 this_load_per_task = cpu_avg_load_per_task(this_cpu);
3349 if (max_load - this_load + 2*busiest_load_per_task >=
3350 busiest_load_per_task * imbn) {
3351 *imbalance = busiest_load_per_task;
3356 * OK, we don't have enough imbalance to justify moving tasks,
3357 * however we may be able to increase total CPU power used by
3361 pwr_now += busiest->__cpu_power *
3362 min(busiest_load_per_task, max_load);
3363 pwr_now += this->__cpu_power *
3364 min(this_load_per_task, this_load);
3365 pwr_now /= SCHED_LOAD_SCALE;
3367 /* Amount of load we'd subtract */
3368 tmp = sg_div_cpu_power(busiest,
3369 busiest_load_per_task * SCHED_LOAD_SCALE);
3371 pwr_move += busiest->__cpu_power *
3372 min(busiest_load_per_task, max_load - tmp);
3374 /* Amount of load we'd add */
3375 if (max_load * busiest->__cpu_power <
3376 busiest_load_per_task * SCHED_LOAD_SCALE)
3377 tmp = sg_div_cpu_power(this,
3378 max_load * busiest->__cpu_power);
3380 tmp = sg_div_cpu_power(this,
3381 busiest_load_per_task * SCHED_LOAD_SCALE);
3382 pwr_move += this->__cpu_power *
3383 min(this_load_per_task, this_load + tmp);
3384 pwr_move /= SCHED_LOAD_SCALE;
3386 /* Move if we gain throughput */
3387 if (pwr_move > pwr_now)
3388 *imbalance = busiest_load_per_task;
3394 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3395 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3398 if (this == group_leader && group_leader != group_min) {
3399 *imbalance = min_load_per_task;
3409 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3412 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3413 unsigned long imbalance, const cpumask_t *cpus)
3415 struct rq *busiest = NULL, *rq;
3416 unsigned long max_load = 0;
3419 for_each_cpu_mask(i, group->cpumask) {
3422 if (!cpu_isset(i, *cpus))
3426 wl = weighted_cpuload(i);
3428 if (rq->nr_running == 1 && wl > imbalance)
3431 if (wl > max_load) {
3441 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3442 * so long as it is large enough.
3444 #define MAX_PINNED_INTERVAL 512
3447 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3448 * tasks if there is an imbalance.
3450 static int load_balance(int this_cpu, struct rq *this_rq,
3451 struct sched_domain *sd, enum cpu_idle_type idle,
3452 int *balance, cpumask_t *cpus)
3454 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3455 struct sched_group *group;
3456 unsigned long imbalance;
3458 unsigned long flags;
3463 * When power savings policy is enabled for the parent domain, idle
3464 * sibling can pick up load irrespective of busy siblings. In this case,
3465 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3466 * portraying it as CPU_NOT_IDLE.
3468 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3469 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3472 schedstat_inc(sd, lb_count[idle]);
3476 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3483 schedstat_inc(sd, lb_nobusyg[idle]);
3487 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3489 schedstat_inc(sd, lb_nobusyq[idle]);
3493 BUG_ON(busiest == this_rq);
3495 schedstat_add(sd, lb_imbalance[idle], imbalance);
3498 if (busiest->nr_running > 1) {
3500 * Attempt to move tasks. If find_busiest_group has found
3501 * an imbalance but busiest->nr_running <= 1, the group is
3502 * still unbalanced. ld_moved simply stays zero, so it is
3503 * correctly treated as an imbalance.
3505 local_irq_save(flags);
3506 double_rq_lock(this_rq, busiest);
3507 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3508 imbalance, sd, idle, &all_pinned);
3509 double_rq_unlock(this_rq, busiest);
3510 local_irq_restore(flags);
3513 * some other cpu did the load balance for us.
3515 if (ld_moved && this_cpu != smp_processor_id())
3516 resched_cpu(this_cpu);
3518 /* All tasks on this runqueue were pinned by CPU affinity */
3519 if (unlikely(all_pinned)) {
3520 cpu_clear(cpu_of(busiest), *cpus);
3521 if (!cpus_empty(*cpus))
3528 schedstat_inc(sd, lb_failed[idle]);
3529 sd->nr_balance_failed++;
3531 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3533 spin_lock_irqsave(&busiest->lock, flags);
3535 /* don't kick the migration_thread, if the curr
3536 * task on busiest cpu can't be moved to this_cpu
3538 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3539 spin_unlock_irqrestore(&busiest->lock, flags);
3541 goto out_one_pinned;
3544 if (!busiest->active_balance) {
3545 busiest->active_balance = 1;
3546 busiest->push_cpu = this_cpu;
3549 spin_unlock_irqrestore(&busiest->lock, flags);
3551 wake_up_process(busiest->migration_thread);
3554 * We've kicked active balancing, reset the failure
3557 sd->nr_balance_failed = sd->cache_nice_tries+1;
3560 sd->nr_balance_failed = 0;
3562 if (likely(!active_balance)) {
3563 /* We were unbalanced, so reset the balancing interval */
3564 sd->balance_interval = sd->min_interval;
3567 * If we've begun active balancing, start to back off. This
3568 * case may not be covered by the all_pinned logic if there
3569 * is only 1 task on the busy runqueue (because we don't call
3572 if (sd->balance_interval < sd->max_interval)
3573 sd->balance_interval *= 2;
3576 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3577 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3583 schedstat_inc(sd, lb_balanced[idle]);
3585 sd->nr_balance_failed = 0;
3588 /* tune up the balancing interval */
3589 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3590 (sd->balance_interval < sd->max_interval))
3591 sd->balance_interval *= 2;
3593 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3594 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3605 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3606 * tasks if there is an imbalance.
3608 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3609 * this_rq is locked.
3612 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3615 struct sched_group *group;
3616 struct rq *busiest = NULL;
3617 unsigned long imbalance;
3625 * When power savings policy is enabled for the parent domain, idle
3626 * sibling can pick up load irrespective of busy siblings. In this case,
3627 * let the state of idle sibling percolate up as IDLE, instead of
3628 * portraying it as CPU_NOT_IDLE.
3630 if (sd->flags & SD_SHARE_CPUPOWER &&
3631 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3634 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3636 update_shares_locked(this_rq, sd);
3637 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3638 &sd_idle, cpus, NULL);
3640 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3644 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3646 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3650 BUG_ON(busiest == this_rq);
3652 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3655 if (busiest->nr_running > 1) {
3656 /* Attempt to move tasks */
3657 double_lock_balance(this_rq, busiest);
3658 /* this_rq->clock is already updated */
3659 update_rq_clock(busiest);
3660 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3661 imbalance, sd, CPU_NEWLY_IDLE,
3663 spin_unlock(&busiest->lock);
3665 if (unlikely(all_pinned)) {
3666 cpu_clear(cpu_of(busiest), *cpus);
3667 if (!cpus_empty(*cpus))
3673 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3674 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3675 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3678 sd->nr_balance_failed = 0;
3680 update_shares_locked(this_rq, sd);
3684 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3685 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3686 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3688 sd->nr_balance_failed = 0;
3694 * idle_balance is called by schedule() if this_cpu is about to become
3695 * idle. Attempts to pull tasks from other CPUs.
3697 static void idle_balance(int this_cpu, struct rq *this_rq)
3699 struct sched_domain *sd;
3700 int pulled_task = -1;
3701 unsigned long next_balance = jiffies + HZ;
3704 for_each_domain(this_cpu, sd) {
3705 unsigned long interval;
3707 if (!(sd->flags & SD_LOAD_BALANCE))
3710 if (sd->flags & SD_BALANCE_NEWIDLE)
3711 /* If we've pulled tasks over stop searching: */
3712 pulled_task = load_balance_newidle(this_cpu, this_rq,
3715 interval = msecs_to_jiffies(sd->balance_interval);
3716 if (time_after(next_balance, sd->last_balance + interval))
3717 next_balance = sd->last_balance + interval;
3721 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3723 * We are going idle. next_balance may be set based on
3724 * a busy processor. So reset next_balance.
3726 this_rq->next_balance = next_balance;
3731 * active_load_balance is run by migration threads. It pushes running tasks
3732 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3733 * running on each physical CPU where possible, and avoids physical /
3734 * logical imbalances.
3736 * Called with busiest_rq locked.
3738 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3740 int target_cpu = busiest_rq->push_cpu;
3741 struct sched_domain *sd;
3742 struct rq *target_rq;
3744 /* Is there any task to move? */
3745 if (busiest_rq->nr_running <= 1)
3748 target_rq = cpu_rq(target_cpu);
3751 * This condition is "impossible", if it occurs
3752 * we need to fix it. Originally reported by
3753 * Bjorn Helgaas on a 128-cpu setup.
3755 BUG_ON(busiest_rq == target_rq);
3757 /* move a task from busiest_rq to target_rq */
3758 double_lock_balance(busiest_rq, target_rq);
3759 update_rq_clock(busiest_rq);
3760 update_rq_clock(target_rq);
3762 /* Search for an sd spanning us and the target CPU. */
3763 for_each_domain(target_cpu, sd) {
3764 if ((sd->flags & SD_LOAD_BALANCE) &&
3765 cpu_isset(busiest_cpu, sd->span))
3770 schedstat_inc(sd, alb_count);
3772 if (move_one_task(target_rq, target_cpu, busiest_rq,
3774 schedstat_inc(sd, alb_pushed);
3776 schedstat_inc(sd, alb_failed);
3778 spin_unlock(&target_rq->lock);
3783 atomic_t load_balancer;
3785 } nohz ____cacheline_aligned = {
3786 .load_balancer = ATOMIC_INIT(-1),
3787 .cpu_mask = CPU_MASK_NONE,
3791 * This routine will try to nominate the ilb (idle load balancing)
3792 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3793 * load balancing on behalf of all those cpus. If all the cpus in the system
3794 * go into this tickless mode, then there will be no ilb owner (as there is
3795 * no need for one) and all the cpus will sleep till the next wakeup event
3798 * For the ilb owner, tick is not stopped. And this tick will be used
3799 * for idle load balancing. ilb owner will still be part of
3802 * While stopping the tick, this cpu will become the ilb owner if there
3803 * is no other owner. And will be the owner till that cpu becomes busy
3804 * or if all cpus in the system stop their ticks at which point
3805 * there is no need for ilb owner.
3807 * When the ilb owner becomes busy, it nominates another owner, during the
3808 * next busy scheduler_tick()
3810 int select_nohz_load_balancer(int stop_tick)
3812 int cpu = smp_processor_id();
3815 cpu_set(cpu, nohz.cpu_mask);
3816 cpu_rq(cpu)->in_nohz_recently = 1;
3819 * If we are going offline and still the leader, give up!
3821 if (cpu_is_offline(cpu) &&
3822 atomic_read(&nohz.load_balancer) == cpu) {
3823 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3828 /* time for ilb owner also to sleep */
3829 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3830 if (atomic_read(&nohz.load_balancer) == cpu)
3831 atomic_set(&nohz.load_balancer, -1);
3835 if (atomic_read(&nohz.load_balancer) == -1) {
3836 /* make me the ilb owner */
3837 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3839 } else if (atomic_read(&nohz.load_balancer) == cpu)
3842 if (!cpu_isset(cpu, nohz.cpu_mask))
3845 cpu_clear(cpu, nohz.cpu_mask);
3847 if (atomic_read(&nohz.load_balancer) == cpu)
3848 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3855 static DEFINE_SPINLOCK(balancing);
3858 * It checks each scheduling domain to see if it is due to be balanced,
3859 * and initiates a balancing operation if so.
3861 * Balancing parameters are set up in arch_init_sched_domains.
3863 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3866 struct rq *rq = cpu_rq(cpu);
3867 unsigned long interval;
3868 struct sched_domain *sd;
3869 /* Earliest time when we have to do rebalance again */
3870 unsigned long next_balance = jiffies + 60*HZ;
3871 int update_next_balance = 0;
3875 for_each_domain(cpu, sd) {
3876 if (!(sd->flags & SD_LOAD_BALANCE))
3879 interval = sd->balance_interval;
3880 if (idle != CPU_IDLE)
3881 interval *= sd->busy_factor;
3883 /* scale ms to jiffies */
3884 interval = msecs_to_jiffies(interval);
3885 if (unlikely(!interval))
3887 if (interval > HZ*NR_CPUS/10)
3888 interval = HZ*NR_CPUS/10;
3890 need_serialize = sd->flags & SD_SERIALIZE;
3892 if (need_serialize) {
3893 if (!spin_trylock(&balancing))
3897 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3898 if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) {
3900 * We've pulled tasks over so either we're no
3901 * longer idle, or one of our SMT siblings is
3904 idle = CPU_NOT_IDLE;
3906 sd->last_balance = jiffies;
3909 spin_unlock(&balancing);
3911 if (time_after(next_balance, sd->last_balance + interval)) {
3912 next_balance = sd->last_balance + interval;
3913 update_next_balance = 1;
3917 * Stop the load balance at this level. There is another
3918 * CPU in our sched group which is doing load balancing more
3926 * next_balance will be updated only when there is a need.
3927 * When the cpu is attached to null domain for ex, it will not be
3930 if (likely(update_next_balance))
3931 rq->next_balance = next_balance;
3935 * run_rebalance_domains is triggered when needed from the scheduler tick.
3936 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3937 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3939 static void run_rebalance_domains(struct softirq_action *h)
3941 int this_cpu = smp_processor_id();
3942 struct rq *this_rq = cpu_rq(this_cpu);
3943 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3944 CPU_IDLE : CPU_NOT_IDLE;
3946 rebalance_domains(this_cpu, idle);
3950 * If this cpu is the owner for idle load balancing, then do the
3951 * balancing on behalf of the other idle cpus whose ticks are
3954 if (this_rq->idle_at_tick &&
3955 atomic_read(&nohz.load_balancer) == this_cpu) {
3956 cpumask_t cpus = nohz.cpu_mask;
3960 cpu_clear(this_cpu, cpus);
3961 for_each_cpu_mask(balance_cpu, cpus) {
3963 * If this cpu gets work to do, stop the load balancing
3964 * work being done for other cpus. Next load
3965 * balancing owner will pick it up.
3970 rebalance_domains(balance_cpu, CPU_IDLE);
3972 rq = cpu_rq(balance_cpu);
3973 if (time_after(this_rq->next_balance, rq->next_balance))
3974 this_rq->next_balance = rq->next_balance;
3981 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3983 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3984 * idle load balancing owner or decide to stop the periodic load balancing,
3985 * if the whole system is idle.
3987 static inline void trigger_load_balance(struct rq *rq, int cpu)
3991 * If we were in the nohz mode recently and busy at the current
3992 * scheduler tick, then check if we need to nominate new idle
3995 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3996 rq->in_nohz_recently = 0;
3998 if (atomic_read(&nohz.load_balancer) == cpu) {
3999 cpu_clear(cpu, nohz.cpu_mask);
4000 atomic_set(&nohz.load_balancer, -1);
4003 if (atomic_read(&nohz.load_balancer) == -1) {
4005 * simple selection for now: Nominate the
4006 * first cpu in the nohz list to be the next
4009 * TBD: Traverse the sched domains and nominate
4010 * the nearest cpu in the nohz.cpu_mask.
4012 int ilb = first_cpu(nohz.cpu_mask);
4014 if (ilb < nr_cpu_ids)
4020 * If this cpu is idle and doing idle load balancing for all the
4021 * cpus with ticks stopped, is it time for that to stop?
4023 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4024 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
4030 * If this cpu is idle and the idle load balancing is done by
4031 * someone else, then no need raise the SCHED_SOFTIRQ
4033 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4034 cpu_isset(cpu, nohz.cpu_mask))
4037 if (time_after_eq(jiffies, rq->next_balance))
4038 raise_softirq(SCHED_SOFTIRQ);
4041 #else /* CONFIG_SMP */
4044 * on UP we do not need to balance between CPUs:
4046 static inline void idle_balance(int cpu, struct rq *rq)
4052 DEFINE_PER_CPU(struct kernel_stat, kstat);
4054 EXPORT_PER_CPU_SYMBOL(kstat);
4057 * Return p->sum_exec_runtime plus any more ns on the sched_clock
4058 * that have not yet been banked in case the task is currently running.
4060 unsigned long long task_sched_runtime(struct task_struct *p)
4062 unsigned long flags;
4066 rq = task_rq_lock(p, &flags);
4067 ns = p->se.sum_exec_runtime;
4068 if (task_current(rq, p)) {
4069 update_rq_clock(rq);
4070 delta_exec = rq->clock - p->se.exec_start;
4071 if ((s64)delta_exec > 0)
4074 task_rq_unlock(rq, &flags);
4080 * Account user cpu time to a process.
4081 * @p: the process that the cpu time gets accounted to
4082 * @cputime: the cpu time spent in user space since the last update
4084 void account_user_time(struct task_struct *p, cputime_t cputime)
4086 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4089 p->utime = cputime_add(p->utime, cputime);
4091 /* Add user time to cpustat. */
4092 tmp = cputime_to_cputime64(cputime);
4093 if (TASK_NICE(p) > 0)
4094 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4096 cpustat->user = cputime64_add(cpustat->user, tmp);
4100 * Account guest cpu time to a process.
4101 * @p: the process that the cpu time gets accounted to
4102 * @cputime: the cpu time spent in virtual machine since the last update
4104 static void account_guest_time(struct task_struct *p, cputime_t cputime)
4107 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4109 tmp = cputime_to_cputime64(cputime);
4111 p->utime = cputime_add(p->utime, cputime);
4112 p->gtime = cputime_add(p->gtime, cputime);
4114 cpustat->user = cputime64_add(cpustat->user, tmp);
4115 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4119 * Account scaled user cpu time to a process.
4120 * @p: the process that the cpu time gets accounted to
4121 * @cputime: the cpu time spent in user space since the last update
4123 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
4125 p->utimescaled = cputime_add(p->utimescaled, cputime);
4129 * Account system cpu time to a process.
4130 * @p: the process that the cpu time gets accounted to
4131 * @hardirq_offset: the offset to subtract from hardirq_count()
4132 * @cputime: the cpu time spent in kernel space since the last update
4134 void account_system_time(struct task_struct *p, int hardirq_offset,
4137 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4138 struct rq *rq = this_rq();
4141 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4142 account_guest_time(p, cputime);
4146 p->stime = cputime_add(p->stime, cputime);
4148 /* Add system time to cpustat. */
4149 tmp = cputime_to_cputime64(cputime);
4150 if (hardirq_count() - hardirq_offset)
4151 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4152 else if (softirq_count())
4153 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4154 else if (p != rq->idle)
4155 cpustat->system = cputime64_add(cpustat->system, tmp);
4156 else if (atomic_read(&rq->nr_iowait) > 0)
4157 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4159 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4160 /* Account for system time used */
4161 acct_update_integrals(p);
4165 * Account scaled system cpu time to a process.
4166 * @p: the process that the cpu time gets accounted to
4167 * @hardirq_offset: the offset to subtract from hardirq_count()
4168 * @cputime: the cpu time spent in kernel space since the last update
4170 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
4172 p->stimescaled = cputime_add(p->stimescaled, cputime);
4176 * Account for involuntary wait time.
4177 * @p: the process from which the cpu time has been stolen
4178 * @steal: the cpu time spent in involuntary wait
4180 void account_steal_time(struct task_struct *p, cputime_t steal)
4182 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4183 cputime64_t tmp = cputime_to_cputime64(steal);
4184 struct rq *rq = this_rq();
4186 if (p == rq->idle) {
4187 p->stime = cputime_add(p->stime, steal);
4188 if (atomic_read(&rq->nr_iowait) > 0)
4189 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4191 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4193 cpustat->steal = cputime64_add(cpustat->steal, tmp);
4197 * This function gets called by the timer code, with HZ frequency.
4198 * We call it with interrupts disabled.
4200 * It also gets called by the fork code, when changing the parent's
4203 void scheduler_tick(void)
4205 int cpu = smp_processor_id();
4206 struct rq *rq = cpu_rq(cpu);
4207 struct task_struct *curr = rq->curr;
4211 spin_lock(&rq->lock);
4212 update_rq_clock(rq);
4213 update_cpu_load(rq);
4214 curr->sched_class->task_tick(rq, curr, 0);
4215 spin_unlock(&rq->lock);
4218 rq->idle_at_tick = idle_cpu(cpu);
4219 trigger_load_balance(rq, cpu);
4223 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
4225 void __kprobes add_preempt_count(int val)
4230 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4232 preempt_count() += val;
4234 * Spinlock count overflowing soon?
4236 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4239 EXPORT_SYMBOL(add_preempt_count);
4241 void __kprobes sub_preempt_count(int val)
4246 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4249 * Is the spinlock portion underflowing?
4251 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4252 !(preempt_count() & PREEMPT_MASK)))
4255 preempt_count() -= val;
4257 EXPORT_SYMBOL(sub_preempt_count);
4262 * Print scheduling while atomic bug:
4264 static noinline void __schedule_bug(struct task_struct *prev)
4266 struct pt_regs *regs = get_irq_regs();
4268 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4269 prev->comm, prev->pid, preempt_count());
4271 debug_show_held_locks(prev);
4273 if (irqs_disabled())
4274 print_irqtrace_events(prev);
4283 * Various schedule()-time debugging checks and statistics:
4285 static inline void schedule_debug(struct task_struct *prev)
4288 * Test if we are atomic. Since do_exit() needs to call into
4289 * schedule() atomically, we ignore that path for now.
4290 * Otherwise, whine if we are scheduling when we should not be.
4292 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4293 __schedule_bug(prev);
4295 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4297 schedstat_inc(this_rq(), sched_count);
4298 #ifdef CONFIG_SCHEDSTATS
4299 if (unlikely(prev->lock_depth >= 0)) {
4300 schedstat_inc(this_rq(), bkl_count);
4301 schedstat_inc(prev, sched_info.bkl_count);
4307 * Pick up the highest-prio task:
4309 static inline struct task_struct *
4310 pick_next_task(struct rq *rq, struct task_struct *prev)
4312 const struct sched_class *class;
4313 struct task_struct *p;
4316 * Optimization: we know that if all tasks are in
4317 * the fair class we can call that function directly:
4319 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4320 p = fair_sched_class.pick_next_task(rq);
4325 class = sched_class_highest;
4327 p = class->pick_next_task(rq);
4331 * Will never be NULL as the idle class always
4332 * returns a non-NULL p:
4334 class = class->next;
4339 * schedule() is the main scheduler function.
4341 asmlinkage void __sched schedule(void)
4343 struct task_struct *prev, *next;
4344 unsigned long *switch_count;
4346 int cpu, hrtick = sched_feat(HRTICK);
4350 cpu = smp_processor_id();
4354 switch_count = &prev->nivcsw;
4356 release_kernel_lock(prev);
4357 need_resched_nonpreemptible:
4359 schedule_debug(prev);
4365 * Do the rq-clock update outside the rq lock:
4367 local_irq_disable();
4368 update_rq_clock(rq);
4369 spin_lock(&rq->lock);
4370 clear_tsk_need_resched(prev);
4372 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4373 if (unlikely(signal_pending_state(prev->state, prev)))
4374 prev->state = TASK_RUNNING;
4376 deactivate_task(rq, prev, 1);
4377 switch_count = &prev->nvcsw;
4381 if (prev->sched_class->pre_schedule)
4382 prev->sched_class->pre_schedule(rq, prev);
4385 if (unlikely(!rq->nr_running))
4386 idle_balance(cpu, rq);
4388 prev->sched_class->put_prev_task(rq, prev);
4389 next = pick_next_task(rq, prev);
4391 if (likely(prev != next)) {
4392 sched_info_switch(prev, next);
4398 context_switch(rq, prev, next); /* unlocks the rq */
4400 * the context switch might have flipped the stack from under
4401 * us, hence refresh the local variables.
4403 cpu = smp_processor_id();
4406 spin_unlock_irq(&rq->lock);
4411 if (unlikely(reacquire_kernel_lock(current) < 0))
4412 goto need_resched_nonpreemptible;
4414 preempt_enable_no_resched();
4415 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4418 EXPORT_SYMBOL(schedule);
4420 #ifdef CONFIG_PREEMPT
4422 * this is the entry point to schedule() from in-kernel preemption
4423 * off of preempt_enable. Kernel preemptions off return from interrupt
4424 * occur there and call schedule directly.
4426 asmlinkage void __sched preempt_schedule(void)
4428 struct thread_info *ti = current_thread_info();
4431 * If there is a non-zero preempt_count or interrupts are disabled,
4432 * we do not want to preempt the current task. Just return..
4434 if (likely(ti->preempt_count || irqs_disabled()))
4438 add_preempt_count(PREEMPT_ACTIVE);
4440 sub_preempt_count(PREEMPT_ACTIVE);
4443 * Check again in case we missed a preemption opportunity
4444 * between schedule and now.
4447 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4449 EXPORT_SYMBOL(preempt_schedule);
4452 * this is the entry point to schedule() from kernel preemption
4453 * off of irq context.
4454 * Note, that this is called and return with irqs disabled. This will
4455 * protect us against recursive calling from irq.
4457 asmlinkage void __sched preempt_schedule_irq(void)
4459 struct thread_info *ti = current_thread_info();
4461 /* Catch callers which need to be fixed */
4462 BUG_ON(ti->preempt_count || !irqs_disabled());
4465 add_preempt_count(PREEMPT_ACTIVE);
4468 local_irq_disable();
4469 sub_preempt_count(PREEMPT_ACTIVE);
4472 * Check again in case we missed a preemption opportunity
4473 * between schedule and now.
4476 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4479 #endif /* CONFIG_PREEMPT */
4481 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4484 return try_to_wake_up(curr->private, mode, sync);
4486 EXPORT_SYMBOL(default_wake_function);
4489 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4490 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4491 * number) then we wake all the non-exclusive tasks and one exclusive task.
4493 * There are circumstances in which we can try to wake a task which has already
4494 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4495 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4497 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4498 int nr_exclusive, int sync, void *key)
4500 wait_queue_t *curr, *next;
4502 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4503 unsigned flags = curr->flags;
4505 if (curr->func(curr, mode, sync, key) &&
4506 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4512 * __wake_up - wake up threads blocked on a waitqueue.
4514 * @mode: which threads
4515 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4516 * @key: is directly passed to the wakeup function
4518 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4519 int nr_exclusive, void *key)
4521 unsigned long flags;
4523 spin_lock_irqsave(&q->lock, flags);
4524 __wake_up_common(q, mode, nr_exclusive, 0, key);
4525 spin_unlock_irqrestore(&q->lock, flags);
4527 EXPORT_SYMBOL(__wake_up);
4530 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4532 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4534 __wake_up_common(q, mode, 1, 0, NULL);
4538 * __wake_up_sync - wake up threads blocked on a waitqueue.
4540 * @mode: which threads
4541 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4543 * The sync wakeup differs that the waker knows that it will schedule
4544 * away soon, so while the target thread will be woken up, it will not
4545 * be migrated to another CPU - ie. the two threads are 'synchronized'
4546 * with each other. This can prevent needless bouncing between CPUs.
4548 * On UP it can prevent extra preemption.
4551 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4553 unsigned long flags;
4559 if (unlikely(!nr_exclusive))
4562 spin_lock_irqsave(&q->lock, flags);
4563 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4564 spin_unlock_irqrestore(&q->lock, flags);
4566 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4568 void complete(struct completion *x)
4570 unsigned long flags;
4572 spin_lock_irqsave(&x->wait.lock, flags);
4574 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4575 spin_unlock_irqrestore(&x->wait.lock, flags);
4577 EXPORT_SYMBOL(complete);
4579 void complete_all(struct completion *x)
4581 unsigned long flags;
4583 spin_lock_irqsave(&x->wait.lock, flags);
4584 x->done += UINT_MAX/2;
4585 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4586 spin_unlock_irqrestore(&x->wait.lock, flags);
4588 EXPORT_SYMBOL(complete_all);
4590 static inline long __sched
4591 do_wait_for_common(struct completion *x, long timeout, int state)
4594 DECLARE_WAITQUEUE(wait, current);
4596 wait.flags |= WQ_FLAG_EXCLUSIVE;
4597 __add_wait_queue_tail(&x->wait, &wait);
4599 if ((state == TASK_INTERRUPTIBLE &&
4600 signal_pending(current)) ||
4601 (state == TASK_KILLABLE &&
4602 fatal_signal_pending(current))) {
4603 timeout = -ERESTARTSYS;
4606 __set_current_state(state);
4607 spin_unlock_irq(&x->wait.lock);
4608 timeout = schedule_timeout(timeout);
4609 spin_lock_irq(&x->wait.lock);
4610 } while (!x->done && timeout);
4611 __remove_wait_queue(&x->wait, &wait);
4616 return timeout ?: 1;
4620 wait_for_common(struct completion *x, long timeout, int state)
4624 spin_lock_irq(&x->wait.lock);
4625 timeout = do_wait_for_common(x, timeout, state);
4626 spin_unlock_irq(&x->wait.lock);
4630 void __sched wait_for_completion(struct completion *x)
4632 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4634 EXPORT_SYMBOL(wait_for_completion);
4636 unsigned long __sched
4637 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4639 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4641 EXPORT_SYMBOL(wait_for_completion_timeout);
4643 int __sched wait_for_completion_interruptible(struct completion *x)
4645 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4646 if (t == -ERESTARTSYS)
4650 EXPORT_SYMBOL(wait_for_completion_interruptible);
4652 unsigned long __sched
4653 wait_for_completion_interruptible_timeout(struct completion *x,
4654 unsigned long timeout)
4656 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4658 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4660 int __sched wait_for_completion_killable(struct completion *x)
4662 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4663 if (t == -ERESTARTSYS)
4667 EXPORT_SYMBOL(wait_for_completion_killable);
4670 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4672 unsigned long flags;
4675 init_waitqueue_entry(&wait, current);
4677 __set_current_state(state);
4679 spin_lock_irqsave(&q->lock, flags);
4680 __add_wait_queue(q, &wait);
4681 spin_unlock(&q->lock);
4682 timeout = schedule_timeout(timeout);
4683 spin_lock_irq(&q->lock);
4684 __remove_wait_queue(q, &wait);
4685 spin_unlock_irqrestore(&q->lock, flags);
4690 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4692 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4694 EXPORT_SYMBOL(interruptible_sleep_on);
4697 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4699 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4701 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4703 void __sched sleep_on(wait_queue_head_t *q)
4705 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4707 EXPORT_SYMBOL(sleep_on);
4709 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4711 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4713 EXPORT_SYMBOL(sleep_on_timeout);
4715 #ifdef CONFIG_RT_MUTEXES
4718 * rt_mutex_setprio - set the current priority of a task
4720 * @prio: prio value (kernel-internal form)
4722 * This function changes the 'effective' priority of a task. It does
4723 * not touch ->normal_prio like __setscheduler().
4725 * Used by the rt_mutex code to implement priority inheritance logic.
4727 void rt_mutex_setprio(struct task_struct *p, int prio)
4729 unsigned long flags;
4730 int oldprio, on_rq, running;
4732 const struct sched_class *prev_class = p->sched_class;
4734 BUG_ON(prio < 0 || prio > MAX_PRIO);
4736 rq = task_rq_lock(p, &flags);
4737 update_rq_clock(rq);
4740 on_rq = p->se.on_rq;
4741 running = task_current(rq, p);
4743 dequeue_task(rq, p, 0);
4745 p->sched_class->put_prev_task(rq, p);
4748 p->sched_class = &rt_sched_class;
4750 p->sched_class = &fair_sched_class;
4755 p->sched_class->set_curr_task(rq);
4757 enqueue_task(rq, p, 0);
4759 check_class_changed(rq, p, prev_class, oldprio, running);
4761 task_rq_unlock(rq, &flags);
4766 void set_user_nice(struct task_struct *p, long nice)
4768 int old_prio, delta, on_rq;
4769 unsigned long flags;
4772 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4775 * We have to be careful, if called from sys_setpriority(),
4776 * the task might be in the middle of scheduling on another CPU.
4778 rq = task_rq_lock(p, &flags);
4779 update_rq_clock(rq);
4781 * The RT priorities are set via sched_setscheduler(), but we still
4782 * allow the 'normal' nice value to be set - but as expected
4783 * it wont have any effect on scheduling until the task is
4784 * SCHED_FIFO/SCHED_RR:
4786 if (task_has_rt_policy(p)) {
4787 p->static_prio = NICE_TO_PRIO(nice);
4790 on_rq = p->se.on_rq;
4792 dequeue_task(rq, p, 0);
4794 p->static_prio = NICE_TO_PRIO(nice);
4797 p->prio = effective_prio(p);
4798 delta = p->prio - old_prio;
4801 enqueue_task(rq, p, 0);
4803 * If the task increased its priority or is running and
4804 * lowered its priority, then reschedule its CPU:
4806 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4807 resched_task(rq->curr);
4810 task_rq_unlock(rq, &flags);
4812 EXPORT_SYMBOL(set_user_nice);
4815 * can_nice - check if a task can reduce its nice value
4819 int can_nice(const struct task_struct *p, const int nice)
4821 /* convert nice value [19,-20] to rlimit style value [1,40] */
4822 int nice_rlim = 20 - nice;
4824 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4825 capable(CAP_SYS_NICE));
4828 #ifdef __ARCH_WANT_SYS_NICE
4831 * sys_nice - change the priority of the current process.
4832 * @increment: priority increment
4834 * sys_setpriority is a more generic, but much slower function that
4835 * does similar things.
4837 asmlinkage long sys_nice(int increment)
4842 * Setpriority might change our priority at the same moment.
4843 * We don't have to worry. Conceptually one call occurs first
4844 * and we have a single winner.
4846 if (increment < -40)
4851 nice = PRIO_TO_NICE(current->static_prio) + increment;
4857 if (increment < 0 && !can_nice(current, nice))
4860 retval = security_task_setnice(current, nice);
4864 set_user_nice(current, nice);
4871 * task_prio - return the priority value of a given task.
4872 * @p: the task in question.
4874 * This is the priority value as seen by users in /proc.
4875 * RT tasks are offset by -200. Normal tasks are centered
4876 * around 0, value goes from -16 to +15.
4878 int task_prio(const struct task_struct *p)
4880 return p->prio - MAX_RT_PRIO;
4884 * task_nice - return the nice value of a given task.
4885 * @p: the task in question.
4887 int task_nice(const struct task_struct *p)
4889 return TASK_NICE(p);
4891 EXPORT_SYMBOL(task_nice);
4894 * idle_cpu - is a given cpu idle currently?
4895 * @cpu: the processor in question.
4897 int idle_cpu(int cpu)
4899 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4903 * idle_task - return the idle task for a given cpu.
4904 * @cpu: the processor in question.
4906 struct task_struct *idle_task(int cpu)
4908 return cpu_rq(cpu)->idle;
4912 * find_process_by_pid - find a process with a matching PID value.
4913 * @pid: the pid in question.
4915 static struct task_struct *find_process_by_pid(pid_t pid)
4917 return pid ? find_task_by_vpid(pid) : current;
4920 /* Actually do priority change: must hold rq lock. */
4922 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4924 BUG_ON(p->se.on_rq);
4927 switch (p->policy) {
4931 p->sched_class = &fair_sched_class;
4935 p->sched_class = &rt_sched_class;
4939 p->rt_priority = prio;
4940 p->normal_prio = normal_prio(p);
4941 /* we are holding p->pi_lock already */
4942 p->prio = rt_mutex_getprio(p);
4947 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4948 * @p: the task in question.
4949 * @policy: new policy.
4950 * @param: structure containing the new RT priority.
4952 * NOTE that the task may be already dead.
4954 int sched_setscheduler(struct task_struct *p, int policy,
4955 struct sched_param *param)
4957 int retval, oldprio, oldpolicy = -1, on_rq, running;
4958 unsigned long flags;
4959 const struct sched_class *prev_class = p->sched_class;
4962 /* may grab non-irq protected spin_locks */
4963 BUG_ON(in_interrupt());
4965 /* double check policy once rq lock held */
4967 policy = oldpolicy = p->policy;
4968 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4969 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4970 policy != SCHED_IDLE)
4973 * Valid priorities for SCHED_FIFO and SCHED_RR are
4974 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4975 * SCHED_BATCH and SCHED_IDLE is 0.
4977 if (param->sched_priority < 0 ||
4978 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4979 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4981 if (rt_policy(policy) != (param->sched_priority != 0))
4985 * Allow unprivileged RT tasks to decrease priority:
4987 if (!capable(CAP_SYS_NICE)) {
4988 if (rt_policy(policy)) {
4989 unsigned long rlim_rtprio;
4991 if (!lock_task_sighand(p, &flags))
4993 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4994 unlock_task_sighand(p, &flags);
4996 /* can't set/change the rt policy */
4997 if (policy != p->policy && !rlim_rtprio)
5000 /* can't increase priority */
5001 if (param->sched_priority > p->rt_priority &&
5002 param->sched_priority > rlim_rtprio)
5006 * Like positive nice levels, dont allow tasks to
5007 * move out of SCHED_IDLE either:
5009 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
5012 /* can't change other user's priorities */
5013 if ((current->euid != p->euid) &&
5014 (current->euid != p->uid))
5018 #ifdef CONFIG_RT_GROUP_SCHED
5020 * Do not allow realtime tasks into groups that have no runtime
5023 if (rt_policy(policy) && task_group(p)->rt_bandwidth.rt_runtime == 0)
5027 retval = security_task_setscheduler(p, policy, param);
5031 * make sure no PI-waiters arrive (or leave) while we are
5032 * changing the priority of the task:
5034 spin_lock_irqsave(&p->pi_lock, flags);
5036 * To be able to change p->policy safely, the apropriate
5037 * runqueue lock must be held.
5039 rq = __task_rq_lock(p);
5040 /* recheck policy now with rq lock held */
5041 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5042 policy = oldpolicy = -1;
5043 __task_rq_unlock(rq);
5044 spin_unlock_irqrestore(&p->pi_lock, flags);
5047 update_rq_clock(rq);
5048 on_rq = p->se.on_rq;
5049 running = task_current(rq, p);
5051 deactivate_task(rq, p, 0);
5053 p->sched_class->put_prev_task(rq, p);
5056 __setscheduler(rq, p, policy, param->sched_priority);
5059 p->sched_class->set_curr_task(rq);
5061 activate_task(rq, p, 0);
5063 check_class_changed(rq, p, prev_class, oldprio, running);
5065 __task_rq_unlock(rq);
5066 spin_unlock_irqrestore(&p->pi_lock, flags);
5068 rt_mutex_adjust_pi(p);
5072 EXPORT_SYMBOL_GPL(sched_setscheduler);
5075 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5077 struct sched_param lparam;
5078 struct task_struct *p;
5081 if (!param || pid < 0)
5083 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5088 p = find_process_by_pid(pid);
5090 retval = sched_setscheduler(p, policy, &lparam);
5097 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5098 * @pid: the pid in question.
5099 * @policy: new policy.
5100 * @param: structure containing the new RT priority.
5103 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5105 /* negative values for policy are not valid */
5109 return do_sched_setscheduler(pid, policy, param);
5113 * sys_sched_setparam - set/change the RT priority of a thread
5114 * @pid: the pid in question.
5115 * @param: structure containing the new RT priority.
5117 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
5119 return do_sched_setscheduler(pid, -1, param);
5123 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5124 * @pid: the pid in question.
5126 asmlinkage long sys_sched_getscheduler(pid_t pid)
5128 struct task_struct *p;
5135 read_lock(&tasklist_lock);
5136 p = find_process_by_pid(pid);
5138 retval = security_task_getscheduler(p);
5142 read_unlock(&tasklist_lock);
5147 * sys_sched_getscheduler - get the RT priority of a thread
5148 * @pid: the pid in question.
5149 * @param: structure containing the RT priority.
5151 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
5153 struct sched_param lp;
5154 struct task_struct *p;
5157 if (!param || pid < 0)
5160 read_lock(&tasklist_lock);
5161 p = find_process_by_pid(pid);
5166 retval = security_task_getscheduler(p);
5170 lp.sched_priority = p->rt_priority;
5171 read_unlock(&tasklist_lock);
5174 * This one might sleep, we cannot do it with a spinlock held ...
5176 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5181 read_unlock(&tasklist_lock);
5185 long sched_setaffinity(pid_t pid, const cpumask_t *in_mask)
5187 cpumask_t cpus_allowed;
5188 cpumask_t new_mask = *in_mask;
5189 struct task_struct *p;
5193 read_lock(&tasklist_lock);
5195 p = find_process_by_pid(pid);
5197 read_unlock(&tasklist_lock);
5203 * It is not safe to call set_cpus_allowed with the
5204 * tasklist_lock held. We will bump the task_struct's
5205 * usage count and then drop tasklist_lock.
5208 read_unlock(&tasklist_lock);
5211 if ((current->euid != p->euid) && (current->euid != p->uid) &&
5212 !capable(CAP_SYS_NICE))
5215 retval = security_task_setscheduler(p, 0, NULL);
5219 cpuset_cpus_allowed(p, &cpus_allowed);
5220 cpus_and(new_mask, new_mask, cpus_allowed);
5222 retval = set_cpus_allowed_ptr(p, &new_mask);
5225 cpuset_cpus_allowed(p, &cpus_allowed);
5226 if (!cpus_subset(new_mask, cpus_allowed)) {
5228 * We must have raced with a concurrent cpuset
5229 * update. Just reset the cpus_allowed to the
5230 * cpuset's cpus_allowed
5232 new_mask = cpus_allowed;
5242 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5243 cpumask_t *new_mask)
5245 if (len < sizeof(cpumask_t)) {
5246 memset(new_mask, 0, sizeof(cpumask_t));
5247 } else if (len > sizeof(cpumask_t)) {
5248 len = sizeof(cpumask_t);
5250 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5254 * sys_sched_setaffinity - set the cpu affinity of a process
5255 * @pid: pid of the process
5256 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5257 * @user_mask_ptr: user-space pointer to the new cpu mask
5259 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
5260 unsigned long __user *user_mask_ptr)
5265 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
5269 return sched_setaffinity(pid, &new_mask);
5272 long sched_getaffinity(pid_t pid, cpumask_t *mask)
5274 struct task_struct *p;
5278 read_lock(&tasklist_lock);
5281 p = find_process_by_pid(pid);
5285 retval = security_task_getscheduler(p);
5289 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
5292 read_unlock(&tasklist_lock);
5299 * sys_sched_getaffinity - get the cpu affinity of a process
5300 * @pid: pid of the process
5301 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5302 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5304 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5305 unsigned long __user *user_mask_ptr)
5310 if (len < sizeof(cpumask_t))
5313 ret = sched_getaffinity(pid, &mask);
5317 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5320 return sizeof(cpumask_t);
5324 * sys_sched_yield - yield the current processor to other threads.
5326 * This function yields the current CPU to other tasks. If there are no
5327 * other threads running on this CPU then this function will return.
5329 asmlinkage long sys_sched_yield(void)
5331 struct rq *rq = this_rq_lock();
5333 schedstat_inc(rq, yld_count);
5334 current->sched_class->yield_task(rq);
5337 * Since we are going to call schedule() anyway, there's
5338 * no need to preempt or enable interrupts:
5340 __release(rq->lock);
5341 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5342 _raw_spin_unlock(&rq->lock);
5343 preempt_enable_no_resched();
5350 static void __cond_resched(void)
5352 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5353 __might_sleep(__FILE__, __LINE__);
5356 * The BKS might be reacquired before we have dropped
5357 * PREEMPT_ACTIVE, which could trigger a second
5358 * cond_resched() call.
5361 add_preempt_count(PREEMPT_ACTIVE);
5363 sub_preempt_count(PREEMPT_ACTIVE);
5364 } while (need_resched());
5367 int __sched _cond_resched(void)
5369 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5370 system_state == SYSTEM_RUNNING) {
5376 EXPORT_SYMBOL(_cond_resched);
5379 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5380 * call schedule, and on return reacquire the lock.
5382 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5383 * operations here to prevent schedule() from being called twice (once via
5384 * spin_unlock(), once by hand).
5386 int cond_resched_lock(spinlock_t *lock)
5388 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5391 if (spin_needbreak(lock) || resched) {
5393 if (resched && need_resched())
5402 EXPORT_SYMBOL(cond_resched_lock);
5404 int __sched cond_resched_softirq(void)
5406 BUG_ON(!in_softirq());
5408 if (need_resched() && system_state == SYSTEM_RUNNING) {
5416 EXPORT_SYMBOL(cond_resched_softirq);
5419 * yield - yield the current processor to other threads.
5421 * This is a shortcut for kernel-space yielding - it marks the
5422 * thread runnable and calls sys_sched_yield().
5424 void __sched yield(void)
5426 set_current_state(TASK_RUNNING);
5429 EXPORT_SYMBOL(yield);
5432 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5433 * that process accounting knows that this is a task in IO wait state.
5435 * But don't do that if it is a deliberate, throttling IO wait (this task
5436 * has set its backing_dev_info: the queue against which it should throttle)
5438 void __sched io_schedule(void)
5440 struct rq *rq = &__raw_get_cpu_var(runqueues);
5442 delayacct_blkio_start();
5443 atomic_inc(&rq->nr_iowait);
5445 atomic_dec(&rq->nr_iowait);
5446 delayacct_blkio_end();
5448 EXPORT_SYMBOL(io_schedule);
5450 long __sched io_schedule_timeout(long timeout)
5452 struct rq *rq = &__raw_get_cpu_var(runqueues);
5455 delayacct_blkio_start();
5456 atomic_inc(&rq->nr_iowait);
5457 ret = schedule_timeout(timeout);
5458 atomic_dec(&rq->nr_iowait);
5459 delayacct_blkio_end();
5464 * sys_sched_get_priority_max - return maximum RT priority.
5465 * @policy: scheduling class.
5467 * this syscall returns the maximum rt_priority that can be used
5468 * by a given scheduling class.
5470 asmlinkage long sys_sched_get_priority_max(int policy)
5477 ret = MAX_USER_RT_PRIO-1;
5489 * sys_sched_get_priority_min - return minimum RT priority.
5490 * @policy: scheduling class.
5492 * this syscall returns the minimum rt_priority that can be used
5493 * by a given scheduling class.
5495 asmlinkage long sys_sched_get_priority_min(int policy)
5513 * sys_sched_rr_get_interval - return the default timeslice of a process.
5514 * @pid: pid of the process.
5515 * @interval: userspace pointer to the timeslice value.
5517 * this syscall writes the default timeslice value of a given process
5518 * into the user-space timespec buffer. A value of '0' means infinity.
5521 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5523 struct task_struct *p;
5524 unsigned int time_slice;
5532 read_lock(&tasklist_lock);
5533 p = find_process_by_pid(pid);
5537 retval = security_task_getscheduler(p);
5542 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5543 * tasks that are on an otherwise idle runqueue:
5546 if (p->policy == SCHED_RR) {
5547 time_slice = DEF_TIMESLICE;
5548 } else if (p->policy != SCHED_FIFO) {
5549 struct sched_entity *se = &p->se;
5550 unsigned long flags;
5553 rq = task_rq_lock(p, &flags);
5554 if (rq->cfs.load.weight)
5555 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5556 task_rq_unlock(rq, &flags);
5558 read_unlock(&tasklist_lock);
5559 jiffies_to_timespec(time_slice, &t);
5560 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5564 read_unlock(&tasklist_lock);
5568 static const char stat_nam[] = "RSDTtZX";
5570 void sched_show_task(struct task_struct *p)
5572 unsigned long free = 0;
5575 state = p->state ? __ffs(p->state) + 1 : 0;
5576 printk(KERN_INFO "%-13.13s %c", p->comm,
5577 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5578 #if BITS_PER_LONG == 32
5579 if (state == TASK_RUNNING)
5580 printk(KERN_CONT " running ");
5582 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5584 if (state == TASK_RUNNING)
5585 printk(KERN_CONT " running task ");
5587 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5589 #ifdef CONFIG_DEBUG_STACK_USAGE
5591 unsigned long *n = end_of_stack(p);
5594 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5597 printk(KERN_CONT "%5lu %5d %6d\n", free,
5598 task_pid_nr(p), task_pid_nr(p->real_parent));
5600 show_stack(p, NULL);
5603 void show_state_filter(unsigned long state_filter)
5605 struct task_struct *g, *p;
5607 #if BITS_PER_LONG == 32
5609 " task PC stack pid father\n");
5612 " task PC stack pid father\n");
5614 read_lock(&tasklist_lock);
5615 do_each_thread(g, p) {
5617 * reset the NMI-timeout, listing all files on a slow
5618 * console might take alot of time:
5620 touch_nmi_watchdog();
5621 if (!state_filter || (p->state & state_filter))
5623 } while_each_thread(g, p);
5625 touch_all_softlockup_watchdogs();
5627 #ifdef CONFIG_SCHED_DEBUG
5628 sysrq_sched_debug_show();
5630 read_unlock(&tasklist_lock);
5632 * Only show locks if all tasks are dumped:
5634 if (state_filter == -1)
5635 debug_show_all_locks();
5638 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5640 idle->sched_class = &idle_sched_class;
5644 * init_idle - set up an idle thread for a given CPU
5645 * @idle: task in question
5646 * @cpu: cpu the idle task belongs to
5648 * NOTE: this function does not set the idle thread's NEED_RESCHED
5649 * flag, to make booting more robust.
5651 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5653 struct rq *rq = cpu_rq(cpu);
5654 unsigned long flags;
5657 idle->se.exec_start = sched_clock();
5659 idle->prio = idle->normal_prio = MAX_PRIO;
5660 idle->cpus_allowed = cpumask_of_cpu(cpu);
5661 __set_task_cpu(idle, cpu);
5663 spin_lock_irqsave(&rq->lock, flags);
5664 rq->curr = rq->idle = idle;
5665 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5668 spin_unlock_irqrestore(&rq->lock, flags);
5670 /* Set the preempt count _outside_ the spinlocks! */
5671 #if defined(CONFIG_PREEMPT)
5672 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5674 task_thread_info(idle)->preempt_count = 0;
5677 * The idle tasks have their own, simple scheduling class:
5679 idle->sched_class = &idle_sched_class;
5683 * In a system that switches off the HZ timer nohz_cpu_mask
5684 * indicates which cpus entered this state. This is used
5685 * in the rcu update to wait only for active cpus. For system
5686 * which do not switch off the HZ timer nohz_cpu_mask should
5687 * always be CPU_MASK_NONE.
5689 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5692 * Increase the granularity value when there are more CPUs,
5693 * because with more CPUs the 'effective latency' as visible
5694 * to users decreases. But the relationship is not linear,
5695 * so pick a second-best guess by going with the log2 of the
5698 * This idea comes from the SD scheduler of Con Kolivas:
5700 static inline void sched_init_granularity(void)
5702 unsigned int factor = 1 + ilog2(num_online_cpus());
5703 const unsigned long limit = 200000000;
5705 sysctl_sched_min_granularity *= factor;
5706 if (sysctl_sched_min_granularity > limit)
5707 sysctl_sched_min_granularity = limit;
5709 sysctl_sched_latency *= factor;
5710 if (sysctl_sched_latency > limit)
5711 sysctl_sched_latency = limit;
5713 sysctl_sched_wakeup_granularity *= factor;
5718 * This is how migration works:
5720 * 1) we queue a struct migration_req structure in the source CPU's
5721 * runqueue and wake up that CPU's migration thread.
5722 * 2) we down() the locked semaphore => thread blocks.
5723 * 3) migration thread wakes up (implicitly it forces the migrated
5724 * thread off the CPU)
5725 * 4) it gets the migration request and checks whether the migrated
5726 * task is still in the wrong runqueue.
5727 * 5) if it's in the wrong runqueue then the migration thread removes
5728 * it and puts it into the right queue.
5729 * 6) migration thread up()s the semaphore.
5730 * 7) we wake up and the migration is done.
5734 * Change a given task's CPU affinity. Migrate the thread to a
5735 * proper CPU and schedule it away if the CPU it's executing on
5736 * is removed from the allowed bitmask.
5738 * NOTE: the caller must have a valid reference to the task, the
5739 * task must not exit() & deallocate itself prematurely. The
5740 * call is not atomic; no spinlocks may be held.
5742 int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask)
5744 struct migration_req req;
5745 unsigned long flags;
5749 rq = task_rq_lock(p, &flags);
5750 if (!cpus_intersects(*new_mask, cpu_online_map)) {
5755 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5756 !cpus_equal(p->cpus_allowed, *new_mask))) {
5761 if (p->sched_class->set_cpus_allowed)
5762 p->sched_class->set_cpus_allowed(p, new_mask);
5764 p->cpus_allowed = *new_mask;
5765 p->rt.nr_cpus_allowed = cpus_weight(*new_mask);
5768 /* Can the task run on the task's current CPU? If so, we're done */
5769 if (cpu_isset(task_cpu(p), *new_mask))
5772 if (migrate_task(p, any_online_cpu(*new_mask), &req)) {
5773 /* Need help from migration thread: drop lock and wait. */
5774 task_rq_unlock(rq, &flags);
5775 wake_up_process(rq->migration_thread);
5776 wait_for_completion(&req.done);
5777 tlb_migrate_finish(p->mm);
5781 task_rq_unlock(rq, &flags);
5785 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5788 * Move (not current) task off this cpu, onto dest cpu. We're doing
5789 * this because either it can't run here any more (set_cpus_allowed()
5790 * away from this CPU, or CPU going down), or because we're
5791 * attempting to rebalance this task on exec (sched_exec).
5793 * So we race with normal scheduler movements, but that's OK, as long
5794 * as the task is no longer on this CPU.
5796 * Returns non-zero if task was successfully migrated.
5798 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5800 struct rq *rq_dest, *rq_src;
5803 if (unlikely(cpu_is_offline(dest_cpu)))
5806 rq_src = cpu_rq(src_cpu);
5807 rq_dest = cpu_rq(dest_cpu);
5809 double_rq_lock(rq_src, rq_dest);
5810 /* Already moved. */
5811 if (task_cpu(p) != src_cpu)
5813 /* Affinity changed (again). */
5814 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5817 on_rq = p->se.on_rq;
5819 deactivate_task(rq_src, p, 0);
5821 set_task_cpu(p, dest_cpu);
5823 activate_task(rq_dest, p, 0);
5824 check_preempt_curr(rq_dest, p);
5828 double_rq_unlock(rq_src, rq_dest);
5833 * migration_thread - this is a highprio system thread that performs
5834 * thread migration by bumping thread off CPU then 'pushing' onto
5837 static int migration_thread(void *data)
5839 int cpu = (long)data;
5843 BUG_ON(rq->migration_thread != current);
5845 set_current_state(TASK_INTERRUPTIBLE);
5846 while (!kthread_should_stop()) {
5847 struct migration_req *req;
5848 struct list_head *head;
5850 spin_lock_irq(&rq->lock);
5852 if (cpu_is_offline(cpu)) {
5853 spin_unlock_irq(&rq->lock);
5857 if (rq->active_balance) {
5858 active_load_balance(rq, cpu);
5859 rq->active_balance = 0;
5862 head = &rq->migration_queue;
5864 if (list_empty(head)) {
5865 spin_unlock_irq(&rq->lock);
5867 set_current_state(TASK_INTERRUPTIBLE);
5870 req = list_entry(head->next, struct migration_req, list);
5871 list_del_init(head->next);
5873 spin_unlock(&rq->lock);
5874 __migrate_task(req->task, cpu, req->dest_cpu);
5877 complete(&req->done);
5879 __set_current_state(TASK_RUNNING);
5883 /* Wait for kthread_stop */
5884 set_current_state(TASK_INTERRUPTIBLE);
5885 while (!kthread_should_stop()) {
5887 set_current_state(TASK_INTERRUPTIBLE);
5889 __set_current_state(TASK_RUNNING);
5893 #ifdef CONFIG_HOTPLUG_CPU
5895 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5899 local_irq_disable();
5900 ret = __migrate_task(p, src_cpu, dest_cpu);
5906 * Figure out where task on dead CPU should go, use force if necessary.
5907 * NOTE: interrupts should be disabled by the caller
5909 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5911 unsigned long flags;
5918 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5919 cpus_and(mask, mask, p->cpus_allowed);
5920 dest_cpu = any_online_cpu(mask);
5922 /* On any allowed CPU? */
5923 if (dest_cpu >= nr_cpu_ids)
5924 dest_cpu = any_online_cpu(p->cpus_allowed);
5926 /* No more Mr. Nice Guy. */
5927 if (dest_cpu >= nr_cpu_ids) {
5928 cpumask_t cpus_allowed;
5930 cpuset_cpus_allowed_locked(p, &cpus_allowed);
5932 * Try to stay on the same cpuset, where the
5933 * current cpuset may be a subset of all cpus.
5934 * The cpuset_cpus_allowed_locked() variant of
5935 * cpuset_cpus_allowed() will not block. It must be
5936 * called within calls to cpuset_lock/cpuset_unlock.
5938 rq = task_rq_lock(p, &flags);
5939 p->cpus_allowed = cpus_allowed;
5940 dest_cpu = any_online_cpu(p->cpus_allowed);
5941 task_rq_unlock(rq, &flags);
5944 * Don't tell them about moving exiting tasks or
5945 * kernel threads (both mm NULL), since they never
5948 if (p->mm && printk_ratelimit()) {
5949 printk(KERN_INFO "process %d (%s) no "
5950 "longer affine to cpu%d\n",
5951 task_pid_nr(p), p->comm, dead_cpu);
5954 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
5958 * While a dead CPU has no uninterruptible tasks queued at this point,
5959 * it might still have a nonzero ->nr_uninterruptible counter, because
5960 * for performance reasons the counter is not stricly tracking tasks to
5961 * their home CPUs. So we just add the counter to another CPU's counter,
5962 * to keep the global sum constant after CPU-down:
5964 static void migrate_nr_uninterruptible(struct rq *rq_src)
5966 struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR));
5967 unsigned long flags;
5969 local_irq_save(flags);
5970 double_rq_lock(rq_src, rq_dest);
5971 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5972 rq_src->nr_uninterruptible = 0;
5973 double_rq_unlock(rq_src, rq_dest);
5974 local_irq_restore(flags);
5977 /* Run through task list and migrate tasks from the dead cpu. */
5978 static void migrate_live_tasks(int src_cpu)
5980 struct task_struct *p, *t;
5982 read_lock(&tasklist_lock);
5984 do_each_thread(t, p) {
5988 if (task_cpu(p) == src_cpu)
5989 move_task_off_dead_cpu(src_cpu, p);
5990 } while_each_thread(t, p);
5992 read_unlock(&tasklist_lock);
5996 * Schedules idle task to be the next runnable task on current CPU.
5997 * It does so by boosting its priority to highest possible.
5998 * Used by CPU offline code.
6000 void sched_idle_next(void)
6002 int this_cpu = smp_processor_id();
6003 struct rq *rq = cpu_rq(this_cpu);
6004 struct task_struct *p = rq->idle;
6005 unsigned long flags;
6007 /* cpu has to be offline */
6008 BUG_ON(cpu_online(this_cpu));
6011 * Strictly not necessary since rest of the CPUs are stopped by now
6012 * and interrupts disabled on the current cpu.
6014 spin_lock_irqsave(&rq->lock, flags);
6016 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6018 update_rq_clock(rq);
6019 activate_task(rq, p, 0);
6021 spin_unlock_irqrestore(&rq->lock, flags);
6025 * Ensures that the idle task is using init_mm right before its cpu goes
6028 void idle_task_exit(void)
6030 struct mm_struct *mm = current->active_mm;
6032 BUG_ON(cpu_online(smp_processor_id()));
6035 switch_mm(mm, &init_mm, current);
6039 /* called under rq->lock with disabled interrupts */
6040 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
6042 struct rq *rq = cpu_rq(dead_cpu);
6044 /* Must be exiting, otherwise would be on tasklist. */
6045 BUG_ON(!p->exit_state);
6047 /* Cannot have done final schedule yet: would have vanished. */
6048 BUG_ON(p->state == TASK_DEAD);
6053 * Drop lock around migration; if someone else moves it,
6054 * that's OK. No task can be added to this CPU, so iteration is
6057 spin_unlock_irq(&rq->lock);
6058 move_task_off_dead_cpu(dead_cpu, p);
6059 spin_lock_irq(&rq->lock);
6064 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6065 static void migrate_dead_tasks(unsigned int dead_cpu)
6067 struct rq *rq = cpu_rq(dead_cpu);
6068 struct task_struct *next;
6071 if (!rq->nr_running)
6073 update_rq_clock(rq);
6074 next = pick_next_task(rq, rq->curr);
6077 migrate_dead(dead_cpu, next);
6081 #endif /* CONFIG_HOTPLUG_CPU */
6083 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6085 static struct ctl_table sd_ctl_dir[] = {
6087 .procname = "sched_domain",
6093 static struct ctl_table sd_ctl_root[] = {
6095 .ctl_name = CTL_KERN,
6096 .procname = "kernel",
6098 .child = sd_ctl_dir,
6103 static struct ctl_table *sd_alloc_ctl_entry(int n)
6105 struct ctl_table *entry =
6106 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6111 static void sd_free_ctl_entry(struct ctl_table **tablep)
6113 struct ctl_table *entry;
6116 * In the intermediate directories, both the child directory and
6117 * procname are dynamically allocated and could fail but the mode
6118 * will always be set. In the lowest directory the names are
6119 * static strings and all have proc handlers.
6121 for (entry = *tablep; entry->mode; entry++) {
6123 sd_free_ctl_entry(&entry->child);
6124 if (entry->proc_handler == NULL)
6125 kfree(entry->procname);
6133 set_table_entry(struct ctl_table *entry,
6134 const char *procname, void *data, int maxlen,
6135 mode_t mode, proc_handler *proc_handler)
6137 entry->procname = procname;
6139 entry->maxlen = maxlen;
6141 entry->proc_handler = proc_handler;
6144 static struct ctl_table *
6145 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6147 struct ctl_table *table = sd_alloc_ctl_entry(12);
6152 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6153 sizeof(long), 0644, proc_doulongvec_minmax);
6154 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6155 sizeof(long), 0644, proc_doulongvec_minmax);
6156 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6157 sizeof(int), 0644, proc_dointvec_minmax);
6158 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6159 sizeof(int), 0644, proc_dointvec_minmax);
6160 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6161 sizeof(int), 0644, proc_dointvec_minmax);
6162 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6163 sizeof(int), 0644, proc_dointvec_minmax);
6164 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6165 sizeof(int), 0644, proc_dointvec_minmax);
6166 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6167 sizeof(int), 0644, proc_dointvec_minmax);
6168 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6169 sizeof(int), 0644, proc_dointvec_minmax);
6170 set_table_entry(&table[9], "cache_nice_tries",
6171 &sd->cache_nice_tries,
6172 sizeof(int), 0644, proc_dointvec_minmax);
6173 set_table_entry(&table[10], "flags", &sd->flags,
6174 sizeof(int), 0644, proc_dointvec_minmax);
6175 /* &table[11] is terminator */
6180 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6182 struct ctl_table *entry, *table;
6183 struct sched_domain *sd;
6184 int domain_num = 0, i;
6187 for_each_domain(cpu, sd)
6189 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6194 for_each_domain(cpu, sd) {
6195 snprintf(buf, 32, "domain%d", i);
6196 entry->procname = kstrdup(buf, GFP_KERNEL);
6198 entry->child = sd_alloc_ctl_domain_table(sd);
6205 static struct ctl_table_header *sd_sysctl_header;
6206 static void register_sched_domain_sysctl(void)
6208 int i, cpu_num = num_online_cpus();
6209 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6212 WARN_ON(sd_ctl_dir[0].child);
6213 sd_ctl_dir[0].child = entry;
6218 for_each_online_cpu(i) {
6219 snprintf(buf, 32, "cpu%d", i);
6220 entry->procname = kstrdup(buf, GFP_KERNEL);
6222 entry->child = sd_alloc_ctl_cpu_table(i);
6226 WARN_ON(sd_sysctl_header);
6227 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6230 /* may be called multiple times per register */
6231 static void unregister_sched_domain_sysctl(void)
6233 if (sd_sysctl_header)
6234 unregister_sysctl_table(sd_sysctl_header);
6235 sd_sysctl_header = NULL;
6236 if (sd_ctl_dir[0].child)
6237 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6240 static void register_sched_domain_sysctl(void)
6243 static void unregister_sched_domain_sysctl(void)
6248 static void set_rq_online(struct rq *rq)
6251 const struct sched_class *class;
6253 cpu_set(rq->cpu, rq->rd->online);
6256 for_each_class(class) {
6257 if (class->rq_online)
6258 class->rq_online(rq);
6263 static void set_rq_offline(struct rq *rq)
6266 const struct sched_class *class;
6268 for_each_class(class) {
6269 if (class->rq_offline)
6270 class->rq_offline(rq);
6273 cpu_clear(rq->cpu, rq->rd->online);
6279 * migration_call - callback that gets triggered when a CPU is added.
6280 * Here we can start up the necessary migration thread for the new CPU.
6282 static int __cpuinit
6283 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6285 struct task_struct *p;
6286 int cpu = (long)hcpu;
6287 unsigned long flags;
6292 case CPU_UP_PREPARE:
6293 case CPU_UP_PREPARE_FROZEN:
6294 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6297 kthread_bind(p, cpu);
6298 /* Must be high prio: stop_machine expects to yield to it. */
6299 rq = task_rq_lock(p, &flags);
6300 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6301 task_rq_unlock(rq, &flags);
6302 cpu_rq(cpu)->migration_thread = p;
6306 case CPU_ONLINE_FROZEN:
6307 /* Strictly unnecessary, as first user will wake it. */
6308 wake_up_process(cpu_rq(cpu)->migration_thread);
6310 /* Update our root-domain */
6312 spin_lock_irqsave(&rq->lock, flags);
6314 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6318 spin_unlock_irqrestore(&rq->lock, flags);
6321 #ifdef CONFIG_HOTPLUG_CPU
6322 case CPU_UP_CANCELED:
6323 case CPU_UP_CANCELED_FROZEN:
6324 if (!cpu_rq(cpu)->migration_thread)
6326 /* Unbind it from offline cpu so it can run. Fall thru. */
6327 kthread_bind(cpu_rq(cpu)->migration_thread,
6328 any_online_cpu(cpu_online_map));
6329 kthread_stop(cpu_rq(cpu)->migration_thread);
6330 cpu_rq(cpu)->migration_thread = NULL;
6334 case CPU_DEAD_FROZEN:
6335 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6336 migrate_live_tasks(cpu);
6338 kthread_stop(rq->migration_thread);
6339 rq->migration_thread = NULL;
6340 /* Idle task back to normal (off runqueue, low prio) */
6341 spin_lock_irq(&rq->lock);
6342 update_rq_clock(rq);
6343 deactivate_task(rq, rq->idle, 0);
6344 rq->idle->static_prio = MAX_PRIO;
6345 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6346 rq->idle->sched_class = &idle_sched_class;
6347 migrate_dead_tasks(cpu);
6348 spin_unlock_irq(&rq->lock);
6350 migrate_nr_uninterruptible(rq);
6351 BUG_ON(rq->nr_running != 0);
6354 * No need to migrate the tasks: it was best-effort if
6355 * they didn't take sched_hotcpu_mutex. Just wake up
6358 spin_lock_irq(&rq->lock);
6359 while (!list_empty(&rq->migration_queue)) {
6360 struct migration_req *req;
6362 req = list_entry(rq->migration_queue.next,
6363 struct migration_req, list);
6364 list_del_init(&req->list);
6365 complete(&req->done);
6367 spin_unlock_irq(&rq->lock);
6371 case CPU_DYING_FROZEN:
6372 /* Update our root-domain */
6374 spin_lock_irqsave(&rq->lock, flags);
6376 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6379 spin_unlock_irqrestore(&rq->lock, flags);
6386 /* Register at highest priority so that task migration (migrate_all_tasks)
6387 * happens before everything else.
6389 static struct notifier_block __cpuinitdata migration_notifier = {
6390 .notifier_call = migration_call,
6394 void __init migration_init(void)
6396 void *cpu = (void *)(long)smp_processor_id();
6399 /* Start one for the boot CPU: */
6400 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6401 BUG_ON(err == NOTIFY_BAD);
6402 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6403 register_cpu_notifier(&migration_notifier);
6409 #ifdef CONFIG_SCHED_DEBUG
6411 static inline const char *sd_level_to_string(enum sched_domain_level lvl)
6424 case SD_LV_ALLNODES:
6433 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6434 cpumask_t *groupmask)
6436 struct sched_group *group = sd->groups;
6439 cpulist_scnprintf(str, sizeof(str), sd->span);
6440 cpus_clear(*groupmask);
6442 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6444 if (!(sd->flags & SD_LOAD_BALANCE)) {
6445 printk("does not load-balance\n");
6447 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6452 printk(KERN_CONT "span %s level %s\n",
6453 str, sd_level_to_string(sd->level));
6455 if (!cpu_isset(cpu, sd->span)) {
6456 printk(KERN_ERR "ERROR: domain->span does not contain "
6459 if (!cpu_isset(cpu, group->cpumask)) {
6460 printk(KERN_ERR "ERROR: domain->groups does not contain"
6464 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6468 printk(KERN_ERR "ERROR: group is NULL\n");
6472 if (!group->__cpu_power) {
6473 printk(KERN_CONT "\n");
6474 printk(KERN_ERR "ERROR: domain->cpu_power not "
6479 if (!cpus_weight(group->cpumask)) {
6480 printk(KERN_CONT "\n");
6481 printk(KERN_ERR "ERROR: empty group\n");
6485 if (cpus_intersects(*groupmask, group->cpumask)) {
6486 printk(KERN_CONT "\n");
6487 printk(KERN_ERR "ERROR: repeated CPUs\n");
6491 cpus_or(*groupmask, *groupmask, group->cpumask);
6493 cpulist_scnprintf(str, sizeof(str), group->cpumask);
6494 printk(KERN_CONT " %s", str);
6496 group = group->next;
6497 } while (group != sd->groups);
6498 printk(KERN_CONT "\n");
6500 if (!cpus_equal(sd->span, *groupmask))
6501 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6503 if (sd->parent && !cpus_subset(*groupmask, sd->parent->span))
6504 printk(KERN_ERR "ERROR: parent span is not a superset "
6505 "of domain->span\n");
6509 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6511 cpumask_t *groupmask;
6515 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6519 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6521 groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6523 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6528 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6537 #else /* !CONFIG_SCHED_DEBUG */
6538 # define sched_domain_debug(sd, cpu) do { } while (0)
6539 #endif /* CONFIG_SCHED_DEBUG */
6541 static int sd_degenerate(struct sched_domain *sd)
6543 if (cpus_weight(sd->span) == 1)
6546 /* Following flags need at least 2 groups */
6547 if (sd->flags & (SD_LOAD_BALANCE |
6548 SD_BALANCE_NEWIDLE |
6552 SD_SHARE_PKG_RESOURCES)) {
6553 if (sd->groups != sd->groups->next)
6557 /* Following flags don't use groups */
6558 if (sd->flags & (SD_WAKE_IDLE |
6567 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6569 unsigned long cflags = sd->flags, pflags = parent->flags;
6571 if (sd_degenerate(parent))
6574 if (!cpus_equal(sd->span, parent->span))
6577 /* Does parent contain flags not in child? */
6578 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6579 if (cflags & SD_WAKE_AFFINE)
6580 pflags &= ~SD_WAKE_BALANCE;
6581 /* Flags needing groups don't count if only 1 group in parent */
6582 if (parent->groups == parent->groups->next) {
6583 pflags &= ~(SD_LOAD_BALANCE |
6584 SD_BALANCE_NEWIDLE |
6588 SD_SHARE_PKG_RESOURCES);
6590 if (~cflags & pflags)
6596 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6598 unsigned long flags;
6600 spin_lock_irqsave(&rq->lock, flags);
6603 struct root_domain *old_rd = rq->rd;
6605 if (cpu_isset(rq->cpu, old_rd->online))
6608 cpu_clear(rq->cpu, old_rd->span);
6610 if (atomic_dec_and_test(&old_rd->refcount))
6614 atomic_inc(&rd->refcount);
6617 cpu_set(rq->cpu, rd->span);
6618 if (cpu_isset(rq->cpu, cpu_online_map))
6621 spin_unlock_irqrestore(&rq->lock, flags);
6624 static void init_rootdomain(struct root_domain *rd)
6626 memset(rd, 0, sizeof(*rd));
6628 cpus_clear(rd->span);
6629 cpus_clear(rd->online);
6631 cpupri_init(&rd->cpupri);
6634 static void init_defrootdomain(void)
6636 init_rootdomain(&def_root_domain);
6637 atomic_set(&def_root_domain.refcount, 1);
6640 static struct root_domain *alloc_rootdomain(void)
6642 struct root_domain *rd;
6644 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6648 init_rootdomain(rd);
6654 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6655 * hold the hotplug lock.
6658 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6660 struct rq *rq = cpu_rq(cpu);
6661 struct sched_domain *tmp;
6663 /* Remove the sched domains which do not contribute to scheduling. */
6664 for (tmp = sd; tmp; tmp = tmp->parent) {
6665 struct sched_domain *parent = tmp->parent;
6668 if (sd_parent_degenerate(tmp, parent)) {
6669 tmp->parent = parent->parent;
6671 parent->parent->child = tmp;
6675 if (sd && sd_degenerate(sd)) {
6681 sched_domain_debug(sd, cpu);
6683 rq_attach_root(rq, rd);
6684 rcu_assign_pointer(rq->sd, sd);
6687 /* cpus with isolated domains */
6688 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
6690 /* Setup the mask of cpus configured for isolated domains */
6691 static int __init isolated_cpu_setup(char *str)
6693 int ints[NR_CPUS], i;
6695 str = get_options(str, ARRAY_SIZE(ints), ints);
6696 cpus_clear(cpu_isolated_map);
6697 for (i = 1; i <= ints[0]; i++)
6698 if (ints[i] < NR_CPUS)
6699 cpu_set(ints[i], cpu_isolated_map);
6703 __setup("isolcpus=", isolated_cpu_setup);
6706 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6707 * to a function which identifies what group(along with sched group) a CPU
6708 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6709 * (due to the fact that we keep track of groups covered with a cpumask_t).
6711 * init_sched_build_groups will build a circular linked list of the groups
6712 * covered by the given span, and will set each group's ->cpumask correctly,
6713 * and ->cpu_power to 0.
6716 init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map,
6717 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6718 struct sched_group **sg,
6719 cpumask_t *tmpmask),
6720 cpumask_t *covered, cpumask_t *tmpmask)
6722 struct sched_group *first = NULL, *last = NULL;
6725 cpus_clear(*covered);
6727 for_each_cpu_mask(i, *span) {
6728 struct sched_group *sg;
6729 int group = group_fn(i, cpu_map, &sg, tmpmask);
6732 if (cpu_isset(i, *covered))
6735 cpus_clear(sg->cpumask);
6736 sg->__cpu_power = 0;
6738 for_each_cpu_mask(j, *span) {
6739 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6742 cpu_set(j, *covered);
6743 cpu_set(j, sg->cpumask);
6754 #define SD_NODES_PER_DOMAIN 16
6759 * find_next_best_node - find the next node to include in a sched_domain
6760 * @node: node whose sched_domain we're building
6761 * @used_nodes: nodes already in the sched_domain
6763 * Find the next node to include in a given scheduling domain. Simply
6764 * finds the closest node not already in the @used_nodes map.
6766 * Should use nodemask_t.
6768 static int find_next_best_node(int node, nodemask_t *used_nodes)
6770 int i, n, val, min_val, best_node = 0;
6774 for (i = 0; i < MAX_NUMNODES; i++) {
6775 /* Start at @node */
6776 n = (node + i) % MAX_NUMNODES;
6778 if (!nr_cpus_node(n))
6781 /* Skip already used nodes */
6782 if (node_isset(n, *used_nodes))
6785 /* Simple min distance search */
6786 val = node_distance(node, n);
6788 if (val < min_val) {
6794 node_set(best_node, *used_nodes);
6799 * sched_domain_node_span - get a cpumask for a node's sched_domain
6800 * @node: node whose cpumask we're constructing
6801 * @span: resulting cpumask
6803 * Given a node, construct a good cpumask for its sched_domain to span. It
6804 * should be one that prevents unnecessary balancing, but also spreads tasks
6807 static void sched_domain_node_span(int node, cpumask_t *span)
6809 nodemask_t used_nodes;
6810 node_to_cpumask_ptr(nodemask, node);
6814 nodes_clear(used_nodes);
6816 cpus_or(*span, *span, *nodemask);
6817 node_set(node, used_nodes);
6819 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6820 int next_node = find_next_best_node(node, &used_nodes);
6822 node_to_cpumask_ptr_next(nodemask, next_node);
6823 cpus_or(*span, *span, *nodemask);
6826 #endif /* CONFIG_NUMA */
6828 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6831 * SMT sched-domains:
6833 #ifdef CONFIG_SCHED_SMT
6834 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6835 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6838 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6842 *sg = &per_cpu(sched_group_cpus, cpu);
6845 #endif /* CONFIG_SCHED_SMT */
6848 * multi-core sched-domains:
6850 #ifdef CONFIG_SCHED_MC
6851 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6852 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6853 #endif /* CONFIG_SCHED_MC */
6855 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6857 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6862 *mask = per_cpu(cpu_sibling_map, cpu);
6863 cpus_and(*mask, *mask, *cpu_map);
6864 group = first_cpu(*mask);
6866 *sg = &per_cpu(sched_group_core, group);
6869 #elif defined(CONFIG_SCHED_MC)
6871 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6875 *sg = &per_cpu(sched_group_core, cpu);
6880 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6881 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6884 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6888 #ifdef CONFIG_SCHED_MC
6889 *mask = cpu_coregroup_map(cpu);
6890 cpus_and(*mask, *mask, *cpu_map);
6891 group = first_cpu(*mask);
6892 #elif defined(CONFIG_SCHED_SMT)
6893 *mask = per_cpu(cpu_sibling_map, cpu);
6894 cpus_and(*mask, *mask, *cpu_map);
6895 group = first_cpu(*mask);
6900 *sg = &per_cpu(sched_group_phys, group);
6906 * The init_sched_build_groups can't handle what we want to do with node
6907 * groups, so roll our own. Now each node has its own list of groups which
6908 * gets dynamically allocated.
6910 static DEFINE_PER_CPU(struct sched_domain, node_domains);
6911 static struct sched_group ***sched_group_nodes_bycpu;
6913 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6914 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
6916 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6917 struct sched_group **sg, cpumask_t *nodemask)
6921 *nodemask = node_to_cpumask(cpu_to_node(cpu));
6922 cpus_and(*nodemask, *nodemask, *cpu_map);
6923 group = first_cpu(*nodemask);
6926 *sg = &per_cpu(sched_group_allnodes, group);
6930 static void init_numa_sched_groups_power(struct sched_group *group_head)
6932 struct sched_group *sg = group_head;
6938 for_each_cpu_mask(j, sg->cpumask) {
6939 struct sched_domain *sd;
6941 sd = &per_cpu(phys_domains, j);
6942 if (j != first_cpu(sd->groups->cpumask)) {
6944 * Only add "power" once for each
6950 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
6953 } while (sg != group_head);
6955 #endif /* CONFIG_NUMA */
6958 /* Free memory allocated for various sched_group structures */
6959 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
6963 for_each_cpu_mask(cpu, *cpu_map) {
6964 struct sched_group **sched_group_nodes
6965 = sched_group_nodes_bycpu[cpu];
6967 if (!sched_group_nodes)
6970 for (i = 0; i < MAX_NUMNODES; i++) {
6971 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6973 *nodemask = node_to_cpumask(i);
6974 cpus_and(*nodemask, *nodemask, *cpu_map);
6975 if (cpus_empty(*nodemask))
6985 if (oldsg != sched_group_nodes[i])
6988 kfree(sched_group_nodes);
6989 sched_group_nodes_bycpu[cpu] = NULL;
6992 #else /* !CONFIG_NUMA */
6993 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
6996 #endif /* CONFIG_NUMA */
6999 * Initialize sched groups cpu_power.
7001 * cpu_power indicates the capacity of sched group, which is used while
7002 * distributing the load between different sched groups in a sched domain.
7003 * Typically cpu_power for all the groups in a sched domain will be same unless
7004 * there are asymmetries in the topology. If there are asymmetries, group
7005 * having more cpu_power will pickup more load compared to the group having
7008 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7009 * the maximum number of tasks a group can handle in the presence of other idle
7010 * or lightly loaded groups in the same sched domain.
7012 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7014 struct sched_domain *child;
7015 struct sched_group *group;
7017 WARN_ON(!sd || !sd->groups);
7019 if (cpu != first_cpu(sd->groups->cpumask))
7024 sd->groups->__cpu_power = 0;
7027 * For perf policy, if the groups in child domain share resources
7028 * (for example cores sharing some portions of the cache hierarchy
7029 * or SMT), then set this domain groups cpu_power such that each group
7030 * can handle only one task, when there are other idle groups in the
7031 * same sched domain.
7033 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
7035 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
7036 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
7041 * add cpu_power of each child group to this groups cpu_power
7043 group = child->groups;
7045 sg_inc_cpu_power(sd->groups, group->__cpu_power);
7046 group = group->next;
7047 } while (group != child->groups);
7051 * Initializers for schedule domains
7052 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7055 #define SD_INIT(sd, type) sd_init_##type(sd)
7056 #define SD_INIT_FUNC(type) \
7057 static noinline void sd_init_##type(struct sched_domain *sd) \
7059 memset(sd, 0, sizeof(*sd)); \
7060 *sd = SD_##type##_INIT; \
7061 sd->level = SD_LV_##type; \
7066 SD_INIT_FUNC(ALLNODES)
7069 #ifdef CONFIG_SCHED_SMT
7070 SD_INIT_FUNC(SIBLING)
7072 #ifdef CONFIG_SCHED_MC
7077 * To minimize stack usage kmalloc room for cpumasks and share the
7078 * space as the usage in build_sched_domains() dictates. Used only
7079 * if the amount of space is significant.
7082 cpumask_t tmpmask; /* make this one first */
7085 cpumask_t this_sibling_map;
7086 cpumask_t this_core_map;
7088 cpumask_t send_covered;
7091 cpumask_t domainspan;
7093 cpumask_t notcovered;
7098 #define SCHED_CPUMASK_ALLOC 1
7099 #define SCHED_CPUMASK_FREE(v) kfree(v)
7100 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
7102 #define SCHED_CPUMASK_ALLOC 0
7103 #define SCHED_CPUMASK_FREE(v)
7104 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
7107 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
7108 ((unsigned long)(a) + offsetof(struct allmasks, v))
7110 static int default_relax_domain_level = -1;
7112 static int __init setup_relax_domain_level(char *str)
7116 val = simple_strtoul(str, NULL, 0);
7117 if (val < SD_LV_MAX)
7118 default_relax_domain_level = val;
7122 __setup("relax_domain_level=", setup_relax_domain_level);
7124 static void set_domain_attribute(struct sched_domain *sd,
7125 struct sched_domain_attr *attr)
7129 if (!attr || attr->relax_domain_level < 0) {
7130 if (default_relax_domain_level < 0)
7133 request = default_relax_domain_level;
7135 request = attr->relax_domain_level;
7136 if (request < sd->level) {
7137 /* turn off idle balance on this domain */
7138 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
7140 /* turn on idle balance on this domain */
7141 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
7146 * Build sched domains for a given set of cpus and attach the sched domains
7147 * to the individual cpus
7149 static int __build_sched_domains(const cpumask_t *cpu_map,
7150 struct sched_domain_attr *attr)
7153 struct root_domain *rd;
7154 SCHED_CPUMASK_DECLARE(allmasks);
7157 struct sched_group **sched_group_nodes = NULL;
7158 int sd_allnodes = 0;
7161 * Allocate the per-node list of sched groups
7163 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
7165 if (!sched_group_nodes) {
7166 printk(KERN_WARNING "Can not alloc sched group node list\n");
7171 rd = alloc_rootdomain();
7173 printk(KERN_WARNING "Cannot alloc root domain\n");
7175 kfree(sched_group_nodes);
7180 #if SCHED_CPUMASK_ALLOC
7181 /* get space for all scratch cpumask variables */
7182 allmasks = kmalloc(sizeof(*allmasks), GFP_KERNEL);
7184 printk(KERN_WARNING "Cannot alloc cpumask array\n");
7187 kfree(sched_group_nodes);
7192 tmpmask = (cpumask_t *)allmasks;
7196 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
7200 * Set up domains for cpus specified by the cpu_map.
7202 for_each_cpu_mask(i, *cpu_map) {
7203 struct sched_domain *sd = NULL, *p;
7204 SCHED_CPUMASK_VAR(nodemask, allmasks);
7206 *nodemask = node_to_cpumask(cpu_to_node(i));
7207 cpus_and(*nodemask, *nodemask, *cpu_map);
7210 if (cpus_weight(*cpu_map) >
7211 SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) {
7212 sd = &per_cpu(allnodes_domains, i);
7213 SD_INIT(sd, ALLNODES);
7214 set_domain_attribute(sd, attr);
7215 sd->span = *cpu_map;
7216 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
7222 sd = &per_cpu(node_domains, i);
7224 set_domain_attribute(sd, attr);
7225 sched_domain_node_span(cpu_to_node(i), &sd->span);
7229 cpus_and(sd->span, sd->span, *cpu_map);
7233 sd = &per_cpu(phys_domains, i);
7235 set_domain_attribute(sd, attr);
7236 sd->span = *nodemask;
7240 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
7242 #ifdef CONFIG_SCHED_MC
7244 sd = &per_cpu(core_domains, i);
7246 set_domain_attribute(sd, attr);
7247 sd->span = cpu_coregroup_map(i);
7248 cpus_and(sd->span, sd->span, *cpu_map);
7251 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
7254 #ifdef CONFIG_SCHED_SMT
7256 sd = &per_cpu(cpu_domains, i);
7257 SD_INIT(sd, SIBLING);
7258 set_domain_attribute(sd, attr);
7259 sd->span = per_cpu(cpu_sibling_map, i);
7260 cpus_and(sd->span, sd->span, *cpu_map);
7263 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
7267 #ifdef CONFIG_SCHED_SMT
7268 /* Set up CPU (sibling) groups */
7269 for_each_cpu_mask(i, *cpu_map) {
7270 SCHED_CPUMASK_VAR(this_sibling_map, allmasks);
7271 SCHED_CPUMASK_VAR(send_covered, allmasks);
7273 *this_sibling_map = per_cpu(cpu_sibling_map, i);
7274 cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map);
7275 if (i != first_cpu(*this_sibling_map))
7278 init_sched_build_groups(this_sibling_map, cpu_map,
7280 send_covered, tmpmask);
7284 #ifdef CONFIG_SCHED_MC
7285 /* Set up multi-core groups */
7286 for_each_cpu_mask(i, *cpu_map) {
7287 SCHED_CPUMASK_VAR(this_core_map, allmasks);
7288 SCHED_CPUMASK_VAR(send_covered, allmasks);
7290 *this_core_map = cpu_coregroup_map(i);
7291 cpus_and(*this_core_map, *this_core_map, *cpu_map);
7292 if (i != first_cpu(*this_core_map))
7295 init_sched_build_groups(this_core_map, cpu_map,
7297 send_covered, tmpmask);
7301 /* Set up physical groups */
7302 for (i = 0; i < MAX_NUMNODES; i++) {
7303 SCHED_CPUMASK_VAR(nodemask, allmasks);
7304 SCHED_CPUMASK_VAR(send_covered, allmasks);
7306 *nodemask = node_to_cpumask(i);
7307 cpus_and(*nodemask, *nodemask, *cpu_map);
7308 if (cpus_empty(*nodemask))
7311 init_sched_build_groups(nodemask, cpu_map,
7313 send_covered, tmpmask);
7317 /* Set up node groups */
7319 SCHED_CPUMASK_VAR(send_covered, allmasks);
7321 init_sched_build_groups(cpu_map, cpu_map,
7322 &cpu_to_allnodes_group,
7323 send_covered, tmpmask);
7326 for (i = 0; i < MAX_NUMNODES; i++) {
7327 /* Set up node groups */
7328 struct sched_group *sg, *prev;
7329 SCHED_CPUMASK_VAR(nodemask, allmasks);
7330 SCHED_CPUMASK_VAR(domainspan, allmasks);
7331 SCHED_CPUMASK_VAR(covered, allmasks);
7334 *nodemask = node_to_cpumask(i);
7335 cpus_clear(*covered);
7337 cpus_and(*nodemask, *nodemask, *cpu_map);
7338 if (cpus_empty(*nodemask)) {
7339 sched_group_nodes[i] = NULL;
7343 sched_domain_node_span(i, domainspan);
7344 cpus_and(*domainspan, *domainspan, *cpu_map);
7346 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
7348 printk(KERN_WARNING "Can not alloc domain group for "
7352 sched_group_nodes[i] = sg;
7353 for_each_cpu_mask(j, *nodemask) {
7354 struct sched_domain *sd;
7356 sd = &per_cpu(node_domains, j);
7359 sg->__cpu_power = 0;
7360 sg->cpumask = *nodemask;
7362 cpus_or(*covered, *covered, *nodemask);
7365 for (j = 0; j < MAX_NUMNODES; j++) {
7366 SCHED_CPUMASK_VAR(notcovered, allmasks);
7367 int n = (i + j) % MAX_NUMNODES;
7368 node_to_cpumask_ptr(pnodemask, n);
7370 cpus_complement(*notcovered, *covered);
7371 cpus_and(*tmpmask, *notcovered, *cpu_map);
7372 cpus_and(*tmpmask, *tmpmask, *domainspan);
7373 if (cpus_empty(*tmpmask))
7376 cpus_and(*tmpmask, *tmpmask, *pnodemask);
7377 if (cpus_empty(*tmpmask))
7380 sg = kmalloc_node(sizeof(struct sched_group),
7384 "Can not alloc domain group for node %d\n", j);
7387 sg->__cpu_power = 0;
7388 sg->cpumask = *tmpmask;
7389 sg->next = prev->next;
7390 cpus_or(*covered, *covered, *tmpmask);
7397 /* Calculate CPU power for physical packages and nodes */
7398 #ifdef CONFIG_SCHED_SMT
7399 for_each_cpu_mask(i, *cpu_map) {
7400 struct sched_domain *sd = &per_cpu(cpu_domains, i);
7402 init_sched_groups_power(i, sd);
7405 #ifdef CONFIG_SCHED_MC
7406 for_each_cpu_mask(i, *cpu_map) {
7407 struct sched_domain *sd = &per_cpu(core_domains, i);
7409 init_sched_groups_power(i, sd);
7413 for_each_cpu_mask(i, *cpu_map) {
7414 struct sched_domain *sd = &per_cpu(phys_domains, i);
7416 init_sched_groups_power(i, sd);
7420 for (i = 0; i < MAX_NUMNODES; i++)
7421 init_numa_sched_groups_power(sched_group_nodes[i]);
7424 struct sched_group *sg;
7426 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg,
7428 init_numa_sched_groups_power(sg);
7432 /* Attach the domains */
7433 for_each_cpu_mask(i, *cpu_map) {
7434 struct sched_domain *sd;
7435 #ifdef CONFIG_SCHED_SMT
7436 sd = &per_cpu(cpu_domains, i);
7437 #elif defined(CONFIG_SCHED_MC)
7438 sd = &per_cpu(core_domains, i);
7440 sd = &per_cpu(phys_domains, i);
7442 cpu_attach_domain(sd, rd, i);
7445 SCHED_CPUMASK_FREE((void *)allmasks);
7450 free_sched_groups(cpu_map, tmpmask);
7451 SCHED_CPUMASK_FREE((void *)allmasks);
7456 static int build_sched_domains(const cpumask_t *cpu_map)
7458 return __build_sched_domains(cpu_map, NULL);
7461 static cpumask_t *doms_cur; /* current sched domains */
7462 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7463 static struct sched_domain_attr *dattr_cur;
7464 /* attribues of custom domains in 'doms_cur' */
7467 * Special case: If a kmalloc of a doms_cur partition (array of
7468 * cpumask_t) fails, then fallback to a single sched domain,
7469 * as determined by the single cpumask_t fallback_doms.
7471 static cpumask_t fallback_doms;
7473 void __attribute__((weak)) arch_update_cpu_topology(void)
7478 * Free current domain masks.
7479 * Called after all cpus are attached to NULL domain.
7481 static void free_sched_domains(void)
7484 if (doms_cur != &fallback_doms)
7486 doms_cur = &fallback_doms;
7490 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7491 * For now this just excludes isolated cpus, but could be used to
7492 * exclude other special cases in the future.
7494 static int arch_init_sched_domains(const cpumask_t *cpu_map)
7498 arch_update_cpu_topology();
7500 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
7502 doms_cur = &fallback_doms;
7503 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
7505 err = build_sched_domains(doms_cur);
7506 register_sched_domain_sysctl();
7511 static void arch_destroy_sched_domains(const cpumask_t *cpu_map,
7514 free_sched_groups(cpu_map, tmpmask);
7518 * Detach sched domains from a group of cpus specified in cpu_map
7519 * These cpus will now be attached to the NULL domain
7521 static void detach_destroy_domains(const cpumask_t *cpu_map)
7526 unregister_sched_domain_sysctl();
7528 for_each_cpu_mask(i, *cpu_map)
7529 cpu_attach_domain(NULL, &def_root_domain, i);
7530 synchronize_sched();
7531 arch_destroy_sched_domains(cpu_map, &tmpmask);
7534 /* handle null as "default" */
7535 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7536 struct sched_domain_attr *new, int idx_new)
7538 struct sched_domain_attr tmp;
7545 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7546 new ? (new + idx_new) : &tmp,
7547 sizeof(struct sched_domain_attr));
7551 * Partition sched domains as specified by the 'ndoms_new'
7552 * cpumasks in the array doms_new[] of cpumasks. This compares
7553 * doms_new[] to the current sched domain partitioning, doms_cur[].
7554 * It destroys each deleted domain and builds each new domain.
7556 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7557 * The masks don't intersect (don't overlap.) We should setup one
7558 * sched domain for each mask. CPUs not in any of the cpumasks will
7559 * not be load balanced. If the same cpumask appears both in the
7560 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7563 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7564 * ownership of it and will kfree it when done with it. If the caller
7565 * failed the kmalloc call, then it can pass in doms_new == NULL,
7566 * and partition_sched_domains() will fallback to the single partition
7569 * Call with hotplug lock held
7571 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new,
7572 struct sched_domain_attr *dattr_new)
7576 mutex_lock(&sched_domains_mutex);
7578 /* always unregister in case we don't destroy any domains */
7579 unregister_sched_domain_sysctl();
7581 if (doms_new == NULL) {
7583 doms_new = &fallback_doms;
7584 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
7588 /* Destroy deleted domains */
7589 for (i = 0; i < ndoms_cur; i++) {
7590 for (j = 0; j < ndoms_new; j++) {
7591 if (cpus_equal(doms_cur[i], doms_new[j])
7592 && dattrs_equal(dattr_cur, i, dattr_new, j))
7595 /* no match - a current sched domain not in new doms_new[] */
7596 detach_destroy_domains(doms_cur + i);
7601 /* Build new domains */
7602 for (i = 0; i < ndoms_new; i++) {
7603 for (j = 0; j < ndoms_cur; j++) {
7604 if (cpus_equal(doms_new[i], doms_cur[j])
7605 && dattrs_equal(dattr_new, i, dattr_cur, j))
7608 /* no match - add a new doms_new */
7609 __build_sched_domains(doms_new + i,
7610 dattr_new ? dattr_new + i : NULL);
7615 /* Remember the new sched domains */
7616 if (doms_cur != &fallback_doms)
7618 kfree(dattr_cur); /* kfree(NULL) is safe */
7619 doms_cur = doms_new;
7620 dattr_cur = dattr_new;
7621 ndoms_cur = ndoms_new;
7623 register_sched_domain_sysctl();
7625 mutex_unlock(&sched_domains_mutex);
7628 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7629 int arch_reinit_sched_domains(void)
7634 mutex_lock(&sched_domains_mutex);
7635 detach_destroy_domains(&cpu_online_map);
7636 free_sched_domains();
7637 err = arch_init_sched_domains(&cpu_online_map);
7638 mutex_unlock(&sched_domains_mutex);
7644 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7648 if (buf[0] != '0' && buf[0] != '1')
7652 sched_smt_power_savings = (buf[0] == '1');
7654 sched_mc_power_savings = (buf[0] == '1');
7656 ret = arch_reinit_sched_domains();
7658 return ret ? ret : count;
7661 #ifdef CONFIG_SCHED_MC
7662 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
7664 return sprintf(page, "%u\n", sched_mc_power_savings);
7666 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
7667 const char *buf, size_t count)
7669 return sched_power_savings_store(buf, count, 0);
7671 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
7672 sched_mc_power_savings_store);
7675 #ifdef CONFIG_SCHED_SMT
7676 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
7678 return sprintf(page, "%u\n", sched_smt_power_savings);
7680 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
7681 const char *buf, size_t count)
7683 return sched_power_savings_store(buf, count, 1);
7685 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
7686 sched_smt_power_savings_store);
7689 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7693 #ifdef CONFIG_SCHED_SMT
7695 err = sysfs_create_file(&cls->kset.kobj,
7696 &attr_sched_smt_power_savings.attr);
7698 #ifdef CONFIG_SCHED_MC
7699 if (!err && mc_capable())
7700 err = sysfs_create_file(&cls->kset.kobj,
7701 &attr_sched_mc_power_savings.attr);
7705 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7708 * Force a reinitialization of the sched domains hierarchy. The domains
7709 * and groups cannot be updated in place without racing with the balancing
7710 * code, so we temporarily attach all running cpus to the NULL domain
7711 * which will prevent rebalancing while the sched domains are recalculated.
7713 static int update_sched_domains(struct notifier_block *nfb,
7714 unsigned long action, void *hcpu)
7716 int cpu = (int)(long)hcpu;
7719 case CPU_DOWN_PREPARE:
7720 case CPU_DOWN_PREPARE_FROZEN:
7721 disable_runtime(cpu_rq(cpu));
7723 case CPU_UP_PREPARE:
7724 case CPU_UP_PREPARE_FROZEN:
7725 detach_destroy_domains(&cpu_online_map);
7726 free_sched_domains();
7730 case CPU_DOWN_FAILED:
7731 case CPU_DOWN_FAILED_FROZEN:
7733 case CPU_ONLINE_FROZEN:
7734 enable_runtime(cpu_rq(cpu));
7736 case CPU_UP_CANCELED:
7737 case CPU_UP_CANCELED_FROZEN:
7739 case CPU_DEAD_FROZEN:
7741 * Fall through and re-initialise the domains.
7748 #ifndef CONFIG_CPUSETS
7750 * Create default domain partitioning if cpusets are disabled.
7751 * Otherwise we let cpusets rebuild the domains based on the
7755 /* The hotplug lock is already held by cpu_up/cpu_down */
7756 arch_init_sched_domains(&cpu_online_map);
7762 void __init sched_init_smp(void)
7764 cpumask_t non_isolated_cpus;
7766 #if defined(CONFIG_NUMA)
7767 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7769 BUG_ON(sched_group_nodes_bycpu == NULL);
7772 mutex_lock(&sched_domains_mutex);
7773 arch_init_sched_domains(&cpu_online_map);
7774 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
7775 if (cpus_empty(non_isolated_cpus))
7776 cpu_set(smp_processor_id(), non_isolated_cpus);
7777 mutex_unlock(&sched_domains_mutex);
7779 /* XXX: Theoretical race here - CPU may be hotplugged now */
7780 hotcpu_notifier(update_sched_domains, 0);
7783 /* Move init over to a non-isolated CPU */
7784 if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0)
7786 sched_init_granularity();
7789 void __init sched_init_smp(void)
7791 sched_init_granularity();
7793 #endif /* CONFIG_SMP */
7795 int in_sched_functions(unsigned long addr)
7797 return in_lock_functions(addr) ||
7798 (addr >= (unsigned long)__sched_text_start
7799 && addr < (unsigned long)__sched_text_end);
7802 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7804 cfs_rq->tasks_timeline = RB_ROOT;
7805 INIT_LIST_HEAD(&cfs_rq->tasks);
7806 #ifdef CONFIG_FAIR_GROUP_SCHED
7809 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7812 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7814 struct rt_prio_array *array;
7817 array = &rt_rq->active;
7818 for (i = 0; i < MAX_RT_PRIO; i++) {
7819 INIT_LIST_HEAD(array->queue + i);
7820 __clear_bit(i, array->bitmap);
7822 /* delimiter for bitsearch: */
7823 __set_bit(MAX_RT_PRIO, array->bitmap);
7825 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7826 rt_rq->highest_prio = MAX_RT_PRIO;
7829 rt_rq->rt_nr_migratory = 0;
7830 rt_rq->overloaded = 0;
7834 rt_rq->rt_throttled = 0;
7835 rt_rq->rt_runtime = 0;
7836 spin_lock_init(&rt_rq->rt_runtime_lock);
7838 #ifdef CONFIG_RT_GROUP_SCHED
7839 rt_rq->rt_nr_boosted = 0;
7844 #ifdef CONFIG_FAIR_GROUP_SCHED
7845 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7846 struct sched_entity *se, int cpu, int add,
7847 struct sched_entity *parent)
7849 struct rq *rq = cpu_rq(cpu);
7850 tg->cfs_rq[cpu] = cfs_rq;
7851 init_cfs_rq(cfs_rq, rq);
7854 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7857 /* se could be NULL for init_task_group */
7862 se->cfs_rq = &rq->cfs;
7864 se->cfs_rq = parent->my_q;
7867 se->load.weight = tg->shares;
7868 se->load.inv_weight = 0;
7869 se->parent = parent;
7873 #ifdef CONFIG_RT_GROUP_SCHED
7874 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7875 struct sched_rt_entity *rt_se, int cpu, int add,
7876 struct sched_rt_entity *parent)
7878 struct rq *rq = cpu_rq(cpu);
7880 tg->rt_rq[cpu] = rt_rq;
7881 init_rt_rq(rt_rq, rq);
7883 rt_rq->rt_se = rt_se;
7884 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7886 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7888 tg->rt_se[cpu] = rt_se;
7893 rt_se->rt_rq = &rq->rt;
7895 rt_se->rt_rq = parent->my_q;
7897 rt_se->my_q = rt_rq;
7898 rt_se->parent = parent;
7899 INIT_LIST_HEAD(&rt_se->run_list);
7903 void __init sched_init(void)
7906 unsigned long alloc_size = 0, ptr;
7908 #ifdef CONFIG_FAIR_GROUP_SCHED
7909 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7911 #ifdef CONFIG_RT_GROUP_SCHED
7912 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7914 #ifdef CONFIG_USER_SCHED
7918 * As sched_init() is called before page_alloc is setup,
7919 * we use alloc_bootmem().
7922 ptr = (unsigned long)alloc_bootmem(alloc_size);
7924 #ifdef CONFIG_FAIR_GROUP_SCHED
7925 init_task_group.se = (struct sched_entity **)ptr;
7926 ptr += nr_cpu_ids * sizeof(void **);
7928 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7929 ptr += nr_cpu_ids * sizeof(void **);
7931 #ifdef CONFIG_USER_SCHED
7932 root_task_group.se = (struct sched_entity **)ptr;
7933 ptr += nr_cpu_ids * sizeof(void **);
7935 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7936 ptr += nr_cpu_ids * sizeof(void **);
7937 #endif /* CONFIG_USER_SCHED */
7938 #endif /* CONFIG_FAIR_GROUP_SCHED */
7939 #ifdef CONFIG_RT_GROUP_SCHED
7940 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7941 ptr += nr_cpu_ids * sizeof(void **);
7943 init_task_group.rt_rq = (struct rt_rq **)ptr;
7944 ptr += nr_cpu_ids * sizeof(void **);
7946 #ifdef CONFIG_USER_SCHED
7947 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7948 ptr += nr_cpu_ids * sizeof(void **);
7950 root_task_group.rt_rq = (struct rt_rq **)ptr;
7951 ptr += nr_cpu_ids * sizeof(void **);
7952 #endif /* CONFIG_USER_SCHED */
7953 #endif /* CONFIG_RT_GROUP_SCHED */
7957 init_defrootdomain();
7960 init_rt_bandwidth(&def_rt_bandwidth,
7961 global_rt_period(), global_rt_runtime());
7963 #ifdef CONFIG_RT_GROUP_SCHED
7964 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7965 global_rt_period(), global_rt_runtime());
7966 #ifdef CONFIG_USER_SCHED
7967 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7968 global_rt_period(), RUNTIME_INF);
7969 #endif /* CONFIG_USER_SCHED */
7970 #endif /* CONFIG_RT_GROUP_SCHED */
7972 #ifdef CONFIG_GROUP_SCHED
7973 list_add(&init_task_group.list, &task_groups);
7974 INIT_LIST_HEAD(&init_task_group.children);
7976 #ifdef CONFIG_USER_SCHED
7977 INIT_LIST_HEAD(&root_task_group.children);
7978 init_task_group.parent = &root_task_group;
7979 list_add(&init_task_group.siblings, &root_task_group.children);
7980 #endif /* CONFIG_USER_SCHED */
7981 #endif /* CONFIG_GROUP_SCHED */
7983 for_each_possible_cpu(i) {
7987 spin_lock_init(&rq->lock);
7988 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
7990 init_cfs_rq(&rq->cfs, rq);
7991 init_rt_rq(&rq->rt, rq);
7992 #ifdef CONFIG_FAIR_GROUP_SCHED
7993 init_task_group.shares = init_task_group_load;
7994 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7995 #ifdef CONFIG_CGROUP_SCHED
7997 * How much cpu bandwidth does init_task_group get?
7999 * In case of task-groups formed thr' the cgroup filesystem, it
8000 * gets 100% of the cpu resources in the system. This overall
8001 * system cpu resource is divided among the tasks of
8002 * init_task_group and its child task-groups in a fair manner,
8003 * based on each entity's (task or task-group's) weight
8004 * (se->load.weight).
8006 * In other words, if init_task_group has 10 tasks of weight
8007 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8008 * then A0's share of the cpu resource is:
8010 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8012 * We achieve this by letting init_task_group's tasks sit
8013 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8015 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
8016 #elif defined CONFIG_USER_SCHED
8017 root_task_group.shares = NICE_0_LOAD;
8018 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
8020 * In case of task-groups formed thr' the user id of tasks,
8021 * init_task_group represents tasks belonging to root user.
8022 * Hence it forms a sibling of all subsequent groups formed.
8023 * In this case, init_task_group gets only a fraction of overall
8024 * system cpu resource, based on the weight assigned to root
8025 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8026 * by letting tasks of init_task_group sit in a separate cfs_rq
8027 * (init_cfs_rq) and having one entity represent this group of
8028 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8030 init_tg_cfs_entry(&init_task_group,
8031 &per_cpu(init_cfs_rq, i),
8032 &per_cpu(init_sched_entity, i), i, 1,
8033 root_task_group.se[i]);
8036 #endif /* CONFIG_FAIR_GROUP_SCHED */
8038 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8039 #ifdef CONFIG_RT_GROUP_SCHED
8040 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8041 #ifdef CONFIG_CGROUP_SCHED
8042 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
8043 #elif defined CONFIG_USER_SCHED
8044 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
8045 init_tg_rt_entry(&init_task_group,
8046 &per_cpu(init_rt_rq, i),
8047 &per_cpu(init_sched_rt_entity, i), i, 1,
8048 root_task_group.rt_se[i]);
8052 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8053 rq->cpu_load[j] = 0;
8057 rq->active_balance = 0;
8058 rq->next_balance = jiffies;
8062 rq->migration_thread = NULL;
8063 INIT_LIST_HEAD(&rq->migration_queue);
8064 rq_attach_root(rq, &def_root_domain);
8067 atomic_set(&rq->nr_iowait, 0);
8070 set_load_weight(&init_task);
8072 #ifdef CONFIG_PREEMPT_NOTIFIERS
8073 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8077 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
8080 #ifdef CONFIG_RT_MUTEXES
8081 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
8085 * The boot idle thread does lazy MMU switching as well:
8087 atomic_inc(&init_mm.mm_count);
8088 enter_lazy_tlb(&init_mm, current);
8091 * Make us the idle thread. Technically, schedule() should not be
8092 * called from this thread, however somewhere below it might be,
8093 * but because we are the idle thread, we just pick up running again
8094 * when this runqueue becomes "idle".
8096 init_idle(current, smp_processor_id());
8098 * During early bootup we pretend to be a normal task:
8100 current->sched_class = &fair_sched_class;
8102 scheduler_running = 1;
8105 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8106 void __might_sleep(char *file, int line)
8109 static unsigned long prev_jiffy; /* ratelimiting */
8111 if ((in_atomic() || irqs_disabled()) &&
8112 system_state == SYSTEM_RUNNING && !oops_in_progress) {
8113 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8115 prev_jiffy = jiffies;
8116 printk(KERN_ERR "BUG: sleeping function called from invalid"
8117 " context at %s:%d\n", file, line);
8118 printk("in_atomic():%d, irqs_disabled():%d\n",
8119 in_atomic(), irqs_disabled());
8120 debug_show_held_locks(current);
8121 if (irqs_disabled())
8122 print_irqtrace_events(current);
8127 EXPORT_SYMBOL(__might_sleep);
8130 #ifdef CONFIG_MAGIC_SYSRQ
8131 static void normalize_task(struct rq *rq, struct task_struct *p)
8135 update_rq_clock(rq);
8136 on_rq = p->se.on_rq;
8138 deactivate_task(rq, p, 0);
8139 __setscheduler(rq, p, SCHED_NORMAL, 0);
8141 activate_task(rq, p, 0);
8142 resched_task(rq->curr);
8146 void normalize_rt_tasks(void)
8148 struct task_struct *g, *p;
8149 unsigned long flags;
8152 read_lock_irqsave(&tasklist_lock, flags);
8153 do_each_thread(g, p) {
8155 * Only normalize user tasks:
8160 p->se.exec_start = 0;
8161 #ifdef CONFIG_SCHEDSTATS
8162 p->se.wait_start = 0;
8163 p->se.sleep_start = 0;
8164 p->se.block_start = 0;
8169 * Renice negative nice level userspace
8172 if (TASK_NICE(p) < 0 && p->mm)
8173 set_user_nice(p, 0);
8177 spin_lock(&p->pi_lock);
8178 rq = __task_rq_lock(p);
8180 normalize_task(rq, p);
8182 __task_rq_unlock(rq);
8183 spin_unlock(&p->pi_lock);
8184 } while_each_thread(g, p);
8186 read_unlock_irqrestore(&tasklist_lock, flags);
8189 #endif /* CONFIG_MAGIC_SYSRQ */
8193 * These functions are only useful for the IA64 MCA handling.
8195 * They can only be called when the whole system has been
8196 * stopped - every CPU needs to be quiescent, and no scheduling
8197 * activity can take place. Using them for anything else would
8198 * be a serious bug, and as a result, they aren't even visible
8199 * under any other configuration.
8203 * curr_task - return the current task for a given cpu.
8204 * @cpu: the processor in question.
8206 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8208 struct task_struct *curr_task(int cpu)
8210 return cpu_curr(cpu);
8214 * set_curr_task - set the current task for a given cpu.
8215 * @cpu: the processor in question.
8216 * @p: the task pointer to set.
8218 * Description: This function must only be used when non-maskable interrupts
8219 * are serviced on a separate stack. It allows the architecture to switch the
8220 * notion of the current task on a cpu in a non-blocking manner. This function
8221 * must be called with all CPU's synchronized, and interrupts disabled, the
8222 * and caller must save the original value of the current task (see
8223 * curr_task() above) and restore that value before reenabling interrupts and
8224 * re-starting the system.
8226 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8228 void set_curr_task(int cpu, struct task_struct *p)
8235 #ifdef CONFIG_FAIR_GROUP_SCHED
8236 static void free_fair_sched_group(struct task_group *tg)
8240 for_each_possible_cpu(i) {
8242 kfree(tg->cfs_rq[i]);
8252 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8254 struct cfs_rq *cfs_rq;
8255 struct sched_entity *se, *parent_se;
8259 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8262 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8266 tg->shares = NICE_0_LOAD;
8268 for_each_possible_cpu(i) {
8271 cfs_rq = kmalloc_node(sizeof(struct cfs_rq),
8272 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8276 se = kmalloc_node(sizeof(struct sched_entity),
8277 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8281 parent_se = parent ? parent->se[i] : NULL;
8282 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent_se);
8291 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8293 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8294 &cpu_rq(cpu)->leaf_cfs_rq_list);
8297 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8299 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8301 #else /* !CONFG_FAIR_GROUP_SCHED */
8302 static inline void free_fair_sched_group(struct task_group *tg)
8307 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8312 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8316 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8319 #endif /* CONFIG_FAIR_GROUP_SCHED */
8321 #ifdef CONFIG_RT_GROUP_SCHED
8322 static void free_rt_sched_group(struct task_group *tg)
8326 destroy_rt_bandwidth(&tg->rt_bandwidth);
8328 for_each_possible_cpu(i) {
8330 kfree(tg->rt_rq[i]);
8332 kfree(tg->rt_se[i]);
8340 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8342 struct rt_rq *rt_rq;
8343 struct sched_rt_entity *rt_se, *parent_se;
8347 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8350 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8354 init_rt_bandwidth(&tg->rt_bandwidth,
8355 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8357 for_each_possible_cpu(i) {
8360 rt_rq = kmalloc_node(sizeof(struct rt_rq),
8361 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8365 rt_se = kmalloc_node(sizeof(struct sched_rt_entity),
8366 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8370 parent_se = parent ? parent->rt_se[i] : NULL;
8371 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent_se);
8380 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8382 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8383 &cpu_rq(cpu)->leaf_rt_rq_list);
8386 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8388 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8390 #else /* !CONFIG_RT_GROUP_SCHED */
8391 static inline void free_rt_sched_group(struct task_group *tg)
8396 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8401 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8405 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8408 #endif /* CONFIG_RT_GROUP_SCHED */
8410 #ifdef CONFIG_GROUP_SCHED
8411 static void free_sched_group(struct task_group *tg)
8413 free_fair_sched_group(tg);
8414 free_rt_sched_group(tg);
8418 /* allocate runqueue etc for a new task group */
8419 struct task_group *sched_create_group(struct task_group *parent)
8421 struct task_group *tg;
8422 unsigned long flags;
8425 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8427 return ERR_PTR(-ENOMEM);
8429 if (!alloc_fair_sched_group(tg, parent))
8432 if (!alloc_rt_sched_group(tg, parent))
8435 spin_lock_irqsave(&task_group_lock, flags);
8436 for_each_possible_cpu(i) {
8437 register_fair_sched_group(tg, i);
8438 register_rt_sched_group(tg, i);
8440 list_add_rcu(&tg->list, &task_groups);
8442 WARN_ON(!parent); /* root should already exist */
8444 tg->parent = parent;
8445 list_add_rcu(&tg->siblings, &parent->children);
8446 INIT_LIST_HEAD(&tg->children);
8447 spin_unlock_irqrestore(&task_group_lock, flags);
8452 free_sched_group(tg);
8453 return ERR_PTR(-ENOMEM);
8456 /* rcu callback to free various structures associated with a task group */
8457 static void free_sched_group_rcu(struct rcu_head *rhp)
8459 /* now it should be safe to free those cfs_rqs */
8460 free_sched_group(container_of(rhp, struct task_group, rcu));
8463 /* Destroy runqueue etc associated with a task group */
8464 void sched_destroy_group(struct task_group *tg)
8466 unsigned long flags;
8469 spin_lock_irqsave(&task_group_lock, flags);
8470 for_each_possible_cpu(i) {
8471 unregister_fair_sched_group(tg, i);
8472 unregister_rt_sched_group(tg, i);
8474 list_del_rcu(&tg->list);
8475 list_del_rcu(&tg->siblings);
8476 spin_unlock_irqrestore(&task_group_lock, flags);
8478 /* wait for possible concurrent references to cfs_rqs complete */
8479 call_rcu(&tg->rcu, free_sched_group_rcu);
8482 /* change task's runqueue when it moves between groups.
8483 * The caller of this function should have put the task in its new group
8484 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8485 * reflect its new group.
8487 void sched_move_task(struct task_struct *tsk)
8490 unsigned long flags;
8493 rq = task_rq_lock(tsk, &flags);
8495 update_rq_clock(rq);
8497 running = task_current(rq, tsk);
8498 on_rq = tsk->se.on_rq;
8501 dequeue_task(rq, tsk, 0);
8502 if (unlikely(running))
8503 tsk->sched_class->put_prev_task(rq, tsk);
8505 set_task_rq(tsk, task_cpu(tsk));
8507 #ifdef CONFIG_FAIR_GROUP_SCHED
8508 if (tsk->sched_class->moved_group)
8509 tsk->sched_class->moved_group(tsk);
8512 if (unlikely(running))
8513 tsk->sched_class->set_curr_task(rq);
8515 enqueue_task(rq, tsk, 0);
8517 task_rq_unlock(rq, &flags);
8519 #endif /* CONFIG_GROUP_SCHED */
8521 #ifdef CONFIG_FAIR_GROUP_SCHED
8522 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8524 struct cfs_rq *cfs_rq = se->cfs_rq;
8529 dequeue_entity(cfs_rq, se, 0);
8531 se->load.weight = shares;
8532 se->load.inv_weight = 0;
8535 enqueue_entity(cfs_rq, se, 0);
8538 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8540 struct cfs_rq *cfs_rq = se->cfs_rq;
8541 struct rq *rq = cfs_rq->rq;
8542 unsigned long flags;
8544 spin_lock_irqsave(&rq->lock, flags);
8545 __set_se_shares(se, shares);
8546 spin_unlock_irqrestore(&rq->lock, flags);
8549 static DEFINE_MUTEX(shares_mutex);
8551 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8554 unsigned long flags;
8557 * We can't change the weight of the root cgroup.
8562 if (shares < MIN_SHARES)
8563 shares = MIN_SHARES;
8564 else if (shares > MAX_SHARES)
8565 shares = MAX_SHARES;
8567 mutex_lock(&shares_mutex);
8568 if (tg->shares == shares)
8571 spin_lock_irqsave(&task_group_lock, flags);
8572 for_each_possible_cpu(i)
8573 unregister_fair_sched_group(tg, i);
8574 list_del_rcu(&tg->siblings);
8575 spin_unlock_irqrestore(&task_group_lock, flags);
8577 /* wait for any ongoing reference to this group to finish */
8578 synchronize_sched();
8581 * Now we are free to modify the group's share on each cpu
8582 * w/o tripping rebalance_share or load_balance_fair.
8584 tg->shares = shares;
8585 for_each_possible_cpu(i) {
8589 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8590 set_se_shares(tg->se[i], shares);
8594 * Enable load balance activity on this group, by inserting it back on
8595 * each cpu's rq->leaf_cfs_rq_list.
8597 spin_lock_irqsave(&task_group_lock, flags);
8598 for_each_possible_cpu(i)
8599 register_fair_sched_group(tg, i);
8600 list_add_rcu(&tg->siblings, &tg->parent->children);
8601 spin_unlock_irqrestore(&task_group_lock, flags);
8603 mutex_unlock(&shares_mutex);
8607 unsigned long sched_group_shares(struct task_group *tg)
8613 #ifdef CONFIG_RT_GROUP_SCHED
8615 * Ensure that the real time constraints are schedulable.
8617 static DEFINE_MUTEX(rt_constraints_mutex);
8619 static unsigned long to_ratio(u64 period, u64 runtime)
8621 if (runtime == RUNTIME_INF)
8624 return div64_u64(runtime << 16, period);
8627 #ifdef CONFIG_CGROUP_SCHED
8628 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8630 struct task_group *tgi, *parent = tg->parent;
8631 unsigned long total = 0;
8634 if (global_rt_period() < period)
8637 return to_ratio(period, runtime) <
8638 to_ratio(global_rt_period(), global_rt_runtime());
8641 if (ktime_to_ns(parent->rt_bandwidth.rt_period) < period)
8645 list_for_each_entry_rcu(tgi, &parent->children, siblings) {
8649 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8650 tgi->rt_bandwidth.rt_runtime);
8654 return total + to_ratio(period, runtime) <=
8655 to_ratio(ktime_to_ns(parent->rt_bandwidth.rt_period),
8656 parent->rt_bandwidth.rt_runtime);
8658 #elif defined CONFIG_USER_SCHED
8659 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8661 struct task_group *tgi;
8662 unsigned long total = 0;
8663 unsigned long global_ratio =
8664 to_ratio(global_rt_period(), global_rt_runtime());
8667 list_for_each_entry_rcu(tgi, &task_groups, list) {
8671 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8672 tgi->rt_bandwidth.rt_runtime);
8676 return total + to_ratio(period, runtime) < global_ratio;
8680 /* Must be called with tasklist_lock held */
8681 static inline int tg_has_rt_tasks(struct task_group *tg)
8683 struct task_struct *g, *p;
8684 do_each_thread(g, p) {
8685 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8687 } while_each_thread(g, p);
8691 static int tg_set_bandwidth(struct task_group *tg,
8692 u64 rt_period, u64 rt_runtime)
8696 mutex_lock(&rt_constraints_mutex);
8697 read_lock(&tasklist_lock);
8698 if (rt_runtime == 0 && tg_has_rt_tasks(tg)) {
8702 if (!__rt_schedulable(tg, rt_period, rt_runtime)) {
8707 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8708 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8709 tg->rt_bandwidth.rt_runtime = rt_runtime;
8711 for_each_possible_cpu(i) {
8712 struct rt_rq *rt_rq = tg->rt_rq[i];
8714 spin_lock(&rt_rq->rt_runtime_lock);
8715 rt_rq->rt_runtime = rt_runtime;
8716 spin_unlock(&rt_rq->rt_runtime_lock);
8718 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8720 read_unlock(&tasklist_lock);
8721 mutex_unlock(&rt_constraints_mutex);
8726 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8728 u64 rt_runtime, rt_period;
8730 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8731 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8732 if (rt_runtime_us < 0)
8733 rt_runtime = RUNTIME_INF;
8735 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8738 long sched_group_rt_runtime(struct task_group *tg)
8742 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8745 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8746 do_div(rt_runtime_us, NSEC_PER_USEC);
8747 return rt_runtime_us;
8750 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8752 u64 rt_runtime, rt_period;
8754 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8755 rt_runtime = tg->rt_bandwidth.rt_runtime;
8757 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8760 long sched_group_rt_period(struct task_group *tg)
8764 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8765 do_div(rt_period_us, NSEC_PER_USEC);
8766 return rt_period_us;
8769 static int sched_rt_global_constraints(void)
8771 struct task_group *tg = &root_task_group;
8772 u64 rt_runtime, rt_period;
8775 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8776 rt_runtime = tg->rt_bandwidth.rt_runtime;
8778 mutex_lock(&rt_constraints_mutex);
8779 if (!__rt_schedulable(tg, rt_period, rt_runtime))
8781 mutex_unlock(&rt_constraints_mutex);
8785 #else /* !CONFIG_RT_GROUP_SCHED */
8786 static int sched_rt_global_constraints(void)
8788 unsigned long flags;
8791 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8792 for_each_possible_cpu(i) {
8793 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8795 spin_lock(&rt_rq->rt_runtime_lock);
8796 rt_rq->rt_runtime = global_rt_runtime();
8797 spin_unlock(&rt_rq->rt_runtime_lock);
8799 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8803 #endif /* CONFIG_RT_GROUP_SCHED */
8805 int sched_rt_handler(struct ctl_table *table, int write,
8806 struct file *filp, void __user *buffer, size_t *lenp,
8810 int old_period, old_runtime;
8811 static DEFINE_MUTEX(mutex);
8814 old_period = sysctl_sched_rt_period;
8815 old_runtime = sysctl_sched_rt_runtime;
8817 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
8819 if (!ret && write) {
8820 ret = sched_rt_global_constraints();
8822 sysctl_sched_rt_period = old_period;
8823 sysctl_sched_rt_runtime = old_runtime;
8825 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8826 def_rt_bandwidth.rt_period =
8827 ns_to_ktime(global_rt_period());
8830 mutex_unlock(&mutex);
8835 #ifdef CONFIG_CGROUP_SCHED
8837 /* return corresponding task_group object of a cgroup */
8838 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8840 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8841 struct task_group, css);
8844 static struct cgroup_subsys_state *
8845 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8847 struct task_group *tg, *parent;
8849 if (!cgrp->parent) {
8850 /* This is early initialization for the top cgroup */
8851 init_task_group.css.cgroup = cgrp;
8852 return &init_task_group.css;
8855 parent = cgroup_tg(cgrp->parent);
8856 tg = sched_create_group(parent);
8858 return ERR_PTR(-ENOMEM);
8860 /* Bind the cgroup to task_group object we just created */
8861 tg->css.cgroup = cgrp;
8867 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8869 struct task_group *tg = cgroup_tg(cgrp);
8871 sched_destroy_group(tg);
8875 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8876 struct task_struct *tsk)
8878 #ifdef CONFIG_RT_GROUP_SCHED
8879 /* Don't accept realtime tasks when there is no way for them to run */
8880 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
8883 /* We don't support RT-tasks being in separate groups */
8884 if (tsk->sched_class != &fair_sched_class)
8892 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8893 struct cgroup *old_cont, struct task_struct *tsk)
8895 sched_move_task(tsk);
8898 #ifdef CONFIG_FAIR_GROUP_SCHED
8899 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8902 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8905 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8907 struct task_group *tg = cgroup_tg(cgrp);
8909 return (u64) tg->shares;
8911 #endif /* CONFIG_FAIR_GROUP_SCHED */
8913 #ifdef CONFIG_RT_GROUP_SCHED
8914 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8917 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8920 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8922 return sched_group_rt_runtime(cgroup_tg(cgrp));
8925 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8928 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8931 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8933 return sched_group_rt_period(cgroup_tg(cgrp));
8935 #endif /* CONFIG_RT_GROUP_SCHED */
8937 static struct cftype cpu_files[] = {
8938 #ifdef CONFIG_FAIR_GROUP_SCHED
8941 .read_u64 = cpu_shares_read_u64,
8942 .write_u64 = cpu_shares_write_u64,
8945 #ifdef CONFIG_RT_GROUP_SCHED
8947 .name = "rt_runtime_us",
8948 .read_s64 = cpu_rt_runtime_read,
8949 .write_s64 = cpu_rt_runtime_write,
8952 .name = "rt_period_us",
8953 .read_u64 = cpu_rt_period_read_uint,
8954 .write_u64 = cpu_rt_period_write_uint,
8959 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8961 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8964 struct cgroup_subsys cpu_cgroup_subsys = {
8966 .create = cpu_cgroup_create,
8967 .destroy = cpu_cgroup_destroy,
8968 .can_attach = cpu_cgroup_can_attach,
8969 .attach = cpu_cgroup_attach,
8970 .populate = cpu_cgroup_populate,
8971 .subsys_id = cpu_cgroup_subsys_id,
8975 #endif /* CONFIG_CGROUP_SCHED */
8977 #ifdef CONFIG_CGROUP_CPUACCT
8980 * CPU accounting code for task groups.
8982 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8983 * (balbir@in.ibm.com).
8986 /* track cpu usage of a group of tasks */
8988 struct cgroup_subsys_state css;
8989 /* cpuusage holds pointer to a u64-type object on every cpu */
8993 struct cgroup_subsys cpuacct_subsys;
8995 /* return cpu accounting group corresponding to this container */
8996 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8998 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8999 struct cpuacct, css);
9002 /* return cpu accounting group to which this task belongs */
9003 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9005 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9006 struct cpuacct, css);
9009 /* create a new cpu accounting group */
9010 static struct cgroup_subsys_state *cpuacct_create(
9011 struct cgroup_subsys *ss, struct cgroup *cgrp)
9013 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9016 return ERR_PTR(-ENOMEM);
9018 ca->cpuusage = alloc_percpu(u64);
9019 if (!ca->cpuusage) {
9021 return ERR_PTR(-ENOMEM);
9027 /* destroy an existing cpu accounting group */
9029 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9031 struct cpuacct *ca = cgroup_ca(cgrp);
9033 free_percpu(ca->cpuusage);
9037 /* return total cpu usage (in nanoseconds) of a group */
9038 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9040 struct cpuacct *ca = cgroup_ca(cgrp);
9041 u64 totalcpuusage = 0;
9044 for_each_possible_cpu(i) {
9045 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9048 * Take rq->lock to make 64-bit addition safe on 32-bit
9051 spin_lock_irq(&cpu_rq(i)->lock);
9052 totalcpuusage += *cpuusage;
9053 spin_unlock_irq(&cpu_rq(i)->lock);
9056 return totalcpuusage;
9059 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9062 struct cpuacct *ca = cgroup_ca(cgrp);
9071 for_each_possible_cpu(i) {
9072 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9074 spin_lock_irq(&cpu_rq(i)->lock);
9076 spin_unlock_irq(&cpu_rq(i)->lock);
9082 static struct cftype files[] = {
9085 .read_u64 = cpuusage_read,
9086 .write_u64 = cpuusage_write,
9090 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9092 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9096 * charge this task's execution time to its accounting group.
9098 * called with rq->lock held.
9100 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9104 if (!cpuacct_subsys.active)
9109 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
9111 *cpuusage += cputime;
9115 struct cgroup_subsys cpuacct_subsys = {
9117 .create = cpuacct_create,
9118 .destroy = cpuacct_destroy,
9119 .populate = cpuacct_populate,
9120 .subsys_id = cpuacct_subsys_id,
9122 #endif /* CONFIG_CGROUP_CPUACCT */